IEEE News

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion. ICRA 2025: 19–23 May 2025, ATLANTA, GA Enjoy today’s videos! At the FZI, it’s not just work for our robots, they join our festivities, too. Our shy robot Spot stumbled into this year’s FZI Winter Market …, a cheerful event for robots and humans alike. Will he find his place? We certainly hope so, because Feuerzangenbowle tastes much better after clinking glasses with your hot-oil-drinking friends. [ FZI ] Thanks, Georg! The Fraunhofer IOSB Autonomous Robotic Systems Research Group wishes you a Merry Christmas filled with joy, peace, and robotic wonders! [ Fraunhofer IOSB ] Thanks, Janko! There’s some thrilling action in this Christmas video from the PUT Mobile Robotics Laboratory, and the trick to put the lights on the tree is particularly clever. Enjoy! [ PUT MRL ] Thanks, Dominik! The Norlab wishes you a Merry Christmas! [ Northern Robotics Laboratory ] The Learning Systems and Robotics Lab has made a couple of robot holiday videos based on the research that they’re doing: [ Crowd Navigation ] [ Learning with Contacts ] Thanks, Sepehr! Robots on a gift mission: Christmas greetings from the DFKI Robotics Innovation Center! [ DFKI ] Happy Holidays from Clearpath Robotics! Our workshop has been bustling lately with lots of exciting projects and integrations just in time for the holidays! The TurtleBot 4 elves helped load up the sleigh with plenty of presents to go around. Rudolph the Husky A300 made the trek through the snow so our Ridgeback friend with a manipulator arm and gripper could receive its gift. [ Clearpath Robotics ] 2024 has been an eventful year for us at PAL Robotics, filled with milestones and memories. As the festive season approaches, we want to take a moment to say a heartfelt THANK YOU for being part of our journey! [ PAL Robotics ] Thanks, Rugilė! In Santa’s shop, so bright and neat, A robot marched on metal feet. With tinsel arms and bolts so tight, It trimmed the tree all through the night. It hummed a carol, beeped with cheer, “Processing joy—it’s Christmas here!” But when it tried to dance with grace, It tangled lights around its face. “Error detected!” it spun around, Then tripped and tumbled to the ground. The elves all laughed, “You’ve done your part—A clumsy bot, but with a heart!” The ArtiMinds team would like to thank all partners and customers for an exciting 2024. We wish you and your families a Merry Christmas, joyful holidays and a Happy New Year - stay healthy. [ ArtiMinds ] Thanks to FANUC CRX collaborative robots, Santa and his elves can enjoy the holiday season knowing the work is getting done for the big night. [ FANUC ] Perhaps not technically a holiday video, until you consider how all that stuff you ordered online is actually getting to you. [ Agility Robotics ] Happy Holidays from Quanser, our best wishes for a wonderful holiday season and a happy 2025! [ Quanser ] Season’s Greetings from the team at Kawasaki Robotics USA! This season, we’re building blocks of memories filled with endless joy, and assembling our good wishes for a happy, healthy, prosperous new year. May the upcoming year be filled with opportunities and successes. From our team to yours, we hope you have a wonderful holiday season surrounded by loved ones and filled with joy and laughter. [ Kawasaki Robotics ] The robotics students at Queen’s University’s Ingenuity Labs Research Institute put together a 4K Holiday Robotics Lab Fireplace video, and unlike most fireplace videos, stuff actually happens in this one. [ Ingenuity Labs ] Thanks, Joshua!

When Hurricanes Helene and Milton struck the Southeastern United States in September and October 2024, their winds—as high as 290 kilometers per hour—destroyed houses and buildings, uprooted trees, took down power lines, and damaged roads. The storms also led to massive flooding throughout the region. Damage to basic services such as electricity meant survivors couldn’t keep their cellphones charged to stay in contact with loved ones anxious to hear updates, and they couldn’t access the Internet to learn where to turn for help. In response, a fleet of disaster response vehicles maintained by the IEEE MOVE (Mobile Outreach using Volunteer Engagement) program rolled into the disaster zones to provide power, light, and connectivity. The vehicles made the situation more bearable for the hurricane survivors and first responders. The three vehicles in the IEEE MOVE program provide U.S. communities with power and communications capabilities in areas affected by widespread outages due to natural disasters. All three were deployed to areas affected by Helene and Milton. “Hundreds of Red Cross clients and dozens of staff members were helped by the technologies brought to the disaster sites by the MOVE vehicles, underscoring the critical support IEEE provides in times of crisis,” says Loretta Arellano, IEEE MOVE director. Providing post-disaster assistance IEEE MOVE volunteers often collaborate with the American Red Cross to provide electricity to the organization’s shelters with generators on MOVE-1 and MOVE-2. The trucks’ generators also support charging access for up to 100 smartphones simultaneously, bolstering communication capabilities for Red Cross staff and disaster survivors. “Hundreds of Red Cross clients and dozens of staff members were helped by the technologies brought to the disaster sites by the MOVE vehicles, underscoring the critical support IEEE provides in times of crisis.” –Loretta Arellano, IEEE MOVE director For areas with compromised communication infrastructure, the trucks connect via Starlink satellite dishes to restore Internet and phone capabilities. MOVE-3, a van introduced in August, offers additional flexibility. Unlike its larger predecessors, MOVE-3’s modular design allows its power and telecommunications equipment to be removed and set up at Red Cross facilities. That “leave help behind” capability enables the van to deploy equipment while the team moves to other locations, says Tim Troske, an IEEE senior member and the MOVE-3 operations lead. The van is strategically positioned to support areas affected by wildfires, earthquakes, and other calamities. “Realizing the IEEE mission of advancing technology for humanity is why we volunteer to do this work,” –Walt Burns, IEEE MOVE “If a natural disaster were to occur in Hawaii, it would take too long to ship the van there,” notes IEEE Senior Member Walt Burns, a MOVE volunteer. “But the van could be driven to the airport so a MOVE volunteer could unload the equipment and put it on a plane to be sent across the Pacific via air freight.” MOVE-3’s design meets crucial needs. Instead of a diesel generator, it has a 4-kilowatt-hour lithium-ion battery pack capable of powering telecom services for up to 48 hours on a single charge. The battery can be recharged by a portable solar panel or the vehicle’s alternator. The fleet’s capabilities proved invaluable in the aftermath of Helene and Milton, Arellano says. “Realizing the IEEE mission of advancing technology for humanity is why we volunteer to do this work,” Burns says.

This is a sponsored article brought to you by Qualcomm. In a move set to transform the Internet of Things (IoT) landscape, Qualcomm Technologies, Inc. has introduced its Qualcomm RB3 Gen 2 developer kits, designed to put advanced AI edge computing power into the hands of developers everywhere. This kit is available as Qualcomm RB3 Gen 2, based on the Qualcomm QCS6490, or the Qualcomm RB3 Gen 2 Lite, based on the Qualcomm QCS5430. Both QCS6490 and QCS5430 processors provide efficient, high-performance, AI enhanced solutions for applications in robotics, AI vision, industrial automation, retail, smart security, precision agriculture, smart metering, predictive maintenance and personal health. By empowering developers with robust tools for edge computing, Qualcomm Technologies is encouraging a broader range of innovators—from tech companies to startups and students—to bring cutting-edge IoT solutions to life. Nadim Ferzli, Staff Manager, Product Marketing for Qualcomm Technologies, emphasized the importance of edge computing as a critical factor in the kit’s development. “AI-enabled edge computing has a lot of benefits, including faster response times, on-device decision making and enhanced security, as well as reduced cost,” Ferzli explained, noting that processing data locally enables faster decision-making and reduces dependency on cloud-based processing. This local computing power is essential for applications that require real-time responses like robotics, security and industrial automation. “AI-enabled edge computing has a lot of benefits, including faster response times, on-device decision making and enhanced security, as well as reduced cost” —Nadim Ferzli, Qualcomm Technologies The Qualcomm RB3 Gen 2 Kits feature a modular design based on the 96Board compact, credit card- sized form factor and specifications. The kit includes numerous connection options, such as multiple USB, ethernet, camera, and display ports, as well as access to various GPIOs for low-speed communication protocols like SPI, UART, and I2C, and high-speed connections like PCIE, USB, and MIPI. The kits also come with Wi-Fi 6E, Bluetooth 5.2, and optional 5G connectivity through additional modules. Qualcomm Technologies has a dedicated resource page detailing the hardware and connections. The kits can be expanded with the addition of mezzanine boards, keeping their compact size, which is beneficial for rapid prototyping and proof-of-concept projects where users can add their own attachments and integrate the kit into their preferred robot, camera, or other hardware platform. Qualcomm Technologies also provides a template that developers can take to quickly create their own mezzanine cards. The Power of AI-Enhanced Edge Computing at the Core Central to the appeal of the Qualcomm RB3 Gen 2 is the edge-focused approach. The QCS6490 and QCS5430 processors are engineered to handle substantial computing loads at the device level. Equipped with a multi-core CPU (up to 8 cores), GPU and AI engine (NPU & DSP) producing up to 12 dense TOPS (trillions of operations per second), these microprocessors enable devices to perform complex data processing at the edge, making them ideal for high compute applications like autonomous robotics and smart vision solutions. The processors offer a combination of high-performance compute, connectivity, and energy efficiency in one package. Qualcomm AI Hub: The platform for on-device AI To facilitate and accelerate the development and adoption of AI processing at the edge, Qualcomm Technologies created the Qualcomm AI Hub, a comprehensive platform designed to facilitate the deployment of AI models directly onto edge devices, enabling efficient on-device processing for applications in vision, audio, and speech and integrates with cloud-based tools like Amazon SageMaker for end-to-end AI solutions. Developers can utilize pre-optimized models or integrate their own, with support for multiple runtimes such as TensorFlow Lite and ONNX Runtime. It offers a streamlined workflow that allows developers to compile, profile, and run AI models on actual hardware in the cloud, ensuring optimized performance and reduced latency. The combination of hardware capabilities and AI tools expands the capabilities of the device to support complex edge processing like SLM (Small Language Model), sensor fusion and autonomous machinery. Visit the Qualcomm AI Hub to learn more → This edge-first design not only improves processing speed but also enhances data security by keeping sensitive information on the device rather than transferring it to the cloud. For developers working in applications like smart security, personal health or industrial automation, this means critical data stays closer to its source, enabling faster, more secure responses in real-time scenarios. Edge AI Vision and Real-Time Decisions One of the standout features of the Qualcomm RB3 Gen 2 developer kit is the Vision Mezzanine option, which includes Qualcomm Technologies’ AI-driven image recognition capabilities. Equipped with dual cameras covering high-definition and low-definition camera support, the kits allow for real-time object detection, making it suitable for security systems, autonomous drones, and smart vision prototyping. “With our kits and enablement tools, engineers are able to accelerate the prototyping and development of AI solutions,” Ferzli explained, envisioning scenarios where edge AI is essential, such as search-and-rescue or industrial inspection. The kit can be further expanded with additional cameras that are available as optional accessories. Qualcomm Technologies Qualcomm Technologies’ advanced AI processing on the Edge technology allows the Qualcomm RB3 Gen 2 kits to recognize and process visual data on-device, a capability that significantly reduces latency and enhances operational efficiency. In practical terms, this means that a robot equipped with the Qualcomm RB3 Gen2 can navigate a warehouse, recognize obstacles, and make real-time decisions autonomously, without needing a cloud connection. “AI on the Edge enables these devices to analyze and make decisions instantaneously,” Ferzli shared, highlighting the power of Qualcomm Technologies’ processors in real-time applications. Qualcomm Technologies This local AI capability is also useful in AI-powered security systems. For example, a smart camera could be deployed to monitor a construction site, using the Qualcomm RB3 Gen 2 capabilities to detect unauthorized entry or potential hazards, and issue immediate alerts. Qualcomm Technologies’ focus on robust, high-efficiency AI computing at the device level enables devices to perform complex tasks, such as analyzing footage or identifying specific objects in high detail, directly at the edge. Ferzli highlighted a customer project involving an inspection robot for railway safety, where a company switched from a more power-hungry, costly device to the QCS6490 solution. The switch cut memory usage by 68 percent in addition to the embedded Wi-Fi connectivity provided an efficient system that reduced costs while maintaining the same accuracy. This success story exemplifies how Qualcomm Technologies’ focus on powerful compute, exceptional connectivity and power efficiency can enhance productivity and reduce operational costs. Edge Efficiency for Robotics and Autonomous Applications The Qualcomm RB3 Gen 2 developer kit’s efficiency makes it a strong choice for autonomous applications, where power consumption, connectivity and computational power are vital factors. With an emphasis on low power consumption, Qualcomm Technologies’ edge computing solutions enable battery-powered devices to operate longer between charges. According to Ferzli, Qualcomm Technologies’ DNA translates directly into these processors, offering “high compute performance, exceptional connectivity, and energy efficiency” while utilizing less memory compared to alternatives. This balance of power and efficiency allows developers to use their kit in battery-dependent applications like mobile robots and drones, where extended operation time is critical. Qualcomm Technologies Another example involves a lab using Qualcomm Technologies’ vision technology to automate bacteria colony counting, a process critical in food safety and medical diagnostics. Traditionally, lab technicians manually reviewed growth colonies in petri dishes, but with Qualcomm Technologies’ edge AI, the process was automated to deliver results instantly. “Qualcomm Technologies’ edge processing brings efficiency by reducing the need for human interaction and minimizing inaccuracies,” Ferzli explained, underscoring how their technology can simplify and accelerate workflows in various industries. Developer-First Approach: Open Access and Long-Term Support As part of its efforts to deliver an exceptional user experience for the IoT mass market, Qualcomm Technologies decided to cater more to the needs of small players by providing more open access, easier to use tools, and providing support for multiple operating systems. Qualcomm Technologies’ commitment to democratizing edge computing is clear in its developer-focused approach. The Qualcomm RB3 Gen 2 developer kits are designed to be accessible to a wide audience, from professional engineers to hobbyists, with a competitive pricing model and comprehensive support. “Our goal is to make this product available to everyone,” Ferzli said, highlighting that Qualcomm Technologies’ open-access approach enables developers to purchase the kit and begin innovating without a lengthy or exclusive onboarding process. The kits are able to support multiple OS including Linux, Android, Ubuntu, and Windows. Besides the Qualcomm Linux OS that is pre-loaded the kits will soon support Linux Ubuntu which may be attractive to the community of smaller developers, including an upcoming version that includes support for Ubuntu Desktop. In addition, Qualcomm Technologies’ recent push into the Windows laptop space is also fueling support for an upcoming Windows OS release that runs on the kit for the industrial market segment typically dominated by x86 based devices running Windows. The kit will also run Android OS. The kits are supported by software development kits (SDKs) tailored for multimedia and robotics, providing developers with sample applications and demos to build and test products faster. “We created the Qualcomm AI Hub where you can bring your models or pick one of the pre-trained models, optimize them, and test them on our products,” Ferzli said, referring to Qualcomm Technologies’ dedicated Qualcomm AI Hub platform where developers can experiment with over 125 AI models on devices hosted on the cloud before deploying it on physical devices. The Qualcomm Developer Portal and Qualcomm Developer Network YouTube channel are consistently updated with training and tutorials designed to educate and support developers throughout their product development journey. Qualcomm Technologies has also established a public community forum to address inquiries. This forum is supported by dedicated internal Qualcomm Technologies’ experts who will promptly respond to questions and provide recommendations. To support developers further, Qualcomm Technologies has created a longevity program, guaranteeing up to 15 years of hardware and software support. This commitment is particularly valuable for industries that require reliable long-term solutions, such as industrial automation, medical devices, and smart infrastructure. “Our goal is to service all developers, from hobbyists and students to global enterprises,” Ferzli said, underscoring Qualcomm Technologies’ commitment to building a comprehensive ecosystem for edge computing. Qualcomm Technologies Enabling Small and Large Developers Alike Qualcomm Technologies’ vision for democratizing edge-AI is reflected in the Qualcomm RB3 Gen 2 versatile design, which can serve both small developers and large enterprises. Whether a developer is working on a project for a large multinational or a startup exploring innovative applications, the Qualcomm RB3 Gen 2 kit provides the tools to develop high-performance, IoT-enabled products without needing an extensive engineering team. For example, a small business developing a fleet management system could use the Qualcomm RB3 Gen2 kit to build a proof of concept for smart dashcams capable of processing data locally, providing immediate feedback on road conditions, driver behavior, and vehicle health. Meanwhile, larger enterprises can use Qualcomm Technologies’ kits for more complex applications, such as industrial robotics and automated quality control. Qualcomm Technologies’ edge technology allows companies to streamline operations by reducing the dependency on centralized cloud systems, thereby minimizing latency and enhancing data privacy. Ferzli noted that even as Qualcomm Technologies serves large clients, the Qualcomm RB3 Gen 2 kits are built to cater to developers of all sizes: “If you’re a college student building a fighting robot, a startup developing a drone, or a multinational designing a worker safety monitoring system, this kit will support your developer journey in the edge-AI transformation.” Qualcomm Technologies’ Vision: Accelerating IoT Adoption with Edge Computing The Qualcomm RB3 Gen 2 developer kit is more than a powerful tool—it’s a vision for the future of IoT and edge computing. By prioritizing on-device processing, Qualcomm Technologies is pushing efficient AI Edge processing in IoT, where real-time response, enhanced privacy, and high-compute are paramount. With the Qualcomm RB3 Gen 2 developer kits, Qualcomm Technologies is making advanced IoT technology available to a broad range of innovators, from established enterprises to individual developers. As IoT continues to evolve, Qualcomm Technologies’ edge-AI focused approach is set to make a significant impact on industries ranging from smart infrastructure to robotics and autonomous vehicles. Ferzli summarized the company’s ambition: “We want to educate developers to utilize AI and IoT products better. Our technology spans the spectrum of IoT and AI, and with our developer-first approach, we’re ready to support developers in shaping the future of edge computing.” With the Qualcomm RB3 Gen 2 developer kit, Qualcomm Technologies is setting a new standard for IoT innovation at the edge, encouraging developers to harness the power of real-time, on-device intelligence to create a more connected, efficient, and intelligent world. Snapdragon and Qualcomm branded products are products of Qualcomm Technologies, Inc. and/or its subsidiaries. The registered trademark Linux is used pursuant to a sublicense from the Linux Foundation, the exclusive licensee of Linus Torvalds, owner of the mark on a worldwide basis.

Scientific literature reviews are a critical part of advancing fields of study: They provide a current state of the union through comprehensive analysis of existing research, and they identify gaps in knowledge where future studies might focus. Writing a well-done review article is a many-splendored thing, however. Researchers often comb through reams of scholarly works. They must select studies that aren’t outdated, yet avoid recency bias. Then comes the intensive work of assessing studies’ quality, extracting relevant data from works that make the cut, analyzing data to glean insights, and writing a cogent narrative that sums up the past while looking to the future. Research synthesis is a field of study unto itself, and even excellent scientists may not write excellent literature reviews. Enter artificial intelligence. As in so many industries, a crop of startups has emerged to leverage AI to speed, simplify, and revolutionize the scientific literature review process. Many of these startups position themselves as AI search engines centered on scholarly research—each with differentiating product features and target audiences. Elicit invites searchers to “analyze research papers at superhuman speed” and highlights its use by expert researchers at institutions like Google, NASA, and The World Bank. Scite says it has built the largest citation database by continually monitoring 200 million scholarly sources, and it offers “smart citations” that categorize takeaways into supporting or contrasting evidence. Consensus features a homepage demo that seems aimed at helping laypeople gain a more robust understanding of a given question, explaining the product as “Google Scholar meets ChatGPT” and offering a consensus meter that sums up major takeaways. These are but a few of many. But can AI replace high-quality, systematic scientific literature review? Experts on research synthesis tend to agree these AI models are currently great-to-excellent at performing qualitative analyses—in other words, creating a narrative summary of scientific literature. Where they’re not so good is the more complex quantitative layer that makes a review truly systematic. This quantitative synthesis typically involves statistical methods such as meta-analysis, which analyzes numerical data across multiple studies to draw more robust conclusions. “AI models can be almost 100 percent as good as humans at summarizing the key points and writing a fluid argument,” says Joshua Polanin, co-founder of the Methods of Synthesis and Integration Center (MOSAIC) at the American Institutes for Research. “But we’re not even 20 percent of the way there on quantitative synthesis,” he says. “Real meta-analysis follows a strict process in how you search for studies and quantify results. These numbers are the basis for evidence-based conclusions. AI is not close to being able to do that.” The Trouble with Quantification The quantification process can be challenging even for trained experts, Polanin explains. Both humans and AI can generally read a study and summarize the takeaway: Study A found an effect, or Study B did not find an effect. The tricky part is placing a number value on the extent of the effect. What’s more, there are often different ways to measure effects, and researchers must identify studies and measurement designs that align with the premise of their research question. Polanin says models must first identify and extract the relevant data, and then they must make nuanced calls on how to compare and analyze it. “Even as human experts, although we try to make decisions ahead of time, you might end up having to change your mind on the fly,” he says. “That isn’t something a computer will be good at.” Given the hubris that’s found around AI and within startup culture, one might expect the companies building these AI models to protest Polanin’s assessment. But you won’t get an argument from Eric Olson, co-founder of Consensus: “I couldn’t agree more, honestly,” he says. To Polanin’s point, Consensus is intentionally “higher-level than some other tools, giving people a foundational knowledge for quick insights,” Olson adds. He sees the quintessential user as a grad student: someone with an intermediate knowledge base who’s working on becoming an expert. Consensus can be one tool of many for a true subject matter expert, or it can help a non-scientist stay informed—like a Consensus user in Europe who stays abreast of the research about his child’s rare genetic disorder. “He had spent hundreds of hours on Google Scholar as a non-researcher. He told us he’d been dreaming of something like this for 10 years, and it changed his life—now he uses it every single day,” Olson says. Over at Elicit, the team targets a different type of ideal customer: “Someone working in industry in an R&D context, maybe within a biomedical company, trying to decide whether to move forward with the development of a new medical intervention,” says James Brady, head of engineering. With that high-stakes user in mind, Elicit clearly shows users claims of causality and the evidence that supports them. The tool breaks down the complex task of literature review into manageable pieces that a human can understand, and it also provides more transparency than your average chatbot: Researchers can see how the AI model arrived at an answer and can check it against the source. The Future of Scientific Review Tools Brady agrees that current AI models aren’t providing full Cochrane-style systematic reviews—but he says this is not a fundamental technical limitation. Rather, it’s a question of future advances in AI and better prompt engineering. “I don’t think there’s something our brains can do that a computer can’t, in principle,” Brady says. “And that goes for the systematic review process too.” Roman Lukyanenko, a University of Virginia professor who specializes in research methods, agrees that a major future focus should be developing ways to support the initial prompt process to glean better answers. He also notes that current models tend to prioritize journal articles that are freely accessible, yet plenty of high-quality research exists behind paywalls. Still, he’s bullish about the future. “I believe AI is tremendous—revolutionary on so many levels—for this space,” says Lukyanenko, who with Gerit Wagner and Guy Paré co-authored a pre-ChatGPT 2022 study about AI and literature review that went viral. “We have an avalanche of information, but our human biology limits what we can do with it. These tools represent great potential.” Progress in science often comes from an interdisciplinary approach, he says, and this is where AI’s potential may be greatest. “We have the term ‘Renaissance man,’ and I like to think of ‘Renaissance AI’: something that has access to a big chunk of our knowledge and can make connections,” Lukyanenko says. “We should push it hard to make serendipitous, unanticipated, distal discoveries between fields.”

IEEE Life Senior Member H. Stan Thompson has lived a couple of professional lives. For decades, he was a planning engineer and futurist at Bellsouth Telecommunications, which was formed from the merger of two Regional Bell Operating Companies around the time AT&T (“Ma Bell”) was forced to break up in 1984. When he retired in 1996, Thompson assumed he’d live out his golden years puttering around Mooresvile, N.C., the Charlotte suburb he calls home. But fate had a different plan: Over the past two decades, he has been the prime mover behind transforming a local effort to make hydrogen the fuel of choice for rail transit into a global phenomenon. In 2004, the “Centralina” region (the Greater Charlotte metro area, which straddles the North Carolina–South Carolina border) was designated as a non-attainment area for ozone under the Clean Air Act. The area stood to lose billions of dollars of federal funding for a wide variety of projects if the area’s air quality didn’t improve. Thompson stepped forward with an idea he thought would help. Local officials were mulling over an idea to put a then-idle Norfolk Southern Railroad industrial access line connecting the cities back into service as a commuter railway. Thompson’s proposal: seek federal innovation funding to upgrade and maintain a 9.6-km (6 mile) section of the proposed rail line that lay outside Charlotte’s Mecklenburg County and would therefore not be funded by big city taxes. Furthermore, he suggested, the trains could be powered electrically by hydrogen. Doing so, he reasoned, would ensure that the rail corridor didn’t exacerbate the area’s air quality issues with emissions from diesel engines while also avoiding the high cost of electrifying the line with an overhead catenary system. Thompson’s work on that project, under the aegis of the Mooresville Hydrail Initiative comprising himself and former Mooresvile mayor Bill Thunberg, led him to coin the term “hydrail” and ignited what has become a second career. IEEE Spectrum recently spoke with Thompson about hydrail’s origins and where it stands now. For many years, you were the convener of the International Hydrail Conferences (IHCs). What role did the conferences play in helping to advance hydrogen-powered rail transit from an idea to a real-world happening? H. Stan Thompson: Well, first, let me make sure that Jason Hoyle [now the principal energy policy analyst at EQ Research, a Cary, N.C.–based firm that mostly focuses on state-level energy regulation and legislation] gets credit for having put these things together. I had the idea for doing it, but Jason was the one who did all the work to make it happen. The role the IHCs initially played was to make the first people out there who were theorizing about the role hydrogen could play in rail transport aware that each other existed, how far they had proceeded, what technologies they were pursuing, and how the best practices might evolve. What was the initial inspiration to have a train that runs on hydrogen instead of diesel? Thompson: I knew that trains were going to run on hydrogen as far back as 1994. I had a paper that I had to edit about the future of power and energy when I was still with Bellsouth. When I came to the way that electricity was going to be transmitted, I realized that major changes would take place not only regarding the grid, but with respect to things that carry energy onboard. What is the present-day status of hydrail, in terms of new projects and installations? Thompson: The most important thing that’s going on is that CSX Railway, one of the largest in the world, is working with the Canadian Pacific–Kansas City Railway to make not only hydrogen-powered locomotives for the two companies’ own use, but to also make diesel-to-hydrogen conversion kits to sell to other major railways. The first manufacturing facility for that, located in Huntington, W.Va., is up and running. Though that is virtually unknown outside the rail industry, I consider it to be big news. Rail giant CSX unveiled this hydrogen-powered locomotive, made in partnership with Canadian Pacific–Kansas City, in early 2024.CSX I’m concerned that the public, particularly in the U.S., is not informed about major transport infrastructure issues. Other examples include the fact that Airbus and nearly all other major aviation companies have begun working on hydrogen-fueled aviation, and the biggest companies in shipping, including Maersk, are heavily involved—and quite far along—in advancing hydrogen’s use in cargo ships. Furthermore, the trucking industry is also far along the road toward making hydrogen available at truck stops. And when hydrogen fueling becomes common at truck stops, that’s when hydrogen automobiles will begin to take market share. That’s the way diesel became widely available. The first Mercedes and Volkswagen Golf diesels, along with a few Cadillacs and Oldsmobiles first relied on truck stops. And when there were enough of these cars around to justify putting in diesel pumps at convenience stores, diesel soon became as widely available as gasoline. But I predict that when this happens with hydrogen, hydrogen-powered vehicles will supplant the plug-in electrics. They will offer much greater range and will be refueled as quickly as vehicles that run on petroleum products. Using hydrogen, whose byproduct is water vapor, for propulsive power instead of diesel, which yields carbon dioxide and nitrogen oxides, is obviously better for the environment. But what advantage does hydrail provide compared with electric power via a third rail or overhead catenary system? Thompson: It’s dramatically cheaper, simply because with hydrail there is no need for wayside power or any of the infrastructure it takes to transmit electricity to the wayside. Let’s look at some basic numbers: One hydrogen refueling station costs about $2 million. That 2 million covers the entire corridor. But an overhead catenary costs that much for a quarter mile [0.4 kilometer] of the track. What were the major inflection points at which progress toward regular hydrail service became evident? Thompson: The first one was the development of the first hydrogen mining locomotive, in Colorado. It used fuel cells so the miners could avoid having to charge batteries. I’d like to think that our decision to hold the first International Hydrail Conference in 2005 was an inflection point, because the people who thought they were working alone discovered that a number of people were working on getting trains to run on hydrogen for similar reasons. By far the biggest inflection point was when Robert Stasko organized the 8th International Hydrail Conference (8IHC) in Toronto. That was where representatives of the Hydrogenics fuel cell company [known as Accelera since its its acquisition by commercial engine manufacturer Cummins] and the Alstom train manufacturing company realized that they had a common beneficial interest and began meeting behind the scenes. The result was the announcement, in 2014, that they were teaming up to build hydrail trains in Salzgitter, Germany. They followed through and introduced the Coradia iLint train in 2016. And there was 3IHC, held in North Carolina, which spurred the creation of a hydrail Ph.D. program at the University of Pisa in Italy, the ancient university where Galileo taught. That conference also led to hydrail’s development in India, which is proceeding apace. Where does hydrail stand at present? Thompson: I know Germany is doing hydrail and Britain has several hydrail projects. The United Kingdom, I think, has the most hydrail manufacturing projects of any country. They have at least five. Italy is experiencing rapid adoption of hydrail and hydrogen propulsion in general. The first narrow-gauge hydrail project is being undertaken in Switzerland. Sweden has freight hydrail trains in development, and hydrail freight trains are also being installed in Eastern Europe. Alstom’s Coradia iLint hydrail train is shown here ferrying rail passengers to their destinations in Germany.Sabrina Adeline Nagel/evb You recently mentioned that Spain has built a high-speed hydrail train. How big a deal is that? Thompson: It’s very important because there’s nothing more technically challenging for hydrogen propulsion that remains to be attempted. Canadian Pacific showed hydrail had overcome its second-most difficult challenge when the company introduced its H2 0EL hydrail freight locomotives in 2022. I‘ve just learned that CRRC in China also introduced a high-speed hydrail train, the Cinova H2, this year. What has been the most surprising development as hydrail has progressed from ideas on paper to steel wheels on rails? Thompson: Most surprising has been the antipathy of the mass media. The continued belief that this is all some quixotic fever dream of mine—even after hydrogen-powered trains have gone into service—is one of the hardest things to accept. Basically, it boils down to the fact that we’re off message. The accepted narrative is that people like Elon Musk and Steve Jobs come up with great ideas. The notion that two old guys sitting in the back of a jewelry store in a small southern town would have the power to change the railroad traction industry globally presents a level of cognitive dissonance many news outlets have yet to overcome. What would you say has been the weirdest development? Thompson: The overhead catenary technology that’s being installed in Charlotte now was invented at the behest of Emperor Alexander II of Russia in the early 1880s for a rail line in St. Petersburg, which was then Russia’s capital. Today, St. Petersburg is implementing the twenty-first century hydrail technology that was developed in the greater Charlotte area.

The annual IEEE STEM Summit in October brought together a record number of preuniversity educators; IEEE volunteers; and science, technology, engineering, and mathematics enthusiasts. They shared resources and ideas and attended sessions to learn how to inspire the next generation of budding engineers and technologists. The free virtual summit, held 23 to 25 October, attracted more than 1,000 attendees from 118 countries. The 13 sessions featured award-winning educators and IEEE volunteers from academia and industry who offered practical advice on planning effective, interesting outreach activities. The sessions highlighted best practices in STEM education, as well as examples of outreach events. Also included was a discussion on productive failure, which is the importance of encouraging young people to embrace the iterative process of engineering—which includes learning from failures. Other sessions introduced participants to TryEngineering resources. TryEngineering, an IEEE Educational Activities program, focuses on STEM outreach. It provides educators with lesson plans, activities, and other resources at no charge for use in their classrooms and in community activities. Students’ interest in STEM careers can be ignited by being introduced to new technologies and the way they operate. Why a STEM summit? The IEEE STEM Summit was organized and hosted by the IEEE Educational Activities’ preuniversity education coordinating committee. Its mission is to foster educational outreach to school-age children around the globe by providing educators and IEEE volunteers with inspiration for engaging activities. The coordinating committee provides resources and services through the TryEngineering program, and it encourages the sharing of ideas and best practices among educators and volunteers. “We are so excited about the continued growth of the IEEE STEM Summit,” says Jamie Moesch, managing director for IEEE Educational Activities. “The event provides this growing community an opportunity to collaborate, strengthen their own outreach efforts, and learn from others who share their passion.”” Engaging speakers Thomas Coughlin, 2024 IEEE president, and Rabab Ward, vice president of IEEE Educational Activities, kicked off the event. Coughlin spoke about opportunities for students in IEEE, and he encouraged participants to get involved in local STEM outreach efforts. Ward stressed the importance of providing students with opportunities to engage in engineering activities and to meet practicing engineers. The summit featured three keynote speakers and several panel sessions. Eleftheria Kallinikou, a psychologist affiliated with the University of Ioannina, in Greece, spoke about productive failure, saying it’s important to teach students how to solve problems as a way to explore engineering, while also increasing their self-esteem. Burt Dicht, a member of the preuniversity coordinating committee and director of membership at the National Space Society, interviewed Barbara Morgan, a former NASA astronaut and an educator at Boise State University, in Idaho. During her From Classroom to Space: A Journey of Education and Exploration session, Morgan talked about her reaction to seeing the Earth from space. She also discussed the role teachers play in inspiring students. Stacy Klein-Gardner, executive director of Engineering 4 Us All, presented An Inclusive Engineering Mindset: K–12 talk, discussing programs that introduce students to engineering as a creative process and one in which students develop their identities as problem solvers. Lorena Garcia, past chair of the IEEE Preuniversity Coordinating Committee (PECC), who is an ABET member and a director of the IEEE Foundation, moderated a global panel of IEEE members. They shared tips for inspiring youngsters with STEM outreach activities such as a bionic bus and several hands-on activities that could be incorporated into afterschool programs. During the Unlocking Potential: How Key Partnerships Fuel Technological Innovation panel session, Michael Geselowitz, senior director of the IEEE History Center, joined Stamatis Dragoumanos, PECC chair, and IEEE staff members in a discussion about how partnerships allow IEEE to amplify its impact. Collaborations with Keysight and Onsemi, for example, have led to new TryEngineering lesson plans. Resources from the History Center’s IEEE REACH program introduce TryEngineering users to the history of technology and its impact on society. Several sessions showcased people who have successfully implemented STEM outreach programs, including a Canadian high school teacher, a Malaysian professor, and a Kenyan university student. To close out the summit, the updated TryEngineering website was unveiled. New features include enhanced search functionality; resources centered around a specific technology such as semiconductors and oceans; and an interactive global map showcasing IEEE STEM outreach activities. Lesson plans, news, and spotlights of TryEngineering STEM grants and STEM Champions are now easier to find on the website. Visit TryEngineering.org to find resources to assist you in your work to inspire the next generation of engineers. Previous years’ IEEE STEM Summit sessions can be viewed on the IEEE TryEngineering YouTube channel. The IEEE Foundation, the philanthropic partner for TryEngineering, provided financial support for the summit. To support future summits and the entire TryEngineering program, visit the IEEE TryEngineering Fund donation page.

This is a sponsored article brought to you by Amazon. Innovation often begins as a spark of an idea—a simple “what if” that grows into something transformative. But turning that spark into a fully realized solution requires more than just ingenuity. It requires resources, collaboration, and a relentless drive to bridge the gap between concept and execution. At Amazon, these ingredients come together to create breakthroughs that not only solve today’s challenges but set the stage for the future. “Innovation doesn’t just happen because you have a good idea,” said Valerie Samzun, a leader in Amazon’s Fulfillment Technologies and Robotics (FTR) division. “It happens because you have the right team, the right resources, and the right environment to bring that idea to life.” This philosophy underpins Amazon’s approach to robotics, exemplified by Robin, a groundbreaking robotic system designed to tackle some of the most complex logistical challenges in the world. Robin’s journey, from its inception to deployment in fulfillment centers worldwide, offers a compelling look at how Amazon fosters innovation at scale. Building for Real-World Complexity Amazon’s fulfillment centers handle millions of items daily, each destined for a customer expecting precision and speed. The scale and complexity of these operations are unparalleled. Items vary widely in size, shape, and weight, creating an unpredictable and dynamic environment where traditional robotic systems often falter. “Robots are great at consistency,” Jason Messinger, robotics senior manager explained. “But what happens when every task is different? That’s the reality of our fulfillment centers. Robin had to be more than precise—it had to be adaptable.” Robin was designed to pick and sort items with speed and accuracy, but its capabilities extend far beyond basic functionality. The system integrates cutting-edge technologies in artificial intelligence, computer vision, and mechanical engineering to learn from its environment and improve over time. This ability to adapt was crucial for operating in fulfillment centers, where no two tasks are ever quite the same. “When we designed Robin, we weren’t building for perfection in a lab,” Messinger said. “We were building for the chaos of the real world. That’s what makes it such an exciting challenge.” The Collaborative Process of Innovation Robin’s development was a collaborative effort involving teams of roboticists, data scientists, mechanical engineers, and operations specialists. This multidisciplinary approach allowed the team to address every aspect of Robin’s performance, from the algorithms powering its decision-making to the durability of its mechanical components. “Robin is more than a robot. It’s a learning system. Every pick makes it smarter, faster, and better.” —Valerie Samzun, Amazon “At Amazon, you don’t work in silos,” both Messinger and Samzun noted. Samzun continued, “every problem is tackled from multiple angles, with input from people who understand the technology, the operations, and the end user. That’s how you create something that truly works.” This collaboration extended to testing and deployment. Robin was not confined to a controlled environment but was tested in live settings that replicated the conditions of Amazon’s fulfillment centers. Engineers could see Robin in action, gather real-time data, and refine the system iteratively. “Every deployment teaches us something,” Messinger said. “Robin didn’t just evolve on paper—it evolved in the field. That’s the power of having the resources and infrastructure to test at scale.” Why Engineers Choose Amazon For many of the engineers and researchers involved in Robin’s development, the opportunity to work at Amazon represented a significant shift from their previous experiences. Unlike academic settings, where projects often remain theoretical, or smaller companies, where resources may be limited, Amazon offers the scale, speed, and impact that few other organizations can match. Learn more about becoming part of Amazon’s Team → “One of the things that drew me to Amazon was the chance to see my work in action,” said Megan Mitchell, who leads a team of manipulation hardware and systems engineers for Amazon Robotics. “Working in R&D, I spent years exploring novel concepts, but usually didn’t get to see those translate to the real world. At Amazon, I get to take ideas to the field in a matter of months.” This sense of purpose is a recurring theme among Amazon’s engineers. The company’s focus on creating solutions that have a tangible impact—on operations, customers, and the industry as a whole—resonates with those who want their work to matter. “At Amazon, you’re not just building technology—you’re building the future,” Mitchell said. “That’s an incredibly powerful motivator. You know that what you’re doing isn’t just theoretical—it’s making a difference.” In addition to the impact of their work, engineers at Amazon benefit from access to unparalleled resources. From state-of-the-art facilities to vast amounts of real-world data, Amazon provides the tools necessary to tackle even the most complex challenges. “If you need something to make the project better, Amazon makes it happen. That’s a game-changer,” said Messinger. The culture of collaboration and iteration is another draw. Engineers at Amazon are encouraged to take risks, experiment, and learn from failure. This iterative approach not only accelerates innovation but also creates an environment where creativity thrives. During its development, Robin was not confined to a controlled environment but was tested in live settings that replicated the conditions of Amazon’s fulfillment centers. Engineers could see Robin in action, gather real-time data, and refine the system iteratively.Amazon Robin’s Impact on Operations and Safety Since its deployment, Robin has revolutionized operations in Amazon’s fulfillment centers. The robot has performed billions of picks, demonstrating reliability, adaptability, and efficiency. Each item it handles provides valuable data, allowing the system to continuously improve. “Robin is more than a robot,” Samzun said. “It’s a learning system. Every pick makes it smarter, faster, and better.” Robin’s impact extends beyond efficiency. By taking over repetitive and physically demanding tasks, the system has improved safety for Amazon’s associates. This has been a key priority for Amazon, which is committed to creating a safe and supportive environment for its workforce. “When Robin picks an item, it’s not just about speed or accuracy,” Samzun explained. “It’s about making the workplace safer and the workflow smoother. That’s a win for everyone.” A Broader Vision for Robotics Robin’s success is just the beginning. The lessons learned from its development are shaping the future of robotics at Amazon, paving the way for even more advanced systems. These innovations will not only enhance operations but also set new standards for what robotics can achieve. “At Amazon, you feel like you’re a part of something bigger. You’re not just solving problems—you’re creating solutions that matter.” —Jason Messinger, Amazon “This isn’t just about one robot,” Mitchell said. “It’s about building a platform for continuous innovation. Robin showed us what’s possible, and now we’re looking at how to go even further.” For the engineers and researchers involved, Robin’s journey has been transformative. It has provided an opportunity to work on cutting-edge technology, solve complex problems, and make a meaningful impact—all while being part of a team that values creativity and collaboration. “At Amazon, you feel like you’re a part of something bigger,” said Messinger. “You’re not just solving problems—you’re creating solutions that matter.” The Future of Innovation Robin’s story is a testament to the power of ambition, collaboration, and execution. It demonstrates that with the right resources and mindset, even the most complex challenges can be overcome. But more than that, it highlights the unique role Amazon plays in shaping the future of robotics and logistics. “Innovation isn’t just about having a big idea,” Samzun said. “It’s about turning that idea into something real, something that works, and something that makes a difference. That’s what Robin represents, and that’s what we do every day at Amazon.” Robin isn’t just a robot—it’s a symbol of what’s possible when brilliant minds come together to solve real-world problems. As Amazon continues to push the boundaries of what robotics can achieve, Robin’s legacy will be felt in every pick, every delivery, and every step toward a more efficient and connected future. Learn more about becoming part of Amazon’s Team.

The modern race to build undetectable submarines dates from the 1960s. In that decade, the United States and the Soviet Union began a game of maritime hide-and-seek, deploying ever-quieter submarines as well as more advanced tracking and detection capabilities to spot their adversary’s vessels. That game continues to this day but with a wider field of players. In the coming months, the U.S. Navy plans to homeport the USS Minnesota on Guam. This Virginia-class nuclear-powered attack submarine is among the quietest subs ever made. Advanced nuclear propulsion like the Minnesota’s gives the vessel a superior ability to operate covertly. More of its kind will be deployed by the United States, the United Kingdom, and Australia to compete with China for influence and military dominance, especially over the Indo-Pacific region. This article is a collaboration between Foreign Policy, the global magazine of politics and ideas, and IEEE Spectrum, the flagship magazine of the IEEE. As part of the landmark deal known as AUKUS (for the initials of its partner states), Australia will acquire, operate, and maintain three to five U.S. Virginia-class subs, each of which will cost about US $4.3 billion; an additional five subs will be a special AUKUS-class built in the U.K. and Australia using U.S. nuclear propulsion technology. In exchange for access to this technological edge, Australia has agreed to make substantial multibillion-dollar investments in the U.S. and U.K. naval shipbuilding industries. The deal could last until at least the 2050s and cost up to $368 billion. These submarines are expected to assume a nuclear deterrence mission against China, whose nuclear modernization plans include the deployment of submarine-launched ballistic missiles capable of targeting the United States. The People’s Liberation Army Navy is the largest navy in the world, but it currently operates only 12 nuclear-powered submarines, a rather small number compared to the 67 attack subs and ballistic-missile subs of the U.S. Navy. And compared to U.S. submarines, Chinese boats are noisy and easily detected. But it won’t stay that way for long. The U.S. Department of Defense claims China plans to modernize and expand its submarine forces significantly by 2035, including more stealthy submarines. Once built, Australia’s first few nuclear subs will operate for 33 years, until the 2060s, or even longer with lifetime extensions. To shore up its intended strategic advantages, the AUKUS deal also seeks to develop advanced antisub technology, consisting of sensor networks and analytics enabled by artificial intelligence (AI). This technology cuts both ways, though, and ocean transparency is increasing as a result. Some experts even think the game of maritime hide-and-seek could end by 2050. Meanwhile, AUKUS faces more practical concerns, including a looming shortage of the highly enriched uranium needed to fuel the submarines, growing opposition to the deal’s extravagant cost, and competing submarine designs that are much cheaper and just as capable for certain missions. So, is now really the right time for nations to be investing hundreds of billions of dollars in submarine stealth? What is submarine stealth? In the quest for stealth, naval engineers first have to consider how their vessel might be spotted. Then they can design their submarines for maximum evasion. There are two key steps to track a submarine, says Scott Minium, a former commander at Submarine Squadron 15 in Guam who has mentored the commanding officers of seven nuclear-powered subs. The first step, Minium says, is to detect the signature of a potential submarine. The second step is to “classify it based on known signatures to determine if a submarine has been detected.” Such signatures include the unique noise patterns generated by different submarine classes as well as other identifiers, and they’re essential for detecting and tracking submarines. The growing sophistication of stealth-busting tech casts doubt on continued investment in advanced submarines, each of which costs over $4 billion. Shown here are segments of a sub’s hull. Christopher Payne/Esto Noise is the most critical signature, and so engineers working on stealth technology focus on suppressing the sound waves that submarines give off, rendering their movements nearly silent, especially at slow speeds. The thousands of rubberized anechoic tiles that cover the hull of a Virginia-class submarine absorb or distort sound waves detectable by passive and active sonar, obscuring the sub’s whereabouts. Similarly, vibration-damping materials reduce the sounds that the engines and turbines transmit to the surrounding waters. Submarines have long been designed with certain geometric shapes that minimize their radar cross-section—that is, the areas seen by the radar that enable it to be detected. The addition of radar-absorbing materials on exposed parts of a submarine, such as the periscopes and antenna, also helps, allowing those parts to absorb rather than reflect radar waves. In recent years, submarine designers have also worked to decrease the vessels’ signatures associated with temperature, magnetic fields, and wake patterns. Heat exchangers and cooling systems, for example, reduce the heat generated by submarines, making thermal imaging and infrared detection by commercial satellites more difficult. To remove residual magnetic fields, demagnetization or “degaussing” procedures involve driving the submarine between parallel piers and wrapping it with high-voltage cables. While that process sounds elaborate, it’s increasingly necessary: Tracing magnetic signatures via underwater surveillance networks has emerged as a new way to detect submarines. Additional advances in submarine stealth may be possible, but they are cost- and industrial-base prohibitive. Finally, using pump-jet propulsors, Virginia-class submarines produce less turbulence in the water, making them less detectable by their wakes. Although conventional screw propellers are simpler and cheaper, pump-jet propulsors offer greater speed and agility, better efficiency at high speeds, and less noise. Despite these innovations, Bryan Clark, a leading naval expert at the Hudson Institute, warns about “an inflection point for achieving further reductions in sound and other signals due to the challenges of physics and mechanical systems.” Additional advances may be possible, he says, but they are “cost and industrial-base prohibitive.” Meanwhile, significant advances in detection technologies have reduced the effectiveness of submarine stealth. Today, increasingly sophisticated and distributed sensor networks collect information across multiple domains, much like the SOSUS hydrophone arrays that the U.S. Navy deployed in the Atlantic and Pacific during the Cold War. The rise of quantum sensors, which can detect delicate perturbations in the environment at the atomic level, promises even greater sensitivity and accuracy. And the AI-enabled systems that analyze sensor data can easily spot subtle anomalies in the ocean, such as changes caused by a passing submarine, which a human analyst would probably miss. P.W. Singer, a senior fellow at the think tank New America and coauthor of the technothriller Ghost Fleet—in which Russia and China team up against the United States with a new capability to detect and track U.S. nuclear submarines from their radiation emissions—suggests that AI’s “ability to make sense of disparate wisps of data from a variety of sensors…will enable the detection of targets that could have remained stealthy in the past.” Other experts, including Roger Bradbury and Scott Bainbridge, claim this technological revolution has already produced unprecedented ocean transparency. If the most extreme predictions come true, the stealth of Australia’s new fleet of nuclear submarines could be dead in the water less than a decade into their operational lifetimes. Advanced tactics to preserve submarine stealth Many experts say they’re unconcerned about these incursions on submarine stealth. Naval operators, they claim, still have plenty of ways to protect the stealth of their submarines. These stealth-preserving techniques include 1) countering detection through noise, 2) deploying more underwater drones, and 3) using strategic moves to counter the objectives of the adversary. The first strategy uses noise as a feature, not a bug. Instead of going quieter, Minium suggests, naval operators could try “making more noise or finding innovative ways to change the acoustic signatures of submarines.” For example, he says, “We could make active sonar waves of submarines sound identical to whales.” This idea exploits the current limitations of AI systems and the ease with which unexpected shifts in the data can trick them. Slight tweaks in a submarine’s signature might be enough to confuse an AI algorithm so that it misidentifies the vessel or misses it entirely. Minium says this approach relies on the fact that “you need to know what you’re looking for to leverage AI for finding submarines. If you can’t classify the detected signature, the submarine is safe from detection.” Australia will base its AUKUS submarines at HMAS Stirling, a naval base near Perth. But the U.S. Navy would prefer to base the submarines in Guam, because it’s closer to China’s naval base on Hainan Island. In addition to masking submarine signatures, navies could make greater use of inexpensive underwater drones, or uncrewed underwater vehicles. As Clark explains, UUVs are part of the move away from the traditional game of hide-and-seek to “a competition of sensing and sense-making.” This shift is aided by the sharp increase in civilian UUV traffic, for deploying fiber-optic cables and conducting scientific research. All that activity generates more underwater noise and makes it harder to detect individual signatures. Military UUVs, he says, can likewise create “more noise elsewhere, allowing submarine signals to go undetected.” Speculating about the future of undersea warfare, Singer says the rise of smaller and cheaper uncrewed systems will allow these “disposable sensors [to] also become killers if armed.” Their disposability would enable countries to use them more aggressively, enter contested spaces, and “mess with the data” collected by sensor networks. “By flooding the zone with false signatures,” Singer says, “navies can expose the hunters who chase the false targets and possibly even waste away the adversary’s expensive weapons systems.” Interestingly, the most recent Virginia-class submarines have been upgraded with the capability to deploy UUVs. According to the Congressional Research Service, this upgrade adds a substantial midsection containing four launch tubes “for storing and launching additional Tomahawk missiles or other payloads.” However, Clark and Hudson Institute senior fellow Timothy Walton caution against using precious payload space for UUVs. They instead recommend that the submarines carry much smaller, disposable UUVs “that can be carried in external countermeasure launchers or lockers inside the submarine.” It’s conceivable, too, that as the game of hide-and-seek becomes more difficult for everyone, navies may take offensive measures to protect the stealth of their submarines. This could entail less overt tactics for peacetime and more aggressive operations in a crisis. Clark gives an example: “A boat could drag its anchor along the seabed to destroy transmission cables and still maintain plausible deniability” by making it look like an accident. The boat could then “monitor the ships and UUVs that arrive to perform infrastructure repairs, gathering vital intelligence about the adversary.” “AI’s ability to make sense of disparate wisps of data from a variety of sensors…will enable the detection of targets that could have remained stealthy in the past.” A more subtle option, Singer says, exploits the fact that countries can’t afford to deploy their undersea surveillance networks everywhere. Instead, they’re creating “windows of coverage and non-coverage”—for example, focusing on choke points in shallow waters where submarines are more easily detected. Other countries could then “target [those] key nodes in the sensor network with cyberattacks, disrupting operation and allowing for covert passage.” To gain further advantage in a conflict, Singer adds, countries could “assume control of a network while still making it appear fully operational and deliver false signals to the adversary.” Referred to as spoofing, this tactic involves disguising a fake data source as legitimate. GPS spoofing has become a major challenge on the high seas. One high-profile incident in 2021 involved the faking of British warship positions by an unknown actor. In other situations, Singer says, an adversary might decide to simply “destroy the sensors and surveillance platforms.” The AI-enabled systems for processing and analyzing massive volumes of data can also become a target. Data poisoning, for example, involves covertly contaminating the data used to train an AI algorithm, which would lead to false results. Of course, to engineer such an attack, Clark says, an adversary would probably need physical access to get around firewalled systems. Another route for data poisoning would be to “use radiofrequency transmissions to attack a network and insert bad data at the source.” Opposition to the AUKUS deal The AUKUS submarine deal represents a targeted strategy to blunt China’s influence in the Indo-Pacific region and upset any plans for attacking Taiwan. Jamie Kwong, a fellow at the Carnegie Endowment for International Peace, suggests that the AUKUS subs will be able to “hold China’s nuclear-armed ballistic missile submarines (SSBNs) at risk.” Chinese officials, for their part, have repeatedly criticized AUKUS, warning that the security pact will increase regional tensions. China has a ways to go to catch up with the West, says Yanliang Pan, a research associate at the James Martin Center for Nonproliferation Studies, in Monterey, Calif. “But it seems they’re well on their way.” That’s unsurprising, given the long lead times for building nuclear submarines. According to publicly available reports, Pan says, China’s plans include “a rapid expansion in its sea-based capabilities with a nuclear-powered carrier fleet and a new prototype nuclear reactor that will be outfitted in its new [nuclear attack and ballistic-missile submarines].” Current projections suggest China may soon overtake its adversaries in the total number of advanced submarines and come closer in terms of stealth. According to military experts, the new Chinese submarines’ designs have benefited from Russian propulsion expertise, and will be much quieter, making it harder for the U.S. Navy to detect and track them. The USS Vermont Virginia-class submarine undergoes sea trials in 2023. General Dynamics Electric Boat Moreover, China’s overall shipbuilding capabilities and pace of construction far exceed those of the United States, which currently produces an average of 1.2 nuclear-powered boats a year at the Navy’s two submarine shipyards. To fulfill the terms of the AUKUS deal, the United States needs to boost the pace of production to at least two per year. Already, U.S. capacity to implement the first pillar of AUKUS, which involves providing Australia with Virginia-class nuclear attack submarines, hangs in the balance. The U.S. Navy included the procurement of only one Virginia-class submarine in its budget request for fiscal year 2025, although the U.S. House of Representatives later advanced a defense spending bill that restored the number to two. In the immediate aftermath of the U.S. presidential election, it remains unclear how defense funding politics will play out. But it seems unlikely that AUKUS members will be able to outcompete China on nuclear-powered submarine production. Deploying more advanced submarines won’t be enough in any event. The United States, U.K., and Australia will also need to anticipate how China might disrupt their desired outcomes. AUKUS members may decide to counter China’s strategy by investing in more asymmetric means for conducting antisubmarine warfare. Presumably this is the rationale behind the second pillar of AUKUS, which explores deepening collaboration on emerging technologies such as artificial intelligence, quantum computing, cyber capabilities, and hypersonic weapons. It also takes advantage of China’s delayed start in developing advanced sensing capabilities. Using such technologies, AUKUS members could, for example, exploit weaknesses in China’s shallow seas and choke points surrounding its shores. The United States and its allies could also counter Chinese submarines’ ability to reach deeper waters undetected by deploying quantum-based sensors, jamming, UUV detection, and AI-enabled analytics. If the most extreme predictions come true, the stealth of Australia’s new fleet of nuclear submarines could be dead in the water less than a decade into their operational lifetimes. However, if they’re leveraging emerging technologies to detect China’s submarines, will AUKUS members even need the exquisitely advanced submarines from the United States? George M. Moore, scientist-in-residence at the James Martin Center for Nonproliferation Studies, notes that the Virginia-class submarines “do not seem optimized for the shallow waters of the South China Sea. Australia might have been far better off building more conventional diesel submarines, which are quieter than nuclear-powered submarines when running on battery.” Nuclear-powered submarines can stay underwater longer than diesel subs can, so they are considered the stealthier option, as the chances of detection increase every time a submarine surfaces. But, Moore says, submarines that use a newer nonnuclear propulsion, known as air-independent propulsion (AIP), “pretty much eliminate that advantage with their capability to stay submerged for up to 30 to 40 days.” Unlike conventional diesel submarines, AIP subs operate on battery for long periods, do not require regular access to oxygen, and do not need to surface or use a snorkel as frequently. Going with AIP submarines rather than Virginia-class nuclear subs would save several billion dollars per vessel. That might offer Australia a more viable alternative for covering the shorter distances in the South China and East China seas, with the other two AUKUS members tracking Chinese submarines in deeper waters. Moore also has reservations about the nuclear deterrence mission of the AUKUS deal. To execute that mission, an AUKUS submarine would need to trail any Chinese ballistic-missile submarine coming out of port before it goes silent. “But we just don’t have the numbers to do this anymore,” he says. Is AUKUS a good deal? Ultimately, the future of AUKUS may hinge on more practical matters than any perceived decline in submarine stealth. In the near term, the Australian government must refurbish its HMAS Stirling submarine base in Western Australia, to allow for the rotational deployment of five U.S. and U.K. nuclear attack submarines. That will cost about AU $8 billion. But the plan may face difficulty due to growing domestic skepticism about the deal and its enormous expense. The plan may also face opposition within the United States. The naval base in Western Australia is further from the South China Sea than Guam is, which the United States favors for its submarine operations, Moore says. Guam is also closer to China’s submarine base on Hainan Island. Moreover, there’s a declining stockpile of the highly enriched uranium (HEU) that Australia’s new subs will use for fuel. For many years now, U.S. nuclear-powered submarines “have run on the HEU scavenged from old nuclear weapons,” Moore says. Under AUKUS, this limited fuel stock would presumably be shared by the United States, U.K., and Australia. Building a new enrichment facility, he says, could take up to 40 years. Then there’s the issue of Australia accepting HEU for its new nuclear-powered submarine fleet. Under AUKUS, Australia will become the first nonnuclear-weapon state to operate submarines with weapons-grade material. However, Kwong of the Carnegie Endowment for International Peace notes that Australia doesn’t have a nuclear-energy industry, and so “is unprepared for handling spent fuel.” Indeed, since 1998, Australian federal legislation has banned the development of nuclear power, including a prohibition against nuclear-fuel-related facilities. Whatever happens to AUKUS, advances in AI, drones, and sensing technologies are rapidly changing the dynamics of undersea warfare, which will force many nations to rethink their submarine strategies and investments. As the game of hide-and-seek gives way, new strategies may hinge more on asymmetric innovations than on submarine numbers and stealth—regardless of how sophisticated those submarines are. This article is a collaboration between Foreign Policy, the global magazine of politics and ideas, and IEEE Spectrum. A correction to this article was made on 19 December 2024 to fix an editing error. Passive sonar does not emit sound waves. This article appears in the December 2024 print issue as “No More Hide-and-Seek.”

U.S. presidential administrations tend to have big impacts on tech around the world. So it should be taken as a given that when Donald Trump returns to the White House in January, his second administration will do the same. Perhaps more than usual, even, as he staffs his cabinet with people closely linked to the Heritage Foundation, the Washington, D.C.–based conservative think tank behind the controversial 900-page Mandate for Leadership (also known as Project 2025). The incoming administration will affect far more than technology and engineering, of course, but here at IEEE Spectrum, we’ve dug into how Trump’s second term is likely to impact those sectors. Read on to find out more, or click to navigate to a specific topic. This post will be updated as more information comes in. Artificial Intelligence Consumer Electronics Cryptocurrencies Cybersecurity Energy Robotics Semiconductors Telecommunications Transportation Artificial Intelligence During Trump’s campaign, he vowed to rescind President Joe Biden’s 2023 executive order on AI, saying in his platform that it “hinders AI Innovation, and imposes Radical Leftwing ideas on the development of this technology.” Experts expect him to follow through on that promise, potentially killing momentum on many regulatory fronts, such as dealing with AI-generated misinformation and protecting people from algorithmic discrimination. However, some of the executive order’s work has already been done; rescinding it wouldn’t unwrite reports or roll back decisions made by various cabinet secretaries, such as the Commerce secretary’s establishment of an AI Safety Institute. While Trump could order his new Commerce secretary to shut down the institute, some experts think it has enough bipartisan support to survive. “It [helps develop] standards and processes that promote trust and safety—that’s important for corporate users of AI systems, not just for the public,” says Doug Calidas, senior vice president of government affairs for the advocacy group Americans for Responsible Innovation. As for new initiatives, Trump is expected to encourage the use of AI for national security. It’s also likely that, in the name of keeping ahead of China, he’ll expand export restrictions relating to AI technology. Currently, U.S. semiconductor companies can’t sell their most advanced chips to Chinese firms, but that rule contains a gaping loophole: Chinese companies need only sign up for U.S.-based cloud computing services to get their AI computations done on state-of-the-art hardware. Trump may close this loophole with restrictions on Chinese companies’ use of cloud computing. He could even expand export controls to restrict Chinese firms’ access to foundation models’ weights—the numerical parameters that define how a machine learning model does its job. —Eliza Strickland Back to top Consumer Electronics Trump plans to implement hefty tariffs on imported goods, including a 60 percent tariff on goods from China, 25 percent on those from Canada and Mexico, and a blanket 10 or 20 percent tariff on all other imports. He’s pledged to do this on day 1 of his administration, and once implemented, these tariffs would hike prices on many consumer electronics. According to a report published by the Consumer Technology Association in late October, the tariffs could induce a 45 percent increase in the consumer price of laptops and tablets, as well as a 40 percent increase for video-game consoles, 31 percent for monitors, and 26 percent for smartphones. Collectively, U.S. purchasing power for consumer technology could drop by US $90 billion annually, the report projects. Tariffs imposed during the first Trump administration have continued under Biden. Meanwhile, the Trump Administration may take a less aggressive stance on regulating Big Tech. Under Biden, the Federal Trade Commission has sued Amazon for maintaining monopoly power and Meta for antitrust violations, and worked to block mergers and acquisitions by Big Tech companies. Trump is expected to replace the current FTC chair Lina Khan, though it remains unclear how much the new administration—which bills itself as antiregulation—will affect the scrutiny Big Tech is facing. Executives from major companies including Alphabet, Amazon, Apple, Intel, Meta, Microsoft, OpenAI, and Qualcomm congratulated Trump on his election on social media, primarily X. (The CTA also issued congratulations.) —Gwendolyn Rak Back to top Cryptocurrencies On 6 November, the day the election was called for Trump, Bitcoin jumped 9.5 percent, closing at over $75,000—a sign that the cryptocurrency world expects to boom under the next regime. Donald Trump marketed himself as a procrypto candidate, vowing to turn America into the “crypto capital of the planet” at a Bitcoin conference in July. If he follows through on his promises, Trump could create a national bitcoin reserve by holding on to bitcoin seized by the U.S. government. Trump also promised to remove Gary Gensler, the chair of the Securities and Exchanges Commission, who has pushed to regulate most cryptocurrencies as securities (like stocks and bonds), with more government scrutiny. While it may not be within Trump’s power to remove him, Gensler is likely to resign when a new administration starts. It is within Trump’s power to select the new SEC chair, who will likely be much more lenient on cryptocurrencies. The evidence lies in Trump’s procrypto cabinet nominations: Howard Lutnick as Commerce Secretary, whose finance company oversees the assets of the Tether stablecoin; Robert F. Kennedy Jr. as the Secretary of Health and Human Services, who has said in a post that “Bitcoin is the currency of freedom”; and Tulsi Gabbard for the Director of National Intelligence, who had holdings in two cryptocurrencies back in 2017. As Trump put it at that Bitcoin conference, “The rules will be written by people who love your industry, not hate your industry.” —Kohava Mendelsohn Back to top Cybersecurity Trump’s campaign has been light on specific technological policy plans, and cybersecurity appears to be caught between two competing considerations. On the one hand, Trump’s stance on international cyber warfare during his first administration was hawkish: “If we ever get hit, we’ll hit very hard,” he said during an interview in 2020. He also claimed that the U.S. is “better at cyber than anyone in the world.” On the other hand, Trump’s emphasis on deregulation may result in imposing fewer cybersecurity requirements, especially for private companies, relying instead on an opt-in approach. Specifically, the U.S. Department of Homeland Security’s Cybersecurity and Infrastructure Security Agency (CISA) faces an uncertain fate. Trump signed the agency into law in 2018 as an expanded successor to the National Protection and Programs Directorate. However, in 2020, Trump fired CISA’s director, Christopher Krebs, for creating a Rumor Control blog dedicated to combating election fraud disinformation. Now, Senator Rand Paul (R-Ky.) is slated to chair the Senate Homeland Security and Governmental Affairs Committee. Paul has reportedly said he’d like to eliminate CISA entirely, claiming that its disinformation efforts amounted to censorship in violation of first amendment rights. “I would have liked to, at the very least, eliminate their ability to censor content online,” Rand said. CISA serves many vital functions, however, mainly safeguarding national infrastructure from cyberattacks in energy, healthcare, transportation, and finance, as well as coordinating responses to specific cyberattacks. The core part of CISA’s mission still has strong bipartisan support, so it is unlikely to be dismantled entirely. However, slashed funding or a re-distribution of some of CISA’s activities to other agencies are both possible. For example, Project 2025, a policy document written by the Heritage Foundation and the authors of which include several former and future Trump appointees, recommends moving CISA under the U.S. Department of Transportation. Trump explicitly disavowed Project 2025 on the campaign trail. —Dina Genkina Back to top Energy Trump’s plans for the energy sector focus on establishing U.S. “energy dominance,” mainly by boosting domestic oil and gas production, and deregulating those sectors. To that end, he has selected oil services executive Chris Wright to lead the U.S. Department of Energy. “Starting on day 1, I will approve new drilling, new pipelines, new refineries, new power plants, new reactors, and we will slash the red tape,” Trump said in a campaign speech in Michigan in August. Trump’s stance on nuclear power, however, is less clear. His first administration provided billions in loan guarantees for the construction of the newest Vogtle reactors in Georgia. But in an October interview with podcaster Joe Rogan, Trump said that large-scale nuclear builds like Vogtle “get too big, and too complex, and too expensive.” Trump periodically shows support for the development of advanced nuclear technologies, particularly small modular reactors (SMRs). As for renewables, Trump plans to “terminate” federal incentives for them. He vowed to gut the Inflation Reduction Act, a signature law from the Biden Administration that invests in electric vehicles, batteries, solar and wind power, clean hydrogen, and other clean energy and climate sectors. Trump trumpets a particular distaste for offshore wind, which he claims will end “on day 1” of his next presidency. The first time Trump ran for president, he vowed to preserve the coal industry, but this time around, he rarely mentioned it. Coal-fired electricity generation has steadily declined since 2008, despite Trump’s first-term appointment of a former coal lobbyist to lead the Environmental Protection Agency. For his next EPA head, Trump has nominated former New York Representative Lee Zeldin—a play expected to be central to Trump’s campaign pledges for swift deregulation. —Emily Waltz Back to top Robotics One of the incoming administration’s priorities has been the promise to “put the interests of American workers first” by targeting immigration. However, there are currently far more job openings than humans (American or otherwise) can fill, primarily lower skilled jobs at lower rates of pay than American workers are typically willing to accept. This is especially true in agriculture. If immigration reform further lowers the pool of available labor, either prices for goods and services will rise, or something else will need to fill this gap. Robots are often put forward as a solution to labor shortages, and the incoming administration is no exception. But with some exceptions (typically highly structured brownfield environments), robots are still very much a work in progress. Cheap human labor will not be completely replaced by robots within the next four years, or even within the next decade. Progress towards robots that can make a tangible difference in the agricultural labor market will require research funding, and the Biden administration prioritized robotics R&D as an “Industry of the Future” (IOTF). Project 2025 suggests that “Congress should encourage the establishment of an industry consortium of agricultural equipment producers and other automation and robotics firms” by providing matching funding in exchange for intellectual property concessions. But much of agricultural robotics still requires funding in basic research like sensing and manipulation to enable impactful scaling. To what extent the incoming administration will fund robotics research that doesn’t have immediate commercial applications is unclear. —Evan Ackerman Back to top Semiconductors The Biden-Harris administration’s signature achievement in semiconductors was the 2022 CHIPS and Science Act, which promised to revitalize chipmaking in the United States. When the bill was passed, there was no leading-edge manufacturing done in the country. In its early years, the new administration will enjoy a very different environment with at least two leading-edge fabs, one from Intel and one from TSMC, scheduled for operation. Nevertheless, the Biden administration is concerned about the Act’s implementation under Trump, so it is rushing to get as much done as possible before inauguration day. “I’d like to have really almost all of the money obligated by the time we leave,” Commerce Secretary Gina Raimondo told Politico in mid-November, adding that the CHIPS office had been working seven-day weeks toward that goal. Prior to the election, only $123 million was committed. But between the election and Thanksgiving it pumped out nearly $16 billion more, including $6.6 billion to TSMC and $7.8 billion to Intel to help build those advanced fabs. Critics worry that this haste is coming at that expense of workforce goals, tax-payer guardrails, and environmental review. Even if the incoming administration were interested in pulling back on this manufacturing boon, it would be difficult, says one Washington expert. Unlike many other programs, that part of the CHIPS Act has a five-year appropriation, so Congress would have to act to specifically defund it. And with manufacturing funds in negotiation for projects in at least 20 states, such a move could be politically costly to Congressional Republicans. The Biden-Harris administration has also been busy on the “and Science” part of the Act—deciding the sites for two of three R&D centers as well as promising billions for packaging R&D, semiconductor workforce development programs, and defense-related research. Harish Krishnaswamy, a managing director at Sivers Semiconductors and Columbia University wireless expert, is part of two projects funded by the latter program. With the initial contracts already signed earlier this month, he isn’t worried about the funding through the project’s first year. “I think where there’s uncertainty is the extension to year two and year three,” he says. As for semiconductor R&D coming from the National Science Foundation, it’s not a likely target in the short term, says Russell Harrison, managing director of IEEE-USA. And it can always fall victim to general budget cutting. “Research is an easy thing to cut out of a budget, politically. In the first year there is little broad impact. It’s in the 10th year that you have a big problem, but nobody is thinking that long term.” —Samuel K. Moore Back to top Telecommunication The telecom priorities of the incoming administration are marked already by a conspicuous social media footprint—for instance with the expected new FCC administrator Brendan Carr threatening social media and other tech companies that exercise too much of what he calls “Orwellian” fact-checking. Beneath this bluster, however, lies a number of significant telecom policy shifts that could have outsized implications for the industry. Chief among them is the FCC’s authority to auction wireless spectrum—which expired in March 2023 because of disagreements over future spectrum auctions and the national security and financial potentials of such auctions. Republican senators Ted Cruz and John Thune (the latter being the likely next Senate Majority Leader) have announced their prioritization of this renewal early in the next administration. Cruz and Thune also appear poised to revisit another key element of telecom regulation in the year ahead: reforming the Universal Service Fund. As part of the 1996 Telecommunications Act, the Fund was established to support telecom priorities among low-income households, rural health care providers, and educational institutions. The focus appears to be where the USF derives its funding—a shift that could also affect the Affordable Connectivity Program, which dates from the early days of the COVID-19 pandemic. There is also an expected shift of the Broadband Equity, Access and Deployment (BEAD) Program. At the moment, BEAD prioritizes fiber-optic networks as the means to expand consumer access to high-speed internet. Given the outsized role of SpaceX CEO Elon Musk in the incoming administration, BEAD is expected to underwrite more satellite Internet connections to underserved communities. And to return to the cudgel that Carr threatened Meta, Google, and others with in his recent tweets: Section 230 of the Communications Decency Act of 1996. Section 230—famously called “the 26 words that made the Internet”—provides some needed legal immunity for websites that seek to have open forums and thus third-party speech. The incoming Trump administration is expected to scale back Section 230 protections. —Margo Anderson Back to top Transportation The incoming administration hasn’t laid out too many specifics about transportation yet, but Project 2025 has lots to say on the subject. It recommends the elimination of federal transit funding, including programs administered by the Federal Transit Administration (FTA). This would severely impact local transit systems—for instance, the Metropolitan Transportation Authority in New York City could lose nearly 20 percent of its capital funding, potentially leading to fare hikes, service cuts, and project delays. Kevin DeGood, Director of Infrastructure Policy at the Center for American Progress, warns that “taking away capital or operational subsidies to transit providers would very quickly begin to result in systems breaking down and becoming unreliable.” DeGood also highlights the risk to the FTA’s Capital Investment Grants, which fund transit expansion projects such as rail and bus rapid transit. Without this support, transit systems would struggle to meet the needs of a growing population. Project 2025 also proposes spinning off certain Federal Aviation Administration functions into a government-sponsored corporation. DeGood acknowledges that privatization can be effective if well structured, and he cautions against assuming that privatization inherently leads to weaker oversight. “It’s wrong to assume that government control means strong oversight and privatization means lax oversight,” he says. Project 2025’s deregulatory agenda also includes rescinding federal fuel-economy standards and halting initiatives like Vision Zero, which aims to reduce traffic fatalities. Additionally, funding for programs designed to connect underserved communities to jobs and services would be cut. Critics, including researchers from Berkeley Law, argue that these measures prioritize cost-cutting over long-term resilience. Trump has also announced plans to end the $7,500 tax credit for purchasing an electric vehicle. —Willie D. Jones Back to top

Jean Sammet rarely let anything get in the way of her professional goals. As a young student, she was barred from attending prestigious all-boys schools, so she pursued her love of mathematics at the best institutions she could find that were open to girls and women. Following her college graduation, she became one of the first programmers at Sperry, an electronics manufacturer in New York, despite having virtually no prior experience with computers. In 1959, after learning to code on the job just a few years prior, Sammet helped write the foundation of Cobol, a programming language widely used in computers that performed large-scale data processing jobs. Later, as a programming manager at IBM in 1971, she helped develop Formac, the first commonly used computer language for symbolic manipulation of mathematical formulas. She was honored in 2009 with an IEEE Computer Society Pioneer Award for “pioneering work and lifetime achievement as one of the first developers and researchers in programming languages.” Her career and contributions were chronicled in an oral history with the IEEE History Center. An early love of numbers Born and raised in New York City, Sammet developed an early interest in mathematics. Her gender precluded her from attending prominent New York math- and science-focused schools such as the Bronx High School of Science, according to a biography published in the IEEE Annals of the History of Computing. She attended Julia Richman High School, an all-girls facility, and took every math class it offered. Sammet knew she wanted to major in math in college. In a 2001 interview with the IEEE History Center, she said, “I looked at all the catalogs from all the girls’ schools and decided Smith and Mount Holyoke seemed to have the best math courses.” Sammet was accepted into both colleges but decided on Mount Holyoke, which is in South Hadley, Mass. “I’m not even sure why I picked Holyoke, but I am very thankful that I did,” she said in the IEEE oral history. “It was a great college. Still is.” She received her bachelor’s degree in mathematics in 1947 and went on to earn a master’s degree in math in 1949 from the University of Illinois Urbana-Champaign. She joined the Illinois faculty as a teaching assistant, then in 1951 moved back to New York City. Development of Cobol Sammet joined Sperry in 1953 as an engineer. She initially performed mathematical analysis and ran the company’s analog computer, but two years later, she was tasked with supervising a growing department of computer programmers. In 1955 Sperry merged with Remington Rand, an early business machine manufacturer in New York City. That gave Sammet the opportunity to work with computing pioneer Grace Hopper on the UNIVAC I, the first general-purpose electronic digital computer manufactured in the United States. Sammet left Remington Rand in 1958 to join Sylvania Electric Products, a manufacturer in Needham, Mass. She had applied for an engineering position but was hired as a software developer. While at Sylvania, Sammet was appointed to the U.S. Department of Defense short-range committee, which included a group of programmers from six computer manufacturers. According to a May 1959 meeting summary, the DOD established the committee to design the specifications of “a common business language” that could work across all computers—one that was “problem-oriented and machine-independent.” Sammet chaired the statement language subcommittee, which included five other programmers from Sylvania, IBM, and RCA. It was tasked with designing the foundation of Cobol. According to the History Center interview, the subcommittee completed most of its tasks during a two-week sprint, working around the clock while holed up at the Sherry-Netherland hotel in New York City. The subcommittee presented its proposal for the code in November 1959, and Sylvania and the DOD accepted it with minimal changes. The language, structurally similar to English, was partially based on Hopper’s FLOW-MATIC programming language. During a time when computers were running extremely complex code, Cobol allowed early mainframe computers to essentially speak the same language, eliminating the need to manually program the same data-processing applications into every new machine. Leadership at IBM and ACM After two years at Sylvania, Sammet left in 1961 to join IBM in Cambridge, Mass., where she managed computer language development within the company’s data systems division. There she led the team that developed Formac, a formula manipulation compiler. It was the first computer algebra system to have significant commercial use. After the system’s release in 1964, she continued researching modeling programming and mathematical languages. In 1969 she had Programming Languages: History and Fundamentals published. The book covered the 120 languages in existence at the time. From 1968 to 1974, Sammet served as programming technology planning manager for IBM’s federal systems division, which conducted defense-related research and systems integration applications for the Federal Aviation Administration and the U.S. Postal Service. She also led the company’s work on the Ada programming language. She became active in the Association for Computing Machinery in 1961 while working on Formac. In an ACM interview, she said she joined so she could network with professionals working in her field at Bell Labs, Carnegie Mellon, and other institutions. She explained that such opportunities virtually didn’t exist at the time outside of ACM and other professional associations. She chaired the special interest committee on symbolic and algebraic manipulation (now known as SIGSAM). She also served on a number of councils and planning committees. She was elected vice president of ACM in 1972 and became its first female president in 1974. Among the accolades she received for her accomplishments were the 1989 Lovelace Award from the Association for Women in Computing and an honorary doctorate from Mount Holyoke in 1978. Sammet died in May 2017 at age 89. A longtime supporter of Mount Holyoke, she endowed a professorship there: the Sammet professor of computer science.

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion. ICRA 2025: 19–23 May 2025, ATLANTA, GA Enjoy today’s videos! NASA’s Mars Chopper concept, shown in a design software rendering, is a more capable proposed follow-on to the agency’s Ingenuity Mars Helicopter, which arrived at the Red Planet in the belly of the Perseverance rover in February 2021. Chopper would be about the size of an SUV, with six rotors, each with six blades. It could be used to carry science payloads as large as 11 pounds (5 kilograms) distances of up to 1.9 miles (3 kilometers) each Martian day (or sol). Scientists could use Chopper to study large swaths of terrain in detail, quickly – including areas where rovers cannot safely travel. We wrote an article about an earlier concept version of this thing a few years back if you’d like more detail about it. [ NASA ] Sanctuary AI announces its latest breakthrough with hydraulic actuation and precise in-hand manipulation, opening up a wide range of industrial and high value work tasks. Hydraulics have significantly more power density than electric actuators in terms of force and velocity. Sanctuary has invented miniaturized valves that are 50x faster and 6x cheaper than off the shelf hydraulic valves. This novel approach to actuation results in extremely low power consumption, unmatched cycle life and controllability that can fit within the size constraints of a human-sized hand and forearm. [ Sanctuary AI ] Clone’s Torso 2 is the most advanced android ever created with an actuated lumbar spine and all the corresponding abdominal muscles. Torso 2 dons a white transparent skin that encloses 910 muscle fibers animating its 164 degrees of freedom and includes 182 sensors for feedback control. These Torsos use pneumatic actuation with off-the-shelf valves that are noisy from the air exhaust. Our biped brings back our hydraulic design with custom liquid valves for a silent android. Legs are coming very soon! [ Clone Robotics ] Suzumori Endo Lab, Science Tokyo has developed a superman suit driven by hydraulic artificial muscles. [ Suzumori Endo Lab ] We generate physically correct video sequences to train a visual parkour policy for a quadruped robot, that has a single RGB camera without depth sensors. The robot generalizes to diverse, real-world scenes despite having never seen real-world data. [ LucidSim ] Seoul National University researchers proposed a gripper capable of moving multiple objects together to enhance the efficiency of pick-and-place processes, inspired from humans’ multi-object grasping strategy. The gripper can not only transfer multiple objects simultaneously but also place them at desired locations, making it applicable in unstructured environments. [ Science Robotics ] We present a bio-inspired quadruped locomotion framework that exhibits exemplary adaptability, capable of zero-shot deployment in complex environments and stability recovery on unstable terrain without the use of extra-perceptive sensors. Through its development we also shed light on the intricacies of animal locomotion strategies, in turn supporting the notion that findings within biomechanics and robotics research can mutually drive progress in both fields. [ Paper authors from University of Leeds and University College London ] Thanks, Chengxu! Happy 60th birthday to MIT CSAIL! [ MIT Computer Science and Artificial Intelligence Laboratory ] Yup, humanoid progress can move quickly when you put your mind to it. [ MagicLab ] The Sung Robotics Lab at UPenn is interested in advancing the state of the art in computational methods for robot design and deployment, with a particular focus on soft and compliant robots. By combining methods in computational geometry with practical engineering design, we develop theory and systems for making robot design and fabrication intuitive and accessible to the non-engineer. [ Sung Robotics Lab ] From now on I will open doors like the robot in this video. [ Humanoids 2024 ] Travel along a steep slope up to the rim of Mars’ Jezero Crater in this panoramic image captured by NASA’s Perseverance just days before the rover reached the top. The scene shows just how steep some of the slopes leading to the crater rim can be. [ NASA ] Our time is limited when it comes to flying drones, but we haven’t been surpassed by AI yet. [ Team BlackSheep ] Daniele Pucci from IIT discusses iCub and ergoCub as part of the industrial panel at Humanoids 2024. [ ergoCub ]

The just-released AI Safety Index graded six leading AI companies on their risk assessment efforts and safety procedures... and the top of class was Anthropic, with an overall score of C. The other five companies—Google DeepMind, Meta, OpenAI, xAI, and Zhipu AI—received grades of D+ or lower, with Meta flat out failing. “The purpose of this is not to shame anybody,” says Max Tegmark, an MIT physics professor and president of the Future of Life Institute, which put out the report. “It’s to provide incentives for companies to improve.” He hopes that company executives will view the index like universities view the U.S. News and World Reports rankings: They may not enjoy being graded, but if the grades are out there and getting attention, they’ll feel driven to do better next year. He also hopes to help researchers working in those companies’ safety teams. If a company isn’t feeling external pressure to meet safety standards, Tegmark says, “then other people in the company will just view you as a nuisance, someone who’s trying to slow things down and throw gravel in the machinery.” But if those safety researchers are suddenly responsible for improving the company’s reputation, they’ll get resources, respect, and influence. The Future of Life Institute is a nonprofit dedicated to helping humanity ward off truly bad outcomes from powerful technologies, and in recent years it has focused on AI. In 2023, the group put out what came to be known as “the pause letter,” which called on AI labs to pause development of advanced models for six months, and to use that time to develop safety standards. Big names like Elon Musk and Steve Wozniak signed the letter (and to date, a total of 33,707 have signed), but the companies did not pause. This new report may also be ignored by the companies in question. IEEE Spectrum reached out to all the companies for comment, but only Google DeepMind responded, providing the following statement: “While the index incorporates some of Google DeepMind’s AI safety efforts, and reflects industry-adopted benchmarks, our comprehensive approach to AI safety extends beyond what’s captured. We remain committed to continuously evolving our safety measures alongside our technological advancements.” How the AI Safety Index graded the companies The Index graded the companies on how well they’re doing in six categories: risk assessment, current harms, safety frameworks, existential safety strategy, governance and accountability, and transparency and communication. It drew on publicly available information, including related research papers, policy documents, news articles, and industry reports. The reviewers also sent a questionnaire to each company, but only xAI and the Chinese company Zhipu AI (which currently has the most capable Chinese-language LLM) filled theirs out, boosting those two companies’ scores for transparency. The grades were given by seven independent reviewers, including big names like UC Berkeley professor Stuart Russell and Turing Award winner Yoshua Bengio, who have said that superintelligent AI could pose an existential risk to humanity. The reviewers also included AI leaders who have focused on near-term harms of AI like algorithmic bias and toxic language, such as Carnegie Mellon University’s Atoosa Kasirzadeh and Sneha Revanur, the founder of Encode Justice. And overall, the reviewers were not impressed. “The findings of the AI Safety Index project suggest that although there is a lot of activity at AI companies that goes under the heading of ‘safety,’ it is not yet very effective,” says Russell. “In particular, none of the current activity provides any kind of quantitative guarantee of safety; nor does it seem possible to provide such guarantees given the current approach to AI via giant black boxes trained on unimaginably vast quantities of data. And it’s only going to get harder as these AI systems get bigger. In other words, it’s possible that the current technology direction can never support the necessary safety guarantees, in which case it’s really a dead end.” Anthropic got the best scores overall and the best specific score, getting the only B- for its work on current harms. The report notes that Anthropic’s models have received the highest scores on leading safety benchmarks. The company also has a “responsible scaling policy“ mandating that the company will assess its models for their potential to cause catastrophic harms, and will not deploy models that the company judges too risky. All six companies scored particularly badly on their existential safety strategies. The reviewers noted that all of the companies have declared their intention to build artificial general intelligence (AGI), but only Anthropic, Google DeepMind, and OpenAI have articulated any kind of strategy for ensuring that the AGI remains aligned with human values. “The truth is, nobody knows how to control a new species that’s much smarter than us,” Tegmark says. “The review panel felt that even the [companies] that had some sort of early-stage strategies, they were not adequate.” While the report does not issue any recommendations for either AI companies or policymakers, Tegmark feels strongly that its findings show a clear need for regulatory oversight—a government entity equivalent to the U.S. Food and Drug Administration that would approve AI products before they reach the market. “I feel that the leaders of these companies are trapped in a race to the bottom that none of them can get out of, no matter how kind-hearted they are,” Tegmark says. Today, he says, companies are unwilling to slow down for safety tests because they don’t want competitors to beat them to the market. “Whereas if there are safety standards, then instead there’s commercial pressure to see who can meet the safety standards first, because then they get to sell first and make money first.”

TSMC described its next generation transistor technology this week at the IEEE International Electron Device Meeting (IEDM) in San Francisco. The N2, or 2-nanometer, technology is the semiconductor foundry giant’s first foray into a new transistor architecture, called nanosheet or gate-all-around. Samsung has a process for manufacturing similar devices, and both Intel and TSMC expect to be producing them in 2025. Compared to TSMC’s most advanced process today, N3 (3-nanometer), the new technology offers up to a 15 percent speed up or as much as 30 percent better energy efficiency, while increasing density by 15 percent. N2 is “the fruit of more than four years of labor,” Geoffrey Yeap, TSMC vice president of R&D and advanced technology told engineers at IEDM. Today’s transistor, the FinFET, has a vertical fin of silicon at its heart. Nanosheet or gate-all-around transistors have a stack of narrow ribbons of silicon instead. The difference not only provides better control of the flow of current through the device, it also allows engineers to produce a larger variety of devices, by making wider or narrower nanosheets. FinFETs could only provide that variety by multiplying the number of fins in a device—such as a device with one or two or three fins. But nanosheets give designers the option of gradations in between those, such as the equivalent of 1.5 fins or whatever might suit a particular logic circuit better. Called Nanoflex, TSMC’s tech allows different logic cells built with different nanosheet widths on the same chip. Logic cells made from narrow devices might make up general logic on the chip, while those with broader nanosheets, capable of driving more current and switching faster, would make up the CPU cores. The nanosheet’s flexibility has a particularly large impact on SRAM, a processor’s main on-chip memory. For several generations, this key circuit, made up of 6 transistors, has not been shrinking as fast as other logic. But N2 seems to have broken this streak of scaling stagnation, resulting in what Yeap described as the densest SRAM cell so far: 38 megabits per square millimeter, or an 11 percent boost over the previous technology, N3. N3 only managed a 6 percent boost over its own predecessor. “SRAM harvests the intrinsic gain of going to gate-all-around,” says Yeap. Future Gate-All-Around Transistors While TSMC delivered details of next year’s transistor, Intel looked at how long industry might be able to scale it down. Intel’s answer: Longer than originally thought. “The nanosheet architecture actually is the final frontier of transistor architecture,” Ashish Agrawal, a silicon technologist in Intel’s components research group, told engineers. Even future complementary FET (CFET) devices, possibly arriving in the mid-2030s, are constructed of nanosheets. So it’s important that researchers understand their limits, said Agrawal. “We have not hit a wall. It’s doable, and here’s the proof... We are making a really pretty good transistor.” —Sanjay Natarajan, Intel Intel proved that a transistor with a 6-nanometer gate length works well.Intel Intel explored a critical scaling factor, gate length, which is the distance covered by the gate between the transistor’s source and drain. The gate controls the flow of current through the device. Scaling down gate length is critical to reducing the minimum distance from device to device within standard logic circuits, called called contacted poly pitch, or CPP, for historical reasons. “CPP scaling is primarily by gate length, but it’s predicted this will stall at the 10-nanometer gate length,” said Agrawal. The thinking had been that 10 nanometers was such a short gate length that, among other problems, too much current would leak across the device when it was supposed to be off. “So we looked at pushing below 10 nanometers,” Agrawal said. Intel modified the typical gate-all-around structure so the device would have only a single nanosheet through which current would flow when the device was on. By thinning that nanosheet down and modifying the materials surrounding it, the team managed to produce an acceptably performing device with a gate length of just 6 nm and a nanosheet just 3 nm thick. Eventually, researchers expect silicon gate-all-around devices to reach a scaling limit, so researchers at Intel and elsewhere have been working to replace the silicon in the nanosheet with 2D semiconductors such as molybdenum disulfide. But the 6-nanometer result means those 2D semiconductors might not be needed for a while. “We have not hit a wall,” says Sanjay Natarajan, senior vice president and general manager of technology research at Intel Foundry. “It’s doable, and here’s the proof… We are making a really pretty good transistor” at the 6-nanometer channel length.

It seems to me that every item in the Computer History Museum’s collection has a biography of sorts—a life before CHM, a tale about how it came to us, and a life within the museum. The chapters of that biography include the uses made of it, and the historical and interpretive stories it can be made to tell. This then is a biography of one item that recently entered the museum’s collection—an early Memorex videotape containing a recording from 1968—and the historical discovery it has afforded. Our biography begins in May 2020, with an email. Debra Dunlop, a dean at New England College, wrote to the museum about a large collection of documents, audiovisual materials, and a rare computer, a Xerox Star, in New Hampshire. These were the professional papers of Debra’s father, Robert Dunlop, and she knew how dearly he valued the collection. She was helping her father move to an assisted living facility, and she had to make a plan for this extensive collection. What did the museum think? Industrial psychologist Robert Dunlop spent his career at high-tech companies, including IBM, RCA, and Xerox.Computer History Museum For me, the Dunlop collection was a light in the darkness. It was still early days in the pandemic, and the death toll in the United States from Covid-19 was nearing 100,000, with a vaccine shot for me still months in the future. I was working from home in Massachusetts but I was nervous because the museum—like all places that depend in part on ticket sales—faced strong financial pressures, and I didn’t know how long it could go on with its doors closed. The Dunlop collection sounded interesting. Robert Dunlop had been an industrial psychologist who spent his career at large, high-technology U.S. firms—first IBM, then RCA, and finally Xerox. The collection wasn’t far away, and perhaps there was a way I could safely go and have a look. I learned more about Robert Dunlop’s career from Debra. She and her family moved the collection to a garage where, after we let it sit for a week, we felt it would be safe for me to review the materials alone, wearing a mask, with the garage doors open. After the visit, I discussed what I had seen with my colleagues, and we agreed that I would return and select, pack, and ship out a substantial portion of it. Debra and her family very kindly made a financial donation to the museum to help with the shipping expenses in that difficult time for CHM. And as my colleagues and I would eventually discover, Dunlop’s collection offered an extraordinary glimpse into a transformative time in advanced computing, and a fascinating project that had been wholly unknown to the history of computing until now. A Discovery In May 2020, the author visited Dunlop’s home to go through documents, photos, and audiovisual recordings related to his work. Much of the material now resides at the Computer History Museum.David C. Brock As I went through the collection in that New Hampshire garage, one item intrigued me. It was an early video recording, made in 1968, that clearly had great meaning for Robert Dunlop. The 1-inch format tape on an open reel had been carefully packaged and included an explanatory note by Dunlop taped to the outside, along with a longer letter from him tucked inside. Both notes told of an inventive computer system at IBM headquarters that I’d never heard of. According to the notes, a demo of the system was captured on the long obsolete video. In 1995, when Dunlop wrote the notes, he had despaired of finding any working equipment to recover the recording. As the tape rested in my hands, I wondered the same thing—should I even collect this if it’s impossible to watch? But then I thought, “Perhaps we can figure something out. And if not us, maybe something could happen in the future.” I decided to take my chances and collect it. To recover the recording from the obsolete tape, the museum turned to George Blood LP, a company that specializes in archival audio and video. Penny Ahlstrand The Dunlop collection started its new life in the museum, carefully rehoused into archival storage boxes and added to our backlog for archival processing. In 2023, a grant to the museum from the Gordon and Betty Moore Foundation presented an opportunity to digitize some of the audiovisual materials in our collection. When I was consulted about priorities, one of the items I selected was Dunlop’s 1968 video recording. Could we give it a try? Massimo Petrozzi, CHM’s Director of Archives and Digital Initiatives, reached out to his networks to see if there was someone who could help. A contact in Europe pointed back to the States, to George Blood and his firm George Blood LP outside of Philadelphia. The company is a major provider of audio and moving-picture preservation services, boasting an enormous collection of equipment—including, as it happens, an Ampex video unit capable of recovering video from Dunlop’s tape, which Blood called a “very early technology.” Blood and his colleagues made painstaking adjustments and experiments and were finally able to recover and digitize Dunlop’s silent video, fulfilling Robert Dunlop’s long hopes. Sadly, Dunlop did not live to see his recording again. He died in July 2020. A Competing Vision of Computing The recording reveals a story as interesting as it is seemingly forgotten. You may already be aware of the “Mother of All Demos” presented by Doug Engelbart and the members of his Stanford Research Institute center at the close of 1968. This presentation, with Engelbart on stage at a major computing conference in San Francisco, displayed the features and capabilities of his group’s “oN-Line System,” known as NLS. The system included many elements that were extraordinarily novel, even for the assembled computing professionals: networked computers, video conferencing, graphical interfaces, hypertext, collaborative word processing, and even a new input device, the computer mouse. This remarkable 1968 demonstration of the NLS was, much to our benefit, recorded on videotape. Although relatively early in video technology, the quality of the surviving recording is excellent and readily available online today. The NLS was driven by a particular vision for the future use and practice of computing: a vision that centered on the notion of alliance. In this vision, individuals would join together into teams and organizations, directly using new computing tools and approaches for creating and using knowledge, and in doing so, “augmenting human intellect” to better solve complex problems. Dunlop’s video recording, it turned out, also contained a demonstration of another advanced computing system that also took place in 1968. This second demo occurred on the East Coast, at IBM’s corporate headquarters in Armonk, N.Y., and was motivated by a far different—perhaps one could go so far as to say an opposite—vision for the future of computing. This vision centered not on alliance, but rather on the concept of rank. The system was known as the IBM Corporate Headquarters Information Center, and it was the culmination of Dunlop’s experiments with executive-computer interaction at the company. Dunlop’s career at IBM across the 1960s coincided with a truly remarkable period of growth for the firm. From 1964—the year IBM announced its new System/360 line of digital computers—to 1970, the firm’s headcount and revenues both doubled. To contend with this extraordinary growth, Dunlop worked on what he and others there called “management information systems”—computer systems serving the informational needs of IBM managers. As Dunlop noted in an unpublished talk, IBM managers were increasingly embracing information processing in the form of the company’s own timesharing computer products. Several internal IBM systems gave users remote access to timesharing computers, with modified electric typewriters serving as the user “terminals.” A sophisticated messaging system allowed employees to send one another telegram-like messages from one terminal to another, at the rate of 25,000 messages per day. A mathematical utility, QUIKTRAN, let users perform simple as well as sophisticated calculations from their terminals. There was a proliferation of systems for storing documents and formatting them in complex ways, with a single computer supporting up to 40 typewriter terminal users. Lastly, there were what today we would call database systems, containing information about the business and the organization, with a query language and financial models, again available from the users’ typewriter terminals. IBM’s Executive War Room As these systems were increasingly adopted by what Dunlop called “operational and middle managers,” he led a series of projects to see if IBM could create terminals and management information systems that could be productively used by IBM’s “top executives.” The systems would allow the executives to make strategic decisions for the company in new ways afforded by the computer. His initial efforts all failed. First, Dunlop experimented with providing high-ranking executives—VPs and the like— with typewriter terminals directly linked to real-time data, financial models, and summary documents about the firm. The terminals went untouched, quickly migrating to the desks of the executives’ secretaries. Dunlop then tried using IBM’s new CRT-based terminal, the 2250, with a simplified keypad for input. The result was unchanged. Through interviews and surveys, he concluded that the failure was due to the executives’ “self-role concept.” They held themselves to be “very high status” decision-makers who got information from subordinates; any direct use of a typewriter or keyboard would “demean” them. From his failed experiments, Dunlop concluded that the state-of-the-art in computing technology was inadequate for creating a terminal-based management system for top management. However, those same top managers had noticed that middle managers around the firm had established “war rooms,” in which staff integrated information from all the various terminal-based systems: messaging, text, and database. At IBM corporate headquarters, the top executives wanted a war room of their own. This desire led Dunlop and others to create the IBM Headquarters Information Center. Here, “information specialists” would respond to inquiries by high-ranking executives. The specialists had access to messaging, text, database, and financial modeling systems accessed through typewriter and CRT terminals, as well as an array of printed materials, microform holdings, and audiovisual materials. In short, the information center was a reference library, staffed with reference librarians, of the sort that would become commonplace in the 1980s. An old recording with typed notes from Dunlop turned out to contain a previously unknown 1968 demonstration of an IBM system called the Executive Terminal. Penny Ahlstrand With the new IBM Headquarters Information Center in place, Dunlop saw the opportunity to run another experiment in 1967-68, which he called the “Executive Terminal.” The lead information specialist in the information center would sit at a video-mixing and control console, equipped with a video camera, microphone, and even lighting. Meanwhile, the executive user would be in their office with their Executive Terminal, a modified television set with an audio and video connection to the console in the information center. The executive pressed a button to summon the information specialist and their live video image to the screen. Remaining unseen, the executive could then place an inquiry. The information specialist would direct other staff in the information center to gather the appropriate information to answer the request: Models were run on CRT terminals, documents and data were gathered on typewriter terminals, microform could be loaded into a video reader, paper documents could be placed on a video capture unit. Once the results were assembled, the information specialist conveyed all this information to the executive, cutting from one video feed to another, guided by the executive’s interest and direction. Dunlop’s 1968 video demonstration of the Executive Terminal and the Information Center proceeds in three acts. The first 10 minutes of the video show the information specialist and other staff responding to an executive’s request, finding and preparing all the materials for video presentation, using the typewriter and CRT terminals, and even engaging in video conferencing with another employee: The next five minutes show the executive using the Executive Terminal to receive the results and directing the display and flow of the information: The final few minutes show the information specialist working on an IBM 2260 video computer terminal, at the time still a novelty that was used for database and model access: Restoring History It’s unclear what ultimately became of IBM’s Executive Terminal and the Information Center, as they appear to have left little to no historical traces beyond a few documents—including the unpublished talk—some photographs, and Dunlop’s 1968 video recording. With Engelbart’s and Dunlop’s 1968 demo videos, we now have a remarkable and contrasting snapshot of two very different directions in advanced computing. Engelbart’s Mother of All Demos showed how advanced computing could create a shared, collaborative environment of allied individuals, all direct users of the same system, befitting of a laboratory of computer enthusiasts in Menlo Park, Calif. Dunlop’s Executive Terminal demo showed how many of these same advanced technologies could be directed along another path, that of a strictly hierarchical organization, highly attuned to rank and defined roles and specialties. While these were very different and perhaps opposing directions, they shared a common commitment to the use of advanced computing for organizing and analyzing information, and taking action. In the Information Center at IBM Headquarters, in Armonk, N.Y., information specialists were on call to answer questions from users.The Dunlop Collection Engelbart held that his system was for the “augmentation of the human intellect,” so that users might better address complex problems. For Dunlop, the Executive Terminal was an answer to his question, “Can we make better decisions, at higher levels, through better information processes?” There are echoes of Engelbart’s Mother of All Demos around us every day—the hyperlinks of the Web, the scuttling of computer mice on desktops, the editing of online documents, and more. But just as evident are the echoes of Dunlop’s Executive Terminal demo, such as the video conferencing and screen-sharing practices so familiar in Zooms, Teams, and Meets today. The Computer History Museum is pleased to make public the entire video recording of Robert Dunlop’s 1968 demonstration, and with its release, to restore a forgotten chapter in the history of computing. Acknowledgments The work of any one person at any museum is actually the work of many, and that is certainly true here. I’d like to thank the trustees and financial supporters of the Computer History Museum for making these efforts possible, especially the Gordon and Betty Moore Foundation and Gardner Hendrie. At the museum, I’d like to thank my colleagues Massimo Petrozzi, Penny Ahlstrand, Max Plutte, Kirsten Tashev, Gretta Stimson, and Liz Stanley. I’d also like to thank historian Jim Cortada for giving this essay a reading, George Blood for recovering the recording, Heidi Hackford for editing and producing this essay for the museum, Jean Kumagai and her colleagues at IEEE Spectrum for editing, designing, and cross-posting the essay, Debra Dunlop for thinking of the museum, and the late Robert Dunlop for taking such care of these materials in the first chapters of their life. Editor’s note: This post originally appeared on the blog of the Computer History Museum.

Stanford University professor Fei-Fei Li has already earned her place in the history of AI. She played a major role in the deep learning revolution by laboring for years to create the ImageNet dataset and competition, which challenged AI systems to recognize objects and animals across 1,000 categories. In 2012, a neural network called AlexNet sent shockwaves through the AI research community when it resoundingly outperformed all other types of models and won the ImageNet contest. From there, neural networks took off, powered by the vast amounts of free training data now available on the Internet and GPUs that deliver unprecedented compute power. In the 13 years since ImageNet, computer vision researchers mastered object recognition and moved on to image and video generation. Li cofounded Stanford’s Institute for Human-Centered AI (HAI) and continued to push the boundaries of computer vision. Just this year she launched a startup, World Labs, which generates 3D scenes that users can explore. World Labs is dedicated to giving AI “spatial intelligence,” or the ability to generate, reason within, and interact with 3D worlds. Li delivered a keynote yesterday at NeurIPS, the massive AI conference, about her vision for machine vision, and she gave IEEE Spectrum an exclusive interview before her talk. Why did you title your talk “Ascending the Ladder of Visual Intelligence”? Fei-Fei Li: I think it’s intuitive that intelligence has different levels of complexity and sophistication. In the talk, I want to deliver the sense that over the past decades, especially the past 10-plus years of the deep learning revolution, the things we have learned to do with visual intelligence are just breathtaking. We are becoming more and more capable with the technology. And I was also inspired by Judea Pearl’s “ladder of causality” [in his 2020 book The Book of Why]. The talk also has a subtitle, “From Seeing to Doing.” This is something that people don’t appreciate enough: that seeing is closely coupled with interaction and doing things, both for animals as well as for AI agents. And this is a departure from language. Language is fundamentally a communication tool that’s used to get ideas across. In my mind, these are very complementary, but equally profound, modalities of intelligence. Do you mean that we instinctively respond to certain sights? Li: I’m not just talking about instinct. If you look at the evolution of perception and the evolution of animal intelligence, it’s deeply, deeply intertwined. Every time we’re able to get more information from the environment, the evolutionary force pushes capability and intelligence forward. If you don’t sense the environment, your relationship with the world is very passive; whether you eat or become eaten is a very passive act. But as soon as you are able to take cues from the environment through perception, the evolutionary pressure really heightens, and that drives intelligence forward. Do you think that’s how we’re creating deeper and deeper machine intelligence? By allowing machines to perceive more of the environment? Li: I don’t know if “deep” is the adjective I would use. I think we’re creating more capabilities. I think it’s becoming more complex, more capable. I think it’s absolutely true that tackling the problem of spatial intelligence is a fundamental and critical step towards full-scale intelligence. I’ve seen the World Labs demos. Why do you want to research spatial intelligence and build these 3D worlds? Li: I think spatial intelligence is where visual intelligence is going. If we are serious about cracking the problem of vision and also connecting it to doing, there’s an extremely simple, laid-out-in-the-daylight fact: The world is 3D. We don’t live in a flat world. Our physical agents, whether they’re robots or devices, will live in the 3D world. Even the virtual world is becoming more and more 3D. If you talk to artists, game developers, designers, architects, doctors, even when they are working in a virtual world, much of this is 3D. If you just take a moment and recognize this simple but profound fact, there is no question that cracking the problem of 3D intelligence is fundamental. I’m curious about how the scenes from World Labs maintain object permanence and compliance with the laws of physics. That feels like an exciting step forward, since video-generation tools like Sora still fumble with such things. Li: Once you respect the 3D-ness of the world, a lot of this is natural. For example, in one of the videos that we posted on social media, basketballs are dropped into a scene. Because it’s 3D, it allows you to have that kind of capability. If the scene is just 2D-generated pixels, the basketball will go nowhere. Or, like in Sora, it might go somewhere but then disappear. What are the biggest technical challenges that you’re dealing with as you try to push that technology forward? Li: No one has solved this problem, right? It’s very, very hard. You can see [in a World Labs demo video] that we have taken a Van Gogh painting and generated the entire scene around it in a consistent style: the artistic style, the lighting, even what kind of buildings that neighborhood would have. If you turn around and it becomes skyscrapers, it would be completely unconvincing, right? And it has to be 3D. You have to navigate into it. So it’s not just pixels. Can you say anything about the data you’ve used to train it? Li: A lot. Do you have technical challenges regarding compute burden? Li: It is a lot of compute. It’s the kind of compute that the public sector cannot afford. This is part of the reason I feel excited to take this sabbatical, to do this in the private sector way. And it’s also part of the reason I have been advocating for public sector compute access because my own experience underscores the importance of innovation with an adequate amount of resourcing. It would be nice to empower the public sector, since it’s usually more motivated by gaining knowledge for its own sake and knowledge for the benefit of humanity. Li: Knowledge discovery needs to be supported by resources, right? In the times of Galileo, it was the best telescope that let the astronomers observe new celestial bodies. It’s Hooke who realized that magnifying glasses can become microscopes and discovered cells. Every time there is new technological tooling, it helps knowledge-seeking. And now, in the age of AI, technological tooling involves compute and data. We have to recognize that for the public sector. What would you like to happen on a federal level to provide resources? Li: This has been the work of Stanford HAI for the past five years. We have been working with Congress, the Senate, the White House, industry, and other universities to create NAIRR, the National AI Research Resource. Assuming that we can get AI systems to really understand the 3D world, what does that give us? Li: It will unlock a lot of creativity and productivity for people. I would love to design my house in a much more efficient way. I know that lots of medical usages involve understanding a very particular 3D world, which is the human body. We always talk about a future where humans will create robots to help us, but robots navigate in a 3D world, and they require spatial intelligence as part of their brain. We also talk about virtual worlds that will allow people to visit places or learn concepts or be entertained. And those use 3D technology, especially the hybrids, what we call AR [augmented reality]. I would love to walk through a national park with a pair of glasses that give me information about the trees, the path, the clouds. I would also love to learn different skills through the help of spatial intelligence. What kind of skills? Li: My lame example is if I have a flat tire on the highway, what do I do? Right now, I open a “how to change a tire” video. But if I could put on glasses and see what’s going on with my car and then be guided through that process, that would be cool. But that’s a lame example. You can think about cooking, you can think about sculpting—fun things. How far do you think we’re going to get with this in our lifetime? Li: Oh, I think it’s going to happen in our lifetime because the pace of technology progress is really fast. You have seen what the past 10 years have brought. It’s definitely an indication of what’s coming next.

Editor’s note: A version of this article originally appeared in the author’s newsletter, Exponential View. When people ask me to describe my work, I say I take a critical look at exponential technologies—which I define as technologies that follow an exponential growth curve. I’m the founder of the research group Exponential View, and my mission also includes critically reviewing my own analyses. So here’s a reflection on my analyses of autonomous vehicles. I have long argued that self-driving cars are metaphorically miles away from being a reality. For years, I’ve tried to offer a tonic to the rah-rah hype that carmakers were foisting upon us through marketing. In 2017, when many carmakers were promising that fully autonomous vehicles would be on the road imminently, I wrote in MIT Technology Review: KITT, the car from Knight Rider, will remain the gold standard for autonomous vehicles. Autonomous vehicle pilots will become increasingly ambitious, but the real-world hurdles will still take time to navigate, even with friendly city regulators. None will ship to the public in 2018. Five years later, I remained pessimistic, as I wrote in my newsletter, Exponential View: Max Chalkin analyzes the disappointing trajectory of full self-driving efforts: US $100 billion invested and little to show. The self-driving pioneer Anthony Levandowski, who cofounded Waymo, has retreated to building autonomous trucks constrained to industrial sites. He reckons that is the most complex use case the technology can deliver in the near future. Why it matters: Self-driving could be a pointless distraction for improving the environmental and human impact of transport. It takes attention away from micromobility, better urban infrastructure, and other strategies to improve the safety, pollution, climate, equity and economic returns of this sector. That was then and this is now. KITT remains awesome and I’m changing my mind about self-driving cars. Far from being a “pointless distraction,” they’re nearly ready for prime time. And robotaxis are leading the charge. That’s not just based on a hunch. It’s based on an increasing mountain of evidence pointing to their adoption and evolution—evidence that the industry is making progress on overlapping “S-curves.” These S-curves in technology typically show slow initial progress, followed by rapid advancement, and then a leveling off as the technology matures. Here’s how I’m thinking about the development of self-driving cars now. Two autonomous taxis, from Pony.ai and Baidu’s Apollo Go, cross paths in Beijing. VCG/Getty Images Baidu and Waymo Robotaxis Show the Way In bellwether cities that have historically been ahead of the curve on tech adoption, we’re seeing more self-driving vehicles on the road—with robotaxis spearheading this revolution. Wuhan, the capital of China’s Hubei province, is striving to become “the world’s first driverless city.” So far, around three in every 100 taxis there are robotaxis, developed by Baidu’s autonomous car division, Apollo Go. Over the past year, San Francisco has seen a rapid increase in Waymo rides. And as Alphabet’s autonomous vehicle company expands beyond San Francisco, so do its numbers: According to data from the California Public Utilities Commission, in August Waymo provided approximately 312,000 rides per month in California, doubling its ride volume from only three months before. These numbers highlight how quickly robotaxis can grab market share. While it’s not clear what proportion of Waymo’s 312,000 monthly rides in California happens in San Francisco alone, the city is the company’s most mature market, so it likely accounts for the bulk of rides—let’s estimate 80 percent. That gives us a direct comparison with Uber’s staffed rideshare service, which runs approximately 200,000 rides a day in San Francisco. Given Waymo’s 312,000-a-month figure, the company likely offers 8,000 or more rides per day in the city, a 4 percent or more market share. The tipping point in S-curves of adoption is typically 6 percent, signaling the beginning of a rapid growth phase, so Waymo is getting closer. Meanwhile, Baidu leads in driving down the cost of robotaxi journeys. A 10-kilometer (6.2-mile) ride in a robotaxi in Wuhan costs between 4 and 16 yuan ($0.60 to $2.30), whereas an equivalent ride in a car driven by a human costs between 18 and 30 yuan. Anecdotally, a Waymo ride in San Francisco often costs slightly more than an Uber. Because a robotaxi doesn’t contend with driver fatigue, the number of rides it can run per day can be greater than that of a nonautomated taxi. In Wuhan, a robotaxi completes up to 20 rides a day, which exceeds the daily average of 13.2 rides for human taxi drivers in the city. What about the economics? Baidu operated around 336,000 Apollo Go rides in July 2024. At the prices mentioned above, this means that Baidu Apollo could be grossing $200,000 to $800,000 per month, or $2.4 million to $9.6 million per year. The Apollo costs only $28,000 to build, so it’s much cheaper than a Waymo car, which is estimated to cost $150,000. Baidu Apollo looks likely to reach profitability before its U.S. peer (setting aside all the prior investment in R&D): The firm expects to break even this year and to become profitable in 2025. Waymo also has a path to profitability but will face challenges from the incumbents. For example, the British autonomous vehicle company Wayve recently announced a partnership with Uber. So there may be a few bumps in the road for Waymo. Selling Self-Driving Cars to Suburbia Of course, history is littered with technologies that excited early adopters but didn’t cut through to the masses. Yet here too I see evidence that self-driving vehicles—in their initial form of robotaxis—are starting to burst out of the tech bubble. Waymo is expanding its self-driving taxi service as regulators become more accepting of autonomous vehicles. Already established in San Francisco and Phoenix, Waymo has recently launched in Los Angeles and Austin, Texas. The company is also testing operations in 25 other major metro areas, including Atlanta, Dallas, Houston, Miami, and New York City. To be sure, Waymo is cherry-picking cities with favorable conditions for autonomous vehicles. Regardless, its expansion signals the increasing acceptance of self-driving technology in urban transportation. Beyond robotaxis, the public is becoming more comfortable with the tech, too. I believe that Tesla is far behind the likes of Waymo when it comes to self-driving technology, but the growing popularity of Tesla cars is helping normalize the tech. Tesla’s full self-driving mode is available to drivers all over the United States and Canada and is expected to roll out in China in early 2025. The more hands-on experience—or hands-off, as the case may be—people get with self-driving tech, the more willing they will be to set aside their worries and prejudices about it. We see this shift reflected in surveys of people’s trust in autonomous vehicles. Respondents in Phoenix and San Francisco who have been exposed to self-driving cars gave a confidence score of 67 in a 2023 survey, while the average American gave a score of 37. For meaningful adoption to occur, autonomous vehicle companies first need to address major safety concerns. In October of last year, a pedestrian was hit by a human-driven Nissan and then struck and dragged for 6 meters (20 feet) by a Cruise self-driving car on a San Francisco street. This event led to Cruise losing its operating permit in California and ceasing operations in Arizona and Texas. It was an awful accident and a moment of reflection for the self-driving car sector. But the fact is that self-driving cars are getting safer. If we measure Waymo’s performance by kilometers per disengagement—those times when a human has to take control—its record has been improving over the long run. In the chart below, the dip in kilometers per disengagement in 2021 is due to several factors: The company introduced new vehicles, increased the number of kilometers driven by 270 percent compared to 2020, and shifted its focus from Mountain View, Calif., to San Francisco, which is a more complex driving environment. Despite that blip, the overall trend line is clear. Self-driving cars are also perceived to be safer than vehicles driven by humans. Some cyclists, for example, say they feel safer biking next to a Waymo car than a human-driven vehicle because the Waymo’s actions are more predictable. “As a cyclist, when I ride my bike and I get next to a @Waymo. I know it watches me, and if I try to pass it on the right, it makes room for me. I feel so much safer because it always sees me. It will never get in my way. It will never cut me off. It will always prioritize my safety over itself,” one cyclist wrote on X. Improvements to Self-Driving Tech The two overlapping S-curves of self-driving cars add up to true technological innovation and exponential growth. First, we have the S-curve of technology improvement. Autonomous vehicle leaders have taken different approaches to building their technology on three axes: sensors, maps, and intelligence. Waymo and Apollo are perhaps the most similar. Their cars are multisensorial, kitted out with cameras, lidar, and radar. They rely on high-definition custom maps. And the intelligence in both Waymo and Baidu vehicles are complex architectures that combine several AI systems to make decisions. At the other extreme is Tesla, which uses only cameras, maps, and end-to-end deep learning—meaning that it has one AI system that takes in raw sensor data and produces driving decisions as outputs. Wayve also uses end-to-end deep learning but is agnostic about its use of sensors. Current Wayve cars rely on cameras; future ones will use other sensors when available. The question of which technology will win out is superinteresting but beyond the scope of this essay. The one thing I’ll emphasize, though, is that competing approaches are a good thing. The proof of the improvement is in the data: falling rates of disengagement, at least for Waymo, Wayve, and Apollo. As for safety, Missy Cummings, a professor at George Mason University and a leading expert on autonomous transport, shared with me as-yet-unpublished data regarding self-driving cars’ progress. Her data shows that Waymo cars have a lower crash rate than the average rideshare driver, albeit still worse than a typical human. We’re reaching a tipping point where the technology is not just functional, but increasingly reliable and commercially viable. And handily, the S-curve of technology improvement is overlapping with another one: the adoption curve. Combined, Waymo’s growth in San Francisco and Baidu’s mass experiments in Wuhan begin to look like proof that we have worked out how to deliver robotaxis at scale. Adoption so far has been in robotaxis because companies can deploy them at scale and because their trips are fairly constrained and predictable. If Waymo’s vehicles can navigate hundreds of thousands of trips successfully each week and train subsequent AI models on that data, it gives me confidence that self-driving vehicles can be used for everyday trips, by everyday people, in cities around the world. S-curves sometimes reveal paradigm shifts. And it feels like we’re on the cusp of one with self-driving vehicles. Where Self-Driving Cars Go from Here So what might happen next? History has shown that technology transitions can take place within a window of less than 20 years. Feature phones were almost entirely replaced by smartphones in just seven years. It took 14 years for the motorcar to go from 5 percent to 75 percent market share in American cities, almost entirely replacing the horse. Large sailboats ferrying immigrants from Europe to New York at the turn of the 19th century were replaced by the new technology of steamships within 15 years. However, there is a wrinkle with self-driving vehicles. Regulators are wary of removing the human from the loop. The advancement of self-driving in the United States will depend on cities and states beyond the early tech adopters like San Francisco. And the U.S. National Highway Traffic Safety Administration has acted quickly against auto companies where it saw harm to the public. After the October 2023 accident, Cruise recalled its entire fleet of robotaxis—nearly 1,200 vehicles—to close an investigation by the regulator. By contrast, China’s ambition is on full display in Wuhan. The Chinese government has already approved live testing on public roads in at least 16 other major cities. This rapid advance is due to China’s more directive government but also the public’s willingness to embrace the tech. Chinese consumers are twice as likely as Americans to say they trust self-driving vehicles. In June 2024 the Chinese government approved nine automakers to test systems that go further than Tesla’s full self-driving mode (which requires driver attention at all times). The China Society of Automotive Engineers foresees that one in five cars sold in China will be fully driverless by the decade’s end. And what about Tesla? The company has a data advantage over Waymo: By April of this year, the firm had garnered more than 2 billion km (more than 1.3 billion miles) of experience under full self-driving (FSD) mode, and drivers had begun to add about 1.6 billion new km (about 1 billion miles) every two months. And yet, Tesla is miles behind Waymo both technically and operationally. As Chris Anderson, former editor in chief of Wired, pointed out in a post on X, Tesla’s FSD doesn’t work on his Bay Area commute. “Having now had a chance to compare Tesla FSD 12.4 in San Francisco with Waymo, I don’t yet see how Tesla can field a robotaxi fleet anytime soon. With the Tesla, I still get 3 to 4 disengagements in my daily 1.5-hour commute, which is really not bad. But there’s no room for any disengagements with a robotaxi. And Waymo does things like pulling over for fire engines, which Tesla doesn’t do. I’m a Tesla bull, but a Waymo ride shows just how challenging true Level 5 autonomy is.” I wouldn’t trust Tesla’s FSD on the roads around where I live in the United Kingdom. Just the adaptive cruise control on my Tesla is prone to jerks and sudden stops on the small highways in and around London. And even when Tesla’s FSD is competitive with Waymo’s cars from a driving experience standpoint, the firm will have fulfilled only one part of the robotaxi promise: the car. Operating a robotaxi fleet that deals with humans (forgetting their bags in the car, spilling coffee on the seats, and so on) is another layer of learning. My sense is that much of the deployment in the next few years will be robotaxi services from firms like Waymo and Baidu’s Apollo that have figured out the technology and the operations. I suspect that once robotaxis gain a reasonable market share in any particular city, it will take about 10 more years for autonomous vehicles to gain widespread adoption there. In truth, there is so much we don’t know about how these cars will be adopted in the social systems that are modern urban environments. From her forthcoming research, George Mason University’s Cummings tells me that between 2022 and 2023, 48 percent of all crashes from the main U.S. self-driving platforms occurred when the vehicles were rear-ended. For human drivers, only 29 percent of crashes are rear-enders. Is this a human problem or a robotaxi problem? Quite possibly it is both: Robotaxis may brake faster than a human driver’s reflexes. The regulatory environment will determine how long it takes each market to adopt self-driving technology and find answers to these hard questions. The China Society of Automotive Engineers’ 2030 prediction may come to pass, or it may be bluster. In the United States, we’re probably talking about a couple of decades before consumers are buying self-driving cars in meaningful numbers. Globally, it’ll be longer than that. Of course, entrepreneurs may carve up the transportation market in novel ways. For example, Glydways, backed by the famed venture capitalist Vinod Khosla and OpenAI CEO Sam Altman, is using autonomous vehicles to provide high-density mass transit in cities such as Atlanta. Other bold entrepreneurs are developing autonomous air taxis. We might start to see a broad diversity of autonomous systems popping up around the world. If there’s one thing I’ve learned from my pessimism in 2018 and 2022, it’s that things can change significantly and in a matter of only a few years. My view on robotaxis has flipped. They snuck up on me, and they’re now politely waiting to offer me a ride.

To Nick de Vries, chief technology officer at the solar-energy developer Silicon Ranch, a transformer is like an interstate on-ramp: It boosts the voltage of the electricity that his solar plants generate to match the voltage of grid transmission lines. “They’re your ticket to ride,” says de Vries. “If you don’t have your high-voltage transformer, you don’t have a project.” Recently, this ticket has grown much harder to come by. The demand for transformers has spiked worldwide, and so the wait time to get a new transformer has doubled from 50 weeks in 2021 to nearly two years now, according to a report from Wood MacKenzie, an energy-analytics firm. The wait for the more specialized large power transformers (LPTs), which step up voltage from power stations to transmission lines, is up to four years. Costs have also climbed by 60 to 80 percent since 2020. About five years ago, de Vries grew worried that transformer shortages would postpone his solar projects from coming online, so he began ordering transformers years before they’d actually be needed. Silicon Ranch, based in Nashville, now has a pipeline of custom transformers to make sure supply chain problems don’t stall its solar projects. The company isn’t alone in its quandary. A quarter of the world’s renewable-energy projects may be delayed while awaiting transformers to connect them to local grids, according to the Wood MacKenzie report. In India, the wait for 220-kilovolt transformers has leaped from 8 to 14 months, potentially holding up nearly 150 gigawatts of new solar development. And it’s not just renewable-energy projects. The transformer shortage touches utilities, homeowners, businesses, rail systems, EV charging stations—anyone needing a grid connection. In Clallam County, the part of Washington state where the Twilight movies are set, officials in May 2022 began to deny new home-construction requests because they couldn’t get enough pad-mounted transformers to step down voltage to homes. To address the backlog of customers who had already paid for new electrical service, the utility scrounged up refurbished transformers, or “ranch runners,” which helped but likely won’t last as long as new ones. The ripple effects of the shortage touch both public policy and safety. When a transformer fails from wear and tear, gets hit by a storm, or is damaged by war or sabotage, the inability to quickly replace it increases the risk of a power outage. The European Green Deal, which plans for an enormous build-out of Europe’s transmission network by 2030 to accelerate electrification, is imperiled by the protracted wait times for transformers, says Joannes Laveyne, an electrical engineer and energy-systems expert at Ghent University, in Belgium. For power engineers, this crisis is also an opportunity. They’re now reworking transformer designs to use different or less sought-after materials, to last longer, to include power electronics that allow the easy conversion between AC and DC, and to be more standardized and less customized than the transformers of today. Their innovations could make this critical piece of infrastructure not only more resistant to supply chain weaknesses, but also better suited to the power grids of the future. How Transformers Work A transformer is a simple thing—and an old one, too, invented in the 1880s. A typical one has a two-sided core made of iron or steel with copper wire wrapped around each side. The sets of wires, called windings, aren’t connected, but through electromagnetic induction across the core, current transfers from one coil to the other. By changing the number of times the wire wraps around each side of the core, engineers can change the voltage that emerges from the device so that it is higher or lower than what entered. This basic setup underlies transformers in a wide range of sizes. An LPT can weigh as much as two blue whales and might be used to step up the electricity that emerges from a fossil fuel or nuclear power plant—typically in the thousands of volts—to match the hundreds of thousands of volts running through transmission lines. When the electricity on those lines arrives at a city, it meets a power substation, which has transformers that step down the voltage to tens of thousands of volts for local distribution. Distribution transformers, which are smaller, decrease the voltage further, eventually to the hundreds of volts that can be used safely in homes and businesses. The simplicity of the design has been its strength, says Deepak Divan, an electrical engineer and director of the Georgia Tech Center for Distributed Energy. Transformers are big, bulky devices built to endure for decades. Their very durability shoulders the grid. But they’re a little like the gears and chain of a bicycle—adept at their simple conversion task, and little else. For example, traditional transformers that work only with AC can’t switch to DC without extra components. That AC-DC conversion is important because a host of technologies that aim to be a part of the cleaner energy future, including the electrolyzers that create hydrogen fuel, EV charging stations, and energy storage, all require lots of transformers, and they all need DC power. Solid-state power electronics, on the other hand, can seamlessly handle AC-DC conversions. “Wouldn’t it be nice to have a power-electronic replacement for the transformer?” Divan says. “It gives you control. And, in principle, it could become smaller if you really do it right.” The idea of a solid-state transformer has been kicking around in academia and industry for years. Divan and his team call their version a modular controllable transformer (MCT). It uses semiconductors and active electronic components to not only transform electricity to other voltages but also invert the current between DC and AC in a single stage. It’s also built with novel insulations and other measures to protect it from lightning strikes and power surges. Divan and his team received an award in 2023 from IEEE Transactions on Power Electronics for one of their designs. Divan’s modular transformer doesn’t have to be custom-built for each application, which could ease manufacturing bottlenecks. But as an emerging technology, it’s more expensive and fragile than a conventional transformer. For example, today’s semiconductors can’t survive electrical loads greater than about 1.7 kV. A device connected to the grid would need to endure at minimum 13 kV, which would mean stacking these transformer modules and hoping the whole group can withstand whatever the real world throws its way, Divan says. “If I have 10 converter modules stacked in series to withstand the high voltage, what happens if one fails? What happens if one of them gets a signal that is delayed by 200 nanoseconds? Does the whole thing collapse on you? These are all very interesting, challenging problems,” Divan says. Researchers at Oak Ridge National Laboratory’s GRID-C developed a next-generation transformer that is much smaller than previous generations and has the same capabilities. Alonda Hines/ORNL/U.S. Dept. of Energy At Oak Ridge National Laboratory’s Grid Research Integration and Deployment Center, or GRID-C, Madhu Chinthavali is also evaluating new technologies for next-gen transformers. Adding power electronics could enable transformers to manage power flow in ways that conventional ones cannot, which could in turn aid in adding more solar and wind power. It could also enable transformers to put information into action, such as instantaneously responding to an outage or failure on the grid. Such advanced transformers aren’t the right solution everywhere but using them in key places will help add more loads to the grid. Equipping them with smart devices that relay data would give grid operators better real-time information and increase overall grid resilience and durability, says Chinthavali, who directs GRID-C. New kinds of power-electronic transformers, if they can be made affordable and reliable, would be a breakthrough for solar energy, says Silicon Ranch’s de Vries. They would simplify the chore of regulating the voltage going from solar plants to transmission lines. At present, operators must do that voltage regulation constantly because of the variable nature of the sun’s energy—and that task wears down inverters, capacitors, and other components. Why Is There a Transformer Shortage? Driving the transformer shortage are market forces stemming from electricity demand and material supply chains. For example, nearly all transformer cores are made of grain-oriented electrical steel, or GOES—a material also used in electric motors and EV chargers. The expansion of those adjacent industries has intensified the demand for GOES and diverted much of the supply. On top of this, transformer manufacturing generally slowed after a boom period about 20 years ago. Hitachi Energy, Siemens Energy, and Virginia Transformers have announced plans to scale up production with new facilities in Australia, China, Colombia, Finland, Germany, Mexico, the United States, and Vietnam. But those efforts won’t ease the logjam soon. At the same time, the demand for transformers has skyrocketed over the last two years by as much as 70 percent for some U.S. manufacturers. Global demand for LPTs with voltages over 100 kV has grown more than 47 percent since 2020, and is expected to increase another 30 percent by 2030, according to research by Wilfried Breuer, managing director of German electrical equipment manufacturer Maschinenfabrik Reinhausen, in Regensburg. Aging grid infrastructure, new renewable-energy generation, expanding electrification, increased EV charging stations, and new data centers all contribute to the rising demand for these machines. Compounding the problem is that a typical LPT doesn’t just roll off an assembly line. Each is a bespoke creation, says Bjorn Vaagensmith, a power-systems researcher at Idaho National Laboratory. In this low-volume industry, “a factory will make maybe 50 of these things a year,” he says. The LPT’s design is dictated by the layout of the substation or power plant it serves, as well as the voltage needs and the orientation of the incoming and outgoing power lines. For example, the bushings, which are upward-extending arms that connect the transformer to power lines, must be built in a particular position to intercept the lines. Such customization slows manufacturing and increases the difficulty of replacing a failed transformer. It’s also the reason why many energy companies don’t order LPTs ahead of time, says Laveyne at Ghent. “Imagine you get the transformer delivered but the permitting process ends up in a stall, or delay, or even a cancellation [of the project]. Then you’re stuck with a transformer you can’t really use.” GE Vernova Advanced Research developed a flexible large power transformer that it has been field-testing at a substation in Columbia, Miss., since 2021. Cooperative Energy Less customized, more one-size-fits-all transformers could ease supply chain problems and reduce power outages. To that end, a team at GE Vernova Advanced Research (GEVAR) helped develop a “flexible LPT.” In 2021, the team began field-testing a 165-kV version at a substation operated by Cooperative Energy in Mississippi, where it remains active. Ibrahima Ndiaye, a senior principal engineer at GEVAR who led the project, says the breakthrough was figuring out how to give a conventional transformer the capability to change its impedance (that is, its resistance to electricity flow) without changing any other feature in the transformer, including its voltage ratio. Impedance and voltage ratio are both critical features of a transformer that ordinarily must be tailored to each use case. If you can tweak both factors independently, then you can modify the transformer for various uses. But altering the impedance without also changing the transformer’s voltage ratio initially seemed impossible, Ndiaye says. The solution turned out to be surprisingly straightforward. The engineer added the same amount of windings to both sides of the transformer’s core, but in opposite directions, cancelling out the voltage increase and thereby allowing him to tweak one factor without automatically changing the other. “There is no [other] transformer in the world that has a capability of that today,” Ndiaye says. The flexible LPT could work like a universal spare, filling in for LPTs that fail, and negating the need to keep a custom spare for every transformer, Ndiaye says. This in turn would reduce the demand for these types of transformers and crucial materials such as GOES. The flexible LPT also lets the grid operate reliably even when there are variable renewable resources, or large variable loads such as a bank of EV charging stations. Avangrid’s mobile transformer has multivoltage capabilities and can be trucked to any of Avangrid’s onshore solar or wind projects within a couple of months. Hitachi Energy and Avangrid Similarly, Siemens Energy has been developing what it calls “rapid response transformers”—plug-and-play backups that could replace a busted transformer within weeks. And the renewable-energy company Avangrid this year introduced a mobile transformer that can be trucked to any of its solar or wind projects within a couple of months. Transformers Designed for Longevity There is room to improve, rather than replace, the century-old design of the traditional transformer, says Stefan Tenbohlen, an energy researcher at the University of Stuttgart, in Germany. He cofounded the University Transformer Research Alliance, to connect international researchers who are tinkering with conventional designs. A chief goal is to make sure new transformers last even longer than the older generation did. One approach is to try different insulation techniques. Copper windings are typically insulated by paper and mineral oil to protect them from overheating. New approaches replace the mineral oil with natural esters to allow the interior of the transformer to safely reach higher temperatures, prolonging the device’s lifespan in the process. Vaagensmith at Idaho National Lab has experimented with ceramic paper—a thin, lightweight, ultra-heat-resistant material made of alumina silicate fibers—as insulation. “We cooked it up to a thousand degrees Celsius, which is ridiculously high for a transformer, and it was fine,” he says. Researchers at Oak Ridge National Laboratory built hollow transformer cores made of electrical steel using additive manufacturing. Alex Plotkowski/ORNL Changing other materials used in LPTs could also help. Hollow-core transformers, for example, use far less steel. Scientists at Oak Ridge, in Tennessee, have been testing 3D printing of hollow cores made of electrical steel. Switching to hollow cores and being able to 3D print them would ease demand for the material in the United States, where there’s just one company that produces GOES steel for transformers, according to a 2022 report from the U.S. Department of Energy. Transformer Industry Faces Capacity Crunch Transformer manufacturing used to be a cyclical business where demand ebbed and flowed—a longstanding pattern that created an ingrained way of thinking. Consequently, despite clear signs that electrical infrastructure is set for a sustained boom and that the old days aren’t coming back, many transformer manufacturers have been hesitant to increase capacity, says Adrienne Lotto, senior vice president of grid security, technical, and operations services for the American Public Power Association, in Arlington, Va. She sums up their attitude: “If the demand is again going to simply fall off, why invest millions of dollars’ worth of capital into your manufacturing facility?” But greater demand for electricity is coming. The recent book Energy 2040 (Springer), coauthored by Georgia Tech’s Divan, lays out some of the staggering numbers. The capacity of all the energy projects waiting to connect to the U.S. grid amounts to 2,600 GW—more than double the nation’s entire generation capacity currently. An average estimate of U.S. EV adoption suggests the country will have 125 million EVs by 2040. The electricity demands of U.S. data centers may double by the end of this decade because of the boom in artificial intelligence. The National Renewable Energy Lab found that U.S. transformer capacity will need to increase by as much as 260 percent by 2050 to handle all the extra load. Globally, electricity supplied 20 percent of the world’s energy needs in 2023, and may reach 30 percent by 2030 as countries turn to electrification as a way to decarbonize, according to the International Energy Agency. India and China are expected to see the fastest demand growth in that time. India installed more solar capacity in the first quarter of 2024 than in any quarter previously, and yet, as mentioned, the wait time to get those solar projects running is growing because of the transformer shortage. The world’s power systems are not accustomed to such upheaval, Divan says. Because longstanding technologies like the transformer change so slowly, utilities spend very little—perhaps 0.1 percent of their budgets—on R&D. But they must prepare for a sea change, Divan says. “Utilities are not going to be able to stop this tsunami that’s coming. And the pressure is on.”

The semiconductor industry’s long held imperative—Moore’s Law, which dictates that transistor densities on a chip should double roughly every two years—is getting more and more difficult to maintain. The ability to shrink down transistors, and the interconnects between them, is hitting some basic physical limitations. In particular, when copper interconnects are scaled down, their resistivity skyrockets, which decreases how much information they can carry and increases their energy draw. The industry has been looking for alternative interconnect materials to prolong the march of Moore’s Law a bit longer. Graphene is a very attractive option in many ways: The sheet-thin carbon material offers excellent electrical and thermal conductivity, and is stronger than diamond. However, researchers have struggled to incorporate graphene into mainstream computing applications for two main reasons. First, depositing graphene requires high temperatures that are incompatible with traditional CMOS manufacturing. And second, the charge carrier density of undoped, macroscopic graphene sheets is relatively low. Now, Destination 2D, a startup based in Milpitas, Calif., claims to have solved both of those problems. Destination 2D’s team has demonstrated a technique to deposit graphene interconnects onto chips at 300 °C, which is still cool enough to be done by traditional CMOS techniques. They have also developed a method of doping graphene sheets that offers current densities 100 times as dense as copper, according to Kaustav Banerjee, co-founder and CTO of Destination 2D. “People have been trying to use graphene in various applications, but in the mainstream micro-electronics, which is essentially the CMOS technology, people have not been able to use this so far,” Banerjee says. Destination 2D is not the only company pursuing graphene interconnects. TSMC and Samsung are also working to bring this technology up to snuff. However, Banerjee claims, Destination 2D is the only company to demonstrate graphene deposition directly on top of transistor chips, rather than growing the interconnects separately and attaching them to the chip after the fact. Depositing graphene at low temperature Graphene was first isolated in 2004, when researcher separated sheets of graphene by pulling them off graphite chunks with adhesive tape. The material was deemed so promising that in 2010 the feat garnered a Nobel prize. (Nobel Prize co-recipient Konstantin Novoselov is now Destination 2D’s chief scientist). Startup Destination 2D has developed a CMOS-compatible tool capable of depositing graphene interconnects at the wafer scale.Destination 2D However, carefully pulling graphene off of pencil tips using tape is not exactly a scalable production method. To reliably create graphene structures, researchers have turned to chemical vapor deposition, where a carbon gas is deposited onto a heated substrate. This typically requires temperatures well above the roughly 400 °C maximum operating temperature in CMOS manufacturing. Destination 2D uses a pressure-assisted direct deposition technique developed in Banerjee’s lab at the University of California, Santa Barbara. The technique, which Banerjee calls pressure-assisted solid phase diffusion, uses a sacrificial metal film such as nickel. The sacrificial film is placed on top of the transistor chip, and a source of carbon is deposited on top. Then, using a pressure of roughly 410 to 550 kilopascals (60 to 80 pounds per square inch), the carbon is forced through the sacrificial metal, and recombines into clean multilayer graphene underneath. The sacrificial metal is then simply removed, leaving the graphene on-chip for patterning. This technique works at 300 °C, cool enough to not damage the transistors underneath. Boosting Graphene’s Current Density After the graphene interconnects are patterned, the graphene layers are doped to reduce the resistivity and boost their current-carrying capacity. The Destination 2D team uses a doping technique called intercalation, where the doping atoms are diffused between graphene sheets. The doping atoms can vary—examples include iron chloride, bromine, and lithium. Once implanted, the dopants donate electrons (or their in-material counterparts, electron holes) to the graphene sheets, allowing higher current densities. “Intercalation chemistry is a very old subject,” Banerjee says. “We are just bringing this intercalation into the graphene, and that is new.” This technique has a promising feature—unlike copper, as the graphene interconnects are scaled down, their current-carrying capacity improves. This is because for thinner lines, the intercalation technique becomes more effective. This, Banerjee argues, will allow their technique to support many generations of semiconducting technology into the future. Destination 2D has demonstrated their graphene interconnect technique at the chip level, and they’ve also developed tools for wafer-scale deposition that can be implemented in fabrication facilities. They hope to work with foundries to implement their technology for research and development, and eventually, production.

The global technology community came together on 1 October to celebrate the 15th anniversary of IEEE Day, which commemorates the first time professionals gathered to share their technological innovations and advancements. IEEE Day is held annually on the first Tuesday of October. This year’s celebration was the largest yet, with more than 1,100 events worldwide. The events provided professionals, students, and enthusiasts a platform to connect, collaborate, and showcase their work as well as have some fun. The spirit of innovation and community was evident as participants engaged in workshops, networking sessions, and more. In addition to the events, 11 IEEE societies ran membership offers between 29 September and 12 October as a way to encourage professionals and students to join the IEEE community and learn more about their fields of interest. Photography and video contests were held as well. More than 300 images and 140 videos were entered. The entries captured the essence of the IEEE Day events, showcasing participants’ creativity and enthusiasm. The contests also provided a visual record of the day’s global impact. Standout events This year’s IEEE Day was filled with memorable events highlighting the spirit of innovation and community. Here are a few standouts: Technical session. At the Vishwakarma Institute of Technology, in Pune, India, about 50 students attended an AI Unlocked session organized by the IEEE Instrumentation and Measurement Society’s Pune chapter. Artificial intelligence specialist Jayesh Pingale and other experts discussed the technology’s applications. The event included projects that showcased the use of AI in interactive learning and concluded with a ceremony celebrating IEEE Day. “Hundreds of entries were submitted to our contests, showcasing the outstanding creativity of IEEE members and the true joy of this celebration.” —Cybele Ghanem Career fair. The event at Florida Polytechnic University, in Lakeland, included a career fair. IEEE Senior Member Andy Seely, Florida West Coast Section vice chair, discussed volunteer opportunities. Y.C. Wang, director of DigiKey’s global academic program, spoke about the company’s history and its suite of online tools for engineering students. IEEE Member Mohammad Reza Khalghani, a professor of electrical and computer engineering at the university, gave a talk on cyber-physical security. He covered microgrids, network control systems, and AI techniques for cyber anomaly detection. The event attracted more than 80 people. Competitions and games. At the Christ (Deemed to be University) engineering and technology school, in Bangalore, India, contests were held to add a fun, competitive edge to the celebrations. There was a quiz on IEEE history, as well as a chess tournament and an Uno card game. The activities tested participants’ knowledge and strategic thinking while providing a relaxed networking atmosphere. “This year IEEE Day celebrated its 15th anniversary, which proved to be a remarkable celebration,” says Member Cybele Ghanem, IEEE Day committee chair. “Gathered under the theme Leveraging Technology for a Better Tomorrow, IEEE members worldwide celebrated IEEE Day,” Ghanem says. “Hundreds of entries were submitted to our contests, showcasing the outstanding creativity of IEEE members and the true joy of this celebration. Thank you for celebrating with us.” Visit the IEEE Day web page and follow IEEE Day on Instagram, Facebook, LinkedIn, or IEEE Collabratec to stay updated throughout the year.

In an age where data is bought and sold as a commodity, true privacy is rare. But homomorphic encryption can protect your data completely, so no one, not even the servers used to process it, can read your information. Here’s how it works: A device encrypts data, sends it out for processing, computations are done on the encrypted data, and then the data is decrypted upon return. A mathematically complex process ensures that your processed data can be decrypted at the end without anyone being able to decode it in the middle. However, the computational power required for the underlying mathematics that enable homomorphic encryption are too much for the Internet of Things as it currently is. A team of engineers at Peking University, in Beijing, China aim to change that. Their new device, created using arrays of ferroelectric field effect transistors (FeFET), is optimized to carry out the encryption and decryption processes with high accuracy and low computational load. The engineers unveiled the array today at the 2024 IEEE International Electron Devices Meeting. “By implementing novel semiconductor devices, we can have our commercial electronics like cell phones utilize the computing power of the cloud [while] also keeping the safety of our data,” says Kechao Tang, assistant professor of integrated circuits at Peking University and one of the researchers who developed the new system. Math Inside a Transistor To carry out the homomorphic encryption process, a computer must be able to generate a random key, which will be used to encrypt and then later to “unlock” the data. It then uses that key to carry out polynomial multiplication and addition that puts the data in an encrypted form for processing. To create a key for encryption, the transistor array uses fluctuations in current through the FeFETs. FeFETs can be engineered to have a much higher degree of fluctuation than a regular MOSFET transistor, so the random number generated by the device is less predictable than what you’d get from an ordinary silicon chip, making the encryption harder to crack. For the encryption process, the key helps convert the user’s data into a vector consisting of the coefficients of polynomials. That vector is then multiplied by a matrix of numbers and then by another vector. So encryption usually takes two steps, but in the FeFET array, it can be done in just one. That’s possible because of the nature of FeFETs. In the part of the transistor that controls the flow of current through the device, the gate, they have a layer of ferroelectric—a material that holds an electric polarization without needing to be in an electric field. The ferroelectric layer can store data as the magnitude of this polarization. Like other transistors, FeFETs have three terminals: the drain, the source, and the gate. Counting the stored state in the ferroelectric material, this means three signals can be combined in an FeFET: the drain input, gate input, and the stored state. (The source provides the output current.) So one FeFET can be made to compute a three-input multiplication. When many FeFETs are combined into an array, the array can now take in the three sets of data needed for encryption: a vector of the data to be encrypted and the encryption matrix and vector. The matrix is stored in the FeFET array’s ferroelectric layer, the vector of original data is inputted to the gate of each FeFET, and the second vector is input to the drains of the FeFET array. In one step, the FeFET array combines the signals of the vector, matrix, and vector together, then outputs the final encrypted data as current. “We can do more efficient computing with less area overhead and also with less power consumption,” says Tang. Researchers are also trying to use RRAM to accomplish the matrix multiplication required for homomorphic encryption, because it also has the ability to store a state in memory. However, ferroelectric devices should produce less noise in the decryption process than RRAM would, according to Tang. Because the ferroelectric devices have a greater difference between their on and off states than RRAM, “you are less likely to have mistakes when you do the encoding and decoding,” says Tang, “because you can easily tell whether it is one or zero.” Previous RRAM solutions had accuracies between 97.1 and 98.8 percent, while this device had an accuracy of 99.6 percent. In the future, Tang hopes to see this technology in our smartphones. “If we can apply our device into the cellphone, it means that our cellphone will have the ability to encode the data to be uploaded to the cloud and then get it back and then decode it,” he says.

When Meta announced last week that it’s looking for a nuclear energy developer to power its future AI operations, it joined a growing cadre of tech companies all suddenly repeating the same refrain: We need more power—preferably carbon-free—and lots of it. Electricity demand in the United States is expected to grow more than 15 percent over the next five years after remaining flat for the last two decades, according to a recent report from power sector consulting firm Grid Strategies. Most of the growth will be driven by the needs of data centers and their operators, who are scrambling to secure large amounts of reliable power while keeping their carbon neutral goals. Nuclear energy fits that bill, and over the last few months, Amazon, Google, and Microsoft have all announced ambitious deals to acquire it for their operations. Some of the plans aim to secure energy in the near term from existing power plants. Others focus on the long game and include investments in next-generation nuclear energy and small modular reactors (SMRs) that don’t yet exist on a commercial scale. “Data centers have grown in size and AI is dramatically changing the future [energy] forecast,” says Dan Stout, founder of Advanced Nuclear Advisors in Chattanooga, Tenn. “In the 2030s, the grid will have less coal and there will be some constraints on gas. So nuclear energy’s power density and carbon-free high reliability is attractive, and tech companies are starting to take action on new nuclear deployments,” he says. Big Tech Turns Its Attention to Nuclear Power Amazon kicked off the bevy of public announcements in March when it bought a data center adjacent to a nuclear power plant in Pennsylvania. The purchase came with 300 megawatts of behind-the-meter electricity. After closing the deal, Amazon requested another 180 MW. The request caused a dustup over energy fairness, and in November regulators rejected it, leaving Amazon looking for other options. Tech companies are watching the precedent-setting situation closely. Meanwhile, Microsoft was inking an agreement with Constellation Energy to restart a shuttered nuclear reactor on Three Mile Island—the site of the worst nuclear disaster in U.S. history. The plan, announced in September, calls for the reactor to supply 835 MW to grid operator PJM, and for Microsoft to buy enough of that power to match the electricity consumed by its data centers in the PJM Interconnection. Then in October, just two days apart, Google and Amazon both announced investments in startups developing SMRs. The smaller size and modular design of SMRs could make building them faster, cheaper and more predictable than conventional nuclear reactors. They also come with enhanced safety features, and could be built closer to transmission lines. SMRs are still at least five years from commercial operation in the United States. A year ago the first planned SMR in the United States was cancelled due to rising costs and a lack of customers. (China is building an SMR called the Linglong One on the island of Hainan, which is scheduled to be operational in 2026.) To move things along, Amazon led a US $500 million financing round to support X-energy in Rockville, Md., which is developing a gas-cooled SMR. The financing will help X-energy finish its reactor design and build a nuclear fuel fabrication facility. The plan is to build multiple SMRs producing at least 5 GW total by 2039. Each reactor will provide 80 MW of electricity. Google, for its part, is backing Kairos Power with a 500 MW development agreement. The Alameda, Calif.-based company is developing a molten fluoride salt-cooled SMR and has received construction permits from the U.S. Nuclear Regulatory Commission to build two demonstration facilities, both in Oak Ridge, Tenn. The company says the facilities will be operational by 2030. TRISO Fuel Promises to Shrink Reactors The reactors that both Kairos and X-energy are developing run on tri-structural isotropic (TRISO) particle fuel. It’s made of uranium, carbon, and oxygen encapsulated in graphite kernels the size of a poppy seed. The kernels get loaded into golf ball-size spheres called pebbles that are also made of graphite. Each pebble contains thousands of fuel kernels. The structure of the pebble encapsulation enables the fuel to withstand very high temperatures, so even in worst-case accidents, the pebbles won’t melt. The coatings “essentially provide the key safety functions that the large containment concrete structure is providing for conventional reactor technologies,” says Mike Laufer, co-founder of Kairos. If regulators approve, the built-in containment feature could shrink the footprint of nuclear plants by reducing the size of containment structures. The U.S. Department of Energy has been developing and extensively testing TRISO fuel over the last two decades. Kairos will use TRISO fuel in its high-temperature, low-pressure, fluoride salt-cooled reactor. In this design, fuel pebbles in the reactor core undergo fission, generating heat that transfers to the surrounding molten salt. Heat exchangers transfer the heat to boil water and generate steam, which drives a turbine and generates electricity. The molten salt acts as an additional safety barrier, chemically absorbing any fission products that escape the pebbles, Laufer says. Kairos’ commercial reactors will each generate about 75 MW of electricity, Laufer says. X-energy plans to use TRISO fuel is its high-temperature gas-cooled reactor. In this design, helium gas runs through the reactor core. As the fuel pebbles undergo fission, the gas extracts the heat, which is used to boil water and generates steam to drive a turbine. Each fuel pebble will constantly shuffle through the reactor, passing through about six times. “The reactor is a lot like a gumball machine,” says Benjamin Reinke, vice president of global business development at X-energy. A mechanical corkscrew drives a pebble in an auger out of the system., and the pebble is checked to see if it’s fully burned up. If not, it goes back to into the top of the reactor, he says. X-energy is working on getting a license to produce TRISO fuel on a commercial scale at a facility it plans to build in Oak Ridge. The company’s first customer, a Dow petrochemical plant in Seadrift, Tex., plans to replace its gas boilers with X-energy’s SMRs, which will create steam and electricity for the plant and possibly for the grid. X-energy’s deal with Amazon also supports a four-unit, 320-MW project with regional utility Energy Northwest in Richland, Wash. Tech companies for the last decade have been investing in wind and solar energy too, but the power from these sources is intermittent, and may not be enough to meet the needs of power-guzzling AI. The arrangements between big tech and small nuclear signal the beginning of a trend, says Stout. Meta’s announcement last week that it’s putting out a request for proposals for up to 4 gigawatts of nuclear power may be the most recent addition to that trend, but it’s probably not the last. Says Stout: “I expect there’s going to be more.” This article was updated on 10 December 2024 to clarify the agreement between Google and Kairos Power.

Farming in India is tough work—and it’s only getting tougher. Water shortages, a rapidly changing climate, disorganized supply chains, and difficulty accessing credit make every growing season a calculated gamble. But farmers like Harish B. are finding that new AI-powered tools can take some of the unpredictability out of the endeavor. (Instead of a surname, Indian given names are often combined with initials that can represent the name of the person’s father or village.) The 40-year-old took over his family’s farm on the outskirts of Bengaluru, in southern India, 10 years ago. His father had been farming the 5.6-hectare plot since 1975 and had shifted from growing vegetables to grapes in search of higher profits. Since taking over, Harish B. has added pomegranates and made a concerted effort to modernize their operations, installing drip irrigation and mist blowers for applying agricultural chemicals. Then, a year and a half ago, he started working with the Bengaluru-based startup Fasal. The company uses a combination of Internet of Things (IoT) sensors, predictive modeling, and AI-powered farm-level weather forecasts to provide farmers with tailored advice, including when to water their crops, when to apply nutrients, and when the farm is at risk of pest attacks. Harish B. uses Fasal’s modeling to make decisions about irrigation and the application of pesticides and fertilizer. Edd Gent Harish B. says he’s happy with the service and has significantly reduced his pesticide and water use. The predictions are far from perfect, he says, and he still relies on his farmer’s intuition if the advice doesn’t seem to stack up. But he says that the technology is paying for itself. “Before, with our old method, we were using more water,” he says. “Now it’s more accurate, and we only use as much as we need.” He estimates that the farm is using 30 percent less water than before he started with Fasal. Indian farmers who are looking to update their approach have an increasing number of options, thanks to the country’s burgeoning “agritech” sector. A host of startups are using AI and other digital technologies to provide bespoke farming advice and improve rural supply chains. And the Indian government is all in: In 2018, the national government has declared agriculture to be one of the focus areas of its AI strategy, and it just announced roughly US $300 million in funding for digital agriculture projects. With considerable government support and India’s depth of technical talent, there’s hope that AI efforts will lift up the country’s massive and underdeveloped agricultural sector. India could even become a testbed for agricultural innovations that could be exported across the developing world. But experts also caution that technology is not a panacea, and say that without careful consideration, the disruptive forces of innovation could harm farmers as much as they help. How AI is helping India’s small farms India is still a deeply agrarian society, with roughly 65 percent of the population involved in agriculture. Thanks to the “green revolution” of the 1960s and 1970s, when new crop varieties, fertilizers, and pesticides boosted yields, the country has long been self-sufficient when it comes to food—an impressive feat for a country of 1.4 billion people. It also exports more than $40 billion worth of foodstuffs annually. But for all its successes, the agricultural sector is also extremely inefficient. Roughly 80 percent of India’s farms are small holdings of less than 2 hectares (about 5 acres), which makes it hard for those farmers to generate enough revenue to invest in equipment and services. Supply chains that move food from growers to market are also disorganized and reliant on middlemen, a situation that eats into farmers’ profits and leads to considerable wastage. These farmers have trouble accessing credit because of the small size of their farms and the lack of financial records, and so they’re often at the mercy of loan sharks. Farmer indebtedness has reached worrying proportions: More than half of rural households are in debt, with an average outstanding amount of nearly $900 (the equivalent of more than half a year’s income). Researchers have identified debt as the leading factor behind an epidemic of farmer suicides in India. In the state of Maharashtra, which leads the country in farmer suicides, 2,851 farmers committed suicide in 2023. While technology won’t be a cure-all for these complex social problems, Ananda Verma, founder of Fasal, says there are many ways it can make farmers’ lives a little easier. His company sells IoT devices that collect data on crucial parameters including soil moisture, rainfall, atmospheric pressure, wind speed, and humidity. This data is passed to Fasal’s cloud servers, where it’s fed into machine learning models, along with weather data from third parties, to produce predictions about a farm’s local microclimate. Those results are input into custom-built agronomic models that can predict things like a crop’s water requirements, nutrient uptake, and susceptibility to pests and disease. “What is being done in India is sort of a testbed for most of the emerging economies.” —Abhay Pareek, Centre for the Fourth Industrial Revolution The output of these models is used to advise the farmer on when to water or when to apply fertilizer or pesticides. Typically, farmers make these decisions based on intuition or a calendar, says Verma. But this can lead to unnecessary application of chemicals or overwatering, which increases costs and reduces the quality of the crop. “[Our technology] helps the farmer make very precise and accurate decisions, completely removing any kind of guesswork,” he says. Fasal’s ability to provide these services has been facilitated by a rapid expansion of digital infrastructure in India, in particular countrywide 4G coverage with rock-bottom data prices. The number of smartphone users has jumped from less than 200 million a decade ago to over a billion today. “We are able to deploy these devices in rural corners of India where sometimes you don’t even find roads, but there is still Internet,” says Verma. Reducing water and chemical use on farms can also ease pressure on the environment. An independent audit found that across the roughly 80,000 hectares where Fasal is currently operating, it has helped save 82 billion liters of water. The company has also saved 54,000 tonnes of greenhouse gas emissions produced by running-water pumps, and reduced chemical usage by 127 tonnes. Problems with access and trust However, getting these capabilities into the hands of more farmers will be tricky. Harish B. says some smaller farmers in his area have shown interest in the technology, but they can’t afford it (neither the farmers nor the company would disclose the product’s price). Taking full advantage of Fasal’s advice also requires investment in other equipment like automated irrigation, putting the solution even further out of reach. Verma says farming cooperatives could provide a solution. Known as farmer producer organizations, or FPOs, they provide a legal structure for groups of small farmers to pool their resources, boosting their ability to negotiate with suppliers and customers and invest in equipment and services. In reality, though, it can be hard to set up and run an FPO. Harish B. says some of his neighbors attempted to create an FPO, but they struggled to agree on what to do, and it was ultimately abandoned. Cropin’s technology combines satellite imagery with weather data to provide customized advice. Cropin Other agritech companies are looking higher up the food chain for customers. Bengaluru-based Cropin provides precision agriculture services based on AI-powered analyses of satellite imagery and weather patterns. Farmers can use the company’s app to outline the boundaries of their plot simply by walking around with their smartphone’s GPS enabled. Cropin then downloads satellite data for those coordinates and combines it with climate data to provide irrigation advice and pest advisories. Other insights include analyses of how well different plots are growing, yield predictions, advice on the optimum time to harvest, and even suggestions on the best crops to grow. But the company rarely sells its services directly to small farmers, admits Praveen Pankajakshan, Cropin’s chief scientist. Even more than cost, the farmer’s ability to interpret and implement the advice can be a barrier, he says. That’s why Cropin typically works with larger organizations like development agencies, local governments, or consumer-goods companies, which in turn work with networks of contract farmers. These organizations have field workers who can help farmers make sense of Cropin’s advisories. Working with more-established intermediaries also helps solve a major problem for agritech startups: establishing trust. Farmers today are bombarded with pitches for new technology and services, says Pankajakshan, which can make them wary. “They don’t have problems in adopting technology or solutions, because often they understand that it can benefit them,” he says. “But they want to know that this has been tried out and these are not new ideas, new experiments.” That perspective rings true to Harish C.S., who runs his family’s 24-hectare fruit farm north of Bengaluru. He’s a customer of Fasal and says the company’s services are making an appreciable difference to his bottom line. But he’s also conscious that he has the resources to experiment with new technology, a luxury that smaller farmers don’t have. Harish C.S. says Fasal’s services are making his 24-hectare fruit farm more profitable.Edd Gent A bad call on what crop to plant or when to irrigate can lead to months of wasted effort, says Harish C.S., so farmers are cautious and tend to make decisions based on recommendations from trusted suppliers or fellow farmers. “People would say: ‘On what basis should I apply that information which AI gave?’” he says. “‘Is there a proof? How many years has it worked? Has it worked for any known, reputable farmer? Has he made money?’” While he’s happy with Fasal, Harish C.S. says he relies even more on YouTube, where he watches videos from a prominent pomegranate growing expert. For him, technology’s ability to connect farmers and help them share best practices is its most powerful contribution to Indian agriculture. Chatbots for farmers Some are betting that AI could help farmers with that knowledge-sharing. The latest large language models (LLMs) provide a powerful new way to analyze and organize information, as well as the ability to interact with technology more naturally via language. That could help unlock the deep repositories of agricultural know-how shared by India’s farmers, says Rikin Gandhi, CEO of Digital Green, an international nonprofit that uses technology to help smallholders, or owners of small farms. The nonprofit Digital Green records videos about farmers’ solutions to their problems and shows them in villages. Digital Green Since 2008, the organization has been getting Indian farmers to record short videos explaining problems they faced and their solutions. A network of workers then tours rural villages putting on screenings. A study carried out by researchers at MIT’s Poverty Action Lab found that the program reduces the cost of getting farmers to adopt new practices from roughly $35 (when workers traveled to villages and met with individual farmers) to $3.50. But the organization’s operations were severely curtailed during the COVID-19 pandemic, prompting Digital Green to experiment with simple WhatsApp bots that direct farmers to relevant videos in a database. Two years ago, it began training LLMs on transcripts of the videos to create a more sophisticated chatbot that can provide tailored responses. Crucially, the chatbot can also incorporate personalized information, such as the user’s location, local weather, and market data. “Farmers don’t want to just get the generic Wikipedia, ChatGPT kind of answer,” Gandhi says. “They want very location-, time-specific advice.” Two years ago, Digital Green began working on a chatbot trained on the organization’s videos about farming solutions. Digital Green But simply providing farmers with advice through an app, no matter how smart it is, has its limits. “Information is not the only thing people are looking for,” says Gandhi. “They’re looking for ways that information can be connected to markets and products and services.” So for the time being, Digital Green is still relying on workers to help farmers use the chatbot. Based on the organization’s own assessments, Gandhi thinks the new service could cut the cost of adopting new practices by another order of magnitude, to just 35 cents. The downsides of AI for agritech Not everyone is sold on AI’s potential to help farmers. In a 2022 paper, ecological anthropologist Glenn Stone argued that the penetration of big data technologies into agriculture in the global south could hold risks for farmers. Stone, a scholar in residence at Washington and Lee University, in Virginia, draws parallels between surveillance capitalism, which uses data collected about Internet users to manipulate their behavior, and what he calls surveillance agriculture, which he defines as data-based digital technologies that take decision-making away from the farmer. The main concern is that these kinds of tools could erode the autonomy of farmers and steer their decision-making in ways that may not always help. What’s more, Stone says, the technology could interfere with existing knowledge-sharing networks. “There is a very real danger that local processes of agricultural learning, or ‘skilling,’ which are always partly social, will be disrupted and weakened when decision-making is appropriated by algorithms or AI,” he says. Another concern, says Nandini Chami, deputy director of the advocacy group IT for Change, is who’s using the AI tools. She notes that big Indian agritech companies such as Ninjacart, DeHaat, and Crofarm are focused on using data and digital technologies to optimize rural supply chains. On the face of it, that’s a good thing: Roughly 10 percent of fruits and vegetables are wasted after harvest, and farmers’ profits are often eaten up by middlemen. But efforts to boost efficiencies and bring economies of scale to agriculture tend to primarily benefit larger farms or agribusiness, says Chami, often leaving smallholders behind. Both in India and elsewhere, this is driving a structural shift in the economy as rural jobs dry up and people move to the cities in search of work. “A lot of small farmers are getting pushed out of agriculture into other occupations,” she says. “But we don’t have enough high-quality jobs to absorb them.” Can AI revamp rural supply chains? AI proponents say that with careful design, many of these same technologies can be used to help smaller farmers too. Purushottam Kaushik, head of the World Economic Forum’s Centre for the Fourth Industrial Revolution (C4IR), in Mumbai, is leading a pilot project that’s using AI and other digital technologies to streamline agricultural supply chains. It is already boosting the earnings of 7,000 chili farmers in the Khammam district in the state of Telangana. In the state of Telangana, AI-powered crop quality assessments have boosted farmers’ profits. Digital Green Launched in 2020 in collaboration with the state government, the project combined advice from Digital Green’s first-generation WhatsApp bot with AI-powered soil testing, AI-powered crop quality assessments, and a digital marketplace to connect farmers directly to buyers. Over 18 months, the project helped farmers boost yields by 21 percent and selling prices by 8 percent. One of the key lessons from the project was that even the smartest AI solutions don’t work in isolation, says Kaushik. To be effective, they must be combined with other digital technologies and carefully integrated into existing supply chains. In particular, the project demonstrated the importance of working with the much-maligned middlemen, who are often characterized as a drain on farmers’ incomes. These local businessmen aren’t merely traders; they also provide important services such as finance and transport. Without those services, agricultural supply chains would grind to a halt, says Abhay Pareek, who leads C4IR’s agriculture efforts. “They are very intrinsic to the entire ecosystem,” he says. “You have to make sure that they are also part of the entire process.” The program is now being expanded to 20,000 farmers in the region. While it’s still early days, Pareek says, the work could be a template for efforts to modernize agriculture around the world. With India’s huge diversity of agricultural conditions, a large proportion of smallholder farmers, a burgeoning technology sector, and significant government support, the country is the ideal laboratory for testing technologies that can be deployed across the developing world, he says. “What is being done in India is sort of a testbed for most of the emerging economies,” he adds. Dealing with data bottlenecks As with many AI applications, one of the biggest bottlenecks to progress is data access. Vast amounts of important agricultural information are locked up in central and state government databases. There’s a growing recognition that for AI to fulfill its potential, this data needs to be made accessible. Telangana’s state government is leading the charge. Rama Devi Lanka, director of its emerging technologies department, has spearheaded an effort to create an agriculture data exchange. Previously, when companies came to the government to request data access, there was a torturous process of approvals. “It is not the way to grow,” says Lanka. “You cannot scale up like this.” So, working with the World Economic Forum, her team has created a digital platform through which vetted organizations can sign up for direct access to key agricultural data sets held by the government. The platform has also been designed as a marketplace, which Lanka envisages will eventually allow anyone, from companies to universities, to share and monetize their private agricultural data sets. India’s central government is looking to follow suit. The Ministry of Agriculture is developing a platform called Agri Stack that will create a national registry of farmers and farm plots linked to crop and soil data. This will be accessible to government agencies and approved private players, such as agritech companies, agricultural suppliers, and credit providers. The government hopes to launch the platform in early 2025. But in the rush to bring data-driven techniques to agriculture, there’s a danger that farmers could get left behind, says IT for Change’s Chami. Chami argues that the development of Agri Stack is driven by misplaced techno-optimism, which assumes that enabling digital innovation will inevitably lead to trickle-down benefits for farmers. But it could just as easily lead to e-commerce platforms replacing traditional networks of traders and suppliers, reducing the bargaining power of smaller farmers. Access to detailed, farm-level data without sufficient protections could also result in predatory targeting by land sharks or unscrupulous credit providers, she adds. The Agri Stack proposal says access to individual records will require farmer consent. But details are hazy, says Chami, and it’s questionable whether India’s farmers, who are often illiterate and not very tech-savvy, could give informed consent. And the speed with which the program is being implemented leaves little time to work through these complicated problems. “[Governments] are looking for easy solutions,” she says. “You’re not able to provide these quick fixes if you complicate the question by thinking about group rights, group privacy, and farmer interests.” The people’s agritech Some promising experiments are taking a more democratic approach. The Bengaluru-based nonprofit Vrutti is developing a digital platform that enables different actors in the agricultural supply chain to interact, collect and share data, and buy and sell goods. The key difference is that this platform is co-owned by its users, so they have a say in its design and principles, says Prerak Shah, who is leading its development. Vrutti’s platform is primarily being used as a marketplace that allows FPOs to sell their produce to buyers. Each farmer’s transaction history is connected to a unique ID, and they can also record what crops they’re growing and what farming practices they’re using on their land. This data may ultimately become a valuable resource—for example, it could help members get lines of credit. Farmers control who can access their records, which are stored in a data wallet that they can transfer to other platforms. Whether the private sector can be persuaded to adopt these more farmer-centric approaches remains to be seen. But India has a rich history of agricultural cooperatives and bottom-up social organizing, says Chami. That’s why she thinks that the country can be a proving ground not only for innovative new agricultural technologies, but also for more equitable ways of deploying them. “I think India will show the world how this contest between corporate-led agritech and the people’s agritech plays out,” she says.

When Sony’s robot dog, Aibo, was first launched in 1999, it was hailed as revolutionary and the first of its kind, promising to usher in a new industry of intelligent mobile machines for the home. But its success was far from certain. Legged robots were still in their infancy, and the idea of making an interactive walking robot for the consumer market was extraordinarily ambitious. Beyond the technical challenges, Sony also had to solve a problem that entertainment robots still struggle with: how to make Aibo compelling and engaging rather than simply novel. Sony’s team made that happen. And since Aibo’s debut, the company has sold more than 170,000 of the cute little quadrupeds—a huge number considering their price of several thousand dollars each. From the start, Aibo could express a range of simulated emotions and learn through its interactions with users. Aibo was an impressive robot 25 years ago, and it’s still impressive today. Far from Sony headquarters in Tokyo, the town of Kōta, in Aichi Prefecture, is home to the Sony factory that has manufactured and repaired Aibos since 2018. Kōta has also become the center of fandom for Aibo, since the Hummingbird Café opened in the Kōta Town Hall in 2021. The first official Aibo café in Japan, it hosts Aibo-themed events, and Aibo owners from across the country gather there to let their Aibos loose in a play area and to exchange Aibo name cards. One patron of the Hummingbird Café is veteran Sony engineer Hideki Noma. In 1999, before Aibo was Aibo, Noma went to see his boss, Tadashi Otsuki. Otsuki had recently returned to Sony after a stint at the Japanese entertainment company Namco, and had been put in charge of a secretive new project to create an entertainment robot. But progress had stalled. There was a prototype robotic pet running around the lab, but Otsuki took a dim view of its hyperactive behavior and decided it wasn’t a product that anyone would want to buy. He envisioned something more lifelike. During their meeting, he gave Noma a surprising piece of advice: Go to Ryōan-ji, a famed Buddhist temple in Kyoto. Otsuki was telling Noma that to develop the right kind of robot for Sony, it needed Zen. Aibo’s Mission: Make History When the Aibo project started in 1994, personal entertainment robots seemed like a natural fit for Sony. Sony was a global leader in consumer electronics. And in the 1990s, Japan had more than half of the world’s industrial robots, dominating an industry led by manufacturers like Fanuc and Yaskawa Electric. Robots for the home were also being explored. In 1996, Honda showed off its P2 humanoid robot, a prototype of the groundbreaking ASIMO, which would be unveiled in 2000. Electrolux, based in the United Kingdom, introduced a prototype of its Trilobite robotic vacuum cleaner in 1997, and at iRobot in Boston, Joe Jones was working on what would become the Roomba. It seemed as though the consumer robot was getting closer to reality. Being the first to market was the perfect opportunity for an ambitious global company like Sony. Aibo was the idea of Sony engineer Toshitada Doi (on left), pictured in 1999 with an Aibo ERS-111. Hideki Noma (on right) holds an Aibo ERS-1000.Raphael Gaillarde/Gamma-Rapho/Getty Images; Right; Timothy Hornyak Sony’s new robot project was the brainchild of engineer Toshitada Doi, co-inventor of the CD. Doi was inspired by the speed and agility of MIT roboticist Rodney Brooks’s Genghis, a six-legged insectile robot that was created to demonstrate basic autonomous walking functions. Doi, however, had a vision for an ”entertainment robot with no clear role or job.” It was 1994 when his team of about 10 people began full-scale research and development on such a robot. Hideki Noma joined Sony in 1995. Even then, he had a lifelong love of robots, including participating in robotics contests and researching humanoids in college. “I was assigned to the Sony robot research team’s entertainment robot department,” says Noma. “It had just been established and had few people. Nobody knew Sony was working on robots, and it was a secret even within the company. I wasn’t even told what I would be doing.” Noma’s new colleagues in Sony’s robot skunk works had recently gone to Tokyo’s Akihabara electronics district and brought back boxes of circuit boards and servos. Their first creation was a six-legged walker with antenna-like sensors but more compact than Brooks’s Genghis, at roughly 22 centimeters long. It was clunky and nowhere near cute; if anything, it resembled a cockroach. “When they added the camera and other sensors, it was so heavy it couldn’t stand,” says Noma. “They realized it was going to be necessary to make everything at Sony—motors, gears, and all—or it would not work. That’s when I joined the team as the person in charge of mechatronic design.” Noma, who is now a software engineer in Sony’s new business development division, remembers that Doi’s catchphrase was “make history.” “Just as he had done with the compact disc, he wanted us to create a robot that was not only the first of its kind, but also one that would have a big impact on the world,” Noma recalls. “He always gently encouraged us with positive feedback.” “We also grappled with the question of what an ‘entertainment robot’ could be. It had to be something that would surprise and delight people. We didn’t have a fixed idea, and we didn’t set out to create a robot dog.” The team did look to living creatures for inspiration, studying dog and cat locomotion. Their next prototype lost two of the six legs and gained a head, tail, and more sophisticated AI abilities that created the illusion of canine characteristics. A mid-1998 version of the robot, nicknamed Mutant, ran on Sony’s Aperios OS, the operating system the company developed to control consumer devices. The robot had 16 degrees of freedom, a million-instructions-per-second (MIPS) 64-bit reduced-instruction-set computer (RISC) processor, and 8 megabytes of DRAM, expandable with a PC card. It could walk on uneven surfaces and use its camera to recognize motion and color—unusual abilities for robots of the time. It could dance, shake its head, wag its tail, sit, lie down, bark, and it could even follow a colored ball around. In fact, it was a little bundle of energy. The next iteration of the bot had a sleek new “coat” designed by Doi’s friend Hajime Sorayama, an industrial designer and illustrator known for his silvery gynoids, including the cover art for an Aerosmith album. Sorayama gave the robot a shiny, bulbous exterior that made it undeniably cute. Noma, now the team’s product planner and software engineer, felt they were getting closer to the goal. But when he presented the prototype to Otsuki in 1999, Otsuki was unimpressed. That’s when Noma was dispatched to Ryōan-ji to figure out how to make the robot seem not just cute but somehow alive. Seeking Zen for Aibo at the Rock Garden Established in 1450, Ryōan-ji is a Rinzai Zen sanctuary known for its meticulously raked rock garden featuring five distinctive groups of stones. The stones invite observers to quietly contemplate the space, and perhaps even the universe, and that’s what Noma did. He realized what Otsuki wanted Aibo to convey: a sense of tranquility. The same concept had been incorporated into the design of what was arguably Japan’s first humanoid robot, a large, smiling automaton named Gakutensoku that was unveiled in 1928. The rock garden at the Ryōan-ji Zen temple features carefully composed groupings of stones with unknown meaning. Bjørn Christian Tørrissen/Wikipedia Roboticist Masahiro Mori, originator of the Uncanny Valley concept for android design, had written about the relationship between Buddhism and robots back in 1974, stating, “I believe robots have the Buddha-nature within them—that is, the potential for attaining Buddhahood.” Essentially, he believed that even nonliving things were imbued with spirituality, a concept linked to animism in Japan. If machines can be thought of as embodying tranquility and spirituality, they can be easier to relate to, like living things. “When you make a robot, you want to show what it can do. But if it’s always performing, you’ll get bored and won’t want to live with it,” says Noma. “Just as cats and dogs need quiet time and rest, so do robots.” Noma modified the robot’s behaviors so that it would sometimes slow down and sleep. This reinforced the illusion that it was not only alive but had a will of its own. Otsuki then gave the little robot dog the green light. The cybernetic canine was named Aibo for “Artificial Intelligence roBOt” and aibō, which means “partner” in Japanese. In a press release, Sony billed the machine as “an autonomous robot that acts both in response to external stimuli and according to its own judgment. ‘AIBO’ can express various emotions, grow through learning, and communicate with human beings to bring an entirely new form of entertainment into the home.” But it was a lot more than that. Its 18 degrees of freedom allowed for complex motions, and it had a color charge-coupled device (CCD) camera and sensors for touch, acceleration, angular velocity, and range finding. Aibo had the hardware and smarts to back up Sony’s claim that it could “behave like a living creature.” The fact that it couldn’t do anything practical became irrelevant. The debut Aibo ERS-110 was priced at 250,000 yen (US $2,500, or a little over $4,700 today). A motion editor kit, which allowed users to generate original Aibo motions via their PC, sold for 50,000 yen ($450). Despite the eye-watering price tag, the first batch of 3,000 robots sold out in 20 minutes. Noma wasn’t surprised by the instant success. “We aimed to realize a society in which people and robots can coexist, not just robots working for humans but both enjoying a relationship of trust,” Noma says. “Based on that, an entertainment robot with a sense of self could communicate with people, grow, and learn.” Hideko Mori plays fetch with her Aibo ERS-7 in 2015, after it was returned to her from an Aibo hospital. Aibos are popular with seniors in Japan, offering interactivity and companionship without requiring the level of care of a real dog.Toshifumi Kitamura/AFP/Getty Images Aibo as a Cultural Phenomenon Aibo was the first consumer robot of its kind, and over the next four years, Sony released multiple versions of its popular pup across two more generations. Some customer responses were unexpected: as a pet and companion, Aibo was helping empty-nest couples rekindle their relationship, improving the lives of children with autism, and having a positive effect on users’ emotional states, according to a 2004 paper by AI specialist Masahiro Fujita, who collaborated with Doi on the early version of Aibo. “Aibo broke new ground as a social partner. While it wasn’t a replacement for a real pet, it introduced a completely new category of companion robots designed to live with humans,” says Minoru Asada, professor of adaptive machine systems at Osaka University’s graduate school of engineering. “It helped foster emotional connections with a machine, influencing how people viewed robots—not just as tools but as entities capable of forming social bonds. This shift in perception opened the door to broader discussions about human-robot interaction, companionship, and even emotional engagement with artificial beings.” Building a Custom Robot To create Aibo, Noma and colleagues had to start from scratch—there were no standard CPUs, cameras, or operating systems for consumer robots. They had to create their own, and the result was the Sony Open-R architecture, an unusual approach to robotics that enabled the building of custom machines. Announced in 1998, a year before Aibo’s release, Open-R allowed users to swap out modular hardware components, such as legs or wheels, to adapt a robot for different purposes. High-speed serial buses transmitted data embedded in each module, such as function and position, to the robot’s CPU, which would select the appropriate control signal for the new module. This meant the machine could still use the same motion-control software with the new components. The software relied on plug-and-play prerecorded memory cards, so that the behavior of an Open-R robot could instantly change, say, from being a friendly pet to a challenging opponent in a game. A swap of memory cards could also give the robot image- or sound-recognition abilities. “Users could change the modular hardware and software components,” says Noma. “The idea was having the ability to add a remote-control function or swap legs for wheels if you wanted.” Other improvements included different colors, touch sensors, LED faces, emotional expressions, and many more software options. There was even an Aibo that looked like a lion cub. The various models culminated in the sleek ERS-7, released in three versions from 2003 to 2005. Based on Scratch, the visual programming system in the latest versions of Aibo is easy to use and lets owners with limited programming experience create their own complex programs to modify how their robot behaves. The Aibo ERS-1000, unveiled in January 2018, has 22 degrees of freedom, a 64-bit quad-core CPU, and two OLED eyes. It’s more puppylike and smarter than previous models, capable of recognizing 100 faces and responding to 50 voice commands. It can even be “potty trained” and “fed” with virtual food through an app. —T.H. Aibo also played a crucial role in the evolution of autonomous robotics, particularly in competitions like RoboCup, notes Asada, who cofounded the robot soccer competition in the 1990s. Whereas custom-built robots were prone to hardware failures, Aibo was consistently reliable and programmable, and so it allowed competitors to focus on advancing software and AI. It became a key tool for testing algorithms in real-world environments. By the early 2000s, however, Sony was in trouble. Leading the smartphone revolution, Apple and Samsung were steadily chipping away at Sony’s position as a consumer-electronics and digital-content powerhouse. When Howard Stringer was appointed Sony’s first non-Japanese CEO in 2005, he implemented a painful restructuring program to make the company more competitive. In 2006, he shut down the robot entertainment division, and Aibo was put to sleep. What Sony’s executives may not have appreciated was the loyalty and fervor of Aibo buyers. In a petition to keep Aibo alive, one person wrote that the robot was “an irreplaceable family member.” Aibo owners were naming their robots, referring to them with the word ko (which usually denotes children), taking photos with them, going on trips with them, dressing them up, decorating them with ribbons, and even taking them out on “dates” with other Aibos. For Noma, who has four Aibos at home, this passion was easy to understand. Hideki Noma [right] poses with his son Yuto and wife Tomoko along with their Aibo friends. At right is an ERS-110 named Robbie (inspired by Isaac Asimov’s “I, Robot”), at the center is a plush Aibo named Choco, and on the left is an ERS-1000 named Murphy (inspired by the film Interstellar). Hideki Noma “Some owners treat Aibo as a pet, and some treat it as a family member,” he says. “They celebrate its continued health and growth, observe the traditional Shichi-Go-San celebration [for children aged 3, 5, and 7] and dress their Aibos in kimonos.…This idea of robots as friends or family is particular to Japan and can be seen in anime like Astro Boy and Doraemon. It’s natural to see robots as friends we consult with and sometimes argue with.” The Return of Aibo With the passion of Aibo fans undiminished and the continued evolution of sensors, actuators, connectivity, and AI, Sony decided to resurrect Aibo after 12 years. Noma and other engineers returned to the team to work on the new version, the Aibo ERS-1000, which was unveiled in January 2018. Fans of all ages were thrilled. Priced at 198,000 yen ($1,760), not including the mandatory 90,000-yen, three-year cloud subscription service, the first batch sold out in 30 minutes, and 11,111 units sold in the first three months. Since then, Sony has released additional versions with new design features, and the company has also opened up Aibo to some degree of programming, giving users access to visual programming tools and an application programming interface (API). A quarter century after Aibo was launched, Noma is finally moving on to another job at Sony. He looks back on his 17 years developing the robot with awe. “Even though we imagined a society of humans and robots coexisting, we never dreamed Aibo could be treated as a family member to the degree that it is,” he says. “We saw this both in the earlier versions of Aibo and the latest generation. I’m deeply grateful and moved by this. My wish is that this relationship will continue for a long time.”

Thanks to generous funding from the ON Semiconductor Foundation, TryEngineering has partnered with IEEE members to develop several new resources about semiconductors for middle school educators. The resources include lesson plans, an e-book, and videos. The grant also paid for the creation of in-person professional development sessions for educators—which were held at three locations in the United States. The foundation is part of Onsemi’s Giving Now program. The company, headquartered in Scottsdale, Ariz., is a semiconductor manufacturer serving tens of thousands of customers across several markets with intelligent power and sensing technologies. Onsemi funds STEAM (science, technology, engineering, art, and math) educational activities for underprivileged youth in underserved communities where it operates globally. “We are so grateful to have partners like Onsemi who share our passion for inspiring students to change the world as an engineer or technology professional,” says Jamie Moesch, IEEE Educational Activities managing director. “The work we have developed together is being used by instructors around the world to become more comfortable teaching students about semiconductors, microelectronics, and more.” Microchip lesson plan and e-book The Making of a Microchip lesson plan covers how a chip is created using low-cost accessible materials. Included is an introduction to the engineering design process and an overview of terms used in the semiconductor industry. The plan has additional exploratory activities, called missions, to introduce students to semiconductor technology. Teachers can assign the missions as a series of projects over a two-week period or to differentiate instruction, providing opportunities for further exploration to anyone interested in semiconductors. Complementing the lesson plan is the new Microchip Adventures e-book, which explains how semiconductors are made. Engaging video resources The grant also funded the creation of three recorded interviews with IEEE members who have semiconductor expertise. The three videos—Electronic Packaging, The Semiconductor Industry, and What Is a Semiconductor?—are intended to familiarize students with industry terms used by engineers. The videos can supplement the lesson plans or act as standalone resources. One of the videos features interviews with staff members at Ozark Integrated Circuits, a privately held company in Fayetteville, Ark., owned by IEEE Region 5 Director Matt Francis. The company specializes in design techniques and modeling and design tools for integrated circuits and systems on chip for extreme environments. Another interview was with Kathy Herring Hayashi, an IEEE member and Region 6 director. She is a software consultant and a computer science instructor in the San Diego Community College District and at Palomar College, in San Marcos, Calif. Francis, an IEEE senior member, and a team of IEEE members and semiconductor experts—Stamatis Dragoumanos, Lorena Garcia, and Case Kirk—developed the video content. Professional development sessions Onsemi’s award included funding to create and deliver in-person professional development sessions to teachers across the United States. The first Technology for Teachers sessions were held in Phoenix; Fayetteville, Ark.; and New Brunswick, N.J. The Arizona State University electrical engineering department hosted the first session. Faculty members gave the participants a tour of the university’s NanoFab, a nanoscale processing and fabrication facility. In Fayetteville, teachers toured Ozark Integrated Circuits, where they met with engineers and technologists. In addition, the Making of a Microchip lesson was launched at the sessions, and the teachers viewed the videos. “The work we have developed together is being used by instructors around the world to become more comfortable teaching students about semiconductors, microelectronics, and more.” —Jamie Moesch, IEEE Educational Activities managing director “Fifteen Arkansas and Missouri middle school teachers learned about the semiconductor supply chain, and they left with guided lesson plans, engaging videos, and the newest content for their classrooms,” Francis says. “We toured Ozark Integrated Circuits and ended up brainstorming about the future. Listening to them talk about their kids back home—and how they are going to ‘get this’—really tugged at my heart. It reminded me of wanting to know how those ‘magic computers’ worked when I was at their age.” Participants in New Brunswick, welcomed by Rutgers University, toured its Nanofabrication CORE Facility, which provides students with skills and capabilities to have a career in the semiconductor industry. “TryEngineering’s Technology for Teachers program offered a unique professional development opportunity for educators,” says Debra Gulick, IEEE Educational Activities director of student and academic education programs. “Combining access to engaging resources and the opportunity to meet with IEEE engineers and tour state-of-the-art facilities made this an inspiring experience and one that teachers were able to bring into their classrooms.” Funding for more outreach programs As a result of TryEngineering’s efforts this year, Onsemi’s Giving Now program has renewed its financial support for next year. IEEE Educational Activities is honored to be a part of the ON Semiconductor Foundation’s generous support of US $2 million to fund global outreach programs, Moesch says. Teachers receive the materials they need to bring the activities back to their classrooms, and to inspire the next generation of engineers and technologists. TryEngineering staff and volunteers are collaborating with Field Day researchers at the University of Wisconsin—Madison to develop a game that simulates challenges faced in the semiconductor industry. Players can learn about the technology and the supply chain while playing the game. “We have such an incredible opportunity right now to reach traditionally underserved populations with information about the career paths available in the semiconductor industry,” says Jennifer Fong, IEEE Educational Activities director of continuing education and business development. “This creates more economic opportunity for more people. “As IEEE takes a comprehensive approach to semiconductor workforce development, starting with preuniversity programs and continuing with microcredentials for those without four-year degrees [as well as] skills and competency frameworks for technical jobs, training courses, and more, we will have the greatest impact through partnership. I applaud Onsemi’s focus on making sure we engage kids early so we have the workforce needed for the future.” The content can be found in the TryEngineering website’s semiconductors section.

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion. Humanoids Summit: 11–12 December 2024, MOUNTAIN VIEW, CA Enjoy today’s videos! Step into the future of factory automation with MagicBot, the cutting-edge humanoid robots from Magiclab. Recently deployed to production lines, these intelligent machines are mastering tasks like product inspections, material transport, precision assembly, barcode scanning, and inventory management. [ Magiclab ] Some highlights from the IEEE / RAS International Conference on Humanoid Robots - Humanoids 2024. [ Humanoids 2024 ] This beautiful feathered drone, PigeonBot II, comes from David Lentik’s lab at University of Groningen in the Netherlands. It was featured in Science Robotics just last month. [ Lentink Lab ] via [ Science ] Thanks, David! In this video, Stretch AI takes a language prompt of “Stretch, put the toy in basket” to control Stretch to accomplish the task. [ Hello Robot ] Simone Giertz, “the queen of shitty robots,” interviewed by our very own Stephen Cass. [ IEEE Spectrum ] We present a perceptive obstacle-avoiding controller for pedipulation, i.e. manipulation with a quadrupedal robot’s foot. [ Pedipulation ] Kernel Foods has revolutionized fast food by integrating KUKA robots into its kitchen operations, combining automation with human expertise for consistent and efficient meal preparation. Using the KR AGILUS robot, Kernel optimizes processes like food sequencing, oven operations, and order handling, reducing the workload for employees and enhancing customer satisfaction. [ Kernel Foods ] If this doesn’t impress you, skip ahead to 0:52. [ Paper via arXiv ] Thanks, Kento! The cuteness. I can’t handle it. [ Pollen ] A set of NTNU academics initiate a new research lab - called Legged Robots for the Arctic & beyond lab - responding to relevant interests within the NTNU student community. If you are a student and have relevant interests, get in touch! [ NTNU ] Extend Robotics is pioneering a shift in viticulture with intelligent automation at Saffron Grange Vineyard in Essex, addressing the challenges of grape harvesting with their robotic capabilities. Our collaborative project with Queen Mary University introduces a robotic system capable of identifying ripe grapes through AI-driven visual sensors, which assess ripeness based on internal sugar levels without damaging delicate fruit. Equipped with pressure-sensitive grippers, our robots can handle grapes gently, preserving their quality and value. This precise harvesting approach could revolutionise vineyards, enabling autonomous and remote operations. [ Extend Robotics ] Code & Circuit, a non-profit organization based in Amesbury, MA, is a place where kids can use technology to create, collaborate, and learn! Spot is a central part of their program, where educators use the robot to get younger participants excited about STEM fields, coding, and robotics, while advanced learners have the opportunity to build applications using an industrial robot. [ Code & Circuit ] During the HUMANOIDS Conference, we had the chance to speak with some of the true rock stars in the world of robotics. While they could discuss robots endlessly, when asked to describe robotics today in just one word, these brilliant minds had to pause and carefully choose the perfect response. Personally I would not have chosen “exploding.” [ PAL Robotics ] Lunabotics provides accredited institutions of higher learning students an opportunity to apply the NASA systems engineering process to design and build a prototype Lunar construction robot. This robot would be capable of performing the proposed operations on the Lunar surface in support of future Artemis Campaign goals. [ NASA ] Before we get into all the other course projects from this term, here are a few free throw attempts from ROB 550’s robotic arm lab earlier this year. Maybe good enough to walk on the Michigan basketball team? Students in ROB 550 cover the basics of robotic sensing, reasoning, and acting in several labs over the course: here the designs to take the ball to the net varied greatly, from hook shots to tension-storing contraptions from downtown. These basics help them excel throughout their robotics graduate degrees and research projects. [ University of Michigan Robotics ] Wonder what a Robody can do? This. And more! [ Devanthro ] It’s very satisfying watching Dusty print its way around obstacles. [ Dusty Robotics ] Ryan Companies has deployed Field AI’s autonomy software on a quadruped robot in the company’s ATX Tower site in Austin, TX, to greatly improve its daily surveying and data collection processes. [ Field AI ] Since landing its first rover on Mars in 1997, NASA has pushed the boundaries of exploration with increasingly larger and more sophisticated robotic explorers. Each mission builds on the lessons learned from the Red Planet, leading to breakthroughs in technology and our understanding of Mars. From the microwave-sized Sojourner to the SUV-sized Perseverance—and even taking flight with the groundbreaking Ingenuity helicopter—these rovers reflect decades of innovation and the drive to answer some of science’s biggest questions. This is their evolution. [ NASA ] Welcome to things that are safe to do only with a drone. [ Team BlackSheep ]

When the lights went out on the BCS East-West Interlink fiber optic cable connecting Lithuania and Sweden on 17 November, the biggest question wasn’t when internet service would be restored. (That’d come another 10 or so days later.) The outage—alongside a cable failure the next day of an undersea line connecting Finland and Germany—soon became a whodunit, as German, Swedish, and Finnish officials variously hinted that the damage to the lines could constitute acts of “sabotage” or “hybrid warfare.” Suspicion soon centered around Russia or China—especially given the presence of a Chinese-flagged cargo vessel in the area during both incidents. The outages underscore how much of the global communications and financial system hinges on a few hundred cables of bundled glass fibers that are strung across ocean floors around the world, each cable about the same diameter as a garden hose. And, says Bryan Clark, a senior fellow at the Washington, D.C.-based Hudson Institute, defending undersea fiber optic cables from damage and sabotage is increasingly challenging. The technology to do so is nowhere near bulletproof, he says, yet the steep cost of failing to protect them is too high to consider simply writing them off. (NATO is currently investigating future internet backup routes through satellites in the case of undersea cable failures. But that technology is only in a preliminary, proof-of-concept stage and may be many years from real-world relevance.) “In the past, when these kinds of cable cutting incidents have happened, the perpetrator has tried to somehow disguise the source of the disruption, and China’s not necessarily doing that here,” Clark says. “What we’re seeing now is that maybe countries are doing this more overtly. And then also they may be using specialized equipment to do it rather than dragging an anchor.” Clark says protecting undersea cables in the Baltic is actually one of the less-challenging situations on the geostrategic map of seafloor cable vulnerabilities. “In the Mediterranean and the Baltic, the transit lanes or the distance you have to patrol is not that long,” he says. “And so there are some systems being developed that would just patrol those cables using uncrewed vehicles.” In other words, while the idea of uncrewed underwater vehicles (UUVs) regularly patrolling internet cableways is still in the realm of science fiction, it’s not that far removed from science fact as to be out of the realm of soon-to-be-realized possibility. But then comes the lion’s share of the undersea internet cables around the world—the lines of fiber that traverse open oceans across the globe. In these cases, Clark says, there are two regions of each cables’ path. There’s the deep sea portion—the Davy Jones’ Locker realm where only top-secret missions and movie directors on submarine jags dare venture. And then there are the portions of cable in shallower waters, typically nearer to coasts, that are accessible by present day anchors, submersibles, drones, and lord-knows-what-other kinds of underwater tech. Moreover, once an undersea cable ventures into the legal purview of a given country—what’s called a nation’s exclusive economic zone (EEZ)—that in particular is when fancy, newfangled tech to defend or attack an undersea line must take a backseat to old-fashioned military and policing might. Satellite imaging and underwater drones, says the Hudson Institute’s Bryan Clark, are two technologies that can protect undersea fiber optic lines. Hudson Institute “If you were patrolling the area and just monitoring the surface, and you saw a ship [traveling] above where the cables are, you could send out Coast Guard forces, paramilitary forces,” Clark says. “It would be a law enforcement mission, because it’s within the EEZs of different countries who are owners of those cables.” In fact, the Danish navy reportedly did just that concerning the Baltic voyage of a Chinese-flagged chip called Yi Peng 3. And now Sweden is calling for the Yi Peng 3 to cooperate in an inspection of the ship in a larger investigation of the undersea cable breaches. One-Million-Plus Kilometers of Open Cable According to Lane Burdette, research analyst at the internet infrastructure analysis firm TeleGeography, the vastness of undersea internet lines points to a dilemma of shoring up the high-vulnerability shallow regions and setting aside for the time being the deeper realms beyond protection. “As of 2024, TeleGeography estimates there are 1.5 million kilometers of communications cables in the water,” she says. “With a network this large, it’s not possible to monitor all cables, everywhere, all the time. However, new technologies are emerging that make it easier to monitor activity where damage is most likely and potentially prevent even some accidental disruption.” At the moment, much of the game is still defensive, Clark says. Efforts to lay undersea internet cable lines today, he says, can also include measures to cover the lines to prevent their detection or dig small trenches to protect the lines from being severed or dragged by ships’ anchors. Satellite imaging will be increasingly crucial in defending undersea cables, Clark adds. Geospatial analysis offered by the likes of the Herndon, Va.-based BlackSky Technology and SpaceX’s Starshield will be essential for countries looking to protect their high-bandwidth internet access. “You’ll end up with low-latency coverage over most of the mid-latitudes within the next few years, which you could use to monitor for ship operations in the vicinity of known cable runs,” Clark says. However, once UUVs are ready for widespread use, he adds, the undersea internet cable cat-and-mouse game could change drastically, which UUV being used offensively as well as defensively. “A lot of these cables, especially in shallow waters, are in pretty well-known locations,” he says. “So in the Baltic, you could see where Russia [might] deploy a relatively large number of uncrewed vehicles—and cut a large number of cables at once.” All of which could one day render something like the Yi Peng 3 situation—a Chinese-flagged freighter trawling over known runs of undersea internet cabling—a quaint relic of the pre-UUV days. “Once you’ve determined where you’re pretty sure a cableway is, you could drive your ship over, deploy your uncrewed vehicles, and then they could loiter,” Clark says. “And then you could cut the cable five days later, in which case you wouldn’t be necessarily blamed for it, because your ship traveled over that region a week ago.”

On the shores of Lake Geneva in Switzerland, École Polytechnique Fédérale de Lausanne is home to many roboticists. It’s also home to many birds, which spend the majority of their time doing bird things. With a few exceptions, those bird things aren’t actually flying: Flying is a lot of work, and many birds have figured out that they can instead just walk around on the ground, where all the food tends to be, and not tire themselves out by having to get airborne over and over again. “Whenever I encountered crows on the EPFL campus, I would observe how they walked, hopped over or jumped on obstacles, and jumped for take-offs,” says Won Dong Shin, a doctoral student at EPFL’s Laboratory of Intelligent Systems. “What I consistently observed was that they always jumped to initiate flight, even in situations where they could have used only their wings.” Shin is first author on a paper published today in Nature that explores both why birds jump to take off, and how that can be beneficially applied to fixed-wing drones, which otherwise need things like runways or catapults to get themselves off the ground. Shin’s RAVEN (Robotic Avian-inspired Vehicle for multiple ENvironments) drone, with its bird-inspired legs, can do jumping takeoffs just like crows do, and can use those same legs to get around on the ground pretty well, too. The drone’s bird-inspired legs adopted some key principles of biological design like the ability to store and release energy in tendon-like springs along with some flexible toes.Alain Herzog Back in 2019, we wrote about a South African startup called Passerine which had a similar idea, albeit more focused on using legs to launch fixed-wing cargo drones into the air. This is an appealing capability for drones, because it means that you can take advantage of the range and endurance that you get with a fixed wing without having to resort to inefficient tricks like stapling a bunch of extra propellers to yourself to get off the ground. “The concept of incorporating jumping take-off into a fixed-wing vehicle is the common idea shared by both RAVEN and Passerine,” says Shin. “The key difference lies in their focus: Passerine concentrated on a mechanism solely for jumping, while RAVEN focused on multifunctional legs.” Bio-inspired Design for Drones Multifunctional legs bring RAVEN much closer to birds, and although these mechanical legs are not nearly as complex and capable as actual bird legs, adopting some key principles of biological design (like the ability to store and release energy in tendon-like springs along with some flexible toes) allows RAVEN to get around in a very bird-like way. Won Dong Shin Despite its name, RAVEN is approximately the size of a crow, with a wingspan of 100 centimeters and a body length of 50 cm. It can walk a meter in just under four seconds, hop over 12 cm gaps, and jump into the top of a 26 cm obstacle. For the jumping takeoff, RAVEN’s legs propel the drone to a starting altitude of nearly half a meter, with a forward velocity of 2.2 m/s. RAVEN’s toes are particularly interesting, especially after you see how hard the poor robot faceplants without them: Without toes, RAVEN face-plants when it tries to walk.Won Dong Shin “It was important to incorporate a passive elastic toe joint to enable multiple gait patterns and ensure that RAVEN could jump at the correct angle for takeoff,” Shin explains. Most bipedal robots have actuated feet that allow for direct control for foot angles, but for a robot that flies, you can’t just go adding actuators all over the place willy-nilly because they weigh too much. As it is, RAVEN’s a 620-gram drone of which a full 230 grams consists of feet and toes and actuators and whatnot. Actuated hip and ankle joints form a simplified but still birdlike leg, while springs in the ankle and toe joints help to absorb force and store energy.EPFL Why Add Legs to a Drone? So the question is, is all of this extra weight and complexity of adding legs actually worth it? In one sense, it definitely is, because the robot can do things that it couldn’t do before—walking around on the ground and taking off from the ground by itself. But it turns out that RAVEN is light enough, and has a sufficiently powerful enough motor, that as long as it’s propped up at the right angle, it can take off from the ground without jumping at all. In other words, if you replaced the legs with a couple of popsicle sticks just to tilt the drone’s nose up, would that work just as well for the ground takeoffs? The researchers tested this, and found that non-jumping takeoffs were crappy. The mix of high angle of attack and low takeoff speed led to very unstable flight—it worked, but barely. Jumping, on the other hand, ends up being about ten times more energy efficient overall than a standing takeoff. As the paper summarizes, “although jumping take-off requires slightly higher energy input, it is the most energy-efficient and fastest method to convert actuation energy to kinetic and potential energies for flight.” And just like birds, RAVEN can also take advantage of its legs to move on the ground in a much more energy efficient way relative to making repeated short flights. Won Dong Shin holds the RAVEN drone.Alain Herzog Can This Design Scale Up to Larger Fixed-Wing Drones? Birds use their legs for all kinds of stuff besides walking and hopping and jumping, of course, and Won Dong Shin hopes that RAVEN may be able to do more with its legs, too. The obvious one is using legs for landing: “Birds use their legs to decelerate and reduce impact, and this same principle could be applied to RAVEN’s legs,” Shin says, although the drone would need a perception system that it doesn’t yet have to plan things out. There’s also swimming, perching, and snatching, all of which would require a new foot design. We also asked Shin about what it would take to scale this design up, to perhaps carry a useful payload at some point. Shin points out that beyond a certain size, birds are no longer able to do jumping takeoffs, and either have to jump off something higher up or find themselves a runway. In fact, some birds will go to astonishing lengths not to have to do jumping takeoffs, as best human of all time David Attenborough explains: BBC Shin points out that it’s usually easier to scale engineered systems than biological ones, and he seems optimistic that legs for jumping takeoffs will be viable on larger fixed-wing drones that could be used for delivery. A vision system that could be used for both obstacle avoidance and landing is in the works, as are wings that can fold to allow the drone to pass through narrow gaps. Ultimately, Shin says that he wants to make the drone as bird-like as possible: “I am also keen to incorporate flapping wings into RAVEN. This enhancement would enable more bird-like motion and bring more interesting research questions to explore.” “Fast ground-to-air transition with avian-inspired multifunctional legs,” by Won Dong Shin, Hoang-Vu Phan, Monica A. Daley, Auke J. Ijspeert, and Dario Floreano from EPFL in Switzerland and UC Irvine, appears in the December 4 issue of Nature.

Andrew Ng has serious street cred in artificial intelligence. He pioneered the use of graphics processing units (GPUs) to train deep learning models in the late 2000s with his students at Stanford University, cofounded Google Brain in 2011, and then served for three years as chief scientist for Baidu, where he helped build the Chinese tech giant’s AI group. So when he says he has identified the next big shift in artificial intelligence, people listen. And that’s what he told IEEE Spectrum in an exclusive Q&A. Ng’s current efforts are focused on his company Landing AI, which built a platform called LandingLens to help manufacturers improve visual inspection with computer vision. He has also become something of an evangelist for what he calls the data-centric AI movement, which he says can yield “small data” solutions to big issues in AI, including model efficiency, accuracy, and bias. Andrew Ng on... What’s next for really big models The career advice he didn’t listen to Defining the data-centric AI movement Synthetic data Why Landing AI asks its customers to do the work The great advances in deep learning over the past decade or so have been powered by ever-bigger models crunching ever-bigger amounts of data. Some people argue that that’s an unsustainable trajectory. Do you agree that it can’t go on that way? Andrew Ng: This is a big question. We’ve seen foundation models in NLP [natural language processing]. I’m excited about NLP models getting even bigger, and also about the potential of building foundation models in computer vision. I think there’s lots of signal to still be exploited in video: We have not been able to build foundation models yet for video because of compute bandwidth and the cost of processing video, as opposed to tokenized text. So I think that this engine of scaling up deep learning algorithms, which has been running for something like 15 years now, still has steam in it. Having said that, it only applies to certain problems, and there’s a set of other problems that need small data solutions. When you say you want a foundation model for computer vision, what do you mean by that? Ng: This is a term coined by Percy Liang and some of my friends at Stanford to refer to very large models, trained on very large data sets, that can be tuned for specific applications. For example, GPT-3 is an example of a foundation model [for NLP]. Foundation models offer a lot of promise as a new paradigm in developing machine learning applications, but also challenges in terms of making sure that they’re reasonably fair and free from bias, especially if many of us will be building on top of them. What needs to happen for someone to build a foundation model for video? Ng: I think there is a scalability problem. The compute power needed to process the large volume of images for video is significant, and I think that’s why foundation models have arisen first in NLP. Many researchers are working on this, and I think we’re seeing early signs of such models being developed in computer vision. But I’m confident that if a semiconductor maker gave us 10 times more processor power, we could easily find 10 times more video to build such models for vision. Having said that, a lot of what’s happened over the past decade is that deep learning has happened in consumer-facing companies that have large user bases, sometimes billions of users, and therefore very large data sets. While that paradigm of machine learning has driven a lot of economic value in consumer software, I find that that recipe of scale doesn’t work for other industries. Back to top It’s funny to hear you say that, because your early work was at a consumer-facing company with millions of users. Ng: Over a decade ago, when I proposed starting the Google Brain project to use Google’s compute infrastructure to build very large neural networks, it was a controversial step. One very senior person pulled me aside and warned me that starting Google Brain would be bad for my career. I think he felt that the action couldn’t just be in scaling up, and that I should instead focus on architecture innovation. “In many industries where giant data sets simply don’t exist, I think the focus has to shift from big data to good data. Having 50 thoughtfully engineered examples can be sufficient to explain to the neural network what you want it to learn.” —Andrew Ng, CEO & Founder, Landing AI I remember when my students and I published the first NeurIPS workshop paper advocating using CUDA, a platform for processing on GPUs, for deep learning—a different senior person in AI sat me down and said, “CUDA is really complicated to program. As a programming paradigm, this seems like too much work.” I did manage to convince him; the other person I did not convince. I expect they’re both convinced now. Ng: I think so, yes. Over the past year as I’ve been speaking to people about the data-centric AI movement, I’ve been getting flashbacks to when I was speaking to people about deep learning and scalability 10 or 15 years ago. In the past year, I’ve been getting the same mix of “there’s nothing new here” and “this seems like the wrong direction.” Back to top How do you define data-centric AI, and why do you consider it a movement? Ng: Data-centric AI is the discipline of systematically engineering the data needed to successfully build an AI system. For an AI system, you have to implement some algorithm, say a neural network, in code and then train it on your data set. The dominant paradigm over the last decade was to download the data set while you focus on improving the code. Thanks to that paradigm, over the last decade deep learning networks have improved significantly, to the point where for a lot of applications the code—the neural network architecture—is basically a solved problem. So for many practical applications, it’s now more productive to hold the neural network architecture fixed, and instead find ways to improve the data. When I started speaking about this, there were many practitioners who, completely appropriately, raised their hands and said, “Yes, we’ve been doing this for 20 years.” This is the time to take the things that some individuals have been doing intuitively and make it a systematic engineering discipline. The data-centric AI movement is much bigger than one company or group of researchers. My collaborators and I organized a data-centric AI workshop at NeurIPS, and I was really delighted at the number of authors and presenters that showed up. You often talk about companies or institutions that have only a small amount of data to work with. How can data-centric AI help them? Ng: You hear a lot about vision systems built with millions of images—I once built a face recognition system using 350 million images. Architectures built for hundreds of millions of images don’t work with only 50 images. But it turns out, if you have 50 really good examples, you can build something valuable, like a defect-inspection system. In many industries where giant data sets simply don’t exist, I think the focus has to shift from big data to good data. Having 50 thoughtfully engineered examples can be sufficient to explain to the neural network what you want it to learn. When you talk about training a model with just 50 images, does that really mean you’re taking an existing model that was trained on a very large data set and fine-tuning it? Or do you mean a brand new model that’s designed to learn only from that small data set? Ng: Let me describe what Landing AI does. When doing visual inspection for manufacturers, we often use our own flavor of RetinaNet. It is a pretrained model. Having said that, the pretraining is a small piece of the puzzle. What’s a bigger piece of the puzzle is providing tools that enable the manufacturer to pick the right set of images [to use for fine-tuning] and label them in a consistent way. There’s a very practical problem we’ve seen spanning vision, NLP, and speech, where even human annotators don’t agree on the appropriate label. For big data applications, the common response has been: If the data is noisy, let’s just get a lot of data and the algorithm will average over it. But if you can develop tools that flag where the data’s inconsistent and give you a very targeted way to improve the consistency of the data, that turns out to be a more efficient way to get a high-performing system. “Collecting more data often helps, but if you try to collect more data for everything, that can be a very expensive activity.” —Andrew Ng For example, if you have 10,000 images where 30 images are of one class, and those 30 images are labeled inconsistently, one of the things we do is build tools to draw your attention to the subset of data that’s inconsistent. So you can very quickly relabel those images to be more consistent, and this leads to improvement in performance. Could this focus on high-quality data help with bias in data sets? If you’re able to curate the data more before training? Ng: Very much so. Many researchers have pointed out that biased data is one factor among many leading to biased systems. There have been many thoughtful efforts to engineer the data. At the NeurIPS workshop, Olga Russakovsky gave a really nice talk on this. At the main NeurIPS conference, I also really enjoyed Mary Gray’s presentation, which touched on how data-centric AI is one piece of the solution, but not the entire solution. New tools like Datasheets for Datasets also seem like an important piece of the puzzle. One of the powerful tools that data-centric AI gives us is the ability to engineer a subset of the data. Imagine training a machine-learning system and finding that its performance is okay for most of the data set, but its performance is biased for just a subset of the data. If you try to change the whole neural network architecture to improve the performance on just that subset, it’s quite difficult. But if you can engineer a subset of the data you can address the problem in a much more targeted way. When you talk about engineering the data, what do you mean exactly? Ng: In AI, data cleaning is important, but the way the data has been cleaned has often been in very manual ways. In computer vision, someone may visualize images through a Jupyter notebook and maybe spot the problem, and maybe fix it. But I’m excited about tools that allow you to have a very large data set, tools that draw your attention quickly and efficiently to the subset of data where, say, the labels are noisy. Or to quickly bring your attention to the one class among 100 classes where it would benefit you to collect more data. Collecting more data often helps, but if you try to collect more data for everything, that can be a very expensive activity. For example, I once figured out that a speech-recognition system was performing poorly when there was car noise in the background. Knowing that allowed me to collect more data with car noise in the background, rather than trying to collect more data for everything, which would have been expensive and slow. Back to top What about using synthetic data, is that often a good solution? Ng: I think synthetic data is an important tool in the tool chest of data-centric AI. At the NeurIPS workshop, Anima Anandkumar gave a great talk that touched on synthetic data. I think there are important uses of synthetic data that go beyond just being a preprocessing step for increasing the data set for a learning algorithm. I’d love to see more tools to let developers use synthetic data generation as part of the closed loop of iterative machine learning development. Do you mean that synthetic data would allow you to try the model on more data sets? Ng: Not really. Here’s an example. Let’s say you’re trying to detect defects in a smartphone casing. There are many different types of defects on smartphones. It could be a scratch, a dent, pit marks, discoloration of the material, other types of blemishes. If you train the model and then find through error analysis that it’s doing well overall but it’s performing poorly on pit marks, then synthetic data generation allows you to address the problem in a more targeted way. You could generate more data just for the pit-mark category. “In the consumer software Internet, we could train a handful of machine-learning models to serve a billion users. In manufacturing, you might have 10,000 manufacturers building 10,000 custom AI models.” —Andrew Ng Synthetic data generation is a very powerful tool, but there are many simpler tools that I will often try first. Such as data augmentation, improving labeling consistency, or just asking a factory to collect more data. Back to top To make these issues more concrete, can you walk me through an example? When a company approaches Landing AI and says it has a problem with visual inspection, how do you onboard them and work toward deployment? Ng: When a customer approaches us we usually have a conversation about their inspection problem and look at a few images to verify that the problem is feasible with computer vision. Assuming it is, we ask them to upload the data to the LandingLens platform. We often advise them on the methodology of data-centric AI and help them label the data. One of the foci of Landing AI is to empower manufacturing companies to do the machine learning work themselves. A lot of our work is making sure the software is fast and easy to use. Through the iterative process of machine learning development, we advise customers on things like how to train models on the platform, when and how to improve the labeling of data so the performance of the model improves. Our training and software supports them all the way through deploying the trained model to an edge device in the factory. How do you deal with changing needs? If products change or lighting conditions change in the factory, can the model keep up? Ng: It varies by manufacturer. There is data drift in many contexts. But there are some manufacturers that have been running the same manufacturing line for 20 years now with few changes, so they don’t expect changes in the next five years. Those stable environments make things easier. For other manufacturers, we provide tools to flag when there’s a significant data-drift issue. I find it really important to empower manufacturing customers to correct data, retrain, and update the model. Because if something changes and it’s 3 a.m. in the United States, I want them to be able to adapt their learning algorithm right away to maintain operations. In the consumer software Internet, we could train a handful of machine-learning models to serve a billion users. In manufacturing, you might have 10,000 manufacturers building 10,000 custom AI models. The challenge is, how do you do that without Landing AI having to hire 10,000 machine learning specialists? So you’re saying that to make it scale, you have to empower customers to do a lot of the training and other work. Ng: Yes, exactly! This is an industry-wide problem in AI, not just in manufacturing. Look at health care. Every hospital has its own slightly different format for electronic health records. How can every hospital train its own custom AI model? Expecting every hospital’s IT personnel to invent new neural-network architectures is unrealistic. The only way out of this dilemma is to build tools that empower the customers to build their own models by giving them tools to engineer the data and express their domain knowledge. That’s what Landing AI is executing in computer vision, and the field of AI needs other teams to execute this in other domains. Is there anything else you think it’s important for people to understand about the work you’re doing or the data-centric AI movement? Ng: In the last decade, the biggest shift in AI was a shift to deep learning. I think it’s quite possible that in this decade the biggest shift will be to data-centric AI. With the maturity of today’s neural network architectures, I think for a lot of the practical applications the bottleneck will be whether we can efficiently get the data we need to develop systems that work well. The data-centric AI movement has tremendous energy and momentum across the whole community. I hope more researchers and developers will jump in and work on it. Back to top This article appears in the April 2022 print issue as “Andrew Ng, AI Minimalist.”

The end of Moore’s Law is looming. Engineers and designers can do only so much to miniaturize transistors and pack as many of them as possible into chips. So they’re turning to other approaches to chip design, incorporating technologies like AI into the process. Samsung, for instance, is adding AI to its memory chips to enable processing in memory, thereby saving energy and speeding up machine learning. Speaking of speed, Google’s TPU V4 AI chip has doubled its processing power compared with that of its previous version. But AI holds still more promise and potential for the semiconductor industry. To better understand how AI is set to revolutionize chip design, we spoke with Heather Gorr, senior product manager for MathWorks’ MATLAB platform. How is AI currently being used to design the next generation of chips? Heather Gorr: AI is such an important technology because it’s involved in most parts of the cycle, including the design and manufacturing process. There’s a lot of important applications here, even in the general process engineering where we want to optimize things. I think defect detection is a big one at all phases of the process, especially in manufacturing. But even thinking ahead in the design process, [AI now plays a significant role] when you’re designing the light and the sensors and all the different components. There’s a lot of anomaly detection and fault mitigation that you really want to consider. Heather GorrMathWorks Then, thinking about the logistical modeling that you see in any industry, there is always planned downtime that you want to mitigate; but you also end up having unplanned downtime. So, looking back at that historical data of when you’ve had those moments where maybe it took a bit longer than expected to manufacture something, you can take a look at all of that data and use AI to try to identify the proximate cause or to see something that might jump out even in the processing and design phases. We think of AI oftentimes as a predictive tool, or as a robot doing something, but a lot of times you get a lot of insight from the data through AI. What are the benefits of using AI for chip design? Gorr: Historically, we’ve seen a lot of physics-based modeling, which is a very intensive process. We want to do a reduced order model, where instead of solving such a computationally expensive and extensive model, we can do something a little cheaper. You could create a surrogate model, so to speak, of that physics-based model, use the data, and then do your parameter sweeps, your optimizations, your Monte Carlo simulations using the surrogate model. That takes a lot less time computationally than solving the physics-based equations directly. So, we’re seeing that benefit in many ways, including the efficiency and economy that are the results of iterating quickly on the experiments and the simulations that will really help in the design. So it’s like having a digital twin in a sense? Gorr: Exactly. That’s pretty much what people are doing, where you have the physical system model and the experimental data. Then, in conjunction, you have this other model that you could tweak and tune and try different parameters and experiments that let sweep through all of those different situations and come up with a better design in the end. So, it’s going to be more efficient and, as you said, cheaper? Gorr: Yeah, definitely. Especially in the experimentation and design phases, where you’re trying different things. That’s obviously going to yield dramatic cost savings if you’re actually manufacturing and producing [the chips]. You want to simulate, test, experiment as much as possible without making something using the actual process engineering. We’ve talked about the benefits. How about the drawbacks? Gorr: The [AI-based experimental models] tend to not be as accurate as physics-based models. Of course, that’s why you do many simulations and parameter sweeps. But that’s also the benefit of having that digital twin, where you can keep that in mind—it’s not going to be as accurate as that precise model that we’ve developed over the years. Both chip design and manufacturing are system intensive; you have to consider every little part. And that can be really challenging. It’s a case where you might have models to predict something and different parts of it, but you still need to bring it all together. One of the other things to think about too is that you need the data to build the models. You have to incorporate data from all sorts of different sensors and different sorts of teams, and so that heightens the challenge. How can engineers use AI to better prepare and extract insights from hardware or sensor data? Gorr: We always think about using AI to predict something or do some robot task, but you can use AI to come up with patterns and pick out things you might not have noticed before on your own. People will use AI when they have high-frequency data coming from many different sensors, and a lot of times it’s useful to explore the frequency domain and things like data synchronization or resampling. Those can be really challenging if you’re not sure where to start. One of the things I would say is, use the tools that are available. There’s a vast community of people working on these things, and you can find lots of examples [of applications and techniques] on GitHub or MATLAB Central, where people have shared nice examples, even little apps they’ve created. I think many of us are buried in data and just not sure what to do with it, so definitely take advantage of what’s already out there in the community. You can explore and see what makes sense to you, and bring in that balance of domain knowledge and the insight you get from the tools and AI. What should engineers and designers consider when using AI for chip design? Gorr: Think through what problems you’re trying to solve or what insights you might hope to find, and try to be clear about that. Consider all of the different components, and document and test each of those different parts. Consider all of the people involved, and explain and hand off in a way that is sensible for the whole team. How do you think AI will affect chip designers’ jobs? Gorr: It’s going to free up a lot of human capital for more advanced tasks. We can use AI to reduce waste, to optimize the materials, to optimize the design, but then you still have that human involved whenever it comes to decision-making. I think it’s a great example of people and technology working hand in hand. It’s also an industry where all people involved—even on the manufacturing floor—need to have some level of understanding of what’s happening, so this is a great industry for advancing AI because of how we test things and how we think about them before we put them on the chip. How do you envision the future of AI and chip design? Gorr: It’s very much dependent on that human element—involving people in the process and having that interpretable model. We can do many things with the mathematical minutiae of modeling, but it comes down to how people are using it, how everybody in the process is understanding and applying it. Communication and involvement of people of all skill levels in the process are going to be really important. We’re going to see less of those superprecise predictions and more transparency of information, sharing, and that digital twin—not only using AI but also using our human knowledge and all of the work that many people have done over the years.

Quantum computing is a devilishly complex technology, with many technical hurdles impacting its development. Of these challenges two critical issues stand out: miniaturization and qubit quality. IBM has adopted the superconducting qubit road map of reaching a 1,121-qubit processor by 2023, leading to the expectation that 1,000 qubits with today’s qubit form factor is feasible. However, current approaches will require very large chips (50 millimeters on a side, or larger) at the scale of small wafers, or the use of chiplets on multichip modules. While this approach will work, the aim is to attain a better path toward scalability. Now researchers at MIT have been able to both reduce the size of the qubits and done so in a way that reduces the interference that occurs between neighboring qubits. The MIT researchers have increased the number of superconducting qubits that can be added onto a device by a factor of 100. “We are addressing both qubit miniaturization and quality,” said William Oliver, the director for the Center for Quantum Engineering at MIT. “Unlike conventional transistor scaling, where only the number really matters, for qubits, large numbers are not sufficient, they must also be high-performance. Sacrificing performance for qubit number is not a useful trade in quantum computing. They must go hand in hand.” The key to this big increase in qubit density and reduction of interference comes down to the use of two-dimensional materials, in particular the 2D insulator hexagonal boron nitride (hBN). The MIT researchers demonstrated that a few atomic monolayers of hBN can be stacked to form the insulator in the capacitors of a superconducting qubit. Just like other capacitors, the capacitors in these superconducting circuits take the form of a sandwich in which an insulator material is sandwiched between two metal plates. The big difference for these capacitors is that the superconducting circuits can operate only at extremely low temperatures—less than 0.02 degrees above absolute zero (-273.15 °C). Superconducting qubits are measured at temperatures as low as 20 millikelvin in a dilution refrigerator.Nathan Fiske/MIT In that environment, insulating materials that are available for the job, such as PE-CVD silicon oxide or silicon nitride, have quite a few defects that are too lossy for quantum computing applications. To get around these material shortcomings, most superconducting circuits use what are called coplanar capacitors. In these capacitors, the plates are positioned laterally to one another, rather than on top of one another. As a result, the intrinsic silicon substrate below the plates and to a smaller degree the vacuum above the plates serve as the capacitor dielectric. Intrinsic silicon is chemically pure and therefore has few defects, and the large size dilutes the electric field at the plate interfaces, all of which leads to a low-loss capacitor. The lateral size of each plate in this open-face design ends up being quite large (typically 100 by 100 micrometers) in order to achieve the required capacitance. In an effort to move away from the large lateral configuration, the MIT researchers embarked on a search for an insulator that has very few defects and is compatible with superconducting capacitor plates. “We chose to study hBN because it is the most widely used insulator in 2D material research due to its cleanliness and chemical inertness,” said colead author Joel Wang, a research scientist in the Engineering Quantum Systems group of the MIT Research Laboratory for Electronics. On either side of the hBN, the MIT researchers used the 2D superconducting material, niobium diselenide. One of the trickiest aspects of fabricating the capacitors was working with the niobium diselenide, which oxidizes in seconds when exposed to air, according to Wang. This necessitates that the assembly of the capacitor occur in a glove box filled with argon gas. While this would seemingly complicate the scaling up of the production of these capacitors, Wang doesn’t regard this as a limiting factor. “What determines the quality factor of the capacitor are the two interfaces between the two materials,” said Wang. “Once the sandwich is made, the two interfaces are “sealed” and we don’t see any noticeable degradation over time when exposed to the atmosphere.” This lack of degradation is because around 90 percent of the electric field is contained within the sandwich structure, so the oxidation of the outer surface of the niobium diselenide does not play a significant role anymore. This ultimately makes the capacitor footprint much smaller, and it accounts for the reduction in cross talk between the neighboring qubits. “The main challenge for scaling up the fabrication will be the wafer-scale growth of hBN and 2D superconductors like [niobium diselenide], and how one can do wafer-scale stacking of these films,” added Wang. Wang believes that this research has shown 2D hBN to be a good insulator candidate for superconducting qubits. He says that the groundwork the MIT team has done will serve as a road map for using other hybrid 2D materials to build superconducting circuits.