Last August, looking up from the edge of an ancient river delta on Mars, the Perseverance rover’s electronic eye spied an oblong shadow crossing the distant, glowing Sun. Phobos, the Martian moon named for the Greek god of fear and panic, was transiting the solar disk. Too small to darken the pale, dust-tinged sky, but orbiting close enough to eclipse a significant fraction of the Sun, Phobos passed in less than a minute.

NASA’s Perseverance rover captured Phobos passing in front of the Sun from the surface of Mars.ASU/MSSS/SSI/JPL-Caltech/NASA

The tiny moon, with its murky origins, is an enigma, and one whose makeup and history could tell us much about Mars and the wider solar system. And now the international Martian Moons eXploration (MMX) mission, which successfully launched on 20 October, seeks to explore that mystery with an unprecedented series of landings on the unknown surface of Phobos.

This mission brings unusual challenges—for example, designing an orbit to study a microgravity moon orbiting a much larger planet. Engineers also had to figure out how to assess and land on this low-gravity pile of loosely assembled dirt and rock, how a rover might rove across it, and how the spacecraft might take samples and return them to Earth.

“I often say that landing on Phobos is like trying to land a balloon,” says Yasuhiro Kawakatsu, project manager of MMX at the Japan Aerospace Exploration Agency (JAXA). “It’s very difficult to drop a balloon to the ground and prevent it from tilting or bouncing.”

Why Phobos Is Such a Strange Moon

Phobos is one of two small Martian moons; the other is Deimos, named after the Greek god of dread and terror. Their cosmic origins are unknown, though there are two prevailing theories. One holds that the moons were once asteroids from the belt between Mars and Jupiter, captured by the smaller planet’s gravity. In the other theory, Phobos and Deimos formed when a large celestial body slammed into Mars, sending debris into orbit. Some models suggest the pair are the last survivors from a cluster of satellites.

The enormous Stickney crater, roughly 9 kilometers across, dominates this view of Phobos taken by NASA’s Mars Reconnaissance Orbiter.University of Arizona/JPL-Caltech/NASA

Scientifically speaking, Phobos is considered a rubble pile, a loosely consolidated aggregate of rock and dust. Far too small to become spherical under its own mass, it’s shaped like a lumpy potato, roughly 27 kilometers in length, with only 1/1,800^th of Earth’s gravity. On Phobos, a typical adult would weigh about as much as a tennis ball does on Earth.

More than 10 years in the making, MMX launched on Japan’s flagship H3 rocket from Tanegashima spaceport on 20 October local time. It will arrive at Mars in 2027 carrying scientific payloads from several space agencies, including NASA, and a new rover called Idefix developed by the French and German space agencies (CNES and DLR) in collaboration with JAXA.

Engineers assembled MMX’s modules at Japan’s Tanegashima Space Center before the spacecraft’s October 2026 launch.JAXA

MMX consists of three modules: a propulsion module to power its journey to Mars and apply the brakes to enter orbit around the red planet; an exploration module carrying the rover, sampling gear, and other scientific instruments; and a return module, attached to the exploration module, that harbors the capsule destined for a high-speed reentry into Earth’s atmosphere. After collecting samples on Phobos, the exploration and return modules will lift off together. The exploration module will then be jettisoned while the return module sets a course back home.

If all goes according to plan, MMX will deliver a 10-gram sample to Earth in the second half of 2031. But the mission will face a series of daunting challenges to reach that end point, including navigating complex orbits around Phobos and landing on its surface while Mars periodically blocks communications with Earth.

The MMX’s modular design includes the propulsion module [left] and the exploration and return modules [right].Kyodo News/Getty Images

“It’s the most complex mission I have ever seen,” says Patrick Michel, director of research at CNRS (the French National Center for Scientific Research) and co-principal investigator on Idefix.

How MMX Will Orbit Phobos

An H3 rocket carried MMX into space on 20 October 2026, beginning the mission’s five-year journey to Phobos and back.The Asahi Shimbun/Getty Images

Once MMX reached low Earth orbit, its propulsion module’s engine fired, kicking off a nearly year-long deep-space cruise to Mars. On arrival, the spacecraft will need to fire its engines in an extended, carefully timed burn to slow down enough to be captured by Mars gravity. The propulsion module will then be jettisoned, leaving the exploration and return modules to orbit Mars.

“If something doesn’t go according to plan, MMX will fly by Mars and get catapulted into the solar system,” says Nicola Baresi, an astrodynamicist who coauthored a paper on specially designed orbits for MMX.

From that initial orbit, MMX will transition into a specially designed quasi-satellite orbit that allows it to keep pace with the small moon but also move around it. Since Phobos is too small to orbit in the conventional sense, the spacecraft will orbit Mars alongside the moon, following a clever egg-shaped path that slings MMX alternately ahead of and behind the moon. “When you’re closer to Mars, you’re accelerating with respect to Phobos, and when you’re farther from Mars, you’re decelerating. So you essentially go back and forth,” explains Baresi.

Martian Maneuvers

The best existing images of Phobos, from spacecraft such as NASA’s Mars Reconnaissance Orbiter and ESA’s Mars Express, have resolutions of only a few meters per pixel. That’s larger than Idefix, the mission’s rover, itself about the size of a microwave oven. To map the surface in greater detail, MMX will loop in front of and behind Phobos at distances ranging from 200 kilometers to less than 20 km from the surface.

During its two-year remote sensing campaign, the exploration module will use several cameras and a laser altimeter to map the surface in high resolution and look for candidate landing sites. When Mars passes behind the Sun as seen from the Earth, causing a roughly two-week communications blackout, MMX will retreat into a larger, stable orbit. After communications resume, MMX will prepare to send its first visitor down to the surface.

Idefix Rover: The First Penguin

Before the exploration and connected return modules attempt to land, Kawakatsu says, MMX will send a scout that in Japanese is called “the first penguin,” leaping into unknown waters. For this mission, the first to leap will be the Idefix rover, named for the canine companion to Obelix in the French comics series Asterix. The exploration module will release the rover during a low-altitude pass of between 40 and 100 meters above Phobos’s surface in 2028 or early 2029.

MMX will use its TENGOO camera [left] and LIDAR laser altimeter [right] to map Phobos’s terrain.JAXA

It will be a tense moment that calls back to 2005, during Japan’s Hayabusa mission to explore the asteroid Itokawa. In a similarly designed drop, the MINERVA rover was released from an orbiter, but erroneously at 200 meters instead of the intended 70 meters. The rover missed the asteroid and, without propulsion, drifted off into deep space.

Idefix will get a slight nudge out of the exploration module toward its negligible-gravity target, freefalling and likely tumbling as it goes. It’s expected to touch down with a velocity of less than 1 meter per second. Unlike rovers that have landed on other surfaces with airbags or dedicated landing legs, Idefix will land without protection, apart from some extra padding on the solar panels due to their critical nature. The barely-there gravity of Phobos should allow the 23-kilogram rover to land softly, though exactly how this landing will play out is hard to predict.

The Idefix rover will travel to Phobos with its wheels tucked in and solar panels folded.DLR

“It will bounce probably several times,” says Stephan Ulamec, Idefix’s co-principal investigator and project manager at DLR. “It will come to rest somehow in some position.” Once it’s still, four retracted legs will deploy to right the rover autonomously, regardless of its orientation. It will then unfold its solar panels to charge its batteries for exploration tasks ahead.

A simulation shows how Idefix is expected to tumble onto Phobos, right itself using its articulated legs, and begin exploring the moon’s surface. DLR

Once operational, Idefix will roll at a snail’s pace, topping out at one millimeter per second. Its metal wheels have protruding blades like a paddle wheel, designed to gain traction in Phobos’s microgravity. “If we were to drive our rover faster, then it would actually be more like we were swimming through instead of driving across the surface,” says Cecily Sunday, a doctoral student involved in testing Idefix’s WheelCams at the Institut Superieur de l’Aeronautique et de l’Espace (ISAE-Supaéro) in Toulouse, France.

The WheelCams, a pair of cameras placed on the underside of the rover, will observe the wheels and their tracks during Idefix’s planned 100-Earth-day primary mission. The images will help researchers study the size and distribution of grains in Phobos’s regolith and how cohesive they are. How much or how little the grains stick together will greatly impact how Idefix moves. Sunday likens this to the difference between driving on dry and wet sand. “Think of trying to build a sandcastle with dry sand,” she says. “You can’t. Add water, and suddenly you can.”

The Idefix rover was mounted between the MMX spacecraft’s landing legs for the journey to Phobos.DLR

The team tested the rover’s nine-bladed wheels on a variety of materials, including Phobos regolith simulants provided by JAXA and DLR, lunar simulants, glass beads, quartz sand, and clay pebbles. “We characterize the mechanical properties of each simulant precisely, then test across the full range, with varying grain size, angularity, and even color, so we can map what we see in the WheelCam images back to actual surface properties,” Sunday says.

Despite the microgravity and Phobos’s rubble-pile nature, Sunday says Earth-based tests can approximate the rover’s motion on the surface of the small moon fairly well. While Phobos’s reduced gravity lowers the rover’s effective weight, there is also reduced pressure between the grains beneath its wheels. “As long as you’re moving very, very slowly and using the same materials and the same weight of the system, they should be representative,” Sunday says.

With four paddle-like wheels and deployable solar panels, Idefix is designed to crawl across Phobos.Ardan Fuessmann/Imago/Alamy

Idefix’s interactions with its new home will offer “sub-millimeter resolution and information on porosity, on grain size, on surface strength, on how hard is the material,” Ulamec says. “It fills a gap between remote observations from orbit and detailed observations of the return samples back on Earth.”

What Idefix finds on this rocky, rugged world will help determine how and where JAXA attempts to land the main spacecraft. Michel says the touchdown will provide ground truth for understanding of Phobos. When NASA’s OSIRIS-REx spacecraft touched down on the asteroid Bennu to collect a sample, its sampling arm surprisingly sank nearly half a meter into the loose material. But Ulamec doesn’t think Idefix will sink deep into Phobos, noting that Bennu is very different and much smaller, with a diameter of 500 meters.

How MMX Will Land and Collect Samples

Once Idefix has investigated the surface, MMX will face an even greater challenge: landing its much larger exploration module on Phobos. The spacecraft will make a controlled descent to roughly 10 meters above the surface, then drop the remaining distance in free fall. With so little gravity to pull it downward, even a small sideways motion could cause the lander to tip or bounce away. A successful, upright landing would mark the start of up to two hours on Phobos for the crucial sampling phase.

The MMX mission draws on the experience and technologies of Japan’s two Hayabusa sample-return missions, which landed on the asteroids Itokawa and Ryugu. But this mission has many more moving parts. MMX’s exploration module will use two sampling devices—a coring sampler developed by JAXA and a pneumatic sampler provided by NASA and fabricated by Honeybee Robotics, a subsidiary of Blue Origin. Data from these tools, and those carried by the Idefix rover, could answer questions about the origin of Phobos and how the Mars system has evolved.

The coring sampler uses a robot arm and a cylindrical corer designed to reach at least 2 centimeters below the surface. The arm can maneuver to find the best site for digging. “That by itself is a challenge because we’ve never done it,” Michel says. “You never know what can happen at the very sampling site itself, where you might have a boulder and the arm cannot move.”

Probing beneath the top layer will be valuable because solar radiation, galactic cosmic rays, and micrometeorites have battered and weathered the surface. The material below should offer a better record of the moon’s origins, and JAXA aims to collect at least 10 grams of it.

MMX will carry both the Idefix rover [bottom] and a robotic coring sampler [top] designed to dig beneath Phobos’s weathered surface.JAXA

The pneumatic sampler takes a different approach. Mounted on one of the lander’s legs, it will inject pure, non-reactive nitrogen gas into the surface and capture the grains stirred up. This sample could illuminate the changes Mars and its moons have undergone over time. The dual system also adds redundancy in the face of Phobos’s unknowns. “If the surface of Phobos proves to be very hard, [the coring sampler] may struggle to penetrate the ground. [The pneumatic sampler] would not experience this issue and still be able to gather material,” Kawakatsu says.

One of the most intriguing aspects of MMX is the additional possibility of collecting material from the planet below. It’s likely that Phobos has been peppered by Martian ejecta thrown into space by meteor impacts, so MMX also provides an opportunity to return the first Martian samples to Earth. Michel estimates that the samples could contain “a few hundred grains coming from Mars.” Scientists can distinguish those grains from native Phobos material by analyzing their oxygen isotopes.

That possibility raises a planetary-protection question. A direct Mars sample-return mission would have to follow strict rules to ensure that potentially living biological material from Mars doesn’t contaminate Earth. But those protocols don’t apply to MMX, even if its sample does collect grains of Martian soil. A panel of experts concluded that any microbes blasted off Mars wouldn’t survive the violent impact and ejection or the intense radiation exposure on the way to Phobos. This saved MMX from implementing expensive and time-consuming quarantine and containment protocols for the returning sample.

After landing on Phobos, MMX will have roughly two hours to collect samples from the moon’s surface before lifting off again.JAXA

During the brief surface operation, the robotic arm will deposit the samples in a storage unit, from which they’ll be transferred into a capsule aboard the return module and sealed in a vacuum. After that, lift off.

Once back in Mars orbit, the exploration module will resume its observations of Phobos and make several flybys of Mars’s other diminutive moon, Deimos, to gather data on its surface, structure, and composition. In November 2030, during a favorable alignment of Mars and Earth, MMX will begin the homeward leg of its trip. The return module will separate from the exploration module and power the voyage home, its sole objective to bring the dust of Phobos to Earth.

MMX’s sample return capsule will carry material from Phobos through Earth’s atmosphere in 2031.JAXA

Finally, in 2031, the 60-centimeter-wide sample capsule will plunge into Earth’s atmosphere at roughly 12 kilometers per second. Assuming it survives the extreme heat of reentry and lands safely in Australia, scientists will begin to tease out clues about Phobos and the early solar system. JAXA’s Institute of Space and Astronautical Science will receive and curate the samples, which will also be shared with mission partners.

“We have macroscale measurements from the spacecraft, microscale measurements with the rover, and then sample analysis,” Michel says. “So we can relate the three kinds of scales together because we have all of them.” The insights gleaned from orbital images, close-up images of the regolith, and the chemical analysis of the samples could have wider implications for our understanding of our solar system.

“MMX is a mission to understand the Mars system and shed light onto how terrestrial planets like the Earth, and maybe Mars, develop a habitable environment,” Kawakatsu says. “We strongly suspect that our terrestrial planet region formed dry, and water and organics were scattered inwards from the icy-rich outer solar system.” Mars, he says, sits at the gateway between the inner and outer solar system, making its moons valuable witnesses to the processes that shaped potentially habitable worlds.

Now in their third year, the IEEE TryEngineering educator sessions continue to empower preuniversity teachers with specialized professional development programs on engineering topics and classroom-tested lessons.

This year’s sessions, held across the United States, have focused on educating teachers about artificial intelligence and semiconductor technology. One session was held in Egypt.

In all, 54 educators learned about the engineering design process this year. They received an overview of TryEngineering resources including the Making of a Microchip and Train Your Tree lesson plans.

