One day in 1982, Joseph “Rod” Canion and two colleagues from Texas Instruments sat down at the House of Pies in Houston and used a napkin to sketch out what would become Compaq’s portable PC. In 1996, Felix Zandman, founder of Vishay Intertechnology, dined at Husker Steak House in Columbus, Neb., grabbed a napkin and drafted a design for a power metal strip resistor, which became crucial to power-management components in industrial, automotive, and consumer applications. Now that wispy piece of paper resides at the Smithsonian’s National Museum of American History. Perhaps most famously, Robert Metcalfe, working at Xerox PARC in 1973, roughed out some early designs for what would become the [?] Ethernet, though contrary to popular belief, no napkin was involved.

Yet another napkin was pressed into service last year, when Kurt Polzin, chief engineer of the space nuclear propulsion project at NASA’s Marshall Space Flight Center, met Robert Schleicher of General Atomics Electromagnetic Systems at a conference and started talking about nuclear rocket design.

“They did the classic let’s-sketch-out-an-idea-on-a-napkin,” says IEEE Spectrum’s Special Projects Editor and our in-house spaceflight expert Stephen Cass. “This rocket engine is still at the paper-planning stage, which, to be fair, is where most of NASA’s humans-to-Mars planning has been for the last 60 years.”

Meanwhile, the world’s richest person is pushing for humans to colonize Mars in time for him to escape our hospitable blue marble for a completely barren red planet. One big problem with this idea: It takes a long time in hostile space to get there.

“Here’s our napkin. Noodle on this with us and tell us what you think.” —Stephen Cass

There’s an old saw, mostly used in the context of road safety, that speed kills. But when it comes to sending humans on interplanetary missions, the faster the better. As Cass pointed out, “Once you get outside the Van Allen belts, there’s so much natural radioactivity—that’s the real killer.” Nuclear electric propulsion could both minimize launch costs and the time fragile human bodies are subjected to microgravity and radiation. And that makes a synchronal bimodal nuclear engine that could both power and propel a spaceship an attractive alternative to a conventional rocket.

Polzin and Schleicher think their nuclear engines could halve the time it takes to get to Mars. There may be ways to tweak the design to go even faster and further. But for now, the idea sits on the drawing board, awaiting feedback and refinement.

“They’re pulling together a lot of fairly mature technology,” says Cass, who edited Polzin and Schleicher’s article, “A Reimagined Nuclear Rocket.” “Electric propulsion is mature. Nuclear thermal propulsion is not, but thanks to [previous efforts], we have a good idea how to do it. The new part is merging them together, and that of course throws up its own challenges.”

Thanks to Cass and illustrator John MacNeill, Polzin and Schleicher’s idea has moved from a napkin to the pages of this month’s issue. Says Cass, “The whole point of the article is to say, ‘Here’s our napkin. Noodle on this with us and tell us what you think.’” We invite you to do so in the comments beneath the web version of this article. Or do it the old-fashioned way and mail the authors your own napkin.

While walking to school as a child, Jernej Barbič would marvel at how beautiful his home was. He was born and raised in a picturesque village in northwestern Slovenia (formerly Yugoslavia), located in the European Alps. Surrounded by alpine and beech trees, Barbic dreamed of replicating their swaying in the wind for others to enjoy.

At the time, he didn’t have the tools or the knowledge to create a system that could do that, but it sparked his interest in computer graphics, he says.

Jernej Barbič

Employer

University of Southern California, in Los Angeles

Title

Professor of computer science

Member grade

Senior member

Alma maters

University of Ljubljana, in Slovenia; Carnegie Mellon

Twenty years later, in 2016, Barbič, a professor of computer science at the University of Southern California, in Los Angeles, made his mark. His Ziva VFX software system allows for the creation of realistic muscle, fat, and skin simulations for 3D digital humans and creatures.

The technology was launched in 2016 by a startup he helped found, Ziva Dynamics, headquartered in Vancouver. It was acquired in 2021 by Unity Technologies of San Francisco.

Ziva VFX has been used in more than 60 movies including Aquaman and the Lost Kingdom; Godzilla x Kong: The New Empire; and Venom: Let There Be Carnage.

For the design and development of Ziva VFX, Barbič, an IEEE senior member, received a 2025 technical achievement Academy Award. It was a “tremendous honor,” he says, as the award recognizes technologies that have had a significant impact on motion picture production.

“Computer graphics and simulation can sometimes feel like a specialized technical field,” he says, “but the award showed that these ideas affect not just science but also art and how stories are told on screen.

“The digital characters enabled by mathematics become important parts of people’s lives.”

Sparking an interest in computer graphics

Barbič says he was inspired to pursue engineering by his father, an engineer who headed a cement factory’s research department and invented a technology that uses magnetic resonance imaging to test the integrity of cement. His father’s work showed him that “mathematics and physics are beautiful on their own, but engineering lets you build something that other people can use,” he says.

It was Barbič’s mother, an elementary school teacher, who introduced him to computer science. When he was 8 years old, the school his mother taught at bought a ZX Spectrum computer. With permission from the principal, she brought it home for her son to play on for two weeks. But he didn’t just play games; he created his own game using the BASIC programming language.

The machine came with a booklet that contained instructions on how to write a computer program, he says.

“At first,” he says, “I copied them verbatim without understanding what they did. But then I started realizing there is structure, and I modified the instructions.”

Of all the creatures brought to life using his technology, Barbič is particularly enamored with King Kong from 2024’s Godzilla vs. Kong.DNEG/Warner Bros. Entertainment Inc./Legendary

By the end of the two weeks, he’d developed a computer game where players guided a snowman along a winding road. It shifted unpredictably to the left or right, and players accumulated points by remaining on the road for as long as possible.

Barbič went on to earn a bachelor’s degree in mathematics in 2000 from the University of Ljubljana, in Slovenia. The following year, he moved to the United States to begin a doctoral program in computer science at Carnegie Mellon. It was a major turning point in his life, he says.

His doctoral research focused on developing simulation methods for objects that can change their shape when an outside force is applied to them, known as “deformable objects.”

That project shaped much of his later research, he says: “I became interested not only in making simulations accurate but also in making them practical: fast enough, robust enough, and controllable enough to be used in real applications.”

After earning his Ph.D. in computer science in 2007, he worked as a postdoctoral researcher at MIT. Two years later, he joined USC as an assistant professor.

Making movie magic possible

It was at USC that Barbič merged his passion for computer science with film. He developed Vega FEM, an open-source software program that allowed people to animate realistic 3D deformable objects. But Vega FEM was narrow in scope and not exactly what filmmakers needed, he says, so he started exploring how to create a version suitable for the movie industry.

“A major theme of my career has been the translation of research ideas into practical tools,” he says. “Academic research often produces beautiful algorithms, but it can be difficult to make those algorithms usable by artists, engineers, or production teams. I have always been interested in that bridge: taking rigorous computational methods and turning them into systems that people can actually use.”

In 2011 he attended an Association for Computing Machinery conference presented by its Special Interest Group on Computer Graphics and Interactive Techniques (SIGGraph). There he met James Jacobs, the creature supervisor at visual effects company Weta FX of Wellington, New Zealand. The company is behind the effects in the Lord of the Rings and Hobbit trilogies and other movies. Jacobs used Barbič’s software to create animals and fantastical creatures.

Two years later, Weta FX offered Barbič a summerlong research position in New Zealand. He accepted and spent the time studying the process of creating visual effects and learning what roadblocks existed in the film industry, he says.

At the time, the technology to create realistic soft-tissue and anatomical simulation for digital characters didn’t exist.

“The visual effects industry had reached a point where surface-level realism was not enough,” Barbič says. “A creature could have beautiful skin textures and detailed geometry, but if the bones, muscles, and fat underneath did not move correctly, the illusion would break.

“The problem was especially difficult for creatures and characters that need to feel alive: animals, monsters, fantasy creatures, or digital doubles. Their bodies may have unfamiliar anatomy, but the audience still anticipates them to move in a way that matches real-world expectations. Muscles should bulge and contract, skin should stretch and slide, fat should have inertia, and tissue should respond to motion and impact.”

In an effort to solve the problem Jacobs in 2014 approached Barbič about founding a startup. In 2015 they launched Ziva Dynamics, where they began what is now Ziva VFX.

The software uses physics-based simulations to model the internal anatomy of a character. It numerically solves the partial differential equations of nonlinear elasticity for musculoskeletal human and creature tissues, Barbič says. The equations describe how muscles, fat, skin, and connective tissue deform, interact with bones, and connect, and how muscles activate. Instead of animating only the outside surface, artists can create a model with underlying muscles, bones, soft tissue, and fat. Each component is assigned material properties, constraints, attachments, and activations. The simulator then computes how they deform and interact over time.

The technology uses ideas from computational mechanics, finite element methods, numerical optimization, contact handling, and computer graphics, Barbič says. Finite element simulation, a method used to predict how a product or structure reacts to heat and other real-world forces, provides a way to model deformable materials volumetrically, not just as surfaces, he says. The tool computes internal elastic forces and solves the equations of motion so the character’s tissues respond plausibly to animation, pose changes, muscle activation, and dynamic motion.

But the system had to be designed for artists, Barbič says. In production, he says, the goal is not only physical realism but also controllable realism.

“Artists need to direct the result, iterate, and fit the simulation into a larger animation pipeline,” he says. “So the technology had to combine scientific simulation with practical controls, robustness, and integration with visual effects workflows.”

Barbič says Ziva VFX has been used in more than 60 movies. Of all the creatures brought to life using his technology, he is particularly enamored with King Kong from 2024’s Godzilla vs. Kong.

“When King Kong is walking, you can see the muscles, how they’re very pronounced, and how they influence the shape of the skin. You can really feel the strength of King Kong,” he says. “And this was made through my software, so I think it’s amazing.”

After Ziva Dynamics was acquired by Unity, Barbič consulted for the company for almost two years.

In 2024 DNEG, a London-based visual effects and computer animation company, acquired the exclusive license to Ziva VFX.

Animating the human hand

Barbič strives to improve visual effects as an entrepreneur and an academic. His most recent research, funded by the U.S. National Science Foundation, focused on the modeling, simulation, and animation of human hands. The goal is to create computer models of hands that can be used to design tools, medical prosthetics, and robotic hands.

“The hand is a fascinating and difficult system,” Barbič says. “It contains many small bones, muscles, tendons, ligaments, skin, fat, and other soft tissues, all packed into a compact structure and interacting mechanically in complex ways.”

He and his team built a digital twin of the human hand.

He aimed to move toward “anatomically meaningful simulation,” he says. He used medical imaging, geometric modeling, finite element methods, and multibody simulation to represent the internal structures of the hand and its motions.

“IEEE lets me place my work not only in the world of images and animation but also in the world of engineering systems that must be accurate, stable, interactive, and useful.”

He worked with Bohan Wang, who at the time was a USC doctoral candidate, and George Matcuk, an associate professor of radiology. Wang is now an assistant professor of computer science at the National University of Singapore.

Barbič, Wang, and Matcuk scanned four people’s hands with an MRI machine. The two men and two women would position their hands in 12 poses, which allowed the team to gather data about how the bones, muscles, and fat move with each pose.

The data sets are available for anyone to use in their own studies.

“This project can help medical doctors learn more about how the hand is moving,” Barbič says. “It’s also great for roboticists to better understand how the human hand actually works, so [the movements] can be replicated.”

IEEE: Integral in interdisciplinary research

Barbič joined IEEE in 2008, when he published his research paper on simulation methods for deformable objects in the inaugural issue of the IEEE Transactions on Haptics. He has since published several papers in the IEEE Transactions on Visualization and Computer Graphics, which he says connected his work to a wider community interested in visual computing and computational methods. You can find his research in the IEEE Xplore Digital Library.

“IEEE recognizes the engineering side of computer science,” he says. “My work is often presented as computer graphics, but at its core, it is also simulation, mechanics, numerical methods, haptics, visualization, and software systems.

“IEEE is a community where that broader identity makes sense. It lets me place my work not only in the world of images and animation but also in the world of engineering systems that must be accurate, stable, interactive, and useful.”

He believes the organization is key in supporting a healthy interdisciplinary research ecosystem at a global scale—which, he says, is why he has served as an associate editor for Transactions on Visualization and Computer Graphics and Transactions on Haptics.

Being a member has made it easier for Barbič to connect with engineers in different fields, he says.

“My research often lives between categories: It is mathematical but also practical; visual but also mechanical; artistic but also engineering-driven,” he says. “IEEE is one of the professional communities where that mixture is understood.”

I never intended to join the cutting edge of electromechanical television. I just wanted to make a nice clock. But sometimes you have to go where the engineering takes you, and in my case it took me to the Scanwheel, a pocket-size wide-screen electromechanical TV with a resolution of 4,096 by 20 pixels. Yup, that’s 4K by 20.

The 3D-printed drum [top] is spun by a motor controlled by a driver board [second row, from top]. The driver board, in turn, is controlled by a Raspberry Pi Pico [middle], which also controls the LEDs [second row, from bottom], which are mounted in the 3D-printed casing [bottom] so that the holes pass over them as the drum turns.James Provost

Electromechanical television was the first form of practical television, developed by John Logie Baird in the 1920s. He used a so-called Nipkow disk, which has a spiral of holes punched through it. As the disk rotates, the holes pass one by one in front of a light source. By varying the brightness of the light as a hole travels across it, you can draw one scan line of a video frame. Spin the disk fast enough, and persistence of vision makes it look like an entire frame is being displayed simultaneously. Commercial electromechanical TV sets were produced in the United Kingdom, with regular broadcasts provided by the BBC in the 1930s.

Although cathode-ray tubes replaced electromechanical televisions in the 1940s, hobbyists have continued to build them and even improve on the original technology. For example, in the June 2022 installment of IEEE Spectrum’s Hands On, Markus Mierse presented a desktop-size 3D-printed color version.

I built an electromechanical display myself some years ago, but it had a traditional design with a Nipkow disk made from a vinyl record with holes drilled in it. Recently I started tinkering with electromechanical TV again as an outgrowth of my YouTube channel. There I’ve been focusing on developing volumetric displays, which create 3D pixels floating within a volume of space. In particular, I was interested in borrowing some ideas from plenoptic cameras, which use pinholes and lenses to capture multidimensional light fields of samples.

I wondered if I could run the process in reverse, to create light fields rather than capture them. I often explore ideas in two dimensions before expanding to the third, so I thought I’d first demonstrate a 2D display. I decided to make an electromechanical device into a clock. After all, you don’t need high resolution to display digits.

How Does the Scanwheel Display Work?

Thinking about the display as a clockface pushed me toward some key ideas. First, instead of having just one display area, I would use five light sources to create multiple areas—four to represent hours and minutes, and a central area for a separator that would blink each second. Second, to align the digits in a readable row rather than have them spread around an arc, I swapped the Nipkow disk for an established alternative: a Nipkow drum.

With a drum, the holes run along the curved cylindrical surface in a stair-step pattern. This means they always trace a straight line from the perspective of a viewer looking from the side, so the clock’s digits would be horizontally aligned.

“In tests, I’ve pushed the horizontal resolution to more than 8,000 pixels.”

These two decisions turned out to be key to achieving both miniaturization and high horizontal resolution. A disk needs a fairly wide diameter so that the scan lines aren’t ridiculously curved. But curvature isn’t a problem with a drum. A drum can be much smaller than a disk that has the same number of scan lines. (And unlike in the 1920s, packing multiple light sources close together inside a small drum isn’t a problem with modern LEDs.) I settled on a 6-centimeter-wide drum, turning the device from desktop-size to something you could carry in your pocket.

I then realized my five display zones could work in concert to create one single wide screen. Because it’s possible to modulate the brightness of an LED at very high rates, the horizontal resolution can also be very high. My system currently has 4K horizontal resolution, and this is primarily limited by the amount of onboard memory I have available. This memory holds the buffer that stores pixel data for each frame before it is read out to the LEDs and displayed. In tests, I’ve pushed the horizontal resolution to more than 8,000 pixels.

Despite the Scanwheel’s low vertical resolution—at 20 pixels, it has fewer scan lines than Baird’s 30-line televisions—its high horizontal resolution makes the legibility of the display surprisingly good: I can display not just crude digits but video streamed into the frame buffer.

Using the RP2040 Chip’s Special Silicon

That frame buffer lives on a Raspberry Pi Pico microcontroller board, based around the RP2040 microcontroller. The RP2040 is ideal for this project because of the chip’s dedicated PIO silicon. PIO stands for programmable input/output, and it’s a block of four coprocessors that uses a very limited instruction set. Each coprocessor can be set up to chew through input/output streams completely independently of the RP2040’s two CPU cores.

The first mechanical TVs used disks, which had to be wide to minimize image distortion but allowed bulky light sources. With small, modern light sources, a smaller drum can create images with minimal distortion.James Provost

It’s thanks to the PIO that I’m able to keep up with the spinning drum and modulate each of the five LEDs simultaneously as holes pass over them, a task complicated by the fact that the center LED is not a monochrome LED, but a color LED with separate red, green, and blue channels. In fact, the PIO does nearly all the work, pulling data from the frame buffer and controlling the LEDs and the spinning of the drum. The code running on the CPU (written in MicroPython) is primarily responsible for setting up the PIO and then leaving well enough alone.

A stepper motor connected to a driver board spins the drum, with power provided by the USB jack on the Pi Pico. All the Pi Pico has to do controlwise is send the board a pulse to incrementally advance the drum’s position once every millisecond. Video data is streamed into the Pi Pico via a network interface. You can set up the Scanwheel to mirror a portion of your computer’s screen, or to act as a separate display.

The casing, including the drum, is 3D printed. Now for a neat bit: In the Scanwheel’s GitHub repository at https://github.com/AncientJames/Scanwheel/tree/main, alongside all the other files you’ll need to make this project yourself, there’s an OpenSCAD file that generates the 3D-print file for the drum based on adjustable parameters. This means you can easily make a taller drum and add more scan lines, or try other customizations for your very own portable electromechanical display. You can even use it as a clock!

The IEEE–Eta Kappa Nu (IEEE-HKN) honor society is preparing to host the Innovating the Future event on 6 November.

The inaugural one-day, in-person event is designed to provide a forum for IEEE and IEEE-HKN undergraduate and graduate student authors to present their original research papers. A keynote address and thematic presentation sessions are planned as well. Student attendees can network with their peers and gain firsthand experience with the academic publishing process.

To present at the conference, students had to submit an abstract of their research before 1 May. Students whose work was accepted were assigned a volunteer IEEE member to mentor them and guide them through the research writing process, including presenting and publishing their original work.

Those whose paper was accepted by 1 August were invited to present at the conference. The conference proceedings will be submitted for publication in the IEEE Xplore Digital Library.

Upholding research integrity in a changing landscape

IEEE Life Fellow Manuel Castro, the conference’s technical program chair, oversees IEEE-HKN’s Innovating the Future program committee. It manages the review process, organizes logistics, and handles the mentoring component.

“This new conference is important to IEEE, as well as to IEEE-HKN,” Castro says, “because it allows student authors to grow in their skills and competencies, and be supported while turning their technical activities into publications.”

“The conference offers me a chance to learn how to communicate my research to a broader audience, gain feedback from other student researchers beyond my institution, and see how my work can be made more accessible.” —David Kwabi-Addo

IEEE Life Fellow Sorel Reisman, a California State University professor emeritus and an IEEE-HKN governor-at-large, says that because the academic research landscape is rapidly shifting, the conference is timely.

“As AI increasingly threatens the integrity of research papers being published in leading journals and conference proceedings, it is essential that future scholars—many of them current IEEE-HKN students—grasp the established standards of legitimate, peer-reviewed research publishing,” Reisman says.

Perspectives from mentors and students

A cornerstone of the conference is its rigorous mentorship initiative, which pairs each author of an accepted abstract with an experienced IEEE volunteer. The mentors provide personalized guidance on organizing the students’ technical content into the correct format for publishing. They also discuss navigating the peer review process, structuring presentations, and preparing the final manuscript for publication.

The impact of the guided process can be valuable for both the mentors and their mentees. IEEE Member Wafa Elmannai, associate professor and chair of the electrical and computer engineering department at Manhattan University, in Riverdale, N.Y., and faculty advisor to the IEEE-HKN Gamma Alpha chapter, serves as a mentor.

“Research is essential to advancing technology and driving innovation,” Elmannai says.

She volunteered to be a mentor, she says, because she has seen how conducting research can transform a student’s future by building their confidence, curiosity, and critical thinking skills.

“Mentoring encourages students to step outside their comfort zones and develop innovative solutions that contribute to society,” she says.

For the students, the conference can be a critical stepping stone. David Kwabi-Addo, an IEEE graduate student member who is researching computational biology at MIT, is president of the IEEE-HKN Beta Theta chapter. He says he views the program as an opportunity to gain experience in producing academic scholarship.

“I submitted an abstract of my research paper because I see the conference as a chance to produce what could become my first conference publication,” Kwabi-Addo says. “The conference offers me a chance to learn how to communicate my research to a broader audience, gain feedback from other student researchers beyond my institution, and see how my work can be made more accessible.”

He says he hopes his participation will highlight the diverse breadth of research that future conferences can showcase.

Workshops on the publishing process

Conference organizers are holding a series of workshops to guide students through every step of the academic publishing process. The workshops are open to anyone and available on the IEEE-HKN YouTube channel.

Topics previously covered are:

• The Art of Crafting a Compelling Abstract (13 March).
• Identifying When a Project Is Mature Enough for Publication (15 May).
• The Mechanics of Writing a Technical Paper (19 June).
• Surviving the Review Cycle and Dealing With Criticism (14 August).

Registration is open to all for this upcoming workshop:

• From Pen to Voice: Adapting a Paper Into a Compelling Conference Talk (2 October).

A launchpad for the next generation

The Innovating the Future program is designed not only to improve the quality of submissions but also to foster long-term professional development and research communication skills to develop the next generation of IEEE authors. The conference is more than a venue for presenting research; it is a launchpad for innovators committed to advancing technology for humanity.

The biggest threat to any crewed expedition to Mars is time. NASA’s shortest blueprint for sending people to the Red Planet and back requires spending 620 days in space and 30 days on Mars. Even setting aside the compounding challenges of building life-support systems that can operate without resupply for that long, or the fact that longer journeys leave more time for unlucky accidents, life in microgravity and solar and cosmic radiation will inexorably exact their cumulative toll on human bodies.

We want to make it possible to dramatically reduce the length of time crews must spend in space—down to just 335 days in transit or less. This will both simplify many engineering challenges and keep astronauts healthier and safer. We believe the key to this time reduction is a new approach to building a holy grail of space exploration, the bimodal nuclear rocket.

In the 1960s, U.S. open-air ground tests demonstrated much of the technology needed for nuclear thermal rockets as part of the NERVA and Rover projects.Nevada State Museum, Las Vegas

Technicians at NASA’s Lewis Research Center test a nozzle design for a nuclear thermal rocket in 1965. GRC/NASA

The prototype SNAP-10A, orbited in 1965, is to date still the only nuclear reactor launched into space by the United States. George Rinhart/Corbis/Getty Images

We are Kurt Polzin, chief engineer of NASA’s space nuclear propulsion project at the Marshall Space Flight Center, with over two decades of experience in advanced propulsion research, and Robert Schleicher, chief engineer for nuclear technologies and materials at General Atomics. And to explain just what a bimodal nuclear rocket is, and why the new version we have conceived together brings it closer to future reality, we first need to take a quick trip to the past.

As early as 1946, researchers realized that nuclear reactors had the potential to become extremely efficient thermal rocket engines. Most rockets are thermal rockets, and they work by expelling hot gases through a nozzle, thrusting the rocket forward. While there are other factors such as nozzle shape, generally speaking, the hotter and faster you make the rocket’s exhaust gases, the more acceleration the rocket will produce for a given mass of propellant. Because a smaller molecule will move faster than a larger one when heated to a given temperature, the smaller the molecular mass of your propellants, the better. By convention, the efficiency of a rocket engine is measured by how long the engine can exert a thrust equal to the initial weight of its propellant, a quantity known as specific impulse.

In a conventional thermal rocket, such as those used in every launch to orbit since Sputnik, the exhaust temperature and speed—and thus the specific impulse—is dictated by the energy released by a chemical reaction and the mass of the reaction’s by-product. The most efficient chemical rockets today combust hydrogen with oxygen, producing water and a specific impulse that tops out around 450 seconds.

But a nuclear rocket is not limited by chemistry. The heart of a nuclear thermal rocket is a nuclear fission reactor, in which chain reactions in uranium fuel release much more energy per kilogram than is possible with chemical combustion. A turbopump forces liquid hydrogen alone—with its very small molecular mass—through the reactor’s core, heating it to temperatures of at least 2,700 kelvin before expelling it, resulting in a specific impulse of 900 seconds or more.

In the 1950s and 1960s, the Rover and NERVA (Nuclear Engine for Rocket Vehicle Applications) programs ­ground-tested nuclear thermal rockets. By the early 1970s, the technology had matured to the point where flight tests were being planned. But changing political and budgetary winds led to nuclear thermal development being shut down in 1973.

Another prong of nuclear propulsion that has also demonstrated considerable promise is nuclear electric propulsion. In electric propulsion, instead of creating a stream of hot rocket exhaust through chemical reactions or exposure to the core of a nuclear reactor, electricity is generated and used to create electromagnetic fields that accelerate an ionized propellant such as xenon or lithium.

Various schemes to do this exist, including some that have already seen considerable time in space, such as the ion thrusters used on the Dawn asteroid mission launched in 2007. So far, these electric thrusters have only been powered by solar panels. But with a nuclear reactor as part of a power plant that supplies the juice, more thrust could be produced. And moving beyond solar power is particularly important in missions to the outer solar system where sparse solar photons would require enormous solar arrays.

With electric thrusters, specific impulses in the range of 2,200 to 4,600 seconds are possible, but currently with very low thrust. With the energy available to a nuclear-powered electric ­propulsion engine, you could have greater acceleration and reduced mission times. The nuclear reactor could also provide electrical power for all the spacecraft systems as well.

The System for Nuclear Auxiliary Power (SNAP) program launched the SNAP-10A in 1965 as a proof of concept, the first—and so far only—U.S. nuclear power reactor in space. It generated about 600 watts of electrical power for 43 days before shutdown and is still in orbit. Subsequent U.S. initiatives for more substantive electric power and nuclear thermal propulsion systems, such as the SP-100, Project Timberwind, and Project Prometheus, along with more recent projects like Demonstration Rocket for Agile Cislunar Operations (DRACO) and Joint Emergent Technology Supplying On-Orbit Nuclear (JETSON), have emerged sporadically over the years. None of these have yet progressed to actual flight.

However, space nuclear power got a huge shot in the arm in March 2026 when NASA Administrator Jared Isaacman announced a new space exploration initiative. As part of that initiative, the agency plans to launch Space Reactor-1 Freedom (SR-1) to deliver a trio of robot-survey helicopters to Mars. Driven by nuclear electric propulsion, SR-1 aims to demonstrate fission technology in deep space and would be the first nuclear-powered interplanetary spacecraft, generating 20 kilowatts of electric power aboard.

This is a bold step for NASA, and brings us up to the present, but the details of the proposed mission also highlight a familiar limitation of nuclear electric propulsion. Even with improved acceleration, electric propulsion still cannot generate the powerful bursts of thrust needed to escape gravity wells, such as those of Earth or Mars, or perform time-critical maneuvers, like course corrections. On the other hand, while not as efficient and unable to supply electrical power for spacecraft systems, nuclear thermal engines are great at delivering high thrust at critical moments.

What is a bimodal nuclear rocket?

Some engineers would suggest we build two separate systems—one reactor for thermal propulsion and another reactor for power and electric propulsion. But since at least the 1990s, it has been the dream of many engineers to combine nuclear thermal and nuclear electric in one package, with one reactor: the bimodal nuclear rocket.

Most previous bimodal proposals depend on complex valve arrangements to integrate the propulsion and power systems. In thermal propulsion mode, the reactor is brought to maximum activity by a set of control drums that ring the core, which is composed of a matrix of long uranium-fuel elements. The drums take the shape of long cylinders made of beryllium, with a 120-degree segment of each cylinder covered with boron carbide. Boron absorbs neutrons, and when that segment faces the reactor, the reactor’s activity is low as neutrons escaping from the core are captured. Rotating the boron segment so that it faces away from the core (leaving only the beryllium exposed) increases nuclear activity as the beryllium reflects escaping neutrons back into the core’s fuel elements, where they can contribute to chain reactions.

This proposed trajectory, developed at NASA’s Glenn Research Center, shows where high-thrust maneuvers [blue dots] are executed by a nuclear thermal engine and additional low-thrust, high-efficiency acceleration and deceleration is performed by electric propulsion [hashed lines show thrust direction].NASA Glenn Research Center

Once the reactor is generating large amounts of heat, liquid hydrogen is pumped through channels that run the length of the core. Turned into an expanding hot gas, the hydrogen blasts from the other end of the core to form the rocket’s powerful exhaust.

In nuclear power mode, the reactor’s activity is damped. Valves seal the channels and a so-called power-conversion fluid—typically a mixture of helium and xenon gas—circulates through the reactor in a closed loop. The reactor is still hot enough to warm this fluid, which drives a turbine connected to an electrical generator.

The key point here is that a single set of flow channels and nuclear-fuel elements are used for both modes. But the valves used to switch modes face the formidable challenge of enduring months, or even years, in a harsh radiation environment while maintaining leak-tight performance.

The core’s activity is controlled by the rotating drums surrounding it. Within the core, low-temperature fuel elements [left in blue, and top right] produce electric power by heating a circulating fluid. High-temperature fuel elements [left in red, and bottom right] heat hydrogen as a propellant. (The taper of the HTFE’s exhaust channel is exaggerated for illustrative purposes. Ways of packaging the HTFE’s uranium fuel other than with particles are possible.)John MacNeill

In addition, the nuclear-fuel elements surrounding the channels must be able to operate for short durations at very high temperatures during thermal thrust maneuvers and for long durations at lower temperatures during the rest of the voyage. It is difficult to build one type of element capable of both. Hence, the complexity and demanding engineering requirements of previous bimodal designs has hindered their practical application.

We propose a simplified approach, a hybrid system we call the synchronal bimodal nuclear rocket (S-BNR). The genesis for this design came about when we were attending a conference together in 2025. One of us (Polzin) had an initial idea, and in time-honored tradition, he sketched it out on a napkin to see if the other (Schleicher) thought there was actually a way to do it. We’ve been working on refining the concept ever since.

How the synchronal bimodal nuclear rocket works

Rather than relying on a complex valve system, the S-BNR uses two hydraulically independent loops within a single reactor core, one open loop (for thermal propulsion) and one closed loop (for electrical power). The core is divided into two zones, one per loop, differentiated by the type of fuel elements in each. Several designs for the fuel elements are possible: In our preliminary design, the high-temperature fuel elements (HTFEs) in the thermal propulsion zone consist of a bed of “pebbles”—uranium fuel encased in zirconium carbide—that surround a central tapering channel and operate at greater than 2,700 K. (One possible alternative for the HTFEs would be a solid fuel design, as with NERVA.) The hydrogen propellant passes through the pebble bed, where the pebbles’ large surface area maximizes the transfer of heat needed for efficient high-thrust propulsion.

The other zone has low-temperature fuel elements (LTFEs), optimized for long-term, efficient production of electricity, which can range from tens of kilowatts to several megawatts. In these elements, the uranium fuel in solid form surrounds a double-walled channel: The power-conversion fluid is pumped down the inside and returns along the outside wall, absorbing heat from the fuel and operating at moderate temperatures (at or above 1,200 K).

The electric-power and nuclear-thrust elements of the core have separate fluid loops, which eliminates the need for valves to switch between closed-loop operation for power generation and open-loop operation for propulsion.John MacNeill

Both the HTFEs and LTFEs contribute the neutrons required to sustain chain reactions. In power-only mode, residual heat moves from the HTFEs into adjoining LTFEs. The physical interface between the elements is designed to moderate this thermal flow to balance two competing needs: It must allow enough heat flow to safely remove the residual heat from the HTFEs, but it must also limit that heat flow so the LTFEs’ temperatures do not go past their allowable limits when the HTFEs operate at high power.

During combined propulsion and power operation, a heat exchanger on the power loop preheats the hydrogen propellant for the thrust loop, aiding the turbopump that feeds the hydrogen through the core. After a propulsion burn is completed and the HTFE chain reactions are damped by the control elements, the power loop removes residual-decay heat coming from the HTFEs as described above, eliminating the requirement in earlier designs for additional propellant flow just to cool down the core while on standby. This dual-loop system also means the engine can produce high thrust whenever needed while allowing the generator to remain active at all times—a significant advantage for crewed missions.

By adopting this dual-loop architecture, the S-BNR removes the need for the problematic mode-switching valves found in earlier concepts. Each fission zone is constructed with materials tailored to its specific temperature and power requirements, ensuring optimal performance and durability. The result is uninterrupted electrical power across all mission stages, making it unnecessary to carry additional liquid hydrogen just to manage decay heat.

The challenges ahead

While significant progress in developing the design of the S-BNR has been made, substantial challenges remain. The reactor must maintain stable control across a wide power range, from modest levels for electricity generation to hundreds of megawatts of thermal power during high-thrust operation. Operating the ­power-generation loop in close proximity to the HTFEs requires very careful management of both temperature and the neutrons emitted by the fuel elements.

And crucially, demonstrating reliable, long-duration performance is particularly demanding: Missions to Mars may require years of continuous power generation. ­Outer-planet probes equipped with S-BNR engines could extend that to a decade or longer.

In the past, nuclear thermal propulsion fuel elements were engineered for extremely high temperatures but only brief operational lifetimes (typically hours), whereas proposed nuclear electric propulsion fuel elements are optimized for lower temperatures and intended to last for years. By using two different types of fuel elements in the S-BNR, we can take advantage of the design heritage of both these development tracks. Fortunately, recent NASA-sponsored research has produced several promising candidates that may meet these demanding requirements.

Ground-testing these systems is also a challenge. Early in the Rover and NERVA era, the exhaust from test engines was blasted into the atmosphere, something now unacceptable. Today, any ground test of an engine must completely capture all potentially radioactive exhaust products. Fortunately, a number of approaches have been developed to capture and scrub the exhaust, although these methods currently carry a significant price tag.

Then there is the ultimate test: flying an S-BNR in space. International regulatory and safety protocols for nuclear launches were developed largely in response to the Soviet Union’s launch of dozens of nuclear-powered Radar Ocean Reconnaissance Satellite (RORSAT) radar spy satellites in the 1970s and 1980s. There were a number of incidents, with the most serious leaving radioactive debris strewn across a swath of Canada in 1978. This history led to a consensus in the space community that might be summarized as “Thou shalt not bring a nuclear reactor to criticality in any Earth orbit that decays faster than dangerous isotopes.”

Thus any S-BNR would be launched atop a conventional chemical rocket, with a completely cold reactor and fresh fuel. Fresh uranium fuel is not in fact very radioactive: The potentially larger concern is the chemical toxicity of this heavy metal, but it can easily be handled by wearing light protective suits, respirators, and gloves. Only after the control elements have been adjusted to permit chain reactions to begin within the core are highly radioactive isotopes able to form from fission fragments. There would be even less cause for concern than when launching a radioisotope thermoelectric generator (RTG), such as the sort that are currently powering the Perseverance rover on Mars and the New Horizons mission in the outer solar system.

Even in the most extreme scenario imaginable—the chemical booster explodes and somehow damages the reactor’s control elements in just the right way to initiate a chain reaction—there wouldn’t be time to produce a large amount of toxic isotopes before the reactor broke apart and reactions ceased. (We can be sure of this because Project Rover actually tested this kind of worst-case scenario in 1965 with the Kiwi-TNT test, where an engine prototype was rigged to produce a runaway chain reaction sufficient to vaporize the reactor core due to the immense internal pressure buildup. Negligible radiation spread outside a radius of two miles (3.2 kilometers), well within the range of safe distances for launching any rocket capable of reaching orbit, and site decontamination was possible after only a few days of radioactive decay.)

Despite all these considerable engineering challenges, the foundation laid by decades of investment in nuclear thermal and electric propulsion and terrestrial nuclear power technologies provides a solid platform for continued advancement. Indeed, much of the foundational work is already underway through ongoing NASA and U.S. Space Force efforts.

The Dawn asteroid mission relied on electric thrusters, demonstrating their utility for long-duration spaceflight.JPL-Caltech/NASA

We envision the following action plan to merge these technology pathways: Modeling must be performed to demonstrate and verify strategies for thermal management and the control of nuclear processes over the full range of operating power levels. Near-term non-nuclear testing will validate fluid loop operation, heat transfer mechanisms, and control strategies. Next, component-level irradiation and thermal trials will qualify new materials. Then, integrated reactor testing will begin, first without nuclear fuel and later with fueled reactors undergoing fission. Finally, initial in-space demonstrations could begin with lower-power systems, eventually scaling up to full bimodal capabilities.

Achieving success will require close collaboration across NASA, the Department of Energy, the Department of Defense, industry partners, and the broader technical community. Progress will depend on advancements in ­high-temperature fuels and materials, improved systems for power conversion and heat transport, and the adoption of innovative manufacturing techniques and methods to control nuclear fission over a wide range of output power. In particular, integrated system testing will be more complex than previous programs such as NERVA, due to the combined functions and distinct operational regimes for thermal propulsion and power generation. We hope engineers and researchers with relevant expertise will be encouraged to contribute to addressing these challenges, whether in the areas of thermal management, reactor modeling and control, extended-duration testing, or safety analysis.

Past ground tests and limited demonstrations have already established the capabilities of space nuclear systems. With architectures like the synchronal bimodal nuclear rocket, the prospect of integrating high-thrust propulsion and sustained power generation becomes increasingly practical and versatile. The next phase is not simply about traveling fast. It’s about building crewed and uncrewed spacecraft that can reliably travel to destinations throughout the solar system that are currently difficult or impossible to reach, with missions potentially lasting years or even decades.

This article appears in the September 2026 print issue as “A Reimagined Nuclear Rocket.”

At the height of the Cold War, one very specialized computer was so secret that the world didn’t know it existed. It ran its jobs up to 200 times as fast as any other computer of its time. It was the U.S. National Security Agency’s main cryptographic processor in operation from the time of the Cuban Missile Crisis in 1962 through the Vietnam War and on past the 1975 Helsinki Accords. The machine stopped running only when its moving parts finally gave out.

The Harvest computer mattered because of what it was as well as when it ran. For 14 years, it was the engine processing the NSA’s most sensitive intercepts at a time when signals intelligence was as close to a strategic weapon as anything short of a warhead.

Designed and built by IBM for the NSA, Harvest was one of the first machines designed to apply operations to enormous datasets rushing past, a precursor to the computers today that manage continuous video streams and security systems in real time. It was also one of the first machines built as an add-on—a specialized helper intended to do one job exceptionally well, bolted onto a general computer. Harvest’s modular design is like a 1960s version of today’s graphics chips that CPUs use to run intensive video-game and AI processing loads.

All that raw processing power meant that Harvest also needed nonstop rivers of data to run on. And that led to another pioneering achievement: the world’s first automated tape library that could robotically fetch any one of hundreds of large cassettes of magnetic tape from the machine’s racks.

Given Harvest’s unprecedented processing and storage capacity, the machine’s designers naturally needed to rethink how their system handled information. So IBM wrote a customized programming language called Alpha to let code breakers rigorously describe cryptographic problems, just as scientists at the time were using the emerging language Fortran to describe equations and data-processing algorithms.

In Fort Meade, Md., an NSA data center hosted one of the world’s fastest computers of its time—although not often discussed, because of its sensitive, high-security code breaking and cipher hunting work. National Cryptologic Museum

The story of Harvest, pieced together from declassified documents and contemporary manuals and technical overviews, provides a new and unexpected vista on the history of computing. It also offers a case study in how national security needs, especially during the Cold War, pushed computer technology beyond the far reaches of what unclassified, civilian computing could achieve. Harvest’s distinctive history reveals a visionary algorithmic, coding, memory, and hardware architecture occasionally decades ahead of its time. But this machine was also built only once, for one singular purpose, and then ultimately quietly retired.

The Heart of NSA’s Secret Machine

IBM’s landmark 1960 transistorized mainframe, the IBM 7030, better known as Stretch, provided the front end for Harvest (which was officially known as the IBM 7950). IBM delivered Stretch to eight or nine customers, mostly scientific research labs, from 1961 through ’63. Designed and prototyped throughout the second half of the 1950s, Stretch introduced the now standard notion of an 8-bit byte. For its first three years of operation, Stretch was the non-classified world’s fastest computer, although it failed to meet IBM’s aggressive goal of running 100 times as fast as Stretch’s predecessor, the IBM 704. While IBM engineers in Poughkeepsie, N.Y., were designing and building Stretch, the company was also quietly discussing a new system that would be built for NSA.

At the time, NSA’s existing cryptanalytic computers—large, batch-processing machines that required human operators to manually stage each tape run—were struggling to keep pace with the sheer volume of intercepted message traffic coming in from around the globe. What the agency needed was a machine that could process an unbroken river of incoming data, automatically, around the clock. That requirement alone profoundly shaped Harvest’s design.

IBM’s Harvest system, custom-built for the NSA for code breaking, paired the IBM 7030 Stretch mainframe with a bespoke data-stream processor. Stretch handled ordinary computing and input/output, including the Tractor automated tape library. Both units shared two kinds of memory: a large main bank and a smaller, faster bank. When Stretch switched to streaming mode, Harvest drew two streams of data, P and Q, from memory, processed them in parallel, and returned the results as a third stream, called R. Chris Philpot

After two failed proposals to NSA, in 1958 IBM finally landed the contract: a Stretch-based machine, augmented by a custom coprocessor, with a revolutionary tape-based storage system, called Tractor.

Stretch’s forte was floating-point math for scientific computations. IBM had designed it primarily for labs working on frontier research like nuclear weapons design and weather prediction. By contrast, the custom coprocessor to be built atop Stretch would help NSA analysts sift through alphanumeric characters—that is, essentially integer data.

Harvest’s coprocessor was the opposite of a general-purpose system. It was, rather, a streaming computer. Instead of executing long series of instructions, it followed one fixed sequence of steps and applied that same sequence to every pair of characters as they streamed past. Harvest shared memory with the main Stretch processor and ran in bursts. Either Stretch was operating, or else it suspended itself while Harvest’s coprocessor shot through data in memory at extreme speeds.

Stretch and Harvest were among the first large computers built entirely from transistors packaged in circuit cards and housed in large, refrigerator-size frames. A 1962 technical manual about Stretch describes the machine’s CPU as divided into functional sections—the instruction unit, the look-ahead unit, the (parallel and serial) arithmetic unit, and the memory bus unit. Harvest inherited Stretch’s basic circuit design but then added something unconventional: Its streaming units processed data in overlapping stages called a pipeline. So while one pair of data bytes was being compared, the next pair was being fetched from memory.

Harvest’s coprocessor operated by fetching two streams of data, called P and Q, from the system’s memory, performing operations on them, then writing the results to memory as a third stream, R. Each stream could be anywhere from 1 to 8 bits wide. Harvest’s memory was bit-addressable, meaning word boundaries could be ignored entirely. For instance, it could fetch just 5 bits rather than filling out a whole byte. Streams P, Q, and R included flexible provisions for looping and addressing data in complex patterns—allowing, for example, repeated fetching of short strings from memory.

Data from P and Q fed into two functional units. The simpler was the logic unit, which performed basic, bitwise operations—the same operations any programmer would recognize today—and wrote its results back to memory. The more complex was a table-lookup unit. It combined incoming data from P and Q to form an address in memory, which could then be used to advance a counter by one, set a specific bit, or retrieve a stored value. The latter unit functioned, in effect, like the rotor wheel inside a cipher-encoding/decoding machine of the era, the kind that electronically substituted one value for another according to the cipher machine’s wiring.

Harvest’s complexity baffled some at the NSA. During employee tours, according to James Bamford’s 2001 NSA history, Body of Secrets (Doubleday), officials would point to the machine and scoff, “It’s beautiful, but it doesn’t work.”

Not everyone at the agency was put off by the monumental device, however. One of the few documented examples of Harvest at work, recounted by Bamford, describes the machine searching 3.5 billion characters of text for any of 7,000 target terms, in just under 4 hours.

In unclassified remarks from 1972, NSA analyst Robert Looney mentions one job Harvest had tackled—though he didn’t specify the end goal or the code-breaking effort behind it. Codenamed “Moretown,” the job involved sifting through 11 million messages spanning 16 years of intercepted traffic against a list of some 8,000 search terms—all in about ten hours.

IBM’s Frances Allen helped design Alpha, Harvest’s custom-built programming language.IBM

IBM’s James H. Pomerene was chief engineer of Harvest, supervising its custom-designed circuits that’d been optimized for algorithms used in many cryptographic jobs.IEEE

IBM’s Fred Brooks Jr. was a key co-architect of Harvest’s hardware system. Computer History Museum

As a unified system, Harvest—that is, Stretch plus IBM’s custom-built streaming processor add-on—streamed 1 byte every 0.3 microseconds, and it boasted about 800 kilobytes of addressable memory.

“Here you see one bank pulled out of its oil bath,” Looney said in his 1972 remarks celebrating Harvest’s tenth anniversary of operations. He held up a photo of Harvest’s magnetic core memory banks—six of them, submerged in oil for cooling.

Factor in the time demands of various data fetches from Tractor’s tape archives, and a single Harvest “instruction” sometimes carried on, without needing any human intervention, for hours.

“It was quite an amazing computer,” recalled IBM Fellow Emerita Frances Allen in a 2001 oral history. “One instruction, for example, could do sorts, and do statistical analysis of the data that was streaming by it.… Everything we were doing at that time was on the cutting edge. There was no question about it.”

Allen, who received the A.M. Turing Award in 2006, was one of the developers who worked on both Stretch and Harvest. At the time she started working on Harvest, Allen noted, the Fort Meade, Md.–based NSA was largely unknown outside of classified intelligence circles. So she at first assumed she was working on an unspecified naval project. “We thought of ourselves as working for the Bureau of Ships, because that was the code name for NSA in the budget!” recalled Allen, who died in 2020.

Other key Harvest designers and early developers wound up becoming influential figures over the course of computing history. Frederick Brooks Jr., recipient of the 1999 Turing Award and a major contributor to the hardware and software for IBM’s System/360, also helped develop Harvest. And James Pomerene, prior to his involvement with Harvest as its chief engineer, had previously helped build the pioneering IAS computer alongside John von Neumann.

How Tractor Stored a World of Data

IBM built the Tractor tape system (IBM 7955) to attach to the same Stretch machine that hosted Harvest, because no existing data storage technologies could keep up with the computer’s staggering throughput. Stretch handled the business of staging tapes from the library to the drives—using Tractor’s automated cassette handler. Stretch also coordinated reading data in from Tractor and writing results back out from Harvest. Harvest, in turn, did all its actual computing on the system’s shared main memory.

In the early 1960s, and even after Tractor and Harvest were installed, hard-drive data storage was in its infancy. For code-breaking jobs of the size Harvest was taking on, disk storage would have been impractical in terms of both cost and sheer floor space. So Tractor had to be based around tape storage.

Each tape was sealed inside a case built like a boombox—twin encased reels under a window, carried by a handle—and, at 6 to 7 kilograms, about as heavy as a bowling ball. Think of a Tractor cassette as an outsize predecessor of the audiocassette, which would come along a decade later, and holding some 120 megabytes of data on a reel of tape 550 meters long. Each storage unit housed up to 160 of these cassettes.

An IBM technician holds one of the data cassettes used with Harvest’s automated Tractor tape drives. IBM

When Harvest launched in 1962, it had three automatic cartridge units, each serving two drives. So the available online storage across the three Tractor units totaled a stunning 44 gigabytes. That’s more than 190 times as much capacity as the IBM 2314 disk storage system, announced in 1965, which held 233 megabytes across its full complement of eight drives.

Tractor had to run continuously, swapping cassettes in and out, 24 hours a day, seven days a week. The system’s tape-handling speed was tuned to keep pace with Harvest’s own appetite for data. The custom-built robotic mechanism for retrieving the cassettes was a servo-driven arm that traversed the system’s storage racks. It fetched a cassette from its slot and delivered it to a handler or received a cassette from one of the handlers and returned it to storage.

Running at 6 meters per second, Tractor’s tapes zipped past the read/write heads faster than the eye could track. Software running on Stretch handled the cassette shuttling as well as reading and writing. For one of Tractor’s drives to move from the completion of processing one tape to reading the next took about 18 seconds, assuming it had already been fetched and was ready to mount. Robotically fetching a cassette from the storage unit and preparing it for reading required no human handling or input whatsoever.

In addition to Tractor, the system had standard reel-to-reel tape drives attached to Stretch. Harvest’s technicians often used the conventional drives for importing and exporting data to and from other systems; there was no other practical way to get large datasets into or out of Harvest. Tractor could also store permanent files and retrieve them directly from its tape libraries when a job required them. In other words, Tractor’s substantial cassette libraries acted both as permanent data storage and as a place to hold transient data for processing by Harvest.

No system in the commercial computing world of 1962 came close to Tractor’s gigabytes of simultaneously accessible data. At most computer centers at the time, “available” data meant physical racks of tape standing somewhere near its drives—accessible only as rapidly as an operator could manually pull a reel and thread it onto a machine, one at a time, over the course of a shift.

Alpha Was Harvest’s Custom-Built Programming Language

Created jointly by IBM and NSA, the Alpha language existed solely to program Harvest’s streaming dataflow engine for code-breaking work. According to a declassified Pentagon history of NSA computers, Alpha stood for Advanced Language for Programming Harvest.

Alpha allowed the programmer to define the alphabet in which code-breaking data would be processed. The language also included two unusual characters with no equivalent in conventional computing until years later, when Multics and Unix introduced wildcard characters. A “scab” (which was represented on Harvest’s input keyboard, a repurposed early IBM Selectric typewriter, by a “?”) stood for a character that was real but unknown. And a “pad” (represented by a blank space) was a null or spacer. These characters provided flexibility of representation for code breaking jobs, in which unknown or uncertain characters were commonplace.

A Harvest operator types on one of the main system consoles, a repurposed IBM Selectric typewriter.IBM

The rules governing Alpha’s operations on strings anticipated other modern rubrics, like “not a number”—a designation describing an unknown value in a dataset that can propagate through calculations, rather than silently corrupting them. Strings in Alpha could also be aggregated into cords, and cords into ropes, giving cryptanalysts a hierarchical vocabulary for describing complex intercepts.

Allen wrote a final technical report on her section of the Harvest software when her part of the project concluded—and just as promptly lost access to it. “I spent the good part of a summer on that,” she recalled in 2001. “And it just disappeared into Fort Meade somewhere.”

Replacing an Irreplaceable Machine

By 1971, according to NSA analyst Looney, the machine was running at its highest utilization ever—115 hours of production a week, or more than two-thirds of the time. Yet the number of jobs it processed had been dropping since 1967. Ordinary data-processing work, Looney noted, was by 1972 migrating to newer, general-purpose machines, leaving Harvest to concentrate on the very large, specialized jobs no other system could handle.

At its tenth anniversary of operations, Looney concluded, Harvest was a machine “conceived in the fifties, born in the sixties, and irreplaceable in the seventies.”

He got the last part wrong.

On 27 February 1976, operators shut down Harvest for the last time. A custom mechanical component in the Tractor tape library had worn out, and the manufacturer of the part was no longer in business. By then Harvest had run continuously for nearly a decade and a half—through the roughest close call in the history of mutually assured destruction and into the age of détente—processing intercepts at a rate no civilian machine could touch. By the time it retired, Harvest had outlived several generations of commercial computing.

A placard commemorates the 1976 decommissioning of IBM’s Harvest computer at the NSA’s headquarters in Fort Meade, Md. National Cryptologic Museum

Somebody at the NSA decided to commemorate the machine with a mock telegram, written under Harvest’s name on the machine’s last day (and now preserved in the agency’s archives). “I first began operations at NSA. Although not widely known, I was probably the largest, fastest, and most technically advanced computer system in the world,” the telegram said. “And now, fourteen years later, the time to retire has come. The cost of my upkeep and operation has been overtaken by more modern equipments and the newer technologies.”

The NSA ultimately replaced Harvest with the landmark Cray-1 supercomputer. The Cray-1 was built from faster, more tightly integrated circuits that could outperform Harvest’s aging transistors at nearly any task, including text processing. Although the Cray was designed primarily for numeric and scientific computing, it sold across many fields—which ultimately made the supercomputer win out once Harvest’s custom-built text-processing hardware was no longer worth the upkeep for just one customer.

The secrecy that shrouded Harvest meant it could claim no lineage of immediate successors. But the ideas it pioneered didn’t disappear—they resurfaced, again and again, in the years that followed.

Tractor’s automated tape library was the forerunner of the robotic storage silos that would become standard in enterprise data centers about 20 years later. Harvest’s pipeline architecture prefigured the dataflow computing movement of the 1980s. The continuous pattern-detecting logic of its match units finds direct echoes in modern hardware packet-inspection intrusion detectors and programmable network switches that today route traffic through the internet at wire speed.

Harvest didn’t found a dynasty. But, in its time, it steadfastly pointed toward the future—in several directions at once.

This article appears in the September 2026 print issue as “The Lost History of IBM’s Cold-War Code Breaker.”

This article is brought to you by Ollobot.

From about 2017, individuals began to truly connect with the initial wave of companion robots. These devices had personality, moved around, joked, and answered when you spoke to them. Most early companion robots, however, were still limited by simple voice-command interactions and narrow functionality. Once the novelty wore off, many ended up sitting unused on shelves. As some of those companies went out of business and turned off their servers, many owners likened it to losing a pet.

What Ollobot describes as “gentle intelligence” is a useful way to think about where the serious work in this category is going. Not toward more powerful assistants, but toward more present ones.

The problem companion robots were trying to solve

Loneliness is not a niche issue. According to one study, nearly one out of three elderly adults resides alone, meaning they do not have daily companions. Research also shows that children whose parents have migrated for work, leaving them in the care of relatives, were 2.5 times more likely to experience loneliness than children whose parents remain with them. Among working adults living alone in urban environments, similar patterns of social isolation emerge, even if they are less visible.

Over the years, technology has time and again attempted to solve this problem via video calls, smart speakers, and messaging apps without much success. Those tools are geared towards communication between people that already have relationships. They do not create presence. They schedule it. That is the gap that a new generation of AI companion robots is being engineered to fill.

Today’s AI robots are different

Today’s companion robots are not just cute and cuddly. They are designed with psychological research, clinical insight and long-term interaction models to be truly useful in real homes.

Three fundamental shifts define the current generation:

1. From reactive to proactive response. Older robots relied on you speaking to them, but modern robots monitor a room with cameras, microphones, and surroundings sensors to initiate interactions without your input, and they can pick up on your emotions.
2. From function-oriented to emotion-oriented design. The original pitch for companion robots was about what they could do. The question driving the serious work now is how they make you feel, which is a harder engineering problem and a more honest framing of what the product is actually for.
3. From standalone hardware to connected ecosystems. Leading brands are creating platforms rather than devices with software included as a built-in layer and remote access from the beginning.

The global AI companion market size was valued at US $36.8 billion in 2025 and is projected to grow from $48 billion in 2026 to $318 billion by 2033, at a compound annual growth rate of 31 percent from 2026 to 2033.

Three household scenarios and interaction models

Ollobot’s advanced AI family companion robot OlloNi SS1 addresses a number of gaps in what existing technology offers.

Elderly individuals living alone. The combination of proactive interaction, fall detection, and persistent presence addresses both safety and companionship without the social overhead of asking family members to check in more frequently.

Children in households where parents work far from home. The SS1 functions as a consistent companion that already knows a child, their preferences, their moods, and their routines. The remote connection features allow parents to stay present without requiring a scheduled call, and the life recording system gives them a passive window into their child’s days that feels less clinical than a monitoring camera.

Single professionals living alone in cities. The SS1 adapts to daily routines, builds up a preference model over time, and provides ambient social presence without demands.

OlloNi SS1 adapts to daily routines over time.Ollobot

What OlloNi SS1 is doing differently?

Ollobot’s goal in building intelligent companion robots is to address the gaps in technology and capability, using innovation not to automate tasks but to fill emotional voids.

Much of the robotics industry has historically pursued human imitation — machines that speak, look, or behave like people. The SS1 is instead designed around familiarity and long-term coexistence rather than realism.

The system integrates multiple subsystems operating in parallel, including visual perception, audio processing, mobility control, and interaction management. It is equipped with a multi-chip AI 4K vision module capable of facial recognition and motion tracking. One small but revealing detail is the inclusion of a physical privacy cover for the camera — a mechanical solution to concerns that software settings alone may not fully resolve.

OlloNi SS1 can actively integrate into family activities, and it can autonomously move closer to capture memorable moments or reposition itself to remain engaged in ongoing interactions.Ollobot

The robot supports advanced mobility across multiple indoor surfaces, including wooden floors, ceramic tiles, and low-pile carpets, with slope climbing capability up to 3.5 degrees. Rather than remaining in a fixed location, it can move naturally throughout the home to stay close to household members as daily activities unfold.

For example, the OlloNi SS1 may greet family members when they arrive home, follow an older adult from the living room to the kitchen while continuing a conversation, remind a child to take a study break after a prolonged period of inactivity, or notice that someone appears unusually quiet and gently check in. During family activities, it can autonomously move closer to capture memorable moments or reposition itself to remain engaged in ongoing interactions.

The robot continues to evolve over time, with over-the-air updates that deliver new features, performance improvements, and AI enhancements

It also incorporates fall detection with optimized accuracy for safety monitoring scenarios. A 6-microphone array enables omnidirectional voice pickup with an effective voice capture range of up to 5 meters, supporting reliable wake-word detection and far-field interaction.

To support continuous companionship, much of the robot’s AI processing takes place directly on the device through its “heart module” architecture, with 16 GB of memory and 64 GB of local storage. This enables the system to retain household memories, recognize familiar faces, and respond with lower latency, making interactions feel more natural even during everyday routines.

Because companion robots are expected to remain available throughout the day rather than only during brief interactions, the SS1 is designed for extended operation, offering up to 12 hours of standby time and around 5 hours of active interaction on a single charge. This allows it to accompany users through meals, conversations, playtime, and other daily activities without frequent interruptions.

To support engaging interactions, much of the robot’s AI processing takes place directly on the device through its “heart module” architecture.Ollobot

Like the relationships it is designed to build, the robot continues to evolve over time. Running on Android OS with over-the-air (OTA) updates, the system continuously receives new features, performance improvements, and AI enhancements, allowing its capabilities to grow alongside the household it serves.

The robot’s behavioral model also improves over time. Rather than reacting to isolated commands, it attempts to establish a baseline understanding of household routines and individuals. Changes in behavior — prolonged quietness, unusual inactivity, or emotional cues — become triggers for interaction.

Presence instead of utility

Several features in the OlloNi SS1 illustrate this emphasis on presence and continuity in its interactions.

The system can identify different household members, including pets, and adapt responses accordingly. Remote communication features allow family members to connect through the device without treating every interaction like a scheduled call. Environmental sensors support contextual reminders tied to weather or room conditions.

Its “2+1” multi-display configuration is also designed around emotional communication. Two circular side displays function as expressive “emotional eyes,” while a separate primary display handles information and structured interaction. The separation allows emotional signaling and functional communication to operate independently, creating more intuitive nonverbal interaction even when no dialogue is taking place.

The robot’s behavioral model improves over time. Rather than reacting to isolated commands, it attempts to establish a baseline understanding of household routines and individuals.Ollobot

The SS1 also includes an automated life-recording system built on facial recognition and behavioral-event detection that can capture moments such as laughter, physical closeness, or group interaction automatically. An integrated AI vlog engine can then organize those moments into edited short-form videos with automated sequencing and soundtrack generation. The design intent is to preserve spontaneous domestic moments without requiring active documentation behavior from users.

An integrated AI vlog engine can organize recorded moments into edited short-form videos with automated sequencing and soundtrack generation

Visual data is processed primarily on the device through the SS1’s on-device AI architecture, with household memories stored locally and managed within Ollobot’s proprietary ecosystem instead of being shared with third-party smart home platforms. Access to recordings and live feeds is restricted to authorized users through the companion app, while encrypted communication helps protect data during remote access. Users also retain direct control over recording preferences, and the physical camera privacy cover provides an additional hardware-level safeguard whenever visual monitoring is not desired.

Remote communication is similarly structured around persistence rather than transaction. Traditional video calls are episodic and screen-bound; the SS1 instead acts as a continuously present interface embedded inside the household environment. Through autonomous mobility, environmental awareness, and persistent household memory, remote family members interact with an ongoing domestic context.

The larger shift to “gentle intelligence”

Ultimately, gentle intelligence is not about making robots behave more like humans — it is about helping them fit more naturally into human lives. Each OlloNi SS1 unit develops a unique behavioral profile based on its household. Two units running in different homes for a year will have become meaningfully different from each other, shaped by the specific people, habits, and rhythms of where they live.

That kind of long-term personalization is what early companion robots never had. It is also what makes the difference between a product that ends up on a shelf and one that actually earns its place in a home.

Learn more at ollobot.com.

For most of my career, my IEEE membership sat quietly in the background—a line on my résumé, a discount code for a conference registration, and access to the IEEE Xplore digital library, which I underutilized. I didn’t think much about the grade of membership available above that of the regular member. I assumed senior membership was reserved for people further along in their career than I was. They published more papers, had more gray hair, and had worked longer in the field.

I was wrong on all three counts. The misunderstanding cost me an important validation of my skills and professional competency.

I suspect a lot of other qualified members are where I was one year ago: eligible but unaware of the benefits of senior membership, and one application away from a meaningful career credential.

The myths that almost stopped me

Here are a few of the misconceptions about senior membership:

It’s mostly for academics and longtime IEEE volunteers. It isn’t. The grade is explicitly built around a person’s professional engineering experience. Plenty of successful applicants have never published a paper. Industry experience counts for a lot.

You need a graduate degree. You don’t. A bachelor’s degree plus enough years of qualifying experience is sufficient on its own. An advanced degree simply offsets some of the required years of experience.

If I’m not well-known in my field, I won’t qualify. Senior membership isn’t a popularity contest. Rather, it hinges on whether you meet specific experience metrics. The requirement is “sustained, significant technical contribution,” not “known beyond your organization.”

I should wait until I have more significant achievements to point to. I believed this for longer than I should have. If you meet the 10-year experience threshold with five years of significant performance, you’re already eligible. Waiting doesn’t strengthen a qualifying application; it just delays getting a credential you’ve already earned.

Why I applied for senior membership

The push to apply came from a practical need. As a senior data scientist at Apple in Austin, Texas, I work in applied machine learning, building large-scale systems that affect customer-support operations. I already had started taking on more peer-review work—checking papers for journals including Neural Networks and IEEE Transactions on Knowledge and Data Engineering, mentoring at Apple, and writing on public platforms such as Medium and SimpleTalk.

I wanted a credential that reflected that shift from “engineer who codes” to “engineer who helps shape the field.”

The IEEE senior member grade turned out to be the validation of my work I was looking for. It’s not an award for a single achievement. You have to apply for it, and it’s a peer-evaluated process that confirms you’ve sustained a meaningful level of professional contributions over time.

That distinction matters. Having a research paper published or being granted a patent proves a moment in time. Senior membership reflects a pattern of continuous contributions.

The benefits to my career happened faster than I expected. It strengthened how search committees, IEEE conference organizers, and IEEE awards panels viewed me. Only senior members can hold certain IEEE leadership positions.

The senior grade also opened doors to editorial and reviewer roles I hadn’t even pursued before. Journal editors and conference organizers often look for reviewers with a track record they can verify quickly, and senior membership gives them that signal without extra vetting on their end. It also gave me a credential I could point to in professional contexts, including, in my case, supporting documentation for a U.S. employment-based immigration petition, where third-party peer recognition carries real evidentiary weight.

Navigating the process

The process for applying for senior membership is easier than the title might suggest. To qualify, you need a combination of professional and academic experience in an IEEE-designated field: engineering, computer science, information technology, physical sciences, mathematics, or technical communications. The two must total at least 10 years, with at least five of them showing significant performance. Crucially, experience isn’t limited to job titles. Graduate research, technical leadership, and progressively responsible engineering work all count toward the total number of years. I’d been quietly accumulating qualifying years without ever framing them that way.

“I suspect a lot of other qualified members are exactly where I was a year ago: eligible but unaware of the benefits of senior membership, and one application away from a meaningful career credential.”

You submit your application through IEEE’s member portal, mapped against the experience requirement, along with three references from current IEEE members—at least two of whom must be senior members or IEEE Fellows who can vouch for the credibility of your work.

The IEEE member grade evaluation committee reviews applications and renders decisions.

How to find references

The part everyone underestimates is references. Applications can stall at this point. References must be IEEE members in good standing, and at least two need to be IEEE senior members—which means you can’t necessarily ask people who know you best. You need to find references who are both willing to vouch for you and are grade-eligible.

My advice is to identify and confirm all three references before you submit your application. It might be difficult to add or swap a reference during the process, and a stalled reference could delay your file.

Where to find references is the part I worried most about. But it turned out to be far easier than I expected.

Here are several sources:

• IEEE Collabratec. This is IEEE’s professional networking platform and, in my opinion, is an underused resource. You can search by technical interest, geography, or society membership and message members directly. I found several of my eventual references this way—colleagues I’d never have thought to ask simply because we hadn’t worked together directly, but ones who knew my technical work through shared communities or conference circles.
• Coworkers and colleagues, current and former. If you’ve worked alongside IEEE members—especially ones senior to you—they’re often the most natural fit because they can speak specifically to your day-to-day technical contributions.
• Former professors. If you did graduate work, your advisor or committee members are usually IEEE members and are well positioned to speak to your research contributions, even years later.
• LinkedIn. A surprising number of my qualifying references came from reconnecting with people on LinkedIn I’d lost touch with professionally. A short, specific, polite message explaining what you’re applying for and why you thought of the person can go a long way.

A pattern I noticed when looking for references is that people are generally glad to be asked. Serving as a reference is a small lift for them and a meaningful one for you. Most senior engineers remember someone doing the same for them and are happy to pay it forward.

If you’re on the fence

If you’ve been in the field for a decade or more, doing real technical work, and IEEE membership has been sitting quietly in the background of your career the way it did in mine, it’s worth 10 minutes to check the eligibility criteria against your history. You might find, as I did, that you qualified for the membership upgrade a while ago.

The AI boom has disrupted the way engineers work, introducing new tools to learn, raising expectations for what teams can achieve in a workday, and making it harder to get hired in the first place. This makes it difficult to advise students on which specific coding languages or technical skills they should learn. So amidst the uncertainty, advice for young professionals often turns to a common refrain: Be adaptable. But what does adaptability look like in practice?

Engineers often operate on the cutting edge of technology, so dealing with change is a normal part of the job, says Samantha Brunhaver, an associate professor of engineering at Arizona State University, in Tempe. Yet university curricula and training in the workplace often don’t prepare students for this.

“We tell engineers that they need to be adaptable when they graduate, but we don’t actually explain what that means, demonstrate what that looks like, [or] help make sure that they’re developing it,” says Brunhaver, who received a National Science Foundation award in 2020 to study how to foster greater workplace adaptability among young engineers. For this ongoing project, she has interviewed engineering managers, early career employees, and undergraduates about their experiences.

Part of the problem, she says, is that every employer has its own idea of what to be adaptable means. Generally, Brunhaver defines adaptability as “the ability to recognize that a change or uncertainty is occurring, and then respond effectively to that change.” But the skill is context-dependent. In software engineering, that might mean responding to turnover in the tools you use on a daily basis, while aerospace or biomedical engineers may need to keep track of changing procedures and regulations. “Managers are all saying adaptability is important,” Brunhaver says, “but defining it in different ways.”

At the same time, engineers are all contending with changes beyond these industry-specific expectations. Jobs in the technology, media, and telecom sectors are experiencing the fastest pace of skill turnover, according to a June 2026 report on the effects of AI from the professional services network PwC. And the World Economic Forum’s most recent Future of Jobs Report, published in 2025, found that employers across all sectors expect 39 percent of workers’ core skills to change by 2030. This uncertainty can be uncomfortable. But with the right mind-set and support from leadership, adaptability can help keep you afloat.

How to Cultivate Adaptability

The AI transition is a big shift—but not an unprecedented one, says Jenna Butler, a research scientist at Microsoft who studies developer well-being and productivity.

During this type of paradigm shift, there is often a “chaos period” when a new normal is being established, Butler says. In AI’s case, it challenges the understanding of what a computer can do. “I think we’re still in this in-between, difficult period that we’ve seen before, but [it] is maybe moving faster than it has historically.” Software engineers—in one of the fields most affected by AI—are now facing a significant increase in code review. “If you ask 20 developers, you get 23 different ways of working with it. Everyone is trying to sort it out,” says Butler, who describes this period as “the uncomfortable middle.”

“We tell engineers that they need to be adaptable when they graduate, but we don’t actually explain what that means, demonstrate what that looks like, [or] help make sure that they’re developing it.”– Samantha Brunhaver, Arizona State University

Brunhaver says one way educators can help prepare students before they enter the workforce is by offering a diversity of real-world experiences, such as internships, team-based projects, community service, and leadership roles. Each of these teach students to adapt to different challenges, easing their transition from school to work.

It’s also important to encourage reflection, Brunhaver adds, noting that metacognition helps individuals use the skill more effectively. “In order to adapt, you have to think that you have agency and the ability to get through a situation.” Ultimately, it comes down to three steps: Perceive a need to adapt, evaluate your options, and act.

For those already in the workforce, that action may mean taking the time to learn new tools and ways of working. Software engineering, for instance, may soon rely more on prompting models and managing agents than coding line by line. “I think people who went into software because they like solving problems are going to have a lot of fun, and people who just enjoy the art of writing code are not,” Butler says.

The More Things Change…

Although the tools engineers use on a daily basis are evolving, the core responsibilities of the job are more stable than they may seem, says Andy Hunt, a software developer who coauthored The Pragmatic Programmer (Addison-Wesley Professional) in 1999. The book outlines practical coding principles, and has been taught in many computer science classrooms. When Hunt was working on the 20th anniversary edition of the book, he was surprised by how much of the advice still applies. And now, seven years later, he maintains that belief.

“The fundamental part of the job is problem solving and communication, and that’s always going to be there,” he says.

Hunt emphasizes the importance of developing systems thinking over particular tools. To him, identifying as a Java programmer, for instance, is “like a carpenter saying, ‘I’m a hammer user,’ or ‘I specialize in cordless drills.’ ”

He acknowledges that today’s hiring process, in which companies often filter résumés for certain languages or years of experience, makes it harder to embrace a more expansive way of relating to your job. Employers, he says, should recognize that “the tech’s not the hard part, and it never has been. Understanding information theory, understanding systems thinking, understanding what constraints you’re up to—that’s still the hard part.”

With this type of misalignment between employers and employees, AI is also intensifying an old source of tension: How can engineers slow down enough to adapt and learn new tools when the pressure to become more productive keeps mounting?

Who’s Responsible for Enabling Change?

Young engineers need to embrace change. However, educators and employers also play a role in building a successful workforce. From the educator’s perspective, Brunhaver says “we need to be more explicit about what [adaptability] means and why it’s important.” Managers, meanwhile, should invest in their employees’ professional development.

Microsoft research scientist Butler often encourages leadership to set aside intentional time for continuous learning for their engineers—even just an hour a week—without any expectation that they will produce code or progress in their daily work. “I realize that’s difficult,” says Butler. “I would encourage people to do it on their own, but I would really encourage organizations and leaders to do it, because you’re not going to get this sudden change in your people if they don’t have time and space to learn how to work differently.”

This also means providing enough instruction, Butler adds. When developers aren’t given enough guidance on adopting something new, while being pressured to increase productivity, they risk doubling down on the tools they already know and burning out.

“I do imagine the next number of years could be challenging,” Butler says. Engineers will have to adapt to find their place in an evolving workforce—but they also have a say in shaping that future.

“Being adaptable sort of implies that you’re going to change based on what’s happening around you, and I would really like people to realize the change that’s happening is somewhat up to us,” she says. All individuals have a choice in how they use AI, for instance, and which models they use. “We need to be adaptable and go with the flow to a degree, but we also need to be directing that flow. The future with AI is absolutely not predetermined.”

This article appears in the September 2026 print issue as “The Adaptable Engineer.”

This article is brought to you by Emerson.

I’ve spent much of my career as an engineer, including years in the semiconductor industry. And one lesson has stayed with me through every major technology shift: innovation always creates new complexity.

In semiconductors, we have seen that repeatedly. Every generation has delivered breakthroughs in performance and capability, but each step forward made it harder to understand system behavior. What used to be easy to validate on the component level with a test bench now needs a much wider view.

“The future of engineering will be defined by who can verify, understand, and improve complex systems fast enough to safely keep innovation moving forward,” says Ritu Favre, President of Emerson’s Test & Measurement business group.Emerson

Chiplet-based designs are a prime example. A chiplet from one supplier, an interposer from another, and a packaging process from a third may all perform perfectly on their own. Yet there is a chance for unexpected behavior when you put them together in a system. More and more often, the hardest engineering challenges are not in the individual components themselves. The problems are found when we start to combine components and have them interact with each other.

These challenges extend far beyond semiconductors. Products are becoming more software-defined and dependent on interactions across different technologies and environments. Think about the interactions needed for a modern car using adaptive cruise control on a bumpy road in a rainstorm. Or a passenger jet adjusting wing flaps and engine speeds in turbulent weather to maintain safety and stability. Both the car and the jet are being guided by complex computer systems with thousands of sensors leading to thousands of interactions every second. And in many cases, there are multiple computer systems working together. We are building systems of remarkable capability but understanding how they will act under real-world conditions is getting harder.

That is why I believe we are entering a new era of test. The defining challenge of modern engineering is no longer simply what we can design and build. It is what we can confidently verify.

Rethinking the Role of Test

In this new era, test can’t be an afterthought. For decades, test was treated as the final checkpoint before release. Design teams developed a product, test teams validated performance, and organizations looked for a final pass/fail to determine whether they were ready to move forward. That model worked fine when systems were more self-contained and predictable. Today, that approach can lead to more risk.

I believe we are entering a new era of test. The defining challenge of modern engineering is no longer simply what we can design and build. It is what we can confidently verify.

Many of the delays and fire drills we face come from issues that were not visible early enough. Problems discovered late in development are more difficult to diagnose, more expensive to fix, and more likely to get you off schedule. The solution is not more testing at the end. The solution is to make test and verification part of the engineering workflow from the start.

When validation is integrated throughout development, teams catch problems early when change is easier. Test also stops being a barrier to release. Instead, it becomes a source of insight, helping us understand how systems behave as they become more connected.

Why Connected Platforms Matter

When confidently verifying technology becomes the key challenge, the tools we choose take on a different level of importance. The tools have a direct impact on how quickly we can diagnose a problem and keep moving forward. In an environment where technology changes rapidly, disconnected tools get in the way of progress. Modern test strategy requires linking information across design, validation, and production, turning measurement data into decisions made quickly enough to keep pace with innovation.

This reminds me of when EDA was first introduced. Before it came along, engineers spent much of their time hand-drawing circuit layouts and placing transistors. EDA eliminated that tedious work by letting teams describe complex behavior in high-level code. It enabled them to focus on overall architecture instead.

A connected test platform does a similar thing for validation. Because a platform can adapt and scale alongside technology, it cuts down on maintenance and downtime, keeping teams from having to rebuild their workflows from scratch as requirements change.

Grounding AI in Engineering Reality

Today, AI is rapidly entering the engineering toolkit to accelerate design and analysis. But in test and measurement, AI cannot reach its potential in isolation.

An AI model is only as effective as the data feeding it. Without context, even the smartest algorithm will struggle to tell the difference between normal hardware variance and a critical failure. A connected platform supplies the structured, traceable data stream AI requires to deliver real insight.

AI can correlate complex multi-system interactions, flag unexpected behavior, and direct an engineer’s attention right at the root cause.

When measurement data flows seamlessly across the workstream, AI moves from being a standalone tool to an active layer of intelligence. It can correlate complex multi-system interactions, flag unexpected behavior, and direct an engineer’s attention right at the root cause.

Every technology shift that accelerates how fast we create new designs also increases the complexity we must verify. AI can help teams keep pace with that complexity. Not by replacing human judgment, but by giving engineers the context we need to act with confidence.

Innovation Demands Confidence

Ultimately, the goal of modern platforms and AI-enabled workflows is to help technical teams spend more time building new things and solving hard problems. Most of us didn’t choose this profession to spend our time searching for data or dealing with last minute surprises. We want to innovate and integrating test directly into development provides a better view of system behavior, allowing teams to focus on that innovation rather than managing complexity.

The future of engineering will not be defined by who can build the most advanced product or technology. It will be defined by who can verify, understand, and improve complex systems fast enough to safely keep innovation moving forward.

That is the new era of test. As the pace of innovation accelerates, every breakthrough creates new paths to failure, and test is how engineers find those failures before the real world does. In an increasingly complex world, that capability is becoming as important as innovation itself.

Innovation has always required great engineering. And now, more than ever, it also requires confidence. Confidence that comes from knowing that we are not only building what is possible, but we are also building technology that people can trust.

How do I touch you
across the ocean,
across cold depths
where light travels through glass.

Not copper—fibers. Optical.

Through liquid glass, through flickering light
that carries you in fragments. Light broken into pulses.

You say: it’s easier this way. What are we missing like this? You smile.
Safe distance.

I say: network.

Signals slide beneath the sea, through cables thinner than trust, faster than touch,
slower than longing.

We stand alone, together. Synchronous, yet apart. Icons replace skin, latency replaces breath.

This distance protects us. Silence that feels intentional.

Everything is under control as long as nothing truly hurts.

And we choose it
because it shields us
from what we might become if we actually met.

You are my counterpoint. My response.
My reflection
at a safe distance.

Beneath the ocean, nodes remember paths.

Packets shake hands without bodies.

If we get lost,
we resend everything, with error,
with noise,
with hope.

Balaji Ingole rarely saw televisions while growing up in Udgir, India. No one in the small Maharashtra village had computers or phones. Only one household owned a television, and neighbors often gathered there to watch shows together.

Ingole never even saw a computer growing up. It wasn’t until he reached middle school that he encountered a computer lab, an experience he says changed his life. Almost immediately, he says, the machine felt like a window into a different scale of possibility for him.

Balaji Ingole

Employer

Amla Commerce in Milwaukee

Title

Project manager

Member grade

Senior member

Alma maters

COEP Technological University and Welingkar Institute of Management, both in India

“I was very studious and not very social, always reading or solving problems in a math textbook,” he says. “At the computer lab, I began learning the C programming language—which was like discovering a whole new world. I was fascinated that you could create something with just a few lines of code.”

His early interest grew into a self-directed education. Outside of Ingole’s formal classwork, he taught himself to build database-backed applications, wire up hardware, write software, and trace error logs.

Today the IEEE senior member similarly splits his time. During the week, he’s a project manager in Milwaukee at B2B e-commerce company Amla, leading AI-driven digital transformation initiatives to help the company’s clients boost their sales. On weekends, he leads a similarly demanding life as an independent researcher. His current projects include developing AI-enabled health care diagnostic tools and assistive technologies to support people with physical disabilities.

“I believe in ‘learn by doing,’” he says. “I really like to test my knowledge and prototype ideas to find out if they truly work.”

A college project becomes an inspiration

Ingole’s tendency to go beyond his coursework continued after he graduated high school in 2004. As a mechanical engineering undergraduate at The College of Engineering, Pune (now COEP Technological University), in India, he participated in several extracurricular activities. One was interviewing entrepreneurs and writing about them for The COEP College Magazine. The experience helped him gain confidence, he says, giving him the push he needed to pursue interviews for the publication with two Indian entrepreneurs he admired: N.R. Narayana Murthy, cofounder of IT giant Infosys; and his wife, philanthropist Sudha Murty. The Murtys cofounded the Infosys Foundation, a nonprofit that runs educational, health care, women’s empowerment, and sustainability programs in underserved areas of India.

“Every week I would fax them: ‘Please give me an interview time,’” Ingole says. Eventually, Sudha Murty’s office offered him a phone interview, but he requested to meet her in person at Infosys’s Bengaluru offices. She agreed, but the offices were 940 kilometers from Pune, and he didn’t have the money to travel or stay overnight in a hotel.

Ingole and a classmate borrowed money from friends and traveled through the night on multiple buses and trains to get to Bengaluru. They freshened up in a public bathroom before heading to the Infosys campus to meet Murty.

Impressed by their persistence, she surprised them by also arranging a brief chat with Narayana Murthy.

“Narayana Murthy handwrote a personal message to the engineering students of [my college]—which we proudly published in our college magazine,” Ingole says. “In his note, Murthy shared that we are at an extraordinary moment in India’s history and that the future looks even brighter. His words encouraged us to work hard and make the most of this time.

“I still have that note,” Ingole says. “They are billionaires, and I was just a regular student. The fact that they took the time to do this really motivated me.”

During the final semester of his engineering studies, Ingole joined the Tata Research Design and Development Center in Pune for a six-month internship. After earning his bachelor’s degree in mechanical engineering in 2008, he became a graduate engineering trainee at Honeywell Automation in Pune.

He left the company in 2009, and during the next 13 years, he held different software engineering and project management positions at IT companies across India.

He earned a master’s degree in business administration from the Welingkar Institute of Management, Mumbai, in 2017.

In 2022 he accepted a project-manager role at Mars IT Solutions in Madison, Wisc. The following year, he left to join Gainwell Technologies, also in Madison, as a senior project manager. At Gainwell, he managed projects for the core IT systems multiple U.S. state health departments use to administer Medicaid benefits, manage provider enrollment, and verify member eligibility. The experience managing projects that directly enabled patients’ access to health care gave Ingole a special appreciation for and interest in this area, he says.

“Health care data is not like other data,” he says. “The stakes are high, compliance requirements are different, and the margin for error is effectively zero.”

Ingole says he enjoyed the rigor of data governance combined with the potential to positively impact lives, and that also applies to his current work at Amla.

Agentic AI in e-commerce

Ingole joined Amla in July 2025. He helps manufacturers and B2B customers modernize their e-commerce operations. He also builds AI tools for them and for his internal team.

For Amla’s customers, he’s developing AI-enabled chatbots that help manufacturers set up and manage large product catalogs in e‑commerce platforms. Such product setup traditionally has been a manual, tedious, error-prone process: Companies upload thousands of products, adjust item names, enter prices, update images, and more.

“Product setup has been one of the most painful processes in e-commerce, and it can take [our] customers two to three months to complete,” Ingole says. “We’re creating an AI agent that will guide them, step-by-step, to get everything set up in two weeks.”

Ingole relies on AI agents for some of his own tasks at Amla. Project managers historically have spent 10 to 12 hours each week assembling and sending status reports to stakeholders. Ingole built an AI agent to handle much of the work.

“It runs every Monday morning and reads through my emails to extract highlights, risks, timelines, and upcoming releases, then sends me a written status report,” Ingole says. The process might sound simple, but the agent’s workflow involves at least a dozen steps including defining parameters, managing temporary files, and integrating with existing tools.

With the information-gathering work handled, it frees up Ingole and his colleagues to spend more time on deeper-thinking work, he says.

Publishing as idea refinery

For nearly a decade, Ingole has spent some of his free time conducting independent research projects in data analytics and AI-enabled applications in health care. He has written more than 40 peer-reviewed papers, which are in the IEEE Xplore Digital Library. He has been granted six patents in the United Kingdom and India. In the U.K., he is a registered coinventor of an AI-powered, cloud-connected wearable device for health monitoring and an AI-based breast cancer detection tool.

Ingole’s patent for the breast cancer detector, he says, reflects his belief that when engineers apply data and AI correctly, they can help doctors diagnose patients more quickly and accurately.

That, he says, is both a power and a responsibility.

He is part of a team helping patients who are paralyzed and nonverbal control items in their environment. His goal, he says, is to develop a brain-computer interface to let patients turn on a fan, switch off a television, and complete similar tasks.

Publishing research requires both academic rigor and peer scrutiny, and Ingole says the function has been critical to improving as both a project manager and a researcher-inventor.

“Lots of research ideas never make it to paper,” he notes. “But when you write for journals or conferences, you’re bombarded with questions from Ph.D.s and experienced researchers. This forces me to refine my methodology, and to combine use cases and technical architecture in a way that stands up to expert review.”

Finding a professional hub

Ingole joined IEEE in 2022, and he says the affiliation has become central to both his research and his professional identity.

“I use the Member Directory often and contact engineers through my IEEE email address, which gives me credibility because they know it’s a genuine research connection,” he says.

The organization has given him a platform to contribute to the research space beyond his own papers, he says. He has served as a conference session chair, keynote speaker, technical program committee member, and peer research reviewer for various conferences and events. His IEEE membership, he says, has opened doors to other communities, helping support his entry into the British Computer Society, which has stringent acceptance criteria.

Those opportunities have helped him build a global network of collaborators with whom to discuss upcoming research, seek advice, and share data, he says.

“IEEE is important for me to continue as an independent researcher,” he says. “It lets me contribute to the community, and I get a lot in return.”

About this Webinar

Turn Yield Excursions into Faster, More Confident Root Cause Analysis

When a yield issue emerges, the answer rarely lives in a single system. Critical clues are spread across metrology data, tool traces, chemical analysis, and facilities systems, while growing data volumes make traditional dashboards slow, fragmented, and difficult to act on.

What You’ll Learn:

Discover how a purpose-built semiconductor analytics platform can help engineers connect insights across domains without moving data. See how Agentic AI, semiconductor-specific visualizations, and push-down compute enable faster investigation of yield excursions and process issues, even across billions of data points. In the session, a live demonstration shows how to conduct a multi-domain root cause investigation using Spotfire® Industry Pro.

Key Takeaways:

• Understand why siloed manufacturing data delays yield recovery and inflates costs
• Learn how Agentic AI automates complex cross-domain analytics and visualization generation
• Explore methods for scaling high-performance analytics across massive fab datasets

Who Should Attend:

Yield, Process, and Integration Engineers; Fab and Manufacturing Operations Managers; Quality and Reliability Engineers; and Data & Analytics leaders supporting wafer fabs, foundries, OSATs, and IDMs who need to identify issues faster while maintaining confidence in decision-making.

Save Your Spot!

Join this webinar to learn how leading semiconductor teams are accelerating root cause investigations, scaling analytics across massive datasets, and transforming disconnected data into actionable manufacturing intelligence. Reserve your seat today.

Register now for this free webinar!

This article is part of our exclusive career advice series in partnership with the IEEE Technology and Engineering Management Society.

Imagine this: You have a strong idea for a new product for your company. Your coworkers encourage you to move forward because they believe it could be the organization’s next big success. The idea clearly falls outside your department’s responsibilities, however, and you have no role in the product line.

What should you do? Sit and wait for “the right group” to pick it up, or push the idea forward without knowing how or what it might mean for your current position?

Such situations occur frequently. Many end up as missed opportunities, even though they could have significantly advanced the company’s technological or market position.

Some organizations actively support such initiatives, allocating specific periods during the workday for employees to focus on developing their own ideas.

Companies known for that include Google and 3M. They allow employees to pursue projects with a portion of their time, such as one day per week. Research that I conducted indicates it pays off for employee performance.

Bootlegging and skunkworks

At some companies, managers know such projects exist, but they deliberately turn a blind eye, allowing them to continue.

Some employees persist through bootlegging or skunkworks projects.

Bootlegging projects have not been approved by a manager or funded by the company.

Skunkworks projects involve a small team within the company that has been given authority and funding to secretly research and develop potentially groundbreaking innovations during their off-hours. The term comes from Lockheed’s Skunk Works division, set up in 1943 in a rented circus tent to build the P-80 fighter jet in secret. It took just 143 days.

The 3M Post-it Note came out of the company’s “15 percent culture,” described as a permitted bootlegging policy. It gives employees paid time off to pursue their own ideas.

The company traces the philosophy to its longtime president and later chairman William L. McKnight. Company scientist Arthur Fry used the policy in 1974 to turn a colleague’s dormant adhesive into the first Post-it prototypes, after his own bookmarks kept falling out of his hymnal.

There are several examples of high-visibility skunkworks projects. At Apple, Steve Jobs pulled roughly 20 people—pirates, as he called them—out of the company to build the original Macintosh computer in a building nicknamed Texaco Towers. In Walter Isaacson’s biography Steve Jobs, he frames the idea as modeled on the skunkworks approach.

Google’s Gmail system is frequently—and incorrectly—cited as a product of the company’s “20% time” policy. In a 2014 interview with Time magazine, the system’s creator, Paul Buchheit, said Gmail was in fact an official assignment. What the Gmail incubation did share with classic skunkworks projects was secrecy: For much of its three years in development, it was kept hidden from most people inside the company.

If you want to drive change in your organization, build a promoter triad around your idea.

At Alphabet, Google X—now known simply as X—operated as a secretive “moonshot” lab, kept hidden from most Google employees, according to a 2011 article in The New York Times. Google’s self-driving car project graduated from X to become Waymo, and Google Glass was likewise incubated there. The X team is now developing the second edition of Glass Enterprise, a successor aimed at industrial rather than consumer use.

Amazon runs a comparable model through Lab126, which, according to an article in Fast Company, evolved from a small skunkworks Amazon subsidiary into a hardware maker with nearly 3,000 employees. Lab126 delivered the Kindle in 2007 and the Echo in 2015.

Then there are so-called submarine projects, which employees work on without permission and despite explicit disapproval. They can lead to disciplinary action and termination.

Innovation management

Innovation management theory offers a more structured and robust approach. It argues that successful organizational change requires support at several levels, according to “Teamwork for Innovation: The ‘Troika’ of Promoters,” published in R&D Management. The promoter theory, developed around 25 years ago, consistently shows that change projects are far more likely to succeed when they are supported on multiple organizational levels. A good idea alone is not enough; you need a network of technology, process, and power promoters to turn a concept into a fully implemented, scalable solution.

First, you need a technology promoter: the person who has the idea, such as a new product, and possesses technical expertise and specific knowledge about the field or industry. Art Fry at 3M would be such an individual.

How can you put that into practice as an individual? Start by clearly formulating your idea into a concise concept paper or one-page summary including benefits, technical feasibility, and potential business impact.

Identify potential technology promoters (experts who can validate and refine your idea), and approach them early to strengthen the technical foundation.

In parallel, map the relevant stakeholders and decision-makers, and identify process promoters who understand how decisions are made in your company. They could be colleagues in innovation, R&D, or business development who understand your idea and how it can benefit the company.

The second is a process promoter: someone who might not know all the technical details but understands the organization’s formal and informal networks and knows how to navigate its processes, committees, and decision-making paths. This person can ensure the idea reaches the right stakeholders at the right time.

In the 3M case, it would be a person from the organizational management department, often called an innovation manager. The key role here is to connect inventors such as Fry with people from other departments needed for further project development, such as manufacturing, quality control, and sales.

Lastly, there’s the power promoter: a person in a leadership position who might not know the technical details but can allocate resources, eliminate obstacles, and maneuver through the company’s political dynamics. This individual has hierarchical power and acts as a sponsor of the idea or project. In the case of Fry, the person could be, say, the chief technology officer, but it also could be a middle manager who has the power for an individual field of action.

The three-level promoter structure applies regardless of whether the change concerns a new product, new service, or internal process innovation.

Engage potential power promoters by presenting a low-risk, small-scale pilot and a clear value proposition. Leaders are more likely to support ideas that are well prepared, vetted for potential risks, and backed by a small coalition.

Building the promoter triad

In short, don’t work in isolation. Systematically build alliances across expertise, networks, and hierarchical levels to create lasting change. If you want to drive change in your organization, build a promoter triad around your idea.

The tech experts and leadership promoters are easier to identify. Process promoters are often found in corporate innovation management, R&D management, or strategy functions, but they also can emerge in line units with strong internal networks.

Innovation management, as the promoter model describes it, looks nothing like the management structure most engineers are trained to expect. Traditional technical management runs on a single reporting line. With the promoter model, influence is spread across three people—technology, process, and power promoters—who may be in different departments, at different levels of seniority, and who might never share a reporting line.

What holds the trio together isn’t a formal structure; it’s the idea itself, for as long as it takes to move the idea forward.

That makes innovation management closer to networked, matrix-style leadership than to the pyramid most engineers picture when they hear the word management. It’s worth understanding both models before you decide which kind of impact you’re actually optimizing for.

The Institute has covered the tension from the individual’s side in “Tips for How to Think Like an Entrepreneur,” “Management Versus Technical Track,” both published in partnership with the IEEE Technology and Engineering Management Society, and “What to Consider Before You Accept a Management Role” from the IEEE Spectrum Career Alert newsletter. All are worth a look if you’re weighing a formal management track against staying close to the technology itself.

Remember: You don’t have to build your promoter network alone or only inside your own company. IEEE societies, sections and chapters, and technical committees, as well as the networking platform IEEE Collabratec, function as a ready-made cross-company network. They are practical places to find technology promoters with deep expertise in a field you don’t fully own yet, or to meet process and power promoters at other organizations who have built a promoter coalition around a similar idea.

For more tips on how to advance your career, check out our Career Advice for Engineers, From Engineers collection.

About 16 percent of the global population—more than 1 billion people—live with some form of disability, according to the World Health Organization. Many of the disabilities affect independence and mobility.

Three high school students working on inventions to help people with disabilities restore movement, translate thoughts, and navigate rough terrain had their work showcased at Regeneron’s International Science and Engineering Fair (ISEF), held in May in Phoenix. Their projects earned them this year’s IEEE Presidents’ Scholarship awards.

IEEE President Mary Ellen Randall presented the awards at a ceremony held during the fair. They also received an IEEE President’s coin, which students said was a highlight of their experience.

Hollie Tang won this year’s IEEE Presidents’ Scholarship of US $10,000 for her wheelchair navigation system. The award is payable over four years of undergraduate university study and includes a complimentary IEEE student membership.

Partap Sidhu, the second-place winner, received a $600 scholarship for his mind-controlled lower-limb exoskeleton. Third-place winner Calvin Shang Hung received a $400 scholarship for his rough-terrain robot. Sidhu and Hung also got complimentary IEEE student memberships.

Established by the IEEE Foundation and administered by IEEE Educational Activities, the Presidents’ Scholarship recognizes high school students who demonstrate an exceptional grasp of electrical engineering, computer science, or another IEEE field of interest.

Controlling movements with a tongue

Holly Tang won the 2026 IEEE Presidents’ Scholarship of US $10,000 for her Tonguage project, which is a noninvasive, computer-vision-based human-machine interface.Lynn Bowlby

Tang, a sophomore at Wilson High School in Hacienda Heights, Calif., secured the top prize for her Tonguage project: a noninvasive, computer-vision-based human-machine interface. Using tongue movements and a standard camera, the interface lets users control a computer and other digital tools as well as assistive technologies including wheelchairs. The tongue pad, one of the system’s core features, allows the user’s tongue to function as a directional cursor, while eye blinks serve as mouse clicks.

Tonguage translates the person’s tongue and eye motions into actionable commands in several ways, such as the tongue’s position inside the mouth and continuous movement patterns. The system’s multimodality combines input from the tongue with other facial cues.

The system includes a face-tracking feature for error prevention that verifies commands are coming from the intended user, disregarding anyone else who moves into the camera’s frame.

That is a critical safety measure for a wheelchair-navigation application, Tang says.

Accessibility was central to Tang’s mission. She built the system to run on relatively affordable, readily available laptop cameras rather than more costly specialized hardware.

“Mobility conditions don’t discriminate,” she says. “They can affect anyone of any income, gender, and socioeconomic status.”

Tang initially imagined Tonguage as a simple substitute for a keyboard and mouse. The more research she did, though, the more she realized that it could offer autonomy through applications such as wheelchair navigation, robotic arm control, and gaming, she says.

“We’re so focused on trying to give people autonomy over just basic human tasks that we often leave out things like gaming,” she says. “They deserve the freedom to play games and enjoy entertainment as well.”

Tang, who plans to pursue biomedical engineering, says a visit to a rehabilitation center solidified her purpose.

“Including empathy in your technological solution is so important,” she says. “Empathy is hard to teach in a classroom, but it can be learned through experience, and through actually meeting people whose lives your work might change.”

Mind-controlled exoskeleton

Sidhu, a junior at Bethpage High School, in New York, took second place for NeuroGait, a mind-controlled, lower-limb exoskeleton. He says he was inspired by his volunteer work at a community center that lacked elevators. He saw individuals with mobility issues struggle to navigate the three flights of stairs.

NeuroGait operates by reading the Bereitschaftspotential (BP), a faint electrical pattern that emerges one to two seconds before a person consciously initiates movement. Using a custom electroencephalogram (EEG) headset and a convolutional neural network (CNN), the system classifies intended movements and sends commands to a 3D-printed exoskeleton. Rather than rigid motors, the suit relies on pneumatic artificial muscles that Sidhu designed to mimic human anatomy.

“The pneumatic artificial muscle in itself is so compliant that it’s able to adjust to the limitations of the human body,” he says.

The technical specifications are striking: The CNN achieves a 99.9 percent accuracy in detecting a person’s intended movement, while the full system—from the brain’s signal to physical movement—operates at 95.2 percent accuracy, according to the results from 500 trials Sidhu conducted.

Perhaps most impressively, Sidhu built the entire system for about $276, less than 1 percent of the $40,000 to $100,000 price tag of commercial exoskeletons, according to a 2025 revenue report from Roots Analysis.

He says he hopes to bring NeuroGait to the community center where the idea for the project began.

He attributes his success to staying current with research from institutions and organizations such as Boston Dynamics and MIT.

“To be successful in research,” he says, “you have to know what’s being done right now.”

A spider-inspired robot

Hung, a sophomore at El Cerrito High School, in California, took third place for Math Into Motion: Robotic Hexapod for Hazardous Environments. The six-legged robot is designed to traverse terrain too unstable for humans or conventional robotic systems.

With only weeks before the science fair deadline for entries and no prior electrical engineering experience, Hung began with an idea inspired by his interest in spaceflight: an insectlike robot. He had spent years watching rovers such as Curiosity and Perseverance struggle on uneven surfaces, leading him to hypothesize that a hexapod design would be better for rugged ground.

As the project progressed, the humanitarian applications for his robot became clearer, he says. Watching news reports of the earthquake that struck Türkiye in 2023, as well as conflicts around the globe, Hung adapted his robot for use in disasters. The hexapod’s stable tripod walking gait, in which three legs stay grounded while the other three move, makes it well suited for navigating in collapsed buildings to locate survivors or to carry sensitive supplies such as insulin in conflict zones.

The current version moves using three mathematical techniques. Inverse kinematics converts a target leg position into the motor angles needed to reach it. Linear interpolation breaks each movement into a series of smaller steps for smoother motion. And Euclidean transformations translate the robot’s travel direction into instructions that each leg can follow, regardless of the way a leg happens to be facing.

Hung taught himself how to design a printed circuit board. He also taught himself 3D modeling, coding, and soldering. Figuring out the complicated mathematical transformations to coordinate legs facing different directions proved to be the toughest hurdle, he says.

After seven months of development and trial and error, a critical circuit board failure in his third version nearly ended the project, he says.

“There was a really strong moment of ‘Should I just give up?’” he recalls.

He simplified the design and rebuilt it from the ground up.

“I just decided to double down,” he says. The fourth version of the robot was the first that successfully walked across his living room floor.

He advises aspiring engineers that “if you find the right project and it truly becomes your passion, designing it almost starts to feel like fun, and that’s what carries you through.”

As the three young innovators demonstrate, the future of engineering goes far beyond technical ingenuity. Much is rooted in empathy and a commitment to human welfare.

Through initiatives such as the IEEE Presidents’ Scholarship, the IEEE Foundation showcases and nurtures bright minds poised to shape the next era of assistive technology and robotics.

For Tang, Sidhu, and Hung, the ISEF stage is just the beginning. They can look forward to impactful careers dedicated to advancing technology for the benefit of humanity.

The impact of AI on software development has been both profound and ever-evolving. Last year, I wrote about AMD’s plans to use AI not just for generating new lines of code, but also for other steps in the software development lifecycle (SDLC), such as triaging problems, debugging code, and testing the software. At the time, we were hoping for a 25 percent productivity boost from AI use over the course of two or three years.

But with each new release, the capabilities of large language models (LLMs) improve dramatically—accelerating software development, increasing the quality of AI-generated code, and fundamentally reshaping how software is engineered. Now, just one year later, we have surpassed our productivity target, achieving a 30 percent overall productivity boost through AI. On top of that, we are rethinking not only how we use AI within the SDLC, but the structure of the SDLC itself.

We believe that the biggest AI revolution in software engineering is still ahead. So far, we have largely been teaching AI how we perform tasks and asking it to mimic existing workflows. In many ways, this constrains AI to human patterns of thinking. The next transformation will come from collaborative swarms of AI agents capable of discovering solutions independently.

Agents of today

AMD began developing AI systems for code generation, testing automation, bug analysis, and code review in 2024. At the time, our objective was to achieve 25 percent AI-generated production code by 2027 while gradually automating larger portions of the SDLC.

Measuring productivity is inherently challenging, but from the outset we have consistently tracked one objective metric: the percentage of source code generated by AI. Importantly, we count only code that passes all reviews and testing and is ultimately included in the final product. While AI-generated code is certainly not the only contributor to productivity gains, it is one of the few metrics that can be measured objectively and consistently.

By this metric, we have crossed the 20 percent mark at the beginning of this year and are now progressing towards 50 percent across entire codebase. In some software components, more than 80 percent of the code is now generated using AI.

Agentic AI has enabled us to include AI in every step of the life cycle: For code analysis and triage, agents are trained to analyze problem reports, identify and group similar requests, and highlight which code snippets are likely to need modification. For debugging and code generation, agents are directed to analyze a bug request and implement required code changes. For testing, the agents generate unit tests, and if those are passed, identify necessary integration and product-level tests. And finally, for the approval and release stage, agents prepare architecture summary, code change review, and full test results for engineers’ review and approval—and, if approved, integrate the changes into the next release.

Agents of tomorrow

Today, engineers create AI agents in their own image: They teach AI what they know about the system, how they would fix an issue, and how they would implement a change. This is already a major technological advancement. Engineers can create multiple “AI versions” of themselves, allowing these agents to work in parallel, scaling their expertise far beyond the limits of individual productivity. The limitation, however, is that these AI agents are still constrained by human thinking and human-defined approaches.

AMD

We believe the next major transformation in software engineering will occur when collaborative AI agent swarms can independently identify and develop solutions, guided by humans on what to solve rather than constrained by human assumptions about how the job should be done. Instead of providing detailed instructions on how to solve a problem, engineers will define the issue, the desired outcome, and the quality, performance, and system constraints, allowing AI agents to determine the optimal path to a solution.

A swarm of AI agents will then work in parallel to generate, evaluate, and refine multiple solution approaches. These agents will automatically validate correctness, measure performance, test trade-offs, and compare alternative implementations against defined success criteria. Finally, AI agents will prepare ranked solution options, along with validation results and performance metrics, for engineer review and approval. The agents won’t be enhancing each step of the SDLC—they will be rewriting the SDLC themselves.

To get to this point, we need to change how agents are trained. Today, improvement occurs one engineer and one agent at a time: An engineer reviews the output, refines the prompt, and repeats the process. To scale beyond this model, agents must continuously learn from one another, reuse successful strategies, and improve collaboratively across projects and teams.

We are already moving in this direction by using multi-agent workflows extensively through agentic harnesses, such as Codex and Claude Code, while simultaneously developing our own internal multi-agent systems to support the next generation of AI-driven software engineering.

A good example is our AI-driven effort to resolve issues in our Radeon Software eXperience (RSX). RSX is a user interface component that allows users to configure and monitor graphics driver behavior. In October 2025, we began using AI agents to automatically debug and fix reported RSX issues. Out-of-the-box AI tools delivered limited results, resolving only 6 percent of issues.

The percentage of software issues fixed automatically by AI agents in AMD’s Radeon Software eXperience (RSX) has been growing steadily, reaching 75 percent in June 2026.

As we analyzed failures and identified ways to improve, we built a learning loop—initially a largely manual process—to understand where the agents were falling short and how to improve them. Rather than retraining the underlying models, we refined the objectives given to the agents, allowing them to iteratively explore multiple approaches, evaluate the results against defined success criteria, and converge on better solutions. At the same time, advances in models and agent run-times further increased effectiveness. Together, these improvements significantly increased our resolution rate from 6 percent to more than 75 percent of RSX issues resolved by agentic loop.

To make agents and agent swarms truly productive, we need a continuous learning loop that feeds errors and human interventions back into future agent workflows. The opportunity is to engineer this loop around clear, measurable goals. Each cycle captures new insights, making the entire AI engineering workflow smarter and more effective. Over time, this self-reinforcing loop—not just the underlying model—will become a key driver of AI progress.

The evolving role of human engineers

At AMD, we view AI as a means of increasing productivity, improving quality, and enabling employees to focus on higher-value work. Our goal is to empower our workforce with AI, not to reduce headcount.

To support this transformation, we are investing heavily in AI education and training across the company. The way we work is evolving rapidly, and we want every AMD employee to be prepared to leverage AI confidently, responsibly, and effectively.

As AI agents continue to improve, engineers will spend less time manually implementing solutions, focusing more on defining specifications, validating outcomes, and making the strategic decisions that drive innovation.

Bede Liu, a digital signal processing pioneer, died on 7 May. He was 91.

Liu was widely regarded as one of the founders of modern digital signal processing, a field that applies mathematical algorithms to analyze, modify, and transmit signals including sound, images, and video.

The IEEE Life Fellow taught electrical engineering at Princeton for more than 50 years. From 1994 to 1997, he chaired the university’s electrical and computer engineering department.

Liu’s research aided the transition from analog to digital processing of sound, images, and video. His work helped establish many of the mathematical and engineering techniques that underpin modern communications, multimedia systems, and consumer electronics.

Although little known outside engineering circles, his work is embedded in technologies used by billions of people. The low-power digital signal processors that make cellphone calls, streaming video, and Internet communications possible can be traced to research he conducted in the 1970s and ‘80s.

Liu received the 2018 IEEE Jack S. Kilby Signal Processing Medal for “sustained contributions to the analysis and the development of low-complexity realizations of digital signal processing algorithms.”

“We stream music and video. We take photos with our phones, and we send them around. We don’t even think about it,” IEEE Life Fellow H. Vincent Poor said in an obituary for Liu. “But it’s all because of the signal processing, image processing, and video processing that’s been developed over the years, as well as other technologies that have grown up beside it and enabled it, like semiconductors. The development of these processing advances was exactly what Bede was a major part of.” Poor is a professor of electrical and computer engineering at Princeton.

An impactful scholar and teacher

Liu was born in Shanghai in 1934. During his childhood, his family relocated to Taiwan amid the upheaval of the Chinese Civil War. His father, Henry Liu Sr., was an electrical engineer.

Liu earned his bachelor’s degree in electrical engineering in 1954 from the National Taiwan University, in Taipei. After graduating, he and his family moved to the United States. Liu and his father attended the Polytechnic Institute of Brooklyn (now the New York University Tandon School of Engineering) together. They earned their master’s degrees in electrical engineering in 1956. Liu continued his studies at the school, earning a doctoral degree in electrical engineering four years later.

In 1959 he was awarded a Bell Labs fellowship and worked at the company’s Murray Hill, N.J., location until he joined Princeton in 1962.

“Liu was a highly impactful scholar and teacher—always thinking ahead of future needs and changing technologies,” said IEEE Life Fellow Peter J. Ramadge, a Princeton professor emeritus of engineering.

Cellphones make use of a considerable amount of digital signal processing, Liu once noted. Many of the field’s advances, he added, involved making sophisticated processing practical on devices with limited computing power—which is the challenge that confronted generations of engineers designing portable electronics.

Liu’s research contributions helped shape both the theory and practice of digital signal processing. With Abe Peled, a former graduate student, he authored the 1976 textbook Digital Signal Processing: Theory, Design, and Implementation, which is a standard reference for engineers. Published before digital signal processing had fully emerged as a distinct discipline, it helped define the subject for practitioners and students around the world.

Liu also published 250 technical papers and was granted 12 U.S. patents. His papers are available to read on the IEEE Xplore Digital Library.

The first patent granted to him and Peled was in 1976 for a hardware design that processed bits in parallel, rather than in sequence. The innovation greatly increased computing efficiency for data including sound and communication signals.

Peled says Liu “demonstrated an openness to new ideas and a willingness to challenge the orthodoxy of the EE department at that time—which leaned heavily toward more theoretical information theory.”

A mentor to well-known engineers

Liu’s influence extended beyond his own research. He advised 53 doctoral students, many of whom went on to distinguished careers in academia and industry, including leadership positions at Google and IBM. One former student, computer scientist Robert Kahn, helped create the architecture of the modern Internet. Kahn, an IEEE Life Fellow, received the 2024 IEEE Medal of Honor.

“His former students were very successful,” Poor said of Liu, “and I think that’s a testament to his skill as a mentor.”

“Liu was a highly impactful scholar and teacher—always thinking ahead of future needs and changing technologies.”—Peter J. Ramadge

Together with several Ph.D. students, Liu developed methods of filtering and compressing digital signals to mitigate errors and dramatically reduce the computation needed for signal processing.

As digital signal processing moved from laboratories into commercial products, the impact of Liu’s ideas spread across industries. His research helped spawn the development of lower-cost and lower-power electronics and contributed to advances in mobile communications, multimedia technology, industrial automation, and biomedical imaging.

A focus on media integrity and copyrights

In the 2000s, Liu turned his attention to media integrity and copyright issues.

“With the increasing accessibility of digital media source material, the protection of ownership and the prevention of unauthorized alteration has become an important concern,” he wrote in his 2002 book, Multimedia Data Hiding. The book, which he co-wrote with his former doctoral student IEEE Fellow Min Wu, discussed the theory, techniques, applications, and security of digital watermarking—hidden signals that could identify a genuine copy of a song, image or video to prevent unauthorized distribution or tampering.

A Princeton team that included Liu, Wu, and another of his doctoral students uncovered serious vulnerabilities in watermarking technologies being considered by an industry consortium. They found that the standardization efforts were immature and would not protect against digital piracy.

“Now nearly every copy of a Hollywood film given to a critic or theater carries a unique digital forensic watermark to prevent unauthorized redistribution,” said Wu.

A force in the community

Liu, an active IEEE volunteer, served on the IEEE Board of Directors in 1984 and 1985. He was the 1982 president of the IEEE Circuits and Systems Society.

He was a member of the U.S. National Academy of Engineering, an academician of China’s Academia Sinica, and a foreign member of the Chinese Academy of Sciences.

Outside the classroom, he was recognized for his humility, humor, enthusiasm, and generosity. When thinking of Liu, IEEE Life Fellow Kenneth Steiglitz says, cheer is the first word that comes to mind.

Liu was “always ready with a positive remark, a quick smile or, maybe, some tips on the right way to cook a duck,” says Steiglitz, professor emeritus of computer science at Princeton.

Liu encouraged his students to take on ambitious, unconventional projects, and he inspired students and colleagues with his adventurous spirit.

Learn how full-wave simulation predicts very low antenna coupling on aircraft-sized platforms, and which three modeling techniques deliver accurate results with fewer computational resources.

Download this free whitepaper now!

If you haven’t already seen a job listing for a “product engineer,” you probably will soon. The job everyone’s suddenly hiring for, this role is like a cross between a product manager and an engineer (as the name suggests). And it’s a hiring trend worth paying attention to.

Companies are opening more of these roles every single month, but they’re struggling to fill them. The reason has almost nothing to do with engineers’ coding skills or years of experience.

The best career move you can make to prepare for these types of roles has almost nothing to do with getting more technical. Instead, it comes down to one of the fluffiest, most overused, and potentially cringiest words in all of tech: mindset.

Stick with me, I promise this goes somewhere useful.

The problem: We were trained to be task-takers

When I started out, my job looked like this:

Drive to an office. Sit through meetings that led to other meetings until a project manager handed me a task they’d already chopped into tiny pieces.

My job was to turn that task into code.

It took years for me to get good at a coding language and tech stack, and once I did, I executed that knowledge against specs that somebody else wrote.

You know what’s freakishly good at that exact job? I’ll give you a hint: It starts with A and ends with I.

Boris Cherny, the creator of Claude Code, recently said: “coding is basically solved,” and “the bottleneck is going to be good ideas.”

So if your entire value is “hand me a task and I’ll build it,” you’re in a footrace with the robots. I don’t like that for you.

The bad news... that is also good news

Many companies are flattening. Middle management is getting stripped out, for better or worse (mostly for worse), which means many of us are doing more with less.

This might sound like purely more work, but it’s also an opening for anyone who cares about what they’re building and can put on their manager hat. Companies are no longer just hunting for the strongest engineer in one narrow domain.

What’s rare, and what actually moves revenue, is an engineer who can spot the thing that’s quietly costing money and either flag it to leadership or just go fix it.

What this actually looks like

Being product-minded has NOTHING to do with your tech stack.

Here’s where to start:

Have an opinion and back it up. As a former engineering manager, the worst thing I ever heard was silence. I’d often ask the team what they thought because I doubted myself and wanted a gut check. I was grateful to the ones who said “nope, bad idea, here’s why.” Pushback is a gift.

Learn the domain, casually. Work for a plumbing company? You don’t need to become a plumber, but spend an hour on Reddit threads where plumbers vent. Now your ideas come from your potential customers.

Make experiments cheap and safe. This is where any engineer has massive leverage. Experiments are not free. A bad one loses customers and frustrates users. Tools like LaunchDarkly and Optimizely let you ship a change to 5 percent of users and roll it back the second it tanks. Learn them, or build a scrappy version yourself. A team that can quickly run safe experiments will out-learn everyone else in the building.

Be data-driven. Stop fighting about button colors. Pick a goal: making money, finding product-market fit, or making the product sticky so people come back. Then measure it. If your gorgeous redesign tanks time-on-site, it failed, no matter how good it looked to you. If the ugly version makes more money, ship the ugly version.

You don’t have to be the ideas person. Maybe you’re not a visionary. That’s fine. Organize a hackathon around an actual company goal. Pull up your company’s quarterly targets and build something against one of them. Don’t know what those targets are? That’s your first assignment.

Good ideas are the new bottleneck—and they always have been

When I was a manager, I asked myself one question every week: What’s the single most impactful thing I could do right now? The answer was almost never “write more code.” It was understanding a gnarly problem nobody had defined yet. Building a deck to spread knowledge that was in one person’s head. Getting the right three people in a room to actually make a decision we’d been putting off.

Code is cheap, and it always has been. We just couldn’t see it, because for decades the typing took so long that it felt like the hard part. It never was. The hard part was always knowing what’s worth building.

— Brian

Siobahn Day Grady Wants Everyone to Be AI Literate

In January 2025, Siobahn Day Grady launched the first AI research institute at a historically Black college or university. The institute aims to help expand AI skills for all students at North Carolina Central University, where Grady is an associate professor, through both AI research opportunities and skills training. Though the institute is the first of its kind, Grady hopes it could serve as a model for other HBCUs.

Read more here.

Should Researchers Write Papers for AI Instead of People?

AI is increasingly used in the scientific research process. So does publishing need to change to keep up? Jiachen Liu recently co-authored a paper published on ArXiv arguing that the PDF should be replaced with an “Agent-Native Research Artifact” designed with AI in mind. In this interview with IEEE Spectrum, Liu lays out a provocative vision of AI-driven research and an infrastructure that captures—and learns from—details that often get left out of today’s papers.

Read more here.

Detect Dark Matter’s Mark From Your Backyard

Astronomers still don’t know exactly what dark matter is, but they can detect it—and so can you. With a small radio telescope and a few other pieces, you can create a DIY setup to gauge how fast hydrogen clouds are moving across the Milky Way. Feed those measurements into a spreadsheet, and you can see the same signals that have baffled the astronomical community for decades.

Read more here.

A survey of over 700 professionals examines how visual and physical AI teams build systems, why models fail, and where data work drives production.

Download this free whitepaper now!

In collaboration with the Kingdom of Bhutan government, IEEE recently introduced its Engineering Education, Research, and Innovation Summit.

Held on 9 and 10 June in Paro, in the eastern Himalayas, the event was designed to help Bhutan navigate its digital transformation by focusing on the critical intersection of digital transformation, engineering education, and sustainable development.

The summit brought together global academic leaders, technology experts, and Bhutanese government officials to discuss how modern engineering curricula can evolve from theory-centric models into application- and skills-based frameworks. Discussions focused on how to build high-value research capabilities in the country, integrate artificial intelligence into higher education, and address foundational infrastructure challenges to ensure equitable, nationwide digital readiness.

“IEEE is proud to collaborate as a catalyst for progress in higher education as AI shifts the technology landscape and Bhutan prepares for its next era of innovation and resilience,” Mary Ellen Randall, 2026 IEEE president and CEO, said at the event. “Our goal is to support local universities and students as they develop trusted, future-ready technology that honors the nation’s commitment to sustainability and human well-being.”

The event featured an address by Bhutanese Princess Chimi Yangzom Wangchuck, who emphasized the importance of aligning technological innovation with the nation’s philosophy of gross national happiness (GNH), which prioritizes well-being, sustainability, and ethics.

“The question before us is not whether technology will shape the future; it certainly will,” the princess said. “The more pressing question is whether we can shape technology according to our values.”

A blueprint for Bhutan’s future

The summit helped establish a collaborative blueprint for a high-value knowledge economy in Bhutan through several key focus areas:

• Workforce readiness: designing industry-driven curriculum modernization and cocreating skills programs to equip graduates with practical, technical competencies.
• AI and research infrastructure: strengthening open science, trusted regional datasets, and global citation impact to prepare universities for AI-enabled learning environments.
• Values-driven innovation: merging GNH principles with technological advancement and helping ensure new engineering practices support climate-resilient infrastructure and green innovation.
• Institutional connectivity: using digital transformation to bridge technical capability gaps between urban and rural institutions; linking classrooms to a global research network.
• Promoting sustainability: convening stakeholders to exchange ideas on green innovation, climate-resilient infrastructure, and engineering education.

Expanding digital access

To help promote the effort, IEEE offered Bhutanese universities, government institutions, and industries a six-month complimentary trial of two key technical resources:

• IEEE Electronic Library. Delivered via the IEEE Xplore Digital Library, the IEL gives users access to more than 7 million documents—including trusted IEEE journals, conference proceedings, standards, and technical papers—to enhance research, teaching, and technology development.
• IEEE eLearning Library. This platform offers online courses developed by experts in engineering, computing, and technology, supporting flexible learning across core and emerging technical fields for professionals, faculty and students.

In a tranquil Boston suburb, on the far edge of a horse farm, where pasture gives way to woods, a crane lowers an enormous electrode into a borehole. The electrode, a half-meter-long cylinder with copper-tipped arms to ensure good contact with the borehole walls, descends—deeper, deeper—through layers of spongy sandstone to the hard, marbled roots of an ancient mountain range hundreds of meters below ground. Here the rock is tight; there are few cracks for water or gases to flow. But that’s about to change.

A stone’s throw away, a second electrode—a twin of the first—has been fixed in another borehole at the same depth. From above ground, a pair of high-voltage generators cabled to the two electrodes fires a series of pulses.

Tsss!…Tsss!…Tsss!…Tsss!…Tsss!….

Each discharge, heard faintly at the surface, is like a miniature, subterranean lightning strike. The rock between the electrodes heats. Pressure builds. Then, suddenly, the rock splits into a spiderweb of fractures.

On a horse farm outside of Boston, a worker sets up the well where Eden’s electrode will be lowered with a winch.Bob O’Connor

Eden GeoPower, the Massachusetts-based startup performing this peculiar field test, calls the technology electrical reservoir stimulation. The company’s tagline: “We break rocks with electricity.”

Eden’s researchers hope their rock-breaking technique will someday aid mineral mining, tap geothermal heat, or create geologic storage areas for carbon. But there’s an even more intriguing use that could create a whole new category of energy production: generating hydrogen underground.

The dream of a hydrogen-powered economy dates back to the 1970s, when petroleum shortages and rising concerns about pollution from fossil fuels sparked visions of cars, ships, planes, and industrial machines running on hydrogen instead of carbon. Hydrogen is often touted as a clean fuel because when it’s burned or consumed in fuel cells, it emits only water and heat. However, it currently takes more energy to make than it yields, and the cheapest and most common way is by reacting steam with methane, a potent greenhouse gas.

How to Break Rocks With Electricity

It’s possible to make zero-carbon hydrogen by splitting water with electrolyzers powered by renewable energy. But in most cases, the process is too expensive to be economical—a reality that burst the hydrogen-hype bubble in the early 2020s. Global demand for hydrogen in 2024 reached approximately 100 million tonnes, containing energy equal to only about 3 percent of the world’s annual energy consumption. Most of it is used as chemical feedstock for petroleum refining and for making fertilizers and plastics.

The frustrations of manufacturing clean hydrogen have convinced many entrepreneurs and scientists to instead seek the element underground. For the past half-decade, dozens of companies around the world have been hunting for buried stores of hydrogen, called natural or geologic hydrogen. But with a commercial-scale operation yet to be proved, Eden and a handful of other startups and research groups are chasing the more audacious scheme of producing geologic hydrogen artificially.

This approach, known as stimulated geologic hydrogen or engineered hydrogen, turns subterranean rock formations into giant hydrogen factories. It typically involves injecting water into iron-rich rock, which oxidizes the iron and releases hydrogen as a by-product. Fracturing the rock, as Eden is doing, creates a network of conduits for the water to reach iron-bearing minerals.

The concept of stimulated hydrogen is so new that few have had a chance to test it. Proponents say that if it works—which is a big “if”—it could provide almost unlimited energy for the indefinite future. There’s one way to find out: Start breaking rocks.

There’s Plenty of Underground Hydrogen

Hydrogen is the simplest and most abundant element in the universe, the stuff of stars and galaxies. Geologists have long known that Earth generates hydrogen gas through natural water-rock reactions, but until recently, the occurrence was regarded as a curiosity. The gas is so light that most experts assumed it all escaped through pores and cracks in Earth’s subsurface and didn’t accumulate in useful quantities.

During a demonstration at Eden’s testing site near Boston, an employee displays a central component of the company’s proprietary electrode. Bob O’Connor

Inklings that they were wrong emerged in the 19th and 20th centuries, when researchers in the former Russian Empire and Soviet Union reported hydrogen seeping from mines and wells. But in the ongoing frenzy for fossil fuels, these observations were largely overlooked or forgotten. Scientists later discovered hydrogen spewing from hydrothermal vents in the seafloor and feeding so-called eternal flames, like those of Türkiye’s Mount Chimaera, where ancient athletes lit torches for the first Olympic games.

Then, in 1987, in the village of Bourakébougou, Mali, people drilling a water well noticed a breeze blowing out of the hole. According to local lore, a worker leaned in for a closer look, a lit cigarette dangling from his mouth. The air instantly ignited, burning a brilliant blue.

The crew capped the well, which stayed sealed for 25 years until, in 2012, a Malian oil and gas prospector confirmed the ground contained a large reservoir of hydrogen. The prospecting company, now called Hydroma, had a small electrical plant constructed to convert the gas into power for the village’s residents. Soon after, startups in Australia, Canada, the United States, and elsewhere began searching for more hydrogen stores. By 2025, large multinational petroleum and mining companies were getting in on the game.

To date, hundreds of exploratory wells have been drilled across the globe. But although researchers have documented widespread hydrogen deposits, none have proved capable of producing the gas at rates and quantities needed for commercialization. “We’ve poked a lot of holes, and nobody has found the gusher—or at least they’re not talking about it,” says Douglas Wicks, a former program director at the United States’ Advanced Research Projects Agency—Energy who now advises companies pursuing geologic hydrogen.

A wellhead guides multiple lines downhole: fluid hose, electric cables, rope, control for a sealing device, and sensor communication. Bob O’Connor

Wicks says that in 2022, while at ARPA-E, he got “dragged into the rabbit hole of geologic hydrogen” by Emily Yedinak, then a Fellow at the agency, who was trying to convince her colleagues to take it seriously. “I was the ultimate doubter,” Wicks says. The astronomical price of electrolyzers had made him skeptical that clean hydrogen was a viable pursuit. Plus, if Earth really did contain vast pools of hydrogen, then surely humanity, which had been digging for natural resources for thousands of years, would have found them by now, he reasoned.

But after talking with geologists—who pointed out that people historically hadn’t found hydrogen because they hadn’t been looking for it—Wicks changed his tune. “I got the epiphany that geologic hydrogen is not just an accumulation; it’s a chemical reaction,” he says. “And if it’s a chemical reaction, then it can be stimulated.”

Finding large accumulations of geologic hydrogen entails stumbling on a Goldilocks set of conditions. You need iron-rich source rocks that have already produced or are producing bountiful hydrogen. You also need porous reservoir rocks that can hold sizable quantities of gas migrating from the source rocks. And you need solid cap rocks above the reservoir that trap the gas underground.

To stimulate hydrogen, however, you don’t need this just-right geology. All you need are iron-rich rocks, and then you can generate the hydrogen yourself.

“These rocks are everywhere,” Wicks says. “If you look at the amount of iron that’s within drilling range of Earth’s crust, you’re talking about quadrillions of tons of hydrogen being accessible. If we’re 1 percent successful just in the United States, we could power the economy for thousands of years.” A back-of-the-envelope calculation convinced him that the cost of stimulated geologic hydrogen could easily compete with hydrogen made from methane. “If we get the technology right,” he concludes, “this could be huge.”

Wicks wasn’t the first person to propose the idea, but he was the first to allocate major funding. In 2024, under his leadership, ARPA-E awarded US $20 million to 16 teams aiming to advance stimulation technologies and research. Winning ideas included fracturing rocks with fluid pressure or mechanical stimuli, exposing them to catalysts to speed hydrogen-generating reactions, and manipulating native microbial communities to enhance production. Eden’s rock-breaking project, the lone electricity-based approach, received $900,000.

Eden GeoPower’s Underground Rock Fracturing

Paris Smalls, Eden’s CEO, founded the company in 2017 as a 23-year-old graduate student at MIT. For his Ph.D. in civil and environmental engineering, he was studying the effects of electricity on rock strength and became interested in enhanced geothermal systems, which require fracturing hot, dry rocks to circulate water through them for extracting heat. This is typically done by hydraulic fracturing, or fracking—a technique borrowed from the oil-and-gas industry that involves injecting high-pressure fluids.

Fracking is controversial because it can cause earthquakes and groundwater contamination, and many regions have banned the practice. From an engineering perspective, it’s also imprecise. The fractures it forms are large and difficult to control. “You can’t get enough fractures where you want because the water ends up just going through the same cracks,” Smalls explains. Electricity, he knew from his Ph.D. work, could create more extensive and finely tuned fracture networks, enabling geothermal systems to produce more heat with less environmental risk.

To determine how permeable its fracture networks are, Eden measures fluid pressure downhole and flow rates at the surface. Bob O’Connor

Smalls immediately grasped that the same rock-breaking strategy could be used for mineral mining, carbon sequestration, and extending the life of oil and gas wells. But he hadn’t considered using it to make hydrogen. So when Wicks invited him to apply for the hydrogen program at ARPA-E, he was confused. “I didn’t get it at all,” Smalls says. “I’m like, ‘I break rocks. How am I going to generate hydrogen?’”

Not long after, Smalls met Alexis Templeton, a geomicrobiologist at the University of Colorado Boulder who had become an expert in geologic hydrogen by studying microbes that consume the gas and the mineralogical transformations that create it. “There was a lot of early interest in whether or not you could engineer the production of hydrogen from rocks,” Templeton recalls. “And the rocks with some of the best potential have all the right chemistry, but they need water. Nobody was excited to do hydraulic fracturing. So everyone was wondering, ‘Well, how are we going to get the water in?’”

Eden’s technology, Templeton understood, could be the answer. She agreed to join the company part-time as its lead geochemist, a position she held from 2023 to 2025. During that time, Eden ran its first pilot experiment, in an oil field in Oman, near where Templeton was already doing her own hydrogen research. The initial setup used DC power to send a steady flow of tens of kilowatts between electrodes in two wells. When Smalls’s team tested it in a petroleum reservoir made of soft, chalky carbonate, the rock fractured readily, increasing oil production by 30 percent.

But when they did the same test in hard rocks, like those needed for hydrogen and geothermal systems, they didn’t fracture much at all. So the team went back to the drawing board and came up with a fix: pulsed power.

Using Pulsed Power for Rock Fracturing

The idea of breaking things using pulsed power—short, concentrated bursts of electrical energy—originated with a mid-20th-century experiment in Soviet-era Russia. As the story goes, a physicist and inventor named Lev Yutkin was out in a thunderstorm when he saw lightning strike a log underwater. Rather than burn, as it would in air, the log exploded, as if blown up by dynamite. Intrigued, Yutkin tried to reproduce the spectacle in his lab. He placed a dinner plate in a water tank, dipped in two wire electrodes, and released a high-voltage pulse. The ensuing spark, he discovered, instantly ionized the water molecules between the electrodes into a plasma channel, which then rapidly expanded, creating a shock wave that shattered the plate.

Yutkin described the phenomenon in his 1955 book Electrohydraulic Effect. He later proposed numerous fanciful uses for it, such as cleaning pipes or breaking up kidney stones, which inspired real tools in use today, including electrohydraulic drills and rock-crushers, and a kidney-stone-busting medical device called a lithotripter. The following decades saw advances in pulsed-power systems and experimental techniques to better understand the complex physical processes involved. By the 2020s, when Smalls’s team began investigating it for subterranean rock fracturing, the technology seemed ripe for use, although that particular application had been little explored outside the laboratory.

“We essentially generate a plasma channel in the rock itself,” says Rafael Villamor-Lora, vice president of R&D at Eden. “This channel then expands very, very rapidly,” fracturing the rock with a shock wave. Bob O’Connor

Eden’s scientists first experimented with pulsed power on thumb-size hard-rock cylinders. Instead of submerging each sample in water, however, they placed a pair of electrodes at opposite ends of the cylinder and delivered pulses directly to the rock. Using this dry-pulse method, drawn from Smalls’s and others’ research, the team found they could form plasma in tiny, moist pockets between mineral grains. “We essentially generate a plasma channel in the rock itself,” explains Rafael Villamor-Lora, Eden’s vice president of research and development. With enough pulses, the fast-swelling channel, as in Yutkin’s investigation, induces a shock wave that fractures the rock.

To bring the technology to the field, Eden needed voltage high enough to break through meters of solid rock. The obvious solution was a Marx generator, which converts low-voltage DC power into high-voltage bursts by slowly charging and then rapidly discharging multiple capacitors in parallel. (Marx generators are commonly used in high-energy physics experiments and to simulate lightning strikes on power lines.) Eden custom-built two devices—named Zeus and Thor after the gods of thunder—which together can release a surge of several hundred kilovolts.

This time, the plan worked. In 2025, in an abandoned gold-and-silver mine in Colorado, Eden used Thor to successfully fracture a hard, igneous column, increasing its permeability tenfold.

Ezra Frank, a mechanical engineer at Eden, works on Zeus, Eden’s custom Marx generator. Bob O’Connor

In March this year, the company began setting up the test site on the Massachusetts horse farm to refine its systems and gather more data on how the technology performs in different geologic environments. Its engineers are also designing more powerful generators to discharge stronger and faster pulses. Because Zeus and Thor consume very little power—akin to running a toaster or two—it takes about a minute to store enough energy to fire a maximal pulse. It then takes around 100 pulses to penetrate around 10 meters of hard rock. So fracturing over longer distances or at multiple depths can take hours to days. That means Eden’s biggest cost is labor, not energy.

Smalls says Eden signed an agreement with a geologic hydrogen startup—he declined to say which one—to demonstrate electrical fracturing in a field pilot of stimulated hydrogen, which could begin late next year. Eden will need to prove its technology can help coax the gas from the ground at a profitable rate and cost.

“It’s no question whether we can produce hydrogen,” Villamor-Lora says. “The question is whether we can produce it fast enough to be economical.” In the lab, Eden researchers found they could generate up to four times more hydrogen from rock samples using the pulsed-power technique, compared with the amount found in unfractured samples. But that may not be enough to make stimulated hydrogen commercially viable without some additional technology.

Other Approaches to Stimulated Geologic Hydrogen

One of the biggest challenges in stimulating hydrogen is that there’s no obvious go-to recipe. Beyond the basic ingredients of water and iron, many factors affect how much hydrogen is generated and for how long, and fractures are only one factor. Laboratory studies have shown, for example, that the ideal temperature for maximizing hydrogen production is around 200 to 300 °C. Acidity, rock and water chemistry, and microbial inhabitants are other important considerations.

Making the puzzle more complex, each rock formation is different and may require different stimulation techniques or a combination of them. “There isn’t a single solution that will work everywhere,” says Alexei Tcherniak, CEO of the hydrogen startup GeoKiln. “You have to know the geology you’re operating in.”

Some promising rock formations, he points out, may already be fractured or porous enough to become saturated with water but too cool to make ample hydrogen naturally. To solve this problem, his company, based in Houston, uses a system of underground heaters originally developed for improving flow in heavy oil reservoirs and converting solid organic matter in young shale rock into extractable oil and gas. The heaters, which are commercially available, can be installed in boreholes drilled into hydrogen source rocks, similar to Eden’s electrodes. Tcherniak says that GeoKiln is ready to start field testing as soon as it can raise the capital.

Other researchers are exploring the use of catalysts—metal or chemical salts that speed hydrogen-generating reactions—which, they say, could replace or complement fracturing or heating to increase hydrogen production at less cost. Vema Hydrogen, for instance, is betting on a mixture of boiler-heated water and proprietary catalysts. “What I can say about our catalysts is basically what they are not, which is not toxic, not expensive, and not dangerous,” says Florian Osselin, Vema’s chief science officer. The company, also headquartered in Houston, has begun drilling pilot wells in Canada to test its mysterious brew. By injecting it into semi-permeable rock, Vema expects to achieve commercial production rates without fracturing. “We’ve done field-scale numerical simulations that give us a lot of confidence,” Osselin says.

Another stimulation method, proposed by the Denver-based startup Koloma, aims to expose more rock surface for generating hydrogen by mimicking natural weathering. The technique involves adding carbon dioxide to water and injecting the fluid at specific times to control for factors like acidity and gas concentrations. The carbon dioxide reacts with the water to form an acid that breaks down mineral chains in rock pores, thereby increasing the pores’ surface area, explains Tom Darrah, the company’s CTO, who studied and patented the method as a professor at Ohio State University. “I call it micro-pitting because the texture goes from smooth to rough,” he says. As with fracturing, more surface area means more hydrogen production—if you can get the formula right.

Rita Esuru Okoroafor, an energy resources engineer at Texas A&M University, is studying the effects of various stimulation approaches, including fracturing, catalysts, and carbon-dioxide injection, on hydrogen generation. Her data, based on laboratory tests of rock samples from around the world and numerical models of stimulated geologic hydrogen systems, suggest that none of these approaches alone will sustain hydrogen production at rates needed for long-term commercial development. “We’re still fine-tuning our models, but they’re telling us that we’re going to need a lot of fracturing, we’re going to need catalysts, and then we’re going to need restimulation,” she says.

The process of generating hydrogen, Okoroafor explains, will eventually consume all the readily available iron in exposed rock surfaces, causing production to plummet. By accelerating hydrogen generation, catalysts also accelerate its decline. “When these reactions happen very fast, they also die very fast,” she says. They also leave behind mineral precipitates that can clog existing cracks. In a recent study, she found that hydrochloric acid helps clear the debris, expose fresh rock surfaces, and reopen water pathways to restore production.

It’s too early to know which technologies will win out in the race for geologic hydrogen and if stimulation will even be needed to make it a viable industry. What’s more, production is just the first step toward commercialization. Many questions remain. Once hydrogen is flowing from the ground, how will the gas be purified? How will it be stored and transported? How will the industry be regulated? What are the environmental risks, and how will they be mitigated? What will be the cost?

“With all these wars and gas prices going up, we need to be preparing for the future,” Smalls says. But as is often the case with nascent technology development, life gets in the way. At the horse farm, fracturing started in June after being delayed for months, first by a snowstorm and then minor equipment failures and other logistical snags. “Everything takes longer than you think,” Smalls says. Still, he’s unfazed, ever the optimist. “I like to go after things that other people are afraid to.”

This article appears in the September 2026 print issue as “How To Get Hydrogen From a Stone.”

The transition from a purely technical expert or individual contributor position to a broader leadership role is one of the most challenging phases in a STEM career. It requires moving away from relying solely on technical excellence toward mastering systems thinking, adaptive leadership, and team alignment.

To help mid-career professionals navigate the shift, the inaugural IEEE International Leadership Conference is designed to provide attendees with practical tools to step into broader responsibility and champion an entrepreneurial mindset.

The ILC event is scheduled for 3 and 4 October in Budapest. Registration is open.

Thinking beyond technical contributions

To successfully step into a leadership role, technical professionals need to look beyond their individual output and focus on “understanding the larger system, and championing innovation by building trust and aligning new ideas with organizational goals,” says IEEE Life Senior Member Daniel Sniezek, cochair of the ILC program committee.

Because engineering decisions don’t exist in a vacuum, navigating the larger system requires recognizing how technical choices intersect with the organization’s broader business, operational, and ethical realities, Sniezek says.

By letting go of the need to be the sole technical expert and focusing instead on collaborative empowerment, he says, engineers can pivot into transformational leaders who align new initiatives with the organization’s strategic vision.

Ultimately, over the span of a career, an individual’s leadership journey evolves far beyond personal advancement to “creating a lasting legacy through the people you develop, the knowledge you share, and the innovations you inspire,” he says.

Solving the intrapreneur’s dilemma

Championing disruptive ideas within established corporate structures—sometimes called the intrapreneur’s dilemma—does not mean working against the organization. Rather, it requires emerging leaders to act like business owners instead of passive task-takers.

To begin thinking like an entrepreneur from within, professionals should shift their focus from merely executing assigned work to proactively identifying hidden areas that would create value for the company, building trust with colleagues, and presenting bold innovations as solutions to the organization’s long-term strategic goals.

That kind of self-starting entrepreneurial mindset is how leaders create opportunities out of institutional constraints.

IEEE Senior Member Deyasini Majumdar, cochair of the ILC program committee, advises professionals to exercise leadership and strategic thinking skills without being asked.

“Within the constraints of established organizational structures you can unearth a treasure trove of opportunities to innovate,” Majumdar says. “Remember: A key trait of an effective leader is to engineer solutions and lead, even in the face of difficulties.”

A multidirectional exchange

Leadership is a multidirectional exchange of ideas across generations—which is one focus of the ILC.

Although emerging leaders can gain invaluable strategic guidance from seasoned executives, the relationship is a dynamic, two-way street.

Modern leadership requires established executives to remain active learners. Addressing what senior leaders can glean from their mid-career counterparts, Majumdar emphasizes, the leaders must maintain “the openness to seek opportunities, to quickly adapt, and to learn and grow with everyone around them.”

A continuous-learning mindset is what keeps leaders agile and effective in a rapidly changing technological landscape, she says.

The mutual openness can serve as a bridge between generations.

Whether an emerging professional is making a mid-career pivot from technical expert to manager, or a senior executive is transitioning into a mentoring and advisory role, the fundamental rule of transformational leadership is similar. Success means shifting your focus from individual achievement to enabling the capability, growth, and legacy of others.

Building influence and impact

To help with the shift toward transformational leadership, the ILC is featuring sessions focused on questions professionals must ask at key career inflection points.

Rather than a single workshop, the distributed sessions aim to address diverse professional transitions, such as evaluating promotions, learning how to influence laterally, pitching innovative projects, and sustaining leadership energy.

Inflection points include evaluating new internal roles; building lateral or upward trust; pitching an idea about a disruptive project; and facing rapidly expanding responsibilities.

The questions include:

• How do I evaluate career transitions without discarding hard-won experience?
• How can I exercise leadership through credibility and collaboration, regardless of formal authority?
• How do I use an entrepreneurial mindset to create value and gain sponsorship for new ideas?
• How can I avoid the early warning signs of burnout while taking on more responsibility?

The conference is designed to equip attendees with systems thinking and communication mastery needed to cultivate influence.

Leadership is not just about reaching the top; it is also about engaging in a collaborative effort to multiply your impact across the ecosystem.

The history of networking is full of tools that repurposed solutions to very different kinds of problems first. Wi-Fi’s origins trace back, in part, to a team of Australian radio astronomers trying to detect signals from evaporating black holes. But the data-processing tools they’d developed also proved capable at extracting clean messages from any chaotic, echoing signal environment. Echoes are echoes, after all, whether from distant star systems or from the far corner of the house.

I research vehicle communications networks, connecting cars to cars and to transportation infrastructure like traffic lights—for tomorrow’s vehicle-to-everthing (V2X) networks.

V2X research has long relied on models that assume “perfect” or “ideal” network conditions, which is a simplifying assumption that makes the math tractable. But this assumption doesn’t reflect how real wireless signals behave in a moving, obstructed, high-density environment. That gap is exactly the kind of real-world unpredictability that open radio access networks (a.k.a. O-RAN)—an open, programmable architecture behind some 4G and 5G cellular networks—were built to manage.

So why has the O-RAN standard—which is open and available to be applied well beyond 5G telecom—never been used for vehicle communications?

Solutions to the vehicle-to-everything (V2X) problem have to date relied on new networking protocols built from scratch—only to discover chicken-and-egg problems, thorny standards wars, and real signal congestion challenges at scale.

By contrast, O-RAN allows V2X engineers to reuse the networking protocols already developed for cellular communications. O-RAN was developed assuming cellphone towers are generally fixed in place. But, as can be seen below, O-RAN accommodates mobile “towers”—cars and trucks, in this case—with little additional effort.

Imagining a New Way to Connect Vehicles

Self-driving vehicle technology has largely been an each-car-for-itself endeavor. Tesla’s approach, for instance, relies heavily on powerful on-board banks of computers and suites of sensors spread around the car.

However, as an alternative to the “data center on wheels” model, this new O-RAN approach to V2X relies on each car’s nearby neighbors, wherever they are on the road. Each O-RAN–connected vehicle can then use a diversity of cars’ sensors and viewing angles for better group coordination and decision-making.

There is, to be clear, no O-RAN V2X test network operating in the world. Not yet.

It was just 10 years ago that the Third-Generation Partnership Project (3GPP) released its initial cellular V2X standard. The 3GPP have refined V2X over three major releases since. In the U.S. and the EU, the FCC and related European agencies have put forward other standards for short-range wireless V2X communication protocols.

However, no consensus standard has yet emerged. So, lacking any clear, unambiguous guidance on the future of V2X networks, autonomous-car makers—like Waymo, Tesla, Zoox, and Cruise—have leaned more on self-reliance, bulking up each vehicle with as many sensors and GPUs as possible.

Here, though, is where O-RAN might be able to help.

A little like APIs (a.k.a. application program interfaces) connect one app to another on your smartphone, O-RAN serves as an API for the network itself. And because of O-RAN’s open standards, a wireless network becomes programmable, vendor-neutral, and open to custom applications called xApps.

To test our proof-of-concept framework, I have been part of a team simulating five minutes of O-RAN V2X network traffic over one square kilometer of urban area, using real buildings and real-world road layouts from OpenStreetMap and traffic patterns generated by the modeling package SUMO. The simulations assumed a traffic density of 50-70 vehicles per kilometer—not rush hour but not light traffic either. In our simulation, we assumed vehicles communicated via a millimeter-wave frequency of 28 gigahertz and that each component of our O-RAN V2X system had its own dedicated xApp.

Taken together, these inputs—real geometry, real traffic, and each vehicle’s live GPS position—constitute what network researchers call a digital twin of the urban environment. That’s a virtual replica detailed enough for the network to reason about the physical world in real time.

This virtual world gave us a real result, too.

The simulations, published recently in IEEE Network, revealed that existing V2X standards—in which cars uncoordinatedly spit out messages into the network—result in signals “talking” over each other some 80-100 percent of the time. However, using O-RAN signal coordination, the message “collision” rate dropped to near zero.

And that matters because a seized-up V2X network doesn’t just fail quietly. It can fail in ways that might make a road turn treacherous.

How O-RAN Can Coordinate V2X Traffic

High-frequency data links between cars are already difficult to maintain, even on a clear day with no buildings or city infrastructure getting in the way.

Yet, in this situation, existing V2X networks leave a car to conduct blind searches for each dropped signal beam. Traveling at highway speeds, that search takes long enough for the surrounding world to change completely.

An O-RAN network continuously tracks signal conditions across the network, and in O-RAN V2X simulations, we also gave the network access to a detailed map of the urban environment—building positions, road geometry, intersection layouts—combined with each vehicle’s GPS trajectory. Together, these parameters let the network’s control layer predict where and when a signal link is about to fail and instruct each car’s antenna to adjust before the connection drops.

Signal pointing is one failure mode. Losing the connection entirely—because no direct path exists at all—is another.

Consider, for instance, a crossroads of two busy streets, with a few alleys and parking lots adding to the list of potential dangers.

If a signal from car A cannot reach car B directly, or if the path length is too far for an individual beam to travel, the signal must find an intermediary car or stationary sensor nearby that can pass along the message. And existing V2X standards are slow and reactive—polling potential relay vehicles one-by-one: Are you available? Can you redirect this message?

By contrast, O-RAN keeps a running graph of optimized message routes, accounting for a range of real-world constraints. So when an O-RAN link fails (whether that link is direct from sender to receiver—or indirect), the system already has a reroute mapped out.

This is partly why we included “multi-hop routing” in the O-RAN V2X simulations.

Multi-hop V2X O-RAN routing complicated three separate elements of the simulation: for each signal’s middleman (some cars may be ideally positioned to relay a signal from car A to car B, but we made the simulation neglect any cars that were also overwhelmed with their own signals and signal-processing needs); for each signal’s strength (we required that every intermediate link be able to maintain a stable network connection, factoring in distance and traffic conditions); and for each signal’s latency (we required a realistic accounting for added signal latency time for each additional hop in a multi-hop routing).

And with each added complication, O-RAN V2X multi-hop routing continued to extend the network’s capacity from 25 percent of nearby cars connected (without multi-hop) to nearly 100 percent (with multi-hop).

These complications, at least at the simulation level, did not slow down the V2X network.

How Could O-RAN Ever Be Scaled Up for the Real World?

We are in touch with potential collaborators and institutions to develop testbeds, prototype hardware, and tester vehicles for potential proving grounds. The Institute of Science Tokyo, for instance, has already expressed interest in working on some of these early-stage problems.

To date, our published research on O-RAN V2X has centered around a computer simulation only. Real-world hardware will undoubtedly surface challenges our simulation could not. So, questions of network latency and the computational overhead needed for O-RAN V2X signaling remain as yet unresolved.

Plus, concerns about full interoperability and realistic security will each demand their own investigations. After all, no one will trust a V2X network to do anything if that network’s cyber vulnerabilities haven’t been anticipated and patched in advance.

Realizing the O-RAN V2X vision will require progress on multiple fronts simultaneously. On the standards side, O-RAN’s vehicular extensions—the interfaces that allow vehicles to participate in the network as managed elements rather than passive users—would ultimately need to be formally adopted by the O-RAN Alliance and recognized by 3GPP’s V2X specifications. That process takes years.

On the industry side, there is a more immediate problem that our architecture is already positioned to solve: interoperability.

Today, a car made by one manufacturer cannot necessarily parse V2X sensor data sent from a car made by another. Firmware is proprietary; data formats differ. But an O-RAN control layer would act as a universal translator—normalizing each vehicle’s data into a common format and accelerating a push toward true multi-platform vehicle-to-vehicle communications. A more widespread and truly universal standard would, by itself, represent a substantial step forward for V2X.

Today’s U.S. electrical grid, among the largest, most complex systems ever built, is operating at its limit. The combination of rapid industrial growth, more frequent extreme weather, and a record surge in electricity use has pushed the grid to its breaking point, according to the U.S. Department of Energy.

Built decades ago for a more predictable world in which power came mostly from centralized coal or gas plants and electricity use grew at a steady pace, the grid faces unanticipated strain due in part to growing demand from data centers. The jobs of professionals managing the infrastructure have evolved from traditional engineering tasks to complex, fast-moving challenges.

Industry reports show that millions of modern digital sensors, smart meters, and grid monitors are generating nonstop waves of information. The sheer volume of data requires instant, automated computer analysis because human operators cannot process it fast enough.

Pressure on utilities stems from two sources: a spike in electricity demand and a shift in how power is generated.

An example of the operational strain can be seen at the regional level. With the recent deployment of artificial intelligence tools and high-performance computing, data centers require immense amounts of energy to operate. The largest power transmission utility in Texas recently reported a staggering 220 gigawatts of new connection requests, driven largely by a surge in AI and cloud-computing facilities, according to a CNBC report.

Alongside the rise in regional demand, global energy networks are absorbing an unpredictable variety of weather-dependent renewable energy such as wind and solar. The switch creates a volatile operating environment wherein supply and demand are balanced, second by second, to prevent blackouts.

The challenges are compounded by the vulnerability of the grid’s physical and digital framework.

More-frequent severe weather events cause costly disruptions, such as the devastating winter freeze that crippled the Texas grid and record-breaking heat waves that have overloaded transformers.

Simultaneously, the energy networks’ digital architecture faces threats. As utilities replace outdated analog equipment with smart meters and control systems, they are increasingly vulnerable to cyberattacks.

To overcome physical and digital vulnerabilities, grid reliability organizations, such as those conducting North American security simulations like GridEx, emphasize that the grid must become smarter, more agile, and completely automated. Energy researchers are noting that the key to this change lies in integrating AI across every layer of utilities’ operations.

The AI imperative

According to energy industry experts, using AI to manage power systems is no longer a futuristic research project; it has become a baseline operational necessity. Grid analysts emphasize that traditional grid-planning methods are too slow to handle rapid energy dynamics or to balance volatile renewable energy in real time within decentralized power systems such as microgrids.

AI can fill the gap by processing vast amounts of data instantly. Machine learning algorithms can quickly analyze information from thousands of sensors, historical usage patterns, and weather forecasts to predict issues before they happen.

An industrial digitization study conducted by McKinsey & Co. indicated that integrating advanced data and automation across infrastructure networks could reduce system design errors, decrease equipment downtime by up to 50 percent through predictive maintenance, and extend the lifespan of power machinery by up to 40 percent.

From forecasting energy spikes to automatically fixing localized voltage drops, AI acts as the digital backbone of a self-healing grid, experts say. Deploying the complex systems requires a new workforce: power engineers who understand data science, as well as data scientists who understand electricity.

Upgrading the Workforce

To bridge the gap between groundbreaking AI research and practical field deployment, IEEE Educational Activities, in partnership with the IEEE Power & Energy Society, has launched the online Artificial Intelligence for Power and Energy Systems course program.

The program explores core challenges threatening modern utilities. Rather than treating AI as an unverified black box that operates without human supervision, the curriculum focuses on safety, asset preservation, and strict reliability standards.

The curriculum is designed to educate power system engineers, utility managers, and data scientists tasked with modernizing the grid. The program was developed by Fangxing “Fran” Li, professor of electrical engineering and computer science at the University of Tennessee in Knoxville and chair of the IEEE Working Group on Machine Learning for Power Systems.

Five learning modules

The program breaks down the technical transition into five modules that bridge high-level theory with real-world solutions:

AI fundamentals. This module teaches engineers how basic machine learning models apply to power grids. It discusses how specialized neural networks solve complex power-flow calculations and how AI models can safely transition from computer simulations to physical, high-voltage equipment.

Accelerating grid control. Learners are taught to leverage deep reinforcement learning, an AI approach that uses trial and error, to accelerate automated grid adjustments during emergency power events.

Forecasting and data analytics. Using predictive modeling, engineers learn how to predict sudden demand surges, variable wind and solar outputs, and fluctuating wholesale electricity market prices to keep power affordable and available.

Physics-informed and safe AI. To address trust—a barrier to utility AI adoption—this course covers AI models hard-coded to obey the laws of physics. The approach is designed to ensure that automated algorithms never make erratic choices that damage grid equipment.

Generative AI and next-generation tech. Learners can explore the frontier of utility technology, including graph neural networks and large language models. This module highlights how generative AI can process complex, interdisciplinary data to streamline utility planning, emergency responses, and regulatory reporting.

The algorithmic literacy and practical execution tools provided by the course program can help convert systemic risks into grid resilience.

For individual access, visit the IEEE Learning Network. If you are looking for customized organizational options, contact a content specialist to discuss volume pricing.

This report examines R&D waste and how AI adoption has outpaced the intelligence needed to make consequential decisions well.

What Attendees will Learn

1. Where R&D budget is lost. More than a third of organizations spend 25 to 40 percent of their R&D budget on projects that never reach market.
2. Why projects fail late. Almost half of teams estimate over one million dollars in wasted investment for each project killed during development or testing.
3. Why AI adoption has not closed the gap. Most organizations apply AI to execution tasks such as data analysis and modeling rather than to decision support.
4. Where intelligence matters most. Respondents say better access to intelligence has the greatest value at early ideation and feasibility before significant investment is committed.

Most hospitals and health care providers use electronic health records instead of paper charts to note patient vaccinations, diagnoses, and procedures. AthenaOne, Epic, and Oracle Health are some of the systems employed around the world. Many patients can access their electronic medical records from home.

The platforms exist thanks to pioneering efforts such as the Medical Information System (MIS-I). The first hospital-wide computer system, it was developed in the 1960s by Lockheed Martin (then known as Lockheed Missiles and Space Co.) in partnership with El Camino Hospital, in Mountain View, Calif. Doctors and nurses used MIS-I (pronounced miss-ONE) to admit patients, order lab work and imaging, schedule follow-up appointments, and issue hospital bills, according to a 1973 article published by Datamation.

Although many people now know how to type on a keyboard, in the 1960s, most did not. Therefore, MIS-I included a light pen, which worked like a stylus for today’s touchscreens.

MIS-I was recognized as an IEEE Milestone during a ceremony on 14 May at El Camino Hospital. The IEEE Santa Clara Valley Section sponsored the Milestone.

“The medical information system was more than a technological breakthrough; it was proof of what can happen when clinicians, engineers, administrators, and community leaders unite around a common goal: improving care for patients,” Dan Woods said at the event. He is chief executive of El Camino Health, the nonprofit organization that maintains the hospital.

“This pioneering work helped establish the foundation for the modern medical informatics industry,” Woods said. “The legacy of Lockheed’s innovation continues to benefit patients and health care providers around the world, making this achievement truly worthy of lasting recognition.”

Bringing technology into clinical care

Prior to Lockheed’s effort, health records remained paper-based. They often were stored in dedicated rooms within the hospital. Files were kept in heavy-duty manila folders organized on mechanized open-shelf filing systems, revolving rotary files, or locked steel filing cabinets, according to EO Johnson Business Technologies. The process of retrieving a patient’s medical history was cumbersome and could hamper decision-making in a life-or-death situation. Paper records were prone to human error, according to an EHR in Practice article. In addition, upkeep could be costly due to administrative expenses such as transcribing doctors’ notes, storing patient charts, adding medical codes, and managing insurance claims.

Companies and universities including General Electric, IBM, and Harvard began exploring how to use computers to improve clinical care. They developed several systems for hospital laboratories to track test orders and results, as explained in the Milestone webpage.

In 1964 Lockheed was looking to diversify its portfolio, Melville Hodge, who helped lead the MIS development, said at the dedication ceremony. The company decided to apply its expertise to health care, and later that year this focus-area became part of a new information systems division. Hodge, who at the time oversaw multiple R&D efforts, became the driving force behind the MIS program.

MIS-I displayed patient information on a 14-inch television purchased from a department store. Below that monitor was a keyboard and a light pen.Ian Thomson/Computer History Museum

In 1966 Lockheed secured a contract with the Mayo Clinic, in Rochester, Minn., to assess its computer system needs and those of its two associated hospitals, Hodge wrote in a paper detailing MIS history.

He and a small team of engineers worked with Mayo Clinic physicians for two years to build the prototype of what would become MIS-I.

The system displayed patient information on a 14-inch television purchased from a department store. Below that monitor was a keyboard, and to its right was a printer. Doctors and nurses would swipe their ID badge to access the system, then use the keyboard to put information into the patient’s file or send a request to a pharmacy or laboratory. They also could print documents.

But one problem kept cropping up: Most doctors didn’t know how to type. The computer mouse was still in its infancy, and Hodge suspected it would not solve the problem, according to a video shown at the dedication ceremony. Instead, he “borrowed technology from a then-secret satellite program,” he said.

That technology was the light pen, which was used with MIT’s Whirlwind Computer in the 1950s.

“The insight that physicians could not and would not learn to type, combined with the innovative solution of light pen interaction, transformed an impossible dream into practical reality,” the Milestone proposers wrote.

To display text, the system used matrix programming, a 2D data structure consisting of rows and columns. Using the light pen, a doctor or nurse would select text from a list of general categories on the monitor. The options included the patient’s personal and medical information, family medical history, current illness, and physical exam findings, according to a 1968 article in the medical journal JAMA. The computer would display the requested information or list the next steps to complete tasks such as sending a prescription to a pharmacy. The keyboard remained part of the setup because it could allow users to input new information and update patient records.

To further develop the system, they submitted a proposal to the U.S. Department of Health, Education, and Welfare (now split into the Departments of Health and Human Services and Education) to secure additional funding, but it was rejected.

Herschel Brown, Lockheed’s executive vice president, and Kenneth Larkin, its director of information systems, decided to fund its commercial development, Hodge wrote.

When the company’s contract with the Mayo Clinic ended, the Lockheed team returned to Sunnyvale, California to refine, test, and deploy the system at El Camino Hospital.

“I admire Ed Hawkins, who was its first administrator, for having the courage to take on this kind of project while running a hospital that was only four years old at the time,” Hodge said at the dedication ceremony.

Making MIS-I a commercial success

Starting in 1968, early prototypes were installed in the hospital’s M.D. lounges and nursing station at El Camino Hospital. The organizations worked to configure the system so it met the hospital’s needs.

By 1969, a number of monitors had been installed, including in admissions, pharmacy, and radiology. The information from all the connected machines was stored in a data center housed in a separate location outside the hospital.

In 1971 Lockheed encountered difficulties with its C-5A and L-1011 aircraft programs, according to Hodge. The company was forced to curtail discretionary new business programs including MIS-I. The program was sold to Technicon of Tarrytown, N.Y., a leader in clinical laboratory automation.

The medical information system business operated independently as a subsidiary unit, and the transition marked the beginning of MIS-I’s commercial expansion.

That same year, MIS-I went live for hospital-wide use. Physicians’ orders were communicated to other departments, test results and radiology reports were retrieved, and nursing care planning and documentation were available, according to the Journal of Nursing Scholarship. MIS-I supported most information handling for nurses, physicians, and other medical personnel in the hospital.

But the change was not welcomed by all, according to the video about the technology. Nurses tended to praise the system, but many doctors had a hard time transitioning from paper to computers. They complained they were “spending more time fighting a machine” than interacting with their patients, according to the video. Some even retired to avoid learning the system. But nurses fought to keep it, emphasizing to doctors how much it improved patient care.

In 1974 El Camino Hospital held a vote of medical staff to determine whether to keep the system or return to paper-based records. About 60 percent of doctors and more than 90 percent of nurses voted in favor of keeping it, according to the Milestone webpage.

“The medical information system was more than a technological breakthrough; it was proof of what can happen when clinicians, engineers, administrators, and community leaders unite around a common goal: improving care for patients.” —Dan Woods, El Camino Health CEO

In 1975 nonprofit Battelle of Columbus, Ohio, evaluated how well the system was working for El Camino Hospital. It found that MIS-I reduced the time nursing staff spent on clerical tasks, improved communications among departments, and facilitated better planning of patient care. The survey also showed that more readily available, complete, and accurate information was being used to administer care and monitor patient progress, according to the report.

By 1993, the technology was installed in more than 200 hospitals in the United States, Canada, and Europe, according to the Milestone entry.

El Camino Hospital used MIS-I for 34 years, until its decommissioning in 2005. It was initially replaced by Eclipsys Sunrise XA and then ultimately by Epic.

Honoring an IEEE Milestone

The dedication ceremony brought together IEEE leaders, hospital staff, and government representatives. Hodge and his family also attended. IEEE President-Elect Jill Gostin made a presentation about IEEE, and Brian Berg of the IEEE History Committee discussed the organization’s Milestone program.

Hodge participated in a Q&A session with Deb Muro, chief information officer of El Camino Health. He told a story about the early days of MIS-I that he said he will never forget. During a hospital board meeting at which physicians were complaining about the system, an announcement was made over the hospital’s public address system that MIS-I wasn’t working.

“I had to ignore it to survive,” Hodge said, laughing. “As physicians got more used to it, and with a phenomenal poking from the nurses, doctors who wouldn’t use it were forced to.”

A bronze plaque recognizing the MIS-I as an IEEE Milestone has been installed in the lobby of the hospital in Mountain View. The plaque reads:

From 1965 to 1974, the first hospital-wide computerized medical information system was created by Lockheed Missiles and Space Co. in partnership with El Camino Hospital. Innovative light-pen terminals enabled physicians and staff across all departments to efficiently and accurately access patient data and enter work orders. By providing immediate feedback and seamless communication, it reduced costs and errors, improved safety and outcomes, and led the way to modern medical and clinical informatics.

Reviewed by the IEEE History Committee and approved by the IEEE Board of Directors, IEEE Milestones recognize outstanding technical developments around the world that are at least 25 years old. The Milestone program is administered by the IEEE History and Heritage group.

To learn more about historical figures in engineering, IEEE Milestones, and IEEE History Center programs and events, check out The Institute’s IEEE Tech History collection. IEEE Spectrum also covers aspects of tech history.

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.