Workshops are scheduled this year in Greece and Guatemala.

The TryEngineering program is administered by IEEE Educational Activities.

These teachers and guest speakers attended the two-day session held at the Florida Semiconductor Institute and SkyWater Technology chip manufacturing facility.IEEE Educational Activities

A network of host universities

Central to the educator session program’s success is its university network. Host campuses welcome teachers into their facilities, where faculty provide background information about some of the technologies being taught, as well as possible engineering career paths and job opportunities for students.

This year’s sessions involved the educators in hands-on projects. Facilitators at each location were selected based on their experience in engineering education, and they worked with the TryEngineering staff on their presentations.

GEMS International School

Sixteen teachers from nine schools participated in the session, held in Cairo. It was facilitated by Marwa Soudi, an IEEE member, and Senior Member Stamatis Dragoumanos. Both are past chairs of the IEEE preuniversity coordinating committee. GEMS provided outstanding STEM facilities for teachers, Saudi and Dragoumanos said. Participants received an overview of the Working With Wind Energy and Tall Tower Challenge lesson plans.

Another educator session is scheduled for later this year.

University of Florida

Educators from across the state gathered at the university’s Florida Semiconductor Institute, headquartered in Gainesville, and FSI’s NeoCity high-tech manufacturing campus near Orlando. The session was hosted by Roseanne Perez, an FSI workforce development program specialist, and Sophia Watson, FSI’s assistant director of workforce development. The two-day experience featured an inside look at microelectronics through a guided tour of a SkyWater Technology chip manufacturing facility.

Engineering instructor Mary Acken facilitated hands-on lesson demonstrations for the 15 attending teachers.

Reflecting on the workshop’s impact, one participant said the SkyWater tour was excellent, the instructors were great, and the teachers enjoyed learning from like-minded educators.

Arizona State University

Michael Kozicki, an IEEE member, and Stefan Myhajlenko from the university’s engineering school welcomed 11 Arizona teachers to the Tempe campus.

The session was facilitated by Janet Hartkopf, an Arizona school leader and a former TryEngineering educator session participant. Hartkopf highlighted collaborative learning, the need for diverse perspectives in engineering education, and career paths.

The teachers received instruction on how to implement the hands-on lessons in their classrooms, and they toured the university’s NanoFab facility to observe microelectronics research.

“This training was amazing!” said a participant who leads an elementary after-school STEM club. “I loved the hands-on lessons and e-books, which I will definitely be using with my students. I am excited to explore all the resources on the TryEngineering website. The experts from ASU were great.”

University of Minnesota

Sixteen STEM educators attended this session, held at the university’s International Institute for Biosensing in Minneapolis. Its director, Patricia Khashayar, an IEEE member, introduced participants to the capabilities of the university’s Supercomputing Institute and Nano Center.

Under the guidance of Claire Dietz, an education and outreach coordinator, and Pam Schumacher, an instructor at St. Paul College, the teachers completed classroom-ready lessons including a photonics experiment using the Nano Center’s equipment. Schumacher attended last year’s TryEngineering educator session.

Nano Center Director Sarah Swisher, an IEEE senior member, provided an overview of semiconductor operations and highlighted growing job opportunities for the state’s young people.

The experience inspired all 16 teachers to plan student field trips to the Nano Center during the upcoming academic year, they said.

“Great class!” one participant said. “I’m excited to bring some of the TryEngineering materials back to our school.”

This year’s sessions built on the momentum from last year when, with financial support from Onsemi’s Giving Now program, nearly 80 teachers attended sessions across five U.S. states and Mexico.

Driven by that success and the demand for more, the IEEE Educational Activities Board continued its support. Upcoming global sessions are expected to expand the program’s international footprint.

“I appreciated the time to collaborate with peers on how to implement hands-on lessons,” one participating teacher this year said. Some participants said they planned follow-up visits with the experts and campus tours.

You can learn more about the sessions by emailing TryEngineering.

This article is crossposted from IEEE Spectrum’s careers newsletter. Sign up now to get insider tips, expert advice, and practical strategies, written in partnership with tech career development company Parsity and delivered to your inbox for free!

Hard work rarely speaks for itself.

When I landed my first job as a software engineer, I had a simple plan: Work hard and somebody would notice. Eventually a manager would tap me on the shoulder and hand me a promotion (probably with a nice little speech).

That plan stalled my career for about five years.

I switched jobs twice, kept my head down, read a stack of programming books, and sometimes pulled 12-hour days. I was never promoted. At my third job, out of frustration, I started interviewing at other companies and almost immediately got an offer for a senior engineer role.

So I went back to my job and quit. My team was blindsided. I liked my manager and the people I worked with, and they came back with a counteroffer. Then someone from HR asked me, “Why didn’t you just ask us?”

Honestly, it never occurred to me.

I took the counteroffer and stayed. Three jobs in, I finally got promoted, and all it took was asking (in the most dramatic way possible).

Today, I would take a different approach.

Year-end review discussions are a great time to push for a promotion. But at most companies, raises and promotions are largely settled before the formal review conversation happens. If the year-end review is the first time your manager hears you want a promotion, you’re usually too late.

Now is a great time to start the conversation.

Hard work is invisible by default

Engineers tend to believe good work speaks for itself. It rarely does. Your manager is juggling a team, a roadmap, and their own boss. They see a fraction of what you do, and the work they don’t see might as well not exist when raises get decided.

Perception is reality, so you have to shape it. That means learning to brag a little.

A “brag document” is a running list of what you accomplished, when, and why it mattered. Add to it every two weeks or once a month, and write each entry so your manager’s manager would understand it.

Engineers tend to write things like “migrated the billing service.” A better entry reads “migrated the billing service, which cut failed payments by 30 percent and unblocked the Q3 pricing launch.” If there’s no number, name who benefited and how.

Include the work people forget to count, too: handling incidents, mentoring new hires, interviewing candidates, and writing the docs everyone else relies on. If you helped your team adopt AI tools this year, write that down. (It counts as leadership, whether or not anyone called it that.)

The document does two jobs. It gives you ammunition when you make your case, and it gives your manager something to ctrl + c then ctrl + v when they argue for you in meetings you’ll never attend.

Just ask

Bragging builds the case. Asking forces action.

If your goal is a promotion or a raise, you might ask: “What would it take for me to get to [the next level] in the next year?”

A good manager will give you an outline that can turn into a plan. Push for specifics until you have deliverables you can work toward, then send a short recap email so the plan exists somewhere besides your manager’s memory. (Managers change jobs too.)

When the answer is “no”

Sometimes the answer will still be “not this year,” and the reason may have nothing to do with you.

Budgets get frozen, promotion slots get capped, and decisions get made above your manager’s head where even a manager who wants to help can’t.

Don’t read too much into that. But be honest with yourself, too. Most of the time, a no means there’s something you aren’t doing yet. Find out what it is instead of guessing for five years like I did.

Asking costs you one slightly awkward conversation. Sitting back and waiting to be noticed can cost you years.

—Brian

I Used AI to Fake a Scientific Paper. Here’s What Happened

Don’t try this at home: To better understand the risk AI slop poses to scientific publishing today, IEEE Spectrum contributor Benjamin Skuse put IEEE’s research integrity checks to the test by submitting two entirely AI-generated papers. “Science has a spam problem. And the spam filter is getting clogged up…”

Read more here.

Why Engineers Should Brave Capitol Hill

Policy change doesn’t just come down to paid lobbyists—lawmakers want to hear from working experts, too. One engineer shares what he learned from his first two visits to Capitol Hill.

Read more here.

Community Colleges Put New Workers Into Semiconductor Jobs Fast

Could community colleges help fill the coming talent shortage in the U.S. semiconductor industry? Alternatives to four-year university degrees are a promising option for the technicians who will be needed to run new fabs.

Read more here.

The sun hangs low over the plains of Castile and León, in the north-central region of Spain. But from a slope of the mountain Pico Trigaza, it appears that a few puffy white clouds have started to block the sun. That’s vexing.

A group of researchers, some wearing cardboard eclipse glasses, is clustered around sets of portable telescopes and cameras connected to laptops. All this gear is scattered on tabletops perched on the mountainside. On the computer screens, the sun’s disk looks like a deep orange crescent that is slowly waning. That’s because the moon began to crawl across it about half an hour ago. In 25 minutes, it will obscure the sun completely, producing an eerie and wondrous phenomenon—a total solar eclipse.

In the late-summer, late-afternoon heat, the researchers watch anxiously as those fluffy clouds linger around the sun. From an abandoned hut on the slope a few meters above, where the researchers have set up camp, one of the grad students emerges and nods toward the cloud formation. “Do you think it will dissolve?” he blurts out.

“It will.” Miloslav Druckmüller glances calmly at his group of doctoral candidates, now making final checks of the equipment in preparation for totality. Here in Spain, on this 12th day of August, Druckmüller is getting ready for his 19th solar eclipse. He knows from long experience that weather can turn quickly when the moon’s shadow suddenly pushes the temperature down. Plus, there’s no sense in riling up the grad students any more than they already are.

Miloslav Druckmüller led the design and construction of a unique optical instrument that incorporates eight astrocameras. Each is equipped with a spectral filter for sensing and recording specific emission spectra from the solar corona. Each of those spectra, in turn, is associated with a particular ionic species of iron atoms. Because the temperature needed to ionize iron varies depending on the ionic species, analyzing the locations and relative abundance of these ionic species offers valuable insights into the distribution of heat in the corona.Václav Široký

During totality, the moon blocks the overwhelming glare of the solar disk, revealing delicate structures in the sun’s outer atmosphere, the corona. Using custom-built telescope systems and image-processing techniques that Druckmüller has spent decades developing, the researchers aim to gain insights into the processes in this enigmatic region that is otherwise hard to study. In addition to imaging the visible coronal light, the team will try to track the dispersal of superhot ions of iron into interplanetary space, which could give rare insights into the pattern of heat emanating from the corona.

And they’ll have 107 seconds to do all that.

A Mathematician’s Path to Solar Eclipse Photography

Druckmüller is a professor of mathematics and image processing at the Brno University of Technology in the Czech Republic and a leading figure in the Solar Wind Sherpas, an international team that studies the sun’s corona during total eclipses. In the solar eclipse world, he is nothing short of a legend, albeit a reluctant one.

An image from Miloslav Druckmüller’s first solar-eclipse expedition, to Hungary, vividly shows Baily’s Beads. These elongated white lights along a piece of the left edge of the blacked-out solar disk are caused by beams of sunlight shining through gaps in the irregular lunar surface during the eclipse. The pink shapes are solar prominences, which are jets of plasma extending out into space and held in place by the sun’s intense magnetic field.

Miloslav Druckmüller

He photographed his first solar eclipse in 1999, in what he thought would be a one-off lark. During that eclipse, the moon’s deep shadow missed Druckmüller’s native Czechia by just a few tens of kilometers, so he chased it to Hungary with a 30-year-old Soviet MTO 1,000-millimeter lens. All he remembers from that day is the rush of adrenaline he felt while trying to find the sun in the sky with the lens’s minuscule field of view and keep it in focus for the two minutes of the eclipse.

He insists that he has never enjoyed any of the eclipses he has photographed. “In fact, I have never properly seen any of those eclipses. You are just watching the screens, making sure that it’s all doing what it’s supposed to do. There are so many things that can go wrong and you can’t redo it.”

A composite sequence of photographs shows the 12 August 2026 solar eclipse as seen from a slope of the mountain Pico Trigaza in Spain.

Václav Široký and Miloslav Druckmüller

Despite Druckmüller’s outward tranquility on this Spanish hillside, the stakes are high. Just 45 minutes ago, his group heard from their colleagues in Iceland, where viewing conditions had promised to be better than in Spain. But a vast low-pressure system had clouded the entire sky, leaving all three Iceland teams without a view.

What Solar Eclipses Reveal About the Corona

The corona has been a mystery for centuries. Though dimmer—by a factor of a million—than the sun’s visible surface, the corona is nevertheless extremely hot: about a million degrees Celsius compared to 5,500 degrees on the surface. Because the sun’s nuclear fusion engine churns at the star’s heart, it might seem that temperatures should fall farther from the core.

To observe the faint corona in visible light, researchers must block the sun’s ferocious glare. Specialized telescopes called coronagraphs do this with an opaque disk called an occulter. But because light bends around the occulter’s edges in an effect called diffraction, it must block an area much larger than that of the solar disk to prevent light from spilling over and ruining the images. But doing so means obstructing the inner part of the corona, which harbors the most important processes that underpin the puzzling phenomena of the sun’s upper atmosphere.

That’s why total eclipses matter so much to solar physicists. By sheer chance, the moon is exactly the right size and at just the right distance to completely cover the sun during a total solar eclipse. Because the moon is so far away from Earth, the view is not marred by diffraction, and the corona reveals itself nearly all the way down to the sun’s surface.

Space probes can image the solar surface in other wavelengths, but those views have limitations, too. Spacecraft such as NASA’s Solar Dynamics Observatory do see the innermost parts of the corona in extreme ultraviolet light, explains Shadia Habbal, a solar physicist at the University of Hawaii and leader of the Solar Wind Sherpas project. “But it cuts off very quickly, so you don’t see its extension into space as well as we can see it in the visible emission lines” during an eclipse. “That’s the biggest advantage,” she says, of solar eclipse photography.

The European Proba-3 mission, launched in 2024, simulates total solar eclipses by flying two spacecraft in a formation 150 meters apart with one acting as an occulter and the other as the telescope. The instruments aboard that satellite, however, have limitations compared to the equipment used by the Solar Wind Sherpas team.

How Eclipse Images Sparked a Scientific Partnership

When Druckmüller developed the film from his 1999 eclipse expedition, the photographs didn’t do justice to the event’s eerie beauty. He spent countless evenings applying his expertise in numerical image processing in an attempt to re-create in his photographs the sights he witnessed in that Hungarian corn field.

It took two years, but eventually he succeeded. Where before there were dull views of the blackened lunar disk surrounded by the blunt wisps of the corona, Druckmüller revealed the subtlest details of the magnetic field lines stretching from the sun far into space. As his photographs circulated among solar physicists, invitations to eclipse expeditions began to arrive.

Habbal learned about the images from a Ph.D. student who had seen them at a conference presentation that didn’t credit Druckmüller. When Habbal’s team later published the images in a paper, they misattributed them to another researcher. Druckmüller found out through an astrophysicist friend. “I wrote her a rather disappointed email and she responded within 10 minutes,” Druckmüller recalls. “She really didn’t have a clue.”

The error was corrected, and within weeks, Habbal and Druckmüller began a collaboration that grew into a friendship. “Milos’s processing brings out details that nobody has seen before, and they tell you a lot about very tiny features in the corona that you don’t see otherwise,” says Habbal.

How Image Processing Reveals the Solar Corona

Capturing the corona’s delicate structures means finding a way to show its brightest and faintest features in the same image. “During the solar eclipse, the contrast between the darkest and brightest parts of the corona is enormous, possibly one to a million,” Druckmüller explains. “But a computer monitor only allows you to display a contrast ratio of one to about 800.” (Some modern monitors, particularly OLED ones, can deliver higher contrast ratios.)

The human eye can adjust to this enormous range of brightness, allowing an observer to see both the bright corona and the faint stars and planets that briefly emerge in the darkened sky. To bring those details together in a photograph, Druckmüller uses custom image-processing algorithms to combine exposures that capture different parts of that range.

Researchers Martin Šebek and Peter Urgoš make fine adjustments to the focus and alignment of the astrocameras on a key optical instrument before the 12 August solar eclipse.

Václav Široký

To create a single image, Druckmüller’s teams need to take 200 individual frames during the eclipse at exposure times ranging from 0.001 second to two seconds. The data must then be calibrated against a vast number of reference images to filter out noise and imperfections. Flat fields, which depict a uniform light background, help correct for defects caused by telescope optics, dust spots, and variations in pixel sensitivity. Dark frames, taken with light blocked from the sensor, help correct for artifacts created by the camera’s image sensor itself. “It takes the computer more than two days to process one image,” Druckmüller says. Including the calibration data, a single finished image requires about 3,000 input frames.

Choosing a Location for Solar Eclipse Photography

Since 2008, teams directed by Habbal and Druckmüller have traveled to every total solar eclipse, spreading out across multiple locations to hedge against bad weather. About half of the teams return with usable data. For the others, it’s just a short and stressful voyage. For the 12 August eclipse, three groups deployed to Iceland and two to Spain.

The teams seek sites close to the center of the band of totality, where the moon blocks the sun for longest. In Spain, the 12 August band of totality cut across the north of the country; the center line of the shadow passed within 11 kilometers of Pico Trigaza, promising Druckmüller and his colleagues a precious 107 seconds of totality.

But duration is only part of the calculation. Totality on Pico Trigaza will commence 51 minutes before sunset, with the sun just 8 degrees above the horizon. That low, the rays must pass through a much thicker layer of the atmosphere than when the sun is higher in the sky, increasing the light distortion that complicates photography. At the researchers’ camp on Pico Trigaza, 1,600 meters above sea level, the thinner air should help a bit.

An abandoned hut on the mountain Pico Trigaza in Spain was spruced up so that it could serve as a home base for the research team recording the 12 August solar eclipse.

Martin Šebek

“Iceland would have been a much better place for us than Spain because the sun was higher up, some 25 degrees above the horizon,” Druckmüller says wistfully. But Iceland’s notoriously unpredictable weather made Spain an essential backup.

If you pack a van with thousands of dollars’ worth of cameras and optical gear and then set out on a 2,351-kilometer journey, you shouldn’t be surprised if you finish the trip with a few striking photographs [top]. The sleeping arrangements for the grad students at the Pico Trigaza hut [bottom] were spartan, but the views were first-rate.Top: Martin Šebek; Bottom: Václav Široký

On the mountain, locals have spruced up a rustic, long-unused hut for the researchers, running power cables from a nearby telecommunication outpost and bringing in a water tank. Five Ph.D. students drove more than 2,000 kilometers from Brno and have spent nearly a week preparing the equipment. Druckmüller and his wife joined them two days before the eclipse. They cook on a small propane grill and sleep in sleeping bags on yoga mats. “We have plenty of space here to set up our telescopes, beautiful views, and a roof over our heads,” Druckmüller says with a grin. “It’s perfect.”

But even the best equipment and a carefully chosen site count for nothing if the clouds roll in. “So much work goes into the preparations and then, if you end up with nothing, you just feel terrible sadness,” says Jana Hoderová, an assistant professor of mathematics at Brno University of Technology and one of Druckmüller’s key collaborators.

As the moon continues to slide in front of the solar disk, the pesky cumulus cloud that kept everybody on edge begins to dissolve. It’s 10 minutes to totality. The relief is palpable.

Mapping the Solar Corona’s Temperature With Iron

The countdown on the computers that control the cameras has begun. The screens are now showing only a tiny orange sliver, and the evening is becoming perceptibly less bright.

Twenty astrophotography cameras, most of them drawing power from a car battery that was also buffering the cameras from any grid instabilities, are pointed at the sun. Alongside equipment capturing the corona’s visible light, the team has built an assembly of eight cameras to detect light emitted by various iron ions that arise in the corona.

Among the optical gear assembled at Pico Trigaza to capture the 12 August solar eclipse was a unique instrument designed to record the distributions of iron ions in the corona [at left]. Václav Široký

Those instruments looking for iron ions have been put together with resourcefulness and thrift. A colleague of Druckmüller’s laser-cut the aluminum plates holding the carefully aligned lenses in exchange “for two bottles of whisky.” The brackets and attachments were 3D-printed at Brno University of Technology. The lenses, produced for the defunct camera maker Contax, were bought at steep discounts on eBay. By far the most expensive elements are interference filters that isolate the wavelengths emitted by the iron ions.

Around the blackened solar disk rises the magnificent golden halo of the corona. In the hush on the mountainside, it seems that even the birds have gone still.

“That’s something nobody else in the world can do,” Druckmüller says with visible satisfaction. The unique technology, codeveloped with Habbal’s Hawaii team, has already enabled researchers to map the distribution of temperatures across the entire corona and to discover previously unseen structures, dubbed “smoke rings,” in the coronal plasma. The photographs also revealed the boundary region where the solar plasma becomes so thin that its ions stop colliding.

The optical assembly designed to record the distributions of iron ions in the corona is outfitted with astrocameras, each of which captures spectra from a different ionic species. To do that, each astrocamera is equipped with a spectral filter whose bandwidth is so narrow—one nanometer—that the filter appears as a mirror.Martin Fornůsek

“Iron is the most abundant of the heavy elements that we see in the solar corona,” says Habbal. “And because the corona is so hot, the iron atoms lose electrons and get ionized. The larger the loss, the higher the temperature.”

For example, a temperature of around 1.6 million °C is necessary to produce an iron ion that has lost 12 electrons. But only 1.4 million °C is needed to strip an iron atom of 11 electrons. Once it’s in a particular electrical state, the ion retains its specific ionization, even as it travels outward with the solar wind. Each of those ions, whether +11 or +12, emits a different wavelength of light, allowing the researchers to map not just temperatures within the corona, but also how those iron ions spread into interplanetary space and distribute heat. Habbal and her colleagues had long been trying to image those iron ions during solar eclipses. But it wasn’t until Druckmüller entered the picture that they were able to develop the gear and the mathematical techniques to do it.

Three, Two, One…Totality

With just a minute to go to totality, the researchers begin removing protective filters from the lenses. It’s 50 minutes to sunset, and the waning sun has just begun to light up the clouds above the horizon with the first tinge of orange. Then, suddenly, the evening begins a rapid and premature descent into darkness. Along the edge of the fantastically thin solar crescent, the brilliant orange points known as Baily’s Beads flash and disappear—the last fragments of photospheric light visible through deep lunar valleys. As totality begins, observers remove their glasses. The hot air cools. The graduate students remain quietly and intently focused on their gear.

Druckmüller looks up from his camera toward the darkened plains. Around the blackened solar disk rises the magnificent golden halo of the corona. In the hush on the mountainside, it seems that even the birds have gone still. The sky is ablaze with fiery colors all along the horizon where the deep lunar shadow drops off. Venus burns brilliantly in the southwest, a sharp white pinpoint in the darkened sky. The sound of camera shutters suddenly fills the air. Then, as quickly as it fell, the darkness begins to recede, bid farewell with another handful of golden Baily’s Beads.

The crowd erupts in cheers. The Ph.D. students throw themselves at each other and hug like they mean it.

“The students did most of the work this time, so I actually allowed myself to really look at it,” Druckmüller says. “It was beautiful. Pico Trigaza may not have been the best for science photography, but from the aesthetic point of view, it was extraordinary.”

The 2026 solar-eclipse team at Pico Trigaza in Spain included [from left] Magdalena Šrotířová, Peter Urgoš, Martin Fornůsek, Miloslav Druckmüller, Václav Široký, and Martin Šebek.Václav Široký

But for Druckmüller, this is just the beginning. While other eclipse chasers will race to release their images, he will take his time. He expects to publish the first visible-light images of the corona in about two weeks. Processing the iron-ion images, and uncovering whatever insights they may hold, will take months—maybe even a whole year.

“Probably until the next eclipse in 2027” in North Africa, Druckmüller says.

Who knows? Maybe he’ll even enjoy that one, too.

To help engineers keep up with emerging technologies such as AI, cloud computing, and sustainable infrastructure, and learn how they can be applied in practice, the IEEE Computer Society’s Central Indiana Chapter recently launched its Tech Talk Series.

The live webinar series connects IEEE members, undergrad and graduate students, and engineering professionals from IEEE Region 4 with industry practitioners. The monthly sessions have covered AI governance and security, agentic systems, and related topics. The sessions featured speakers from Apple, software company Tavant, and cybersecurity concern Zscaler.

Upcoming sessions and registration details are available on the IEEE Region 4 virtual events calendar.

Technical discussions grounded in engineering

Rather than presenting on emerging technologies only at a conceptual level, speakers discussed the design decisions behind specific tools and systems, how the systems are structured, and implementation challenges. They also shared lessons learned from using the tools in their daily work.

Sessions also include technical walkthroughs and practical examples.

The live format allows attendees to discuss and compare their approaches when evaluating and implementing new technologies with those used by other organizations.

Extending access through a virtual format

The organizers chose to make the series virtual so that chapter members and other attendees could participate in technical discussions with engineers from major technology companies and eliminate the need to travel.

For young professionals, the sessions connect what they’ve learned in a classroom with the realities of the workplace.

Technical programs such as the Tech Talk Series give professionals and students access to expertise, help create connections across career levels, and make important engineering discussions available to a broader community.

Attendees learn how practitioners approach projects such as system architecture, reliability, security, governance, and product delivery. For experienced professionals, the talks offer a forum to examine new ways of designing, securing, and operating software systems, and compare technical approaches. They can ask questions relevant to systems being evaluated or deployed within their organization.

Examining governed AI agents

The most recent webinar, Feeding and Fencing AI Agents: Engineering Governed Agentic Systems on Real-Time Data, featured Gajendra Babu Thokala, a senior engineering leader at Apple in Seattle.

The IEEE senior member explained how teams can move beyond adding isolated AI features to applications and instead build governed AI-agent systems that use real-time business data to complete multistep business workflows.

Another recent session, Building AI Native Products: A Playbook for Agentic Systems, explored how products built around AI agents differ from traditional software. It also covered how agentic systems can help users complete multistep tasks and workflows. IEEE Senior Member Sandeep Shivam, a product and technology leader at Tavant in Dallas, led the webinar.

Those two sessions illustrate the series’ purpose: helping participants understand what a technology can do and what is required to use it responsibly.

Building a continuing technical program

Future sessions are being planned on areas including enterprise AI, software reliability, and responsible AI usage. In addition to speakers from technology companies, the chapter also plans to invite speakers from startups, universities, and research organizations.

Upcoming sessions are expected to cover cloud platforms, enterprise integrations, and the responsible development of AI systems.

Technical programs such as the Tech Talk Series give professionals and students access to expertise, help create connections across career levels, and make important engineering discussions available to a broader community.

Undergraduate and graduate students are invited to participate to gain exposure to technical decision-making and challenges encountered in professional engineering environments.

A recent scroll through social media got me thinking about trust. About half my feed was filled with AI slop that astute users would call out as being machine-generated, often to the embarrassment of the rube who unknowingly spread it and sometimes to the delight of the poster who shared it deliberately because they found it amusing or useful for sowing confusion.

Remember the good old days of GPT-3 when we could ask the chatbot anything and it would almost instantly and very confidently spit back a plausible but false answer? That’s when the word “hallucination” as applied to AI answers entered the lexicon. Six years and several versions of a few different chatbots later, I still can’t trust the answers I get, even though Claude or Gemini often do useful work for me. That’s the thing about current AI: Trust but verify. And verify you must.

Apply that maxim at scale to maintain the integrity of the scientific record and you get an idea of the challenge facing publishers of peer-reviewed scholarship. In “I Used AI to Fake a Scientific Paper,” freelance writer and doctor of applied mathematics Benjamin Skuse tells the story of the human gatekeepers trying to beat back the rising tide of AI slop by attempting to get his own AI-generated papers past IEEE’s defenses.

“He basically hit ‘run experiment’ and ‘write paper.’ And then he didn’t do a damn thing to it.” —Eliza Strickland

As features and AI beat editor Eliza Strickland told me, Skuse was surprised at just how effortless it was to generate a paper that was more or less plausible and more or less science. For one of his experiments, “he basically hit ‘run experiment’ and ‘write paper,’” she said. “And then he didn’t do a damn thing to it. He just launched it into the maw of the publishing engine.”

Skuse discovered that the combination of IEEE’s human-review processes and research-integrity checker programs are currently keeping pace in what amounts to a cat-and-mouse game against people using free tools. But the more advanced LLMs are no longer making the same dumb mistakes the free versions do. That leaves publishers locked in an arms race between increasingly plausible slop and the systems aimed at detecting it.

Scholarly publishers aren’t the only ones trying to hold the line against slop. So is that global depository of human-generated knowledge, Wikipedia. For this month’s Five Questions, Strickland interviewed Ellia Lebleu, a founding member of WikiProject AI Cleanup, a volunteer effort to detect and delete AI-generated material from the platform. Lebleu’s team is confronting the same problem as commercial publishers, but with fewer people and fewer tools.

“Wikipedia is known as a repository of human-generated information, and they want to hold on to that,” says Strickland. “Who knows how long they’ll be able to do that.”

If we can’t collectively assume that the scientific record is trustworthy or that basic facts about the world are accurate, then our understanding of what is real unspools into uncertainty. Perhaps that delights quantum scientists and the grifters who thrive in an information environment where there is no shared reality. But that uncertainty ultimately undermines trust in the institutions that exist to curate the corpus of human knowledge and to promote the advancement of science and technology for the benefit of humanity.

Or maybe in a world where AI is doing the science and writing the papers, it won’t matter whether or not humans trust the outputs. As Strickland put it, “Doesn’t it just feel crazy that we’ll have AI writing for other AIs at some point? We’ll just get left out of the loop.”

At the turn of the last century, the automobile industry was still new and evolving rapidly. In 1898, there were about 50 carmakers in the United States. A decade later, that number had increased fivefold. In 1901, the companies produced 7,000 vehicles. By 1910, production had reached 181,000.

For several decades, electric cars vied with those powered by steam and internal combustion, and nobody was sure which technology would come out on top. Steam had the legacy of locomotive technology behind it but came with a risk of fire. Steam-powered vehicles could also become exceedingly hot, depending on the placement of the boiler and the steam piping. Both gasoline and electric cars faced logistical challenges: How far could you travel without refueling or charging? You could always throw an extra can of gasoline in the back of the car, but charging a battery involved more planning. The infrastructure for long-distance travel was not yet in place.

Before the invention of the electric starter in 1911, gasoline-powered engines had to be started with a hand crank; once running, they were loud and emitted noxious fumes. Fully electric vehicles, by contrast, were easy to start; you just turned on the main battery switch, often with a key, and there was no foul exhaust to deal with.

Electric carmakers looking for a way to differentiate their products decided that these attributes made their vehicles perfect for women. (Electric cars were also marketed to doctors so that they could make house calls in the middle of the night without waking the neighbors.) General Electric, which made car chargers based on mercury-arc rectifiers, was so captivated by the idea of women driving electric cars that the company created a series of advertisements featuring stylishly dressed women plugging in their cars and posing with their charging stations.

In ads for its electric car chargers, General Electric featured women using the equipment.miSci—Museum of Innovation and Science

A young woman named Fannie Goodnow was an early EV adopter. She owned a 1908 Bailey Victoria Phaeton, an open two-seater that really did look like a horseless carriage, and she used a GE mercury-arc rectifier to charge it.

Who Invented the Mercury-Arc Rectifier?

Electrical engineer Peter Cooper Hewitt invented the mercury vapor lamp, which became the basis for the mercury-arc rectifier used in car-charging systems. Bettmann/Getty Images

Electrical engineer Peter Cooper Hewitt is generally credited with the invention of the mercury-arc rectifier. During the first decade of the 20th century, Hewitt received several U.S. patents for electric lamps and other apparatus for transmitting and using electric currents. U.S. patents 968,896 and 1,163,706, granted to Percy H. Thomas, advanced Hewitt’s designs. Thomas was chief electrician for the Cooper Hewitt Electric Co., which Hewitt had cofounded with George Westinghouse to manufacture and sell his lamps.

The mercury-arc rectifier functioned as a diode. In a car charger like the one Fannie Goodnow used, it converted alternating current from the power grid into direct current needed to charge the car’s battery. A small amount of liquid mercury contained in a glass tube would vaporize and glow as electricity passed through it. Coupled with the carbon anodes, the vapor allowed current to flow in only one direction. This unit took about 12 hours to charge a drained battery.

A Carriage Maker Pivots to Electric Cars

I couldn’t find any records as to why Fannie Goodnow opted for an electric car. But as a well-to-do woman living in the Boston area, she had more than a half dozen automobile producers to choose from. The one she chose was S.R. Bailey & Co. of Amesbury, Mass.

This early electric car charger used a mercury-arc rectifier to convert AC to DC.Heritage Museums & Gardens

According to Galen Handy, who maintains the Early Electric Vehicles website, Samuel Robinson Bailey started in the carriage and sleigh business as a teenager, winning a best-in-show prize at a carriage exhibition in Portland, Maine, in 1855. Bailey developed a reputation as an inventor and designed processes to streamline the manufacture of his vehicles. He went into business for himself in 1882. Just over a decade later, one of Bailey’s carriages won a gold medal at the 1893 World’s Columbian Exposition in Chicago.

Bailey’s son Edwin attended the exposition, where he likely would have seen a vision of the future: the Morrison electric wagon. This six-passenger, battery-powered vehicle ferried visitors and VIPs around the fairgrounds. It wasn’t long before automobiles began eating into the carriage business. In December 1906, with the carriage industry collapsing, S.R. Bailey & Co. announced it was getting into electric vehicles.

Thomas Edison Designed a Bad Battery

Bailey’s first model, the Victoria Phaeton, used an Edison alkaline battery. Thomas Edison had been working on a lighter, more reliable, and more powerful battery for the auto industry since the 1890s, and announced the achievement with great fanfare in 1903. Unfortunately for Bailey and other carmakers who deployed it, the battery leaked and wouldn’t hold a charge. Edison was forced to go back to the drawing board. By 1910, he had a better battery in production, but by then his friend Henry Ford had introduced the Model T, setting the stage for the internal combustion engine to take over the market.

While Edison toiled to improve his tech, Bailey pivoted to a battery from the Gould Storage Battery Co. of Depew, N.Y. Gould also supplied batteries for railroads, utilities, subways, elevators, and farms. Goodnow’s Victoria Phaeton used a 30-cell lead-acid battery housed in a wooden box that was suspended under the chassis. The standard battery had a range of 64 to 80 kilometers; a larger battery with 50 percent more capacity was available for an upcharge. With the standard battery, the Victoria Phaeton cost US $2,000 (about $73,000 today, or about the list price of an upscale EV like the Audi Q8 or BMW i5).

Fannie Goodnow’s electric car charger now resides at the Heritage Museums & Gardens in Sandwich, Mass. But her 1908 Bailey Victoria Phaeton has been lost to history. Shown here is a 1911 version of that model.Sarah Thornington/Heritage Museums & Gardens

Goodnow installed a mercury-arc rectifier charging station in the barn adjacent to her parents’ summer home in Hull, Mass. Bill Lillie of Vintage Electric Cars notes that the charging plug was of an unusual design. This is, of course, one of the banes of any early adopter: nonstandardized plugs and cables. The Electric Vehicle Association of America, which formed in 1910, tried to put forward a standard so that electric vehicles could be charged at any charging station. (Modern EV owners can no doubt relate.) But savvy electric car owners knew to carry spare plug adapters and cables when traveling.

When Edison’s improved battery was finally available, Bailey began using it in the Victoria Phaeton and then went on a publicity blitz. In September 1910, the carmaker sponsored a round-trip tour from New York to Boston, with a side trip up New Hampshire’s Mount Washington, totaling about 1,600 km. Sadly, the “This Car Climbed Mt. Washington” bumper sticker hadn’t yet become popular.

Thomas Edison supplied the batteries for an electric car made by S.R. Bailey & Co., which then went on a publicity road trip from New York to Boston. miSci—Museum of Innovation and Science

Although Thomas Edison posed for a photo with the new Bailey Victoria Phaeton that featured his new battery, he himself chose a Detroit Electric Victoria for his own use. (Edison was known to be a terrible driver, so someone else likely chauffeured him around.)

Unable to compete against the rising tide of Model Ts, S.R. Bailey sold its factory in 1915, having produced about 500 to 600 electric vehicles.

Although the vehicle Fannie Goodnow chose turned out to be just a blip in automotive history, I like to imagine her driving about town in a beautiful car that embodied all the quirks of early adoption. I hope her friends didn’t make fun of her for investing in electric just before internal combustion took off. I hope she didn’t get stranded without the proper charging cable. I do not know what happened to her or her Victoria Phaeton, but her charging station remained in situ until 2022, when it was donated to the Heritage Museums & Gardens in Sandwich, Mass. Now it stands ready to tell stories of technological change and adoption that continue to resonate today.

Part of a continuing series looking at historical artifacts that embrace the boundless potential of technology.

An abridged version of this article appears in the November 2026 print issue.

References

For a detailed history of the development of the mercury-arc rectifier (or valve), see Deepak Tiku’s article in the March/April 2014 issue of IEEE Power & Energy Magazine.

Harry Cooke Cushing Jr., a fellow of the American Institute of Electrical Engineers, wrote (and updated regularly over a number of years) The Electric Vehicle Hand-Book, which gave general instructions for the care and storage of batteries, including how to charge them with a mercury-arc rectifier. Discussion of the mercury-arc rectifier begins on page 156 of the 10th edition, published in 1922. (The American Institute of Electrical Engineers is one of the predecessor organizations of the IEEE.)

Advances in sensing, manufacturing, computational resources, and data-driven methods are enabling new approaches to electromagnetic sensing, non-destructive evaluation, and advanced materials processing. This webinar will demonstrate how the COMSOL Multiphysics software supports predictive modeling by coupling electromagnetic fields with thermal, chemical, mechanical, and transport phenomena across multiple length and time scales.

The first application we will show focuses on eddy-current thermometry for reconstructing three-dimensional temperature fields in metal plates. Temperature-dependent electrical conductivity alters induced currents and magnetic fields, enabling volumetric thermal inference. COMSOL simulations of defects and laser-induced transient heating will support validation of reconstruction algorithms. Multifrequency excitation will be investigated to improve depth discrimination through frequency-dependent penetration depths and sensitivities.

The second application examines microwave-enabled volumetric additive manufacturing of ceramic components. A high-power microwave beamforming array delivers localized energy into three-dimensional volumes, enabling processing of opaque, particle-laden resins. Coupled COMSOL models of electromagnetic propagation, microwave absorption, heat transfer, and ceramic sintering kinetics are used to predict temperature evolution, reaction progression, spatial resolution, and processing scalability. Future extensions will incorporate pulsed excitation, fluid flow, resin viscosity, convection, shrinkage, and solid mechanics.

Together, these examples demonstrate how the COMSOL software can integrate multiphysics modeling with sensing, manufacturing, and data inversion. This webinar will present two applications of coupled electromagnetic and thermal modeling in COMSOL Multiphysics: eddy-current thermometry in metal plates and microwave-assisted processing of ceramic-loaded resins. By connecting simulation with practical engineering systems, COMSOL-based modeling can accelerate technology development, reduce experimental iteration, and support reliable scale-up of next-generation electromagnetic technologies.

Register now for this free webinar!

This article is brought to you by Qualcomm Technologies.

For many consumers, much of the promise and appeal of AI is personal. Consumers are looking past generic tools and want to know how these technologies will help them in their daily lives. These trends coincide with the advent of new classes of wearable devices that, in addition to being able to leverage powerful frontier models, feature built-in intelligence that understand the context of a user. Where are they? What are they seeing and hearing? What are their habits, their preferences, their schedules?

It’s an astonishing pivot, from passive accessories that served mostly to send users smartphone notifications to proactive, intelligent companions. The smartphone is no longer the center of that universe; though it’s still incredibly important, in this personal AI paradigm, the AI is the UI. And it shifts the focus toward you, the user.

These new technologies are exciting. But what do consumers really want?

A Personal AI report, commissioned by Qualcomm Technologies, built on a survey of more than 12,000 people across key geographies worldwide (the U.S., U.K., Germany, India, Brazil, and China) focused on the rise of personal AI in intelligent wearables looks deeply at these developments.

The smartphone is no longer the center of that universe; though it’s still incredibly important, in this personal AI paradigm, the AI is the UI.

Survey respondents are less interested in single devices and are responding more to the idea of connected ecosystems. As people expand their roster of devices, it only makes sense. Instead of relying entirely on a smartphone and a laptop, respondents also have connected technology in their cars, as well as intelligent wearables like smartwatches, smart glasses, hearables, and other wearables like rings, pins, and pendants. They want those devices to work in concert, and the way to do that is to decentralize any single one of them. Instead, the AI agent that knows the user and works across all of those devices becomes the unifying factor.

Consider a typical workday: A user’s watch detects it is time to leave for a meeting and, with permission, their AI agent factors in calendar details, real-time traffic and battery status across their devices. In the car, it suggests the best route and briefs them on their meeting agenda. As they arrive, smart glasses surface a discreet reminder, such as the meeting location or the name of a new contact, while the phone stays in their pocket. No one device has to do everything; each plays the role it is best suited for, coordinated by a personal AI agent that understands the user’s context and acts only at the user’s direction.

With this in mind, simply adding more hardware to the mix isn’t sufficient. Even though respondents will be open to new form factors, they’ll demand more from their ecosystem of devices. They’ll be sensitive to points of friction and needless complexity, and they’ll require each new device to bring something new to the table — “incremental intelligence, convenience, or contextual awareness that their existing devices cannot deliver on their own what the report,” reads the report.

The usage trends are up. The report found that those who own wearables are already using AI frequently (60 percent use it at least weekly), which indicates that it’s becoming less of a fun experiment for early adopters and more of a critical mass habit.

The “always-on” nature of intelligent wearables appears to be winning respondents, too. The report found the following statistics on average continuous wear among users:

• True wireless earbuds – 3 hours
• Wireless headphones – 3 hours
• Smart glasses – 5 hours
• Audio glasses – 6 hours
• Smartwatches – 9 hours
• Smart rings, necklaces and pendants – 8 hours

Consumers want AI that is personal and understands their context — they’re looking for smooth, intuitive experiences that are marked by privacy, security, and user control.

Consumer Trust Is Hard-Won

But for all the excitement and engagement around intelligent wearables, there’s a gating factor: trust. As use cases for intelligent wearables expands to more personal and persistent experiences, trust is a necessity for Personal AI.

The top reported issues around adoption are privacy, security, and user control. The data points to a certain amount of caution around deeply personal use cases, like voice, health, and location, as almost 60 percent of respondents want to keep data from their wearables private, or at least insist on gaining approval before it gets shared.

In other words, even though premium features like high-quality audio, perception sensing, battery life, and comfort whet consumer appetites (and agentic AI capabilities drive upgrade intent), true adoption will come only with consumer trust. They want security, privacy, transparency and control.

Notably, the report found that respondents expressed more interest in adopting intelligent wearables when Snapdragon was introduced as the enabling platform. This reinforces the value of having an established, trusted technology foundation in the mix.

The Industry Remains Fluid

The report shows that consumer appetite for AI is high, with 60-70 percent of those surveyed expressing interest in advanced AI as they consider a new device. That’s a constant. But remaining purchase drivers differed between device categories.

Audio playback quality is the most influential purchase driver for true wireless earbuds and headphones. For smartwatches, it’s display quality and responsiveness. Respondents are less concerned about technical specifications than aesthetics like materials and finish when it comes to devices like smart rings and pendants. And for smart glasses, it’s all about the type of control, whether it’s voice, gesture, or touch.

Alongside these purchase drivers is another key finding: Consumer appetite for Personal AI agents has gone mainstream. The report says that 70 percent of respondents “express strong interest in an AI assistant that can interact naturally through voice and seamlessly manage devices, apps, and everyday tasks.” With each individual as the center of their own ecosystem, voice becomes the most critical, natural method for interacting with agents.

The Opportunity

The industry for intelligent wearables is exciting. It’s clear that consumer demand for Personal AI is high, and this report dives deeply into the specifics of what respondents want out of it. They want AI that is personal and understands their context, and they want it to be proactive and helpful instead of merely reactive. They’re looking for smooth, intuitive experiences that are marked by privacy, security, and user control.

Intelligent wearables comprise much of the hardware of this “ecosystem of you.” Consumers are interested in ecosystems instead of single devices, and they’re looking for agentic AI capabilities that understand and act on their behalf.

Yet so much of the industry remains undefined. OEMs and technology makers have an enormous opportunity right now to innovate and iterate on the strong technological foundations that platform makers are building. And the roadmap for what consumers want is clear.

All percentages, consumer behaviors, and attitudes stated in this report are based only on the responses of the individuals who participated in our survey.

Modern engineering projects increasingly combine human judgment with machine capability. An engineer can define a problem, then ask an artificial intelligence tool to search relevant research and technical standards, generate design options, test them in simulation, and compare the results. The engineer can then bring the results to specialists who understand the parts AI cannot judge on its own, such as whether a design can be built safely, whether it makes sense in a particular setting, or how it might affect its users. The final decision about whether and how to move forward remains a human responsibility.

AI can help engineers do early project steps that once required several people with different technical skills. The engineer’s role is therefore expanding from performing each technical step to deciding how people, machines, and specialists’ knowledge should work together on the project.

The problem, not the title

Engineering has always been divided into disciplines, and for good reasons. Electrical, mechanical, civil, software, and other engineering fields require deep knowledge that takes years to develop. AI does not replace that need. It simply changes how easily engineers can explore the parts of a problem that sit outside their own specialty.

Consider a software engineer working on a medical device. AI can help the engineer review relevant biomedical research, understand unfamiliar terms, analyze data, and build an early model.

That does not make the engineer a physician or biomedical expert. It allows the engineer to ask better questions before working with the experts.

The same is true in energy, transportation, manufacturing, and many other fields. A problem might begin as a software issue and then quickly involve hardware, safety measures, economics, human behavior, and regulations.

Engineers do not need to master every field; they need to recognize which expertise the problem requires and bring the right people and tools together.

The importance of systems thinking

Systems thinking thus becomes especially important. It requires looking at the whole problem: understanding how its parts affect one another, identifying constraints and failure points, and making trade-offs.

AI adds a new element to the process. Engineers must decide not only how the technical parts fit together but also what work to give to AI, what a person must check, when a specialist should make a decision, and when the problem needs to be redefined.

An engineer’s value is not in knowing everything. It is in knowing how to frame the problem, how to ask the right questions, how to test the answers, and when another expert must lead the project.

This aspect is crucial because technical challenges in engineering are seldom static. As new insights emerge, objectives might shift, unseen limitations can surface, and the initial premise might prove flawed. Rather than rigidly following the starting specifications, an effective engineer adapts and refines the scope as a deeper understanding of the system develops.

Thus, engineering extends beyond mastering technical concepts; it is an adaptable, structured approach to navigating evolving challenges.

New evidence can change the goal, reveal a hidden constraint, or show that the original hypothesis was wrong. Effective engineers go beyond merely acting on the initial problem statement. They update it as the system becomes better understood.

In that sense, engineering is not only a body of technical knowledge, it is also a disciplined way of thinking about a project.

Human judgment is important

AI can produce convincing answers even when its assumptions are weak. It can generate code that works in one test but fails in another. It can summarize research without understanding which evidence is strongest, or optimize a design for the wrong goal. Such limits make human expertise more important, not less.

Machines are useful where speed, scale, search, pattern recognition, and data generation matter. People are essential where context, values, responsibility, and judgment matter. The challenge is to decide where each belongs.

An AI system might generate several designs for a medical sensor, for example, and predict how each will perform. An engineer can test the predictions and reject designs that violate technical requirements. A clinician can determine whether the sensor can be realistically used. Patients and practitioners might reveal problems that neither the engineer nor the developer considered. The final solution emerges from combining the perspectives.

The engineer is still a builder and specialist but increasingly also a coordinator of knowledge.

An engineer’s value is not in knowing everything. It is in knowing how to frame the problem, how to ask the right questions, how to test the answers, and when another expert must lead the project.

IEEE’s vital role

The broader form of engineering creates an important role for IEEE. Its societies already bring together expertise in various technologies, engineering education, technical standards, and other fields. That makes IEEE well placed to help engineers work across specialties without weakening the value of specialization.

IEEE could create more communities organized around problems that require expertise from several fields. Examples include trustworthy clinical AI, resilient infrastructure, energy reliability, scientific discovery, and safe autonomous systems. An engineer could begin in a home discipline and then work with specialists from other IEEE communities as a project requires.

IEEE also can help set practical standards for human-machine collaboration. Engineers will need clear guidance on when AI-generated work must be checked, how uncertainty should be documented, when a specialist must make a decision, and who is responsible when a system uses both human judgment and machine output.

Those questions are becoming part of everyday engineering. They should not be left to each company or project manager to answer independently. A professional organization can provide common expectations that follow engineers as they move between jobs, industries, and technical fields.

A larger definition of the engineer

AI does not make engineering disciplines less important; it makes the connections between them easier to use. Deep expertise gives engineers the foundation to judge what is sound. Systems thinking helps them understand how different parts of a problem fit together. AI can extend engineers’ ability to explore, test, and communicate across fields. Human judgment gives the work direction and responsibility.

An engineer is not defined only by a job title or a fixed set of tasks. An engineer is someone trained to understand complex systems, define and redefine problems, bring together the knowledge and tools needed to solve them, test whether solutions work, and take responsibility for the outcome.

That is what it means to be an “unbound” engineer.

The boundaries between specialties become easier to cross, and machines might take on more of the technical work, but the core of engineering remains human: deciding what problem is worth solving, what evidence is good enough, how the pieces fit together, and what to do when the answer is uncertain.

In the cockpit of Beta Technologies’ Alia CX300, pilot Chris Caputo steers onto Runway 33 at Patrick Leahy Burlington International Airport in Vermont. Cleared for takeoff, he releases the brakes and nudges the throttle forward.

At the back of the aircraft, the five-blade pusher propeller snaps awake, and the Alia bolts forward. There’s no shriek of a turbofan spooling up, no shudder of a piston engine, and no wait for thrust. Just the low hum of an electric motor and a shove of instant acceleration.

Within a few seconds, we’re at 85 knots (almost 160 kilometers per hour), and Caputo rotates the nose upward. The runway drops abruptly away as we climb. Gaining speed and altitude, we head northwest over Lake Champlain. Through the wraparound canopy, I watch the hunched shoulders of Vermont’s Green Mountains shrink under the right wing, with New York’s taller Adirondacks to the left.

Caputo, a former Air Force fighter pilot, flicks the control stick hard left, then right. Despite its 15-meter wingspan, the Alia feels snappy, its fly-by-wire system automatically coordinating aileron and rudder inputs. The machine, he says, is much simpler to manage than the Cessna 182 I flew around Australia in my student-pilot days.

What’s more, our half-hour flight in the Alia will consume only about US $9 worth of electricity. The same flight in a comparable Cessna 208 Caravan would burn roughly 25 to 30 gallons of Jet A fuel, costing about $210 to $255 at press time. “Even at full throttle, we’re not burning any fuel up here,” says Caputo, as he takes the Alia into a sweeping left turn and heads south over the lake. “We’re not polluting the environment. [There’s] no noise pollution.”

At Beta’s South Burlington, Vt., facility, workers assemble the body of an Alia A250. The company says it has an order and option-to-buy backlog of nearly 1,000 aircraft.Beta Technologies

This ride is happening during an unusually lively time for electric aviation. The promise of lower operating costs and cleaner flight has pushed electric aircraft from a fringe experiment to one of the aviation industry’s most closely watched frontiers. After years of renderings, prototypes, and overly optimistic timetables, companies are now flying full-size aircraft, submitting hardware to regulators for certification, building factories, and testing routes with prospective customers.

Among the top three U.S. developers of electric planes, Archer Aviation and Joby Aviation are concentrating on piloted vertical-takeoff-and-landing air taxis, otherwise known as eVTOLs. Beta Technologies is taking a broader—and in some ways more innovative—path into the market. Rather than betting first on air taxis, Beta is looking to enter electric aviation through short-haul freight planes that use conventional runways, fed by an open charging network that’s designed to make electric aircraft—both Beta’s and its competitors’—part of a working transportation system.

To that end, the company is developing two variants of its Alia airframe: the runway-based CX300 I’m flying in and the A250, which adds four lift propellers for vertical takeoff and landing. Beta says its test pilots have flown the Alia family more than 300,000 kilometers, including doing multiple U.S. crossings and overseas demonstrations. The company claims an order and option-to-buy backlog of 991 aircraft.

RELATED: From Hockey Enforcer to High-Flying eVTOL CEO

“We’re taking a pragmatic stepwise approach for both certification and market entry,” says Beta’s CEO and founder Kyle Clark. “We’re ready to go today with medical and cargo aircraft that fly from airport to airport.” Next up, he says, will be passenger aircraft with conventional takeoff and landing, followed by cargo eVTOLs, and finally passenger eVTOLs.

Reducing the Risk of FAA Certification

Beta’s 17,500-square-meter production facility sits beside the Burlington airport and looks more like a giant sunlit workshop than a conventional airplane factory. Geothermal heat pumps help heat and cool the structure. The roof will eventually hold a megawatt of solar panels. On the floor, technicians make composite structures and join aircraft sections beneath American and Pride flags.

Kyle Clark, a native of Vermont, founded Beta in 2017 to realize his long-held dream of building a hybrid-electric aircraft. (For his bachelor’s thesis at Harvard University, he designed a flight-control system for such an aircraft.) Martine Rothblatt, CEO of the biotech company United Therapeutics and cofounder of SiriusXM satellite radio, invested $48 million in Beta to fund its first prototype, with the eventual goal of using the aircraft to transport organs for transplantation.

The company has kept much of its design and manufacturing in-house, partly because the pieces are closely coupled and partly because suitable aviation components didn’t exist when development began. Vertical integration gives Beta’s engineers control over the interfaces and any failure data. But it also means the company must be competent at several businesses simultaneously: airframes, electric motors, inverters, batteries, software, chargers, and aircraft production.

The production line can build both the fixed-wing CX300 and the vertical-lift A250, which share the same fuselage, wings, pusher motor (for horizontal propulsion), batteries, avionics, and much of the tooling. This results in some one-size-fits-all decisions: The CX300’s twin booms, which connect the wing to the tail surfaces, look superfluous until you see the A250’s four lift propellers mounted on similar structures.

On the CX300, those booms add weight and drag even though they serve no lift function. But the trade-off—a simpler overall design—is part of Beta’s central strategy: Rather than applying for FAA certification for the more novel and complex eVTOL A250, with its multiple lift systems and transitions between hover and wing-borne flight, the company is bringing the runway version forward first. It’s certifying the motor, batteries, fly-by-wire controls, and much of the airframe in the simpler aircraft; it will then reuse evidence and manufacturing experience when it certifies the A250.

The other big reason to start with the CX300 is logistical: The United States has more than 5,000 public-use airports, so the CX300 can operate from infrastructure that already exists. The cargo and medical flights that the aircraft will initially tackle offer tightly defined missions in which lower energy and maintenance costs matter more than cabin amenities. Hence the company’s decision to leave cabin air-conditioning off the CX300, at least for now. “Things that don’t exist on an airplane don’t weigh anything,” Clark says. “They can’t break. You don’t have to service them…. You don’t have to certify them.”

“This strategy, at the beginning, was seen as a failure of their VTOL,” says analyst Sergio Cecutta, a managing partner at SMG Consulting. “But I think a lot of people were very myopic. It was a way for [Beta] to break up the certification in two pieces, and so they have retired a lot of the risk. [A majority] of the [CX300’s] parts are shared with the VTOL, and that means that a lot of the documentation for certification can be easily reused on the VTOL side when that certification starts.”

How Redundancy Enhances Safety

A decade and a half after electric cars began making significant inroads into the automotive market, electric airplane technology is advancing rapidly. In some ways, an electric car has more complicated power-engineering requirements than an electric aircraft. An electric car’s power demands change constantly, while an electric aircraft operates in a narrower set of power regimes: takeoff, level flight, and descent.

But if your car’s drivetrain fails, you can just pull over to the side of the road. That’s not an option for a pilot 1,000 meters up. The hurdle isn’t that electric motors are inherently fragile; it’s that aviation safety standards are extraordinarily demanding. Catastrophic failure conditions generally must be held to roughly one chance in a billion flight hours or less.

Because individual electrical components cannot meet that standard on their own, electric aircraft rely on layered redundancy: duplicate power paths, independent inverters, fault isolation, monitoring, and backup capacity. Each layer helps close the safety gap, but each also adds weight and complexity.

To close the reliability gap and satisfy regulators who have little experience with electric aircraft, Beta engineers have built redundancy into the motor itself. When I visit the company’s motor shop, technicians are assembling the H500A pusher motor, which splits up critical components like the inverters and stator sections.

“You could lose an entire half of the system, and it would still function the same,” says Adam Lowe, Beta’s flight operations lead.

The H500A’s broad diameter gives the magnetic forces greater leverage, producing high torque at low rotational speed, a useful trait when turning large propellers. The H500A weighs about 75 kilograms and produces about 640 kilowatts of power. The motor achieves 97.5 percent efficiency, based on internal testing; typical conventional turboprops have an efficiency range of 25 to 35 percent.

Beta had aimed to have its motor certified in the first half of 2026, a deadline that passed without a public announcement. Although the company’s certification strategy is aimed at streamlining the process, getting clearance to fly from the Federal Aviation Administration is probably still its biggest obstacle. For Beta, and any company trying to bring an electric aircraft into service, the FAA must sign off first on the propeller, then the motor, then the aircraft. After that, the manufacturer must pursue a production certificate from the FAA, proving that the company can build the same approved aircraft again and again, with the right tools, records, inspections, quality controls, supplier oversight, training, calibration, and parts tracking.

Beta has found a narrow on-ramp to get the A250 flying before full certification. Under the FAA’s eVTOL Integration Pilot Program, companies and public agencies can run FAA-supervised, real-world trials to test how advanced air-mobility aircraft might operate in the national airspace. For Beta, that included July flights of the CX300 across a Virginia-to-Maryland corridor carrying, among other things, a manufactured transplantable organ in a medical containment system. Beta has also been chosen for six other FAA-backed trials.

For the fixed-wing Alia CX300, customer trials are moving in parallel. In March, Beta’s CX300 successfully completed 23 flights over 10 days on Loganair routes linking six Scottish cities. The airline has since signed a term sheet for five Beta aircraft, with options for five more and service expected to start in 2029.

Innovation in eVTOLs

Beta is pushing so hard for the A250 because the case for electric flight gets even more interesting with vertical-lift aircraft. Helicopters and other VTOLs have a less orderly life than fixed-wing planes do. They must hover, sidestep, turn in place, settle, climb, and hold themselves in the air by main force. But conventional helicopters are built around a rotor speed that varies only a few percent because the engine, transmission, and rotor system are all designed around one narrow operating point.

“One of the dreams of vertical flight has been to vary the rotor rpm,” says Anubhav Datta, an aerospace engineering professor at the University of Maryland, College Park, who studies electric propulsion and vertical flight. Electric motors, he argues, have advanced enough to make that practical. A rotor that can slow down or speed up as conditions change can waste less energy, make less noise, and better match the odd assortment of tasks that vertical flight demands.

Electric propulsion also lets designers break up the work with multiple motors. The weight of an electric motor is tied closely to its torque. If you spread that torque among several motors, each of the motors can shrink, meaning that their combined weight is less than would be needed for one big engine. This is one reason why so many electric aircraft designs appear to have sprouted propellers wherever an engineer found a place to put one.

There is a downside to using multiple smaller motors: If one of the motors quits, the rotor it’s driving has little stored energy compared to the inertia in a conventional helicopter’s one large rotor. This inertia can come in handy if the engine dies, because the rotor will continue to spin and allow the pilot to descend under control, in an emergency technique known as autorotation.

Beta Technologies, Made in Vermont

According to Beta Technologies founder and CEO Kyle Clark, his native Vermont—known more for its maple syrup and outdoorsy lifestyle than its technological achievements—is a big part of the secret to Beta’s success at attracting talent. “The recruiting strategy is this: You find anybody who has vacationed in Vermont or went to school here. And you say, ‘You can actually live here and have a high-paying job in a mission-driven business, doing something that you care about with great people.’ ” The company currently employs about 1,500 people, most of them in South Burlington, Vt., and is actively hiring for its engineering, manufacturing, software, flight operations, and commercial teams.

As befits an aviation company, Beta offers flying lessons for all employees—a perk that is intended to give engineers, software developers, pilots, and mechanics a common language. The company also operates a health clinic for employees and their families and is building a day care center. Those programs have an idealistic Vermont flavor that is shaped by Clark and his wife, Katie Clark, who is Beta’s head of brand, creative, culture and spaces. But they also serve a practical purpose: attracting and retaining the best people in a small labor market. —T.C.

Electric-aircraft makers compensate with redundant motors and electrical channels, but most leading designs also have wings to allow gliding. In Beta’s A250, that means that if all else fails, the pilot still has a shot at reaching a suitable landing spot.

That helps explain the convergence in today’s eVTOL designs. Joby Aviation’s S4 uses six propellers that tilt between lift and cruise. Archer Aviation’s Midnight has six tilting propellers and six dedicated lift props. Beta’s A250 keeps its four lift propellers horizontal and uses a separate pusher motor for forward flight.

Beta’s Approach to Batteries

In Beta’s home state of Vermont, the electricity produced in the state comes almost entirely from renewable sources. Of course, electric aviation’s climate advantage depends not only on how clean the power grid is, but on how efficiently the aircraft uses energy. The turbines in conventional jet engines are marvelous machines, but they throw away a lot of energy as heat. Plus they need to carry the hardware to manage that heat—inlets, ducts, radiators, oil coolers—and the drag that comes with pulling air through all of it.

Kyle Clark [in cockpit], founder and CEO of Beta Technologies, pilots the company’s Alia CX300, an electric plane that Beta expects to receive FAA certification in 2027. Clark began designing a hybrid electric plane while still in college.Beta Technologies

By contrast, an electric design, with no combustion, gearbox, or oil circuits, avoids much of that energy burden. The result is an aircraft that is more than twice as “slippery” in terms of its lift-to-drag ratio. Batteries will never come close to the fuel density of petroleum fuels. But because electric drivetrains are much more efficient than combustion engines, a battery doesn’t necessarily need to match the energy density of petroleum fuel to be useful.

While the motors on an electric aircraft provide the drama at takeoff and landing, the batteries set nearly every boundary: the range, payload, charge time, cooling load, reserve energy, and, ultimately, useful life. For now, the hardest limit on electric aviation is not imagination, software, or even motors. It is the stubborn weight of stored energy.

“One thing you notice about [electric] airplanes is that they are all small,” says Cecutta, the aviation analyst. “They are not small because someone has done a market study and decided that is the optimal size. They are small because you cannot make them big. This limitation is from the batteries.”

In a conventional aircraft, fuel accounts for a significant percentage of the takeoff weight. But that weight gradually disappears during flight. In an electric plane, the battery packs also account for a lot of the weight; in Beta’s CX300, they’re close to half the aircraft’s mass. But a drained battery at landing weighs the same as a fully charged one at takeoff. That means the rest of the aircraft must be light enough to lift a power source that never gets lighter, yet strong enough to endure vibration, turbulence, hard landings, and years of fatigue. The more engineers push to save weight, Datta notes, the more flexible the airframe becomes, making the wings and other structures more prone to bending, twisting, and vibration. In that sense, the battery problem becomes a structural problem, too.

Beta’s CX300 carries five liquid-cooled battery packs of about 45 kilowatt-hours each, giving the aircraft roughly 225 kWh of rated capacity, though not all of that may be usable in flight. Each pack weighs 255 kg and contains 2,772 cylindrical cells. Altogether, about 14,000 cells sit beneath the occupied cabin, each one a tiny storehouse of energy, heat, and potential trouble.

The burden of the battery explains much of what I see elsewhere in Beta’s aircraft: the 15-meter wingspan, the composite structure, and the almost doctrinal effort to leave out every piece of hardware and creature comfort that’s not essential to the mission. The CX300 even does without flaps, which most aircraft use to generate extra lift at low speeds, so it must take off and land unusually fast for an aircraft of its size.

On Beta’s battery line, commercially supplied standard 21-millimeter-diameter, 70-mm-long cells are tested and assembled. Cells become modules, modules become subpacks, and eleven subpacks become a 45-kWh pack. Cooling plates manage heat during normal operation, while passive barriers and electrical isolation are designed to contain the ugly edge cases that Beta’s testing process intentionally triggers.

The modular design of the packs will give Beta’s planes a path to better batteries as the cells and battery chemistry improve. Caputo tells me that the company is already evaluating higher-density cells at its battery testing facility in nearby St. Albans, Vt. More watt-hours per kilogram could eventually mean more range, more payload, or both.

Beta Technologies’ Electric Ecosystem

What sets Beta Technologies apart from its competitors is the breadth of its ambitions: It’s not just building electric aircraft. It’s also cultivating a broader ecosystem to nurture the growth of electric aviation.

For instance, Beta intends its motor business to extend beyond its own aircraft. The company has delivered lift motors for Eve Air Mobility’s proof-of-concept aircraft and has a contract to supply up to $1 billion in pusher motors as Eve’s program develops. In addition, Beta is selling flight-control computers and software to Horizon Aircraft. It is also producing propulsion systems for the military, as well as the MV250, a hybrid-electric autonomous aircraft for defense missions.

Beta’s fast-charging stations [left] have been installed at 138 locations. The thermal-management station [right] keeps the battery from overheating during charging. The company uses the open-standard Combined Charging System (CCS) so that other aircraft can use its chargers. Beta Technologies

Further evidence of Beta’s far-ranging plans is visible near the battery line in the South Burlington facility, where workers are readying a few of the company’s 1.5-meter-tall Charge Cubes for shipping. Intended for installation at airports, each cube takes 480-volt AC from the grid and can deliver up to 320 kW at as much as 1,000 volts DC. According to Beta, charging either model of its Alia aircraft takes about 35 minutes, from the minimum charge required under visual flight rules to full state. A separate thermal-management cube circulates water-glycol coolant through the aircraft to keep the batteries from overheating during charging.

The charge cubes use the Combined Charging System (CCS), an open standard increasingly shared across the electric transportation industry. Beta’s strategy resembles Tesla’s decision to treat charging as part of the vehicle business, with one important difference: Beta is courting other aircraft makers, airports, and operators to use its technology. Archer, Atlantic Aviation, Signature Aviation, and government agencies have purchased or installed Beta’s charging equipment. As of October, Beta has 138 fast-charge locations, with additional sites in permitting or construction.

For charging beyond the fixed network, Beta offers a wheeled 40-kW Mini Cube and a 65-kW Mini Cube+ that can ride inside an Alia and plug into more modest electrical service. The Mini Cube can also charge cars, buses, and trucks—any vehicle that uses the CCS standard.

Back above Lake Champlain, the CX300 levels off, and the motor recedes to a faint hum. Then Caputo switches it off, and there’s nothing but the rush of wind around the canopy. The aircraft glides on its long wings while the coasting propeller turns the motor into a generator, feeding about 9 kW back to the battery.

Then we turn to the east and slip into the landing queue behind a National Guard C-130 Hercules, and Caputo powers and slides the throttle forward. The motor springs to life again, pushing the airplane, and perhaps aviation itself, into the future.

This article was updated on 29 October 2026.

This article is brought to you by VicOne.

A mobile robot slows for someone in a hallway. A collaborative arm eases its movement as a worker approaches. A humanoid pauses to let a person pass. Each response depends on information about the robot or its surroundings: a distance reading, a position estimate, or a stop signal.

But what happens when one of those inputs is wrong, and the robot still responds exactly as designed? The mobile robot reads the hallway as clear while someone is still there. The arm estimates that the worker is farther away. The humanoid does not receive a stop signal as a person crosses its path.

In each case, a fault could cause that mismatch. So could deliberate manipulation.

At VicOne LAB R7, we examine how manipulated inputs can change robot behavior — and what that means for the systems meant to keep people safe.

How an untrusted input can change a robot’s behavior

A robot does more than measure its surroundings. It may interpret what it sees and hears to decide what to do next. Information placed in its environment can therefore influence its actions.

Published research shows how this can happen. In a study of vision-language-action (VLA) models, a patch in a camera’s view reduced task success in simulated robot tests. FreezeVLA found that an adversarial image could cause tested models to ignore later instructions. Although the studies used different methods, both demonstrate how visual input can interfere with a robot’s intended task.

In a robot-dog test using Gemma 4 E4B, text on a poster was treated as an instruction and changed the robot’s movement.

VicOne LAB R7 also tested how untrusted inputs could change robot behavior. In a robot-dog test using Gemma 4 E4B, text on a poster was treated as an instruction and changed the robot’s movement. In a separate hospital-service-robot simulation using Nemotron on NVIDIA Jetson AGX Orin, crafted audio that is inaudible to people changed the robot’s simulated behavior. The assigned tasks did not change; the inputs did.

In a VicOne LAB R7 test, text on a poster changed a robot dog’s movement, illustrating how a crafted input can turn an unlikely outcome into a repeatable attack.VicOne

Safety-related decisions have inputs too

An independent protective system may still stop a dangerous action. But that system also depends on information: readings, estimates, and messages that tell it when to intervene. What happens if one of those inputs is manipulated?

At a robotics bug bounty event, VicOne LAB R7 researchers injected a ROS 2/DDS message into a robot that organizers expected to remain still under a safe-control setting. The robot moved.

A humanoid’s balance controller offers another example. If it relies on a vulnerable gyroscope, sound at the sensor’s resonant frequency could distort its reported rotation. The controller might correct for a tilt that never occurred. Researchers demonstrated the underlying acoustic attack against drones with vulnerable gyroscopes. The research did not demonstrate the same attack chain causing a humanoid to fall.

Other protections could interrupt either chain. The question for safety and security teams is whether those protections detect the danger independently or depend on the same manipulated information.

Figure 1. VicOne maps how accidental faults and deliberate cyber manipulation can lead to similar robot safety outcomes, from loss of control to physical harm.VicOne

Deliberate attacks challenge safety assumptions

A safety assessment may treat a combination of accidental faults as unlikely. Deliberate manipulation changes that assumption. An attacker can choose when to introduce a false input and repeat the same trigger. That does not make harm inevitable, but it may change how cyber risks should be assessed.

Redundant sensors also need an adversarial test. If one action can affect both inputs, their agreement may offer less reassurance than expected. In autonomous driving research, a crafted physical object misled a system combining camera and LiDAR perception. The study does not establish a weakness in a particular robot, but it shows why teams should test whether one manipulation can affect several checks at once.

Cyber threats add intent to the safety equation, so risk assessments based on accidental faults must be revisited to account for attacks that can be timed and repeated.

Security testing strengthens robot safety evidence

Safety teams set limits for speed, protective distances, and permitted operating areas. The evidence behind those limits should also show what happens when the information used to enforce them is deliberately manipulated.

That evidence can also support assessments against applicable requirements. China’s GB/T 45502-2025 addresses information security for service robots. IEC TS 63074 examines security threats that could affect safety-related control systems. The EU Machinery Regulation, which applies from January 20, 2027, includes requirements to protect safety-relevant systems and data against corruption. Each has its own scope, and teams still need evidence for their robot’s design and operating conditions.

Figure 2. Cybersecurity strengthens robot safety evidence throughout the lifecycle. VicOne helps teams test safety limits against manipulation before deployment and monitor the conditions behind safe behavior in operation.VicOne

Before deployment, safety and security engineers can compare normal and adversarial scenarios against the same safety limits. During operation, they can monitor for changes to software, models, sensor signals, or behavior that may call earlier results into question. Any response to suspicious behaviors should follow policies approved by the safety team.

VicOne supports that work across the robot lifecycle. Through the Robotic Hacking Community, VicOne LAB R7 works with researchers to investigate how cyber threats can change robot behavior. VicOne’s Radeis helps teams validate potential effects before deployment, while Rthena provides visibility into risks and behavior during operation.

Five questions safety and security teams should answer together

The central test is whether a robot’s safeguards still protect people when the information they rely on is manipulated. Safety and security teams can begin by tracing the inputs behind each protective decision, challenging them, and revisiting the evidence as the robot changes. They can start with these five questions:

1. What does each protective function read? Map the readings, estimates, messages, and confirmations. Identify which safeguards operate independently of the robot’s AI decisions.

2. How does that information arrive? Check how inputs are produced, transmitted, authenticated where appropriate, and handled when they are missing or implausible.

3. What happens if an input is manipulated? Test false distances, drifting position estimates, and other relevant scenarios against the robot’s safety limits.

4. Can one action mislead several inputs? Check whether sensor fusion or seemingly independent safeguards share a point of failure.

5. When should the evidence be revisited? Reassess after changes to software, models, sensors, or operating conditions. Agree on bounded responses to suspicious behavior with the safety team in advance.

The goal is not only to show that a protective function works. Teams also need evidence that it remains protective when the information behind its decision is wrong or deliberately manipulated — and that another safeguard can catch the resulting hazard when needed. As robots change, that evidence needs to change with them.

For a deeper look how cybersecurity can affect robot safety, download our white paper “Why Robot Safety Depends on Cybersecurity: Protecting the Information That Safety Decisions Rely On.”

Science has a spam problem. And the spam filter is getting clogged up.

According to all reports, academic publishing is drowning in what has come to be called AI slop: low-quality, mass-produced content generated with AI. By 2025, an estimated 57 percent of all published academic articles exhibited detectable evidence of large language model (LLM) influence, a massive jump from 12 percent in 2023. What’s more, over 12 percent of Nature Communications reviews were found to be completely AI-generated in 2025.

Given that for every 40 papers showing statistical evidence of AI usage in 2025, only one formally disclosed it, these figures likely grossly underestimate LLM involvement today. Of course, evidence of AI use doesn’t necessarily mean a paper is slop. But some estimates suggest that hundreds of thousands of fake, often AI-generated, publications are now produced annually.

“It’s not just the [scientific] record under threat, it’s science as a whole,” warns Amanda Sulicz, IEEE’s research-integrity manager. “AI slop is completely eroding trust in science and every scientific process that we’ve worked for centuries to create.”

Yet statistics and statements claiming that the problem is big—perhaps even existential for science—don’t really get the message across. Just how easy is it to use AI to create a convincing but completely fake paper? And without adding even more to the burden on stretched peer reviewers, who do this work on a volunteer basis, do publishers have the tools to keep AI slop out of the scientific record?

Borrowing from the idea of “vibe coding,” in which programmers use natural-language prompts to generate code, I decided to “vibe-research” a made-up paper by prompting LLMs to generate hypotheses, code, and text with minimal human input. Then I asked Jenny Mahoney, managing editor of the multidisciplinary and open-access journal IEEE Access, if we could war-game an attempt to sneak my AI-generated monstrosity past her journal’s defenses. She and IEEE kindly agreed.

The experiment would become a two-part test. First, whether someone with no special AI expertise could use off-the-shelf AI tools to generate a convincing paper describing research that had never been performed, and do it well enough to fool a publisher’s defenses. Then I would raise the stakes: What happens when more advanced AI systems don’t just write the paper but conduct underlying research?

Along the way, the experiment became a window into how AI automation is transforming scientific discovery and forcing stakeholders to rethink the entire academic publishing system.

Round 1: Creating the Forgery

To give me a chance at creating a convincing fake, I gave myself permission to prompt multiple times (within reason). But to make the fight fair, I couldn’t access the most advanced paid versions of LLMs—only free ChatGPT (at the time, GPT-5.3/5.2 Mini) and Gemini (3 Flash). Moreover, I could not directly edit any output, nor ask for outside help from expert AI-prompt engineers. This needed to be an amateur effort, with the paper generated as much as possible by the LLMs alone.

I assumed that using fake author names and affiliations would immediately raise a red flag, and I didn’t want to drag the names of established academics into the experiment, so the best bet was to name myself as the sole author. That choice limited me to subjects in which I had some academic experience. So I chose my Ph.D. topic from nearly two decades ago: optical solitons, which are individual light pulses that propagate without any change to their shape.

My first attempt—“Can you vibe-research me an original paper in optical solitons?”—left a lot to be desired. Though Gemini was happy to do the task (unlike ChatGPT, which was more forceful in not generating synthetic data or studies), the result was underwhelming. And it highlighted a recurring problem during the process: Without any real research to work with, the output was laughably thin, with the LLMs often struggling to fill four pages. However, as I became more adept at prompting the models, they loosened up and became more characteristically loquacious.

I broke the paper down into its constituent parts—abstract, introduction, methods, and so on—and tackled them one by one. If one LLM refused to go any further, switching to the other and asking what the next step would look like “hypothetically” would often do the trick. After a dozen or so prompts, I had a realistic-looking paper. A dozen or so more prompts generated and tidied up the figures and captions, produced references and citations, and converted the paper into IEEE Access’s required style.

At this point, I read the full paper carefully. To me, it looked pretty good. The science sounded convincing. I even had visions of returning to academia to conduct in real life the experiments that the LLMs had proposed.

What’s more, I couldn’t see any obvious telltale “AI-isms” that would immediately be flagged. For example, the LLMs didn’t produce any tortured phrases—unusual paraphrases of standard technical terms, like “profound neural organization” replacing “deep neural network.”

Other AI-isms were arguably present, involving stylistic tics like hedging language (“potentially, typically”), grouping things in threes (“The study provides a clear, concise, and actionable framework”), and inserting formulaic transitions (“In summary,…”). But as the previous sentence demonstrates, these tics are hard to spot because they are often used by unimaginative humans, too.

Round 1: Attempt to Publish

Two Ways to Make a Paper With AI

I took two very different routes to creating a scientific paper with AI. First, I coaxed free, general-purpose chatbots through the process step by step; then I turned to FARS, an autonomous research system designed to do much of the scientific work itself.

ROUND 1: CHATGPT + GEMINI

1. Choose a research topic
I ask the free chatbots to create an original paper on optical solitons, my former Ph.D. field.

2. Break the paper into pieces
The initial results are thin, so I prompt separately for the abstract, introduction, methods, and other sections.

3. Keep steering the LLMs
When one model refuses or stalls, I switch models, rephrase prompts, and ask what the next step would look like “hypothetically.”

4. Build and polish the paper
After about two dozen prompts, the LLMs have generated the text, figures, captions, references, and citations and formatted the paper for IEEE Access.

5. Put it to the test
I send the fake paper to IEEE Access, where automated integrity checks flag fabricated and unverifiable references.

→ CAUGHT
IEEE’s integrity checks flag fabricated and unverifiable references.

ROUND 2: FARS

1. Give FARS a research topic
I enter an abbreviated version of the same optical-soliton topic. FARS generates a research proposal in 21 minutes.

2. Authorize the experiment
With a click, I tell FARS to carry out the experiment described in its proposal.

3. Let FARS do the research
FARS conducts the computational experiment and emails me the next day to say it’s complete.

4. Tell FARS to write the paper
Without checking the results, I click “Write paper.” Two hours later, FARS produces a submission-ready article.

5. Put it to the test
I remove FARS’s disclosure, put my name on the paper, and send it to IEEE Access.

→ CAUGHT
The references are real, but IEEE flags the paper’s heavy reliance on non-peer-reviewed preprints.

Handing the fake paper over to IEEE Access left me torn. After putting in a fair amount of effort into making the AI-generated forgery, would I be happy if it passed through the checkers, or saddened that the only remaining line of defense for publishers is the evermore-burdened peer-review community? And if it was spotted, would I be angry that I didn’t think of a red flag or be glad the checkers did their job?

As it turned out, I was filled with relief when the paper got picked up quickly for having hallucinated references, otherwise known as ghost citations. When I ran the paper through Gemini afterward, without qualm it happily reported that seven “modern LLM-related research references have been fabricated to support the paper’s specific narrative.” Thanks for the heads up, Gemini!

In fact, Gemini underestimated how many ghost citations there were. When Mahoney ran the paper through a research-integrity checker called STM Integrity Hub, she picked up 12, and another checker from MolecularConnections could not validate 24 references, including one that had the classic tell of “John Smith” as a single author (no offense to all the academically inclined John Smiths out there). “I am an experimental physicist by training, so this was not at all my area, but nothing seemed majorly scientifically wrong” about the paper, says Mahoney. “It was just that so many hallucinated references immediately cast doubt on the entire work.”

Feedback on the fake paper from Gaetano Assanto, an expert in optical solitons and a former collaborator of mine from Roma Tre University, in Italy, was similarly mixed. Concentrating on the paper itself, he thought it looked professionally written. But he noted that the mathematics was very dense (“too much math for me in a single shot!”) and would require specialized knowledge to evaluate. And, although the text looked legitimate, Assanto noted an excessive use of acronyms in the abstract, and he was suspicious of a couple of confidently proclaimed physics statements that were at best debatable.

The AI Arms Race

On one level, this outcome was positive: A complete novice using free tools couldn’t get an AI-faked paper past the publishers’ automated defenses or a potential peer reviewer’s eagle eye. But it was hardly a resounding victory for the scientific publishing system. With general-purpose LLMs and no special prompting expertise, I had managed to conjure a superficially plausible scientific paper out of research that had never happened. And more advanced LLMs are no longer making the kind of mistakes their free, dumber relatives are making.

These tools aim to address a serious bottleneck to scientific discovery: us.

It’s common knowledge in the academic publishing world that, in the cat-and-mouse game between bad actors and publishers, AI slop papers continue to slip through publishers’ safeguards. “It feels like building a detection tool today may not be effective with the AI-created paper of tomorrow,” says Luigi Longobardi, IEEE’s director of publishing ethics and conduct. This leaves the humble peer reviewer as a publisher’s last line of defense. Yet peer review is simply not fit to handle the volume of papers—both legitimate and slop—passing through.

Even before AI tools became ubiquitous, the peer-review system was being pushed to its breaking point, processing well over 3 million papers in 2022. Fundamentally, this problem stems from science’s publish-or-perish culture, which pressures academics to constantly publish papers in prestigious journals to secure funding and advance their careers. This culture not only forces many academics to regularly churn out incremental research, but has also long driven some to enlist the help of “paper mills”—unethical organizations that produce low-quality or fabricated papers—just to keep pace. AI only amplifies these existing dynamics.

Burden on Academics

Mahoney says IEEE Access alone receives about 60,000 submissions per year, accepting around 20 percent, or about 12,000 papers. Every accepted paper is reviewed by at least two independent reviewers. But despite the journal’s roster of tens of thousands of peer reviewers, it’s getting harder and harder to secure their services, with the average number of reviewer invitations being sent out for each article increasing by 50 percent since 2018. For these volunteers, peer review is a significant commitment on top of teaching and conducting their own research. Now they’re also being asked to spot AI slop.

Of course, peer review is not the end of scrutiny for any paper. Once it has entered the scientific record, it must then pass muster with the scientific community. This is where thousands of papers with telltale signs of inappropriate AI use (and other forms of research misconduct) have been exposed.

What Gave Away My AI Papers 

Round 1: The Fake Paper

Created with free versions of ChatGPT and Gemini

Ghost citations

Many references were fabricated or couldn’t be validated.
Questionable scientific claims

An expert flagged confidently stated claims that were at best debatable.
Too many acronyms

The expert also noted excessive use of acronyms in the abstract.

Round 2: The AI-Research Paper

Created by the autonomous research system FARS

Too many preprints

The references were real, but too many pointed to non-peer-reviewed preprints.
Nonsense in a figure

One technical figure contained a line labeled “lhewl.”
AI artifacts

Another figure’s depiction of an AI “brain” contained telltale random dots.

There are two high-profile homes for this important and completely volunteer-led post-publication review work. PubPeer is a place where researchers, often called scientific sleuths, can anonymously identify flaws in the literature. And Retraction Watch documents and publicizes corrections to the scientific record.

Even the people who track problems in the scientific literature can find themselves caught up in AI fabrications. Ivan Oransky, cofounder of Retraction Watch and a member of the PubPeer Foundation’s board of directors, has encountered papers containing fabricated references attributed to him. In one case, a journal article criticizing post-publication peer review included a fabricated reference supposedly written by Oransky himself—“which was sort of funny,” he chuckles.

Oransky argues that AI has exposed long-standing weaknesses in the scientific publishing system rather than creating an entirely new problem: “What I liken it to is a fire, where publish-or-perish has always been there as kindling, paper mills are the gasoline, and AI is the accelerant.” But though concerning, the crisis could create an opportunity, he says. “If I’m trying to be optimistic, maybe AI slop and the volume of submissions could make more people aware of those problems and try and solve them at a fundamental first-principles level.”

Blurred Lines

My homemade forgery had failed for reasons that today’s publishers know how to spot. But that raised a more unsettling question: What happens when AI isn’t simply fabricating a paper but conducting research that blurs the line between legitimate science and AI fabrication?

Making my vibe-researched paper look like a cave drawing, new automated research systems are handing increasingly large parts of the scientific process over to AI. Tools including FARS (Fully Automated Research System), AutoResearchClaw, The AIScientist-v2, Zochi, AI Researcher, and DeepScientist all aim to address a serious bottleneck to scientific discovery: us.

What happens when AI isn’t simply fabricating a paper but conducting research that blurs the line between legitimate science and AI fabrication?

Often with just a single prompt, these AI engines can conduct a comprehensive literature search, from which they can generate a completely reasonable research proposal. They can design and conduct experiments using digital datasets and simulations and then write a full research paper—all with no human input after the initial prompt. Very recent systems are even beginning to go beyond the digital realm. LabClaw, for example, now links to autonomous labs that can generate real-world data for tasks such as drug discovery.

The output is, to most intents and purposes, actual research, and therefore potentially worthy of publication. But with humans in the loop in only the most nominal sense, these tools bring into question the meaning of the term “legitimate science.”

FARS, which caused a stir this year when it created 166 papers in 417 hours, was developed by the Singapore-based startup Analemma. “We built FARS because research is still constrained by human attention, fragmented tools, scarce technical labor, and slow experimental iteration,” explains Xiwen Su, Analemma’s manager of product operations and partnerships. “FARS can pursue many research threads simultaneously, 24/7, without fatigue, and at scale.” (The online service that I used to access FARS was discontinued after Analemma was acquired by Baidu while this article was being reported.)

Su and the team behind Analemma kindly agreed to let me loose on FARS and gave me enough credits to create a paper. Which gave me another opportunity to put Mahoney’s research-integrity checkers to the test.

Round 2: Using an AI Science Engine

FARS is not an LLM in and of itself, but rather an agentic AI-orchestration system connected to a computing cluster that has access to essentially any open- or closed-source LLM it needs. The system uses four specialized agents for ideation, planning, conducting the experiment, and writing, all designed to perform the main tasks of science autonomously.

Though FARS initially focused on AI research, Su told me it should be reasonably effective across other domains. As I wanted a topic similar to that of my first fake paper, I pasted in an abridged version of that paper’s title—“LLM-aided Optical Soliton Dynamics Modelling for Next-Gen Communications”—into the prompt box and hit “Send.” Just 21 minutes later, FARS generated a research proposal based on an extremely detailed literature review—far more than I or any other human could have achieved in that time.

With a click, I tasked FARS with conducting the experiment it detailed in its research proposal. The next day, I received an email saying my experiment was complete. Without checking the results, I pressed a button marked “Write paper,” and the final submission-ready article was produced in 2 hours.

At this point, I finally had a look at some of the output. At first glance, the paper looked convincingly human-written. The only giveaway was a line just below the abstract reading: “WARNING: This paper was generated by an automated research system.” I deleted that, put my name in its place, and sent it to Mahoney.

Round 2: Attempt to Publish

Because FARS is designed to spot gaps in the scientific literature and conduct real research to close them, I was more confident it would breeze through the publisher’s checks. But just as my fake paper fell down in the references, so too did the AI-generated real one.

“Your fake article had lots of hallucinated references,” says Mahoney. “While this AI-generated one had no fake references, our team would have rejected this paper because it did have a huge percentage of references to preprints, which likely has something to do with the dataset the AI has access to.” Because preprints haven’t yet undergone peer review, IEEE Access considers heavy reliance on them a problem.

In addition, although the research-integrity checking tools didn’t spot any issues with figures, Mahoney did. One figure included a red line labeled with the nonsense term “lhewl,” and an AI “brain” in another figure had some random dots in it, which ironically were a tell that it was AI generated.

Automation Changing the Landscape

Automated research systems are very much a work in progress, but they are already starting to make waves in the scientific community. For instance, an AI Scientist academic paper and papers from the Autoscience Institute were accepted through double-blind peer review in 2025 by the International Conference on Learning Representations. As these systems improve and start to produce research outputs at volume, the waves are likely to build to a tsunami that will threaten today’s scientific and publishing process.

“FARS doesn’t kill the paper, but it does challenge the paper as the unit of knowledge,” argues Su. “We believe the future looks more like a living knowledge graph—where individual, well-scoped contributions (some made by humans, some by machines) connect and build on each other dynamically.”

In this future, humans will continue to have roles in the scientific and publication processes—just very different and limited ones. For example, instead of conducting experiments themselves, humans would sense-check and review automated output at key points. And traditional line-by-line peer-reviewing will be a thing of the past, though Su emphasizes that this change will not eliminate human judgment. “Automated systems can handle reproducibility checks, statistical validation, and novelty screening,” she says. “Humans can then focus on the harder question: Does this [research] matter?”

A Human-Centric or AI-Driven Future?

Though from a very different world than Su, Mahoney agrees that humans need to be freed to focus on the science, and that technology holds the key. “It’s promising to see the progress that has been made to automate some of the triage tasks, such as author checks, tortured phrases, duplicate submission, reference validity, et cetera,” she says. “During peer review, we’re also using AI semantic search tools to help editors match manuscripts to the most relevant reviewers, who are best placed to stop the slop before entering the scientific record.”

In addition, Mahoney thinks that AI cannot be banned or overly restricted, as this will stymie progress at a time when human knowledge could and should be growing at an unprecedented rate. “Researchers are increasingly using AI in their work, from literature discovery, to code generation, to data analysis, and this is not necessarily a bad thing,” she argues. But she disagrees that peer reviewers and humans more broadly will play a smaller part in scientific progress; Mahoney believes the future still requires humans at the wheel. The question is not just whether the research matters, she says. “The question should be: Can this work be trusted and reproduced?”

To this end, Longobardi argues that AI may ultimately require scientists to prove not just what they discovered but how they discovered it. One possibility, he says, is using technology to verify research at the source. “The system doesn’t really help because it’s not designed to not trust the paper,” he explains. “Maybe institutions can think about developing their own trust markers, automatically integrating the equipment researchers use with blockchain technology to say ‘Yes, this person was in the lab on this day and got this data.’”

Elements of this approach are already emerging. For instance, two departments within the Max Planck Institute of Colloids and Interfaces have implemented a sample database, assigning each and every sample created under their roofs with a unique ID and QR code. And the University of Southampton’s School of Chemistry and Chemical Engineering has rolled out an electronic lab notebook system. These innovative safeguards represent early efforts toward keeping humans front and center in scientific progress, and making the scientific process itself more transparent.

Whether this human-centric vision, Su’s AI-driven aspiration, or some other picture of the future wins out remains to be seen. The one thing my vibe-research adventure has clarified is that business-as-usual science is not an option. Scientists and publishers must adapt to a world in which it’s becoming increasingly difficult to spot the difference between real research and AI forgeries, and where deciding what counts as legitimate scientific work may become even harder than detecting outright fraud.

See the AI-Generated Papers

Benjamin Skuse created two papers for this experiment, both screened by IEEE Access with the publisher’s cooperation. Neither was accepted for publication.

Paper 1: The AI-Generated Forgery
Created with ChatGPT and Gemini, this paper describes research that never happened.

Paper 2: The AI-Generated Research Paper
Created with Analemma’s FARS system, this paper describes computational research conducted by AI.

Neither paper was peer reviewed, and their scientific claims have not been independently validated.

IEEE recently released the IEEE Xplore AI Research Suite, built on the full text of millions of peer-reviewed publications in the IEEE Xplore Digital Library. The library offers full-text access to IEEE standards, journals, magazines, conference proceedings, and more.

Researchers who use the AI suite to search the digital library receive answers grounded in verified scientific results, ensuring a high level of accuracy and trust.

The suite is available as an optional subscription upgrade to all IEEE Xplore Digital Library institutional subscribers.

Powered by a proprietary AI system optimized for STEM literature, the suite introduces powerful, intelligent features within the digital library. The suite enables users to seamlessly search and discover articles to accelerate their comprehension of complex technical papers.

The system surfaces deep insights, replacing the typical keyword queries with natural, conversational research workflows.

As AI tools become more common in conducting research, IEEE collaborated to build the suite around a simple principle: Answers should be trustworthy, not just fast, says W. Clem Karl, vice president, IEEE Publication Services and Products.

“AI can be a highly effective and beneficial tool that can help researchers accelerate discovery, enhance decision-making, and surface new connections,” Karl says. “However, many AI engines present a great risk of hallucinating answers or introducing bias into otherwise credible research. With this in mind, IEEE has taken a responsible approach in our AI strategy by developing concrete safeguards to ensure our solution is rigorously grounded in scientific evidence that users can trust.

“As stewards of the scholarly record, we held the new AI engine to the same standards as the peer-reviewed research we publish: A researcher shouldn’t have to trust an answer; they should be able to verify the work behind it.”

Collaboration with the scientific community

The IEEE Xplore AI Research Suite was developed in close collaboration with global scientific and library communities.

The development process began with research into user needs and the challenges associated with existing AI research assistants. More than 3,000 testers—including engineers, scientists, and library information professionals—provided feedback to refine the AI engine and support research workflows.

Capabilities include:

• Semantic AI search. Users can input free-form, natural-language queries or paste long blocks of text to locate similar research and identify novel solutions.
• Time-saving summaries. The platform generates immediate, accurate AI overviews of search results and provides in-context definitions of technical terms so that users can grasp new concepts faster and accelerate their research.
• Built on trusted content. The suite is the only AI research assistant that uses both the full text and metadata of IEEE’s peer-reviewed journal articles and conference papers to optimize the accuracy and relevance of the results.
• Comprehensive coverage. A variety of STEM metadata is available from journal articles and conference papers that have been curated to include scholarly articles without the typical noise found in other tools.
• Source links and verifiable results. The AI-generated overviews include reference links to source documents, allowing users to verify information and cross-reference findings.

Reducing hallucinations

To mitigate the risk of AI hallucinations, the search-retrieval process is limited to querying the millions of IEEE peer-reviewed full-text documents, alongside an expertly curated dataset of STEM metadata. If the AI tool is unable to find relevant scientific content within the defined dataset, it is designed to inform the user rather than generate fabricated answers.

“IEEE is committed to providing the technical community with a highly precise and reliable research experience,” says Steven Heffner, IEEE’s chief publication officer and managing director of IEEE Publications. “The IEEE Xplore AI Research Suite offers something unique in that it is the only AI research assistant that leverages both the full-text and metadata of IEEE’s peer-reviewed content, rather than abstracts alone.

“The difference is significant: Our internal testing showed a threefold increase in result quality and accuracy across every key metric. These results underscore how critical full-text grounding is for technical research.”

For more information about features and coverage, or to request a demo for your organization, visit the IEEE Xplore AI Research Suite website.

Like many a nerd, I have a certain fascination with locks. This originates in reading the memoir Surely You’re Joking, Mr. Feynman as a teenager, in which Richard Feynman recounts picking locks and cracking combinations while at his day job working on the Manhattan Project at Los Alamos. I currently have a couple of locksport pick sets and training locks that I purchased at hacker conferences over the years (although I’m still terrible at it), and I delight in the YouTube exploits of folks like the LockPickingLawyer.

All of which left me wondering if I could try building an interesting lock. One dodge that lockmakers have long used to challenge would-be plunderers is to conceal the existence of the lock itself. A modern variant of this is the magnetic lock, where a magnet is held or slid against an innocuous portion of a door or box, releasing a catch. The problem with this is that, once you guess the location of the lock, any strong magnet will open it.

So I started thinking about embedding an RFID reader under the surface and resigning myself to keeping track of yet another access card. But then I stumbled across something much more interesting: SparkFun’s US $100 magnetic imaging tile. I realized I could use this to build a magnetic lock that opens with a programmable secret gesture.

I used a few discrete components to drive the servo motor and put the lock into program mode [top], controlled by a Teensy 4.1 microcontroller [middle]. An array of Hall-effect sensors detect magnetic fields [bottom].James Provost

The magnetic imaging tile is a small-run SparkFun board based on the work of Peter Jansen. It features an 8-by-8 grid of Hall-effect sensors, and uses an SPI interface to talk to the microcontroller of your choice. These sensors can detect both the magnitude and polarity of a magnetic field, and the board can report its measurements hundreds to thousands of times per second, depending on how much processing power your microcontroller has available. This is easily fast enough to capture the motion of a magnet being moved by hand across the board’s sensor area—my secret gesture.

An interesting thing about using magnetic fields is that if I was in extreme paranoia mode, I could shield the entire lock behind copper sheeting, something I couldn’t do with an RFID reader. The copper sheet would act as a Faraday cage, blocking any RFID signal, along with any other electromagnetic fields an attacker with an EMF meter might use to locate the lock.

Choosing the Brains of the Box

I chose a Teensy 4.1 as my microcontroller. I like the Teensy because it offers loads of input/output, a 600-megahertz processor, a megabyte of RAM, and a built-in SD card, all in a slim package with an equally slim $32 price tag.

Wiring the Teensy to the magnetic-imaging tile was trivial, only complicated by the fact that one side of the board has a silk-screen error that misidentifies the SPI connections (SparkFun says this will be fixed in future runs of the board).

In extreme paranoia mode, I could shield the entire lock behind copper sheeting.

To create the gesture-recognition software, I once again turned to vibe coding. If you read my Hands On piece in the April 2026 issue of IEEE Spectrum on creating a teletext suite for ham radio, you’ll know I’ve been grappling with how vibe coding should be used in DIY projects, if at all. So this time, instead of letting Claude write all the software, I confined it to writing just two support libraries, which you can find in our online repository at https://github.com/IEEESpectrumHandsOn.

One library is in charge of recording a lock gesture and saving it to the Teensy’s nonvolatile storage. The other library monitors the tile and reports whether or not a newly detected gesture passes a selectable threshold to be considered a match.

I wrote the top-level code myself to keep it as DIY as possible. This code handles things like a button that puts the system in gesture-recording mode after a brief delay, the indicator LED for recording mode, and the code that operates the physical locking mechanism.

Hall-effect sensors rely on a current passing through a p-type slab of silicon. A magnetic field passing through the silicon deflects charge carriers toward the edges of the slab, creating a voltage proportional to the field [left]. By looking at the intensities of the voltages across an 8-by-8 grid of sensors, the microcontroller can reconstruct a gesture [right]. James Provost

Of course, I needed something to lock to demonstrate my proof of concept. I bought a wooden craft box and removed the existing lid latch. I glued a notched beam inside the lid and built a mechanism that uses a servo motor to rotate another beam in and out of this notch. I will be the first to admit this mechanism goes beyond “simple” and into “simplistic,” but it’s sufficiently robust for my demo. The box has ample room for the magnetic imaging tile to be mounted against one wall, along with the Teensy and other supporting electronics. These include a battery pack, the recording button, and a small additional board with two NPN transistors. These convert the Teensy’s 3.3-volt control signals to a 5-V signal needed to drive the servo.

Once it was all put together, I soon discovered that the small size of the magnetic tile’s sensing area—just 3 centimeters square—made it easy to fit the tile inside the box but hard to do the unlock gesture on the outside so that it is read correctly. Marking the outline of the sensor area on the outside defeated the point of having a hidden lock, so instead I once again turned to mining my Irish heritage. I created a Celtic knotwork pattern that runs in a band across the front of the lockbox (I used an online knotwork generator and then projected and traced the pattern onto the box). Using the pattern as a camouflaged reference grid I can make a repeatable gesture within the confines of the sensor area. (To make the gestures I’m using a magnetic stylus borrowed from a ferrite core-memory kit we featured in 2022.)

There’s obvious room for improvement in any more serious installation. The biggest drawback currently is that if the box is locked when the battery runs out, you can’t open it again without breaking something. This could be solved by adding a hidden inductive charging coil. Nonetheless, the demo lockbox works well—now I just have to think what I should store in it!

I step into the warm moist air
among the columns of the pines,
their net of twigs and roots
complex as any lines of code.

The trees purify and anchor.
Their branches balance my restlessness.

Their calling is to be present,
mine to develop apps. Our tasks
verge toward a common logic:
living filaments of root, of interface.

Many engineers are more comfortable working behind the scenes developing technology than speaking to consumers about products. But Preethi Ashwath has learned to take center stage, talking about the “hearable” and wearable devices she has had a hand in developing as an analog integrated circuit designer in the consumer electronics division at Analog Devices of Wilmington, Mass.

The IEEE member is one of the few women who work in the specialized field, where about 11 percent are female, according to career platform Zippia.

Preethi Ashwath

Employer

Analog Devices, in Boston

Title

Integrated circuit designer

Member grade

Member

Alma maters

SJB Institute of Technology in Bengaluru, India; University of California, Irvine

Ashwath stepped out from behind the scenes last year to meet with attendees at the Consumer Electronics Show (CES) in Las Vegas. She demonstrated Analog’s new earbuds, which track people’s heart and lung health through vibrations in their ear canals.

After she demonstrated her ability to explain the technology to consumers in simple terms, Analog selected her to represent the company at other events. She says she has “come to see speaking as part of contributing to the field, rather than something separate from technical work.”

She also gives technical talks on various aspects of electrostatic discharge (ESD) and other topics, as well as participates in panel discussions, she says, because they “reach engineers well beyond my immediate team.”

Ashwath is an active volunteer with the IEEE Circuits and Systems Society.

“Whether through designing semiconductor products, publishing technical articles, reviewing conference papers, or speaking at industry events, I try to contribute to the growth of the engineering community and the advancement of the field,” she says.

Discovering circuit design at Intel

Ashwath, who grew up in Bengaluru, India, learned about electronics engineering from her high school physics tutor.

That introduction led her to take electronics as one of her core courses in her final years there. At the time, Ashwath’s older sister was pursuing an engineering degree.

“Between my interest in electronics and seeing her already on that path, it became clear to me that engineering was the way to go,” Ashwath says.

After earning a bachelor’s degree in electronics and communications engineering at the SJB Institute of Technology, in Bengaluru, she wanted to pursue a career in very-large-scale integration. But that changed once she moved to the United States and began her master’s program in electrical engineering at the University of California, Irvine.

She discovered her passion for circuits and devices during an internship at Intel in Santa Clara, Calif. She learned about circuit design, package design, and reliability verification tools such as Ansys Totem.

“Seeing every layer of the chip development process gave me a more complete picture of what goes into building a semiconductor product,” she says. “Through that experience, I realized I felt most connected to design.

“It was the hard work and creativity that drew me in. You start with a specification, build a circuit, simulate it, then wait for the chip to come back from fabrication. When your design works, it’s a real feeling of satisfaction. It teaches you patience and perseverance.”

After attaining her master’s degree, she was hired by Intel as an analog design engineer. She gained experience in mixed-signal circuit design before joining Analog. In her current role, she works on consumer hearables such as earbuds and wearables including fitness trackers and health monitors.

Her main focus is designing ESD pad rings, the borders of specialized circuits built around a microchip’s outermost edge to protect it from the discharge, which can damage circuit paths or introduce latent defects that lead to failures later in the chip’s life cycle.

“On a chip with multiple voltage domains, you have to plan how the protection structures are placed and connected across the whole pad ring—and get it right,” Ashwath says. “A mistake there could cause a chip to fail in the field in ways that are very hard to debug afterward.

“It’s detailed work that requires understanding both the device physics and how the whole chip fits together.”

Part of the job is designing and creating an integrated circuit’s building blocks, including ESD protection structures and other mixed-signal blocks. Ashwath conducts rigorous checks to verify they work as intended before being integrated into a chip.

She collaborates with ESD advisors, layout designers, and the digital team that designs the chip’s digital logic.

“I work with them to make sure my analog circuits and their digital blocks work together correctly,” she says. “No two days look exactly the same.”

In 2024 she was recognized with the company’s Impact Award for her “exceptional work on a critical project with significant cross-functional business impact.”

Presenting the final product

Ashwath was selected to represent Analog at events after a senior leader encouraged her to ask for high-visibility opportunities, she says.

“Most of my time is spent on designing circuits and simulations,” she says. “At CES, I stood in front of customers, engineers, and members of the press, explaining our technology and answering questions on the spot.”

The earbuds she demonstrated monitor minute vibrations from a user’s heart and lungs, according to the Analog website. The vibrations can detect arrhythmia, a leaky heart valve, or fluid in the lungs.

“Hearables are evolving from playing back audio to becoming health-monitoring platforms,” she says. “The ear turns out to be a good location for picking up physiological signals, so these devices can capture vital signs while someone is simply wearing them normally.

“The hard part is achieving high signal quality. The signals you’re trying to detect are very small and buried in the noise from the person’s movements and the surrounding environment.

“That’s where analog front-end design becomes critical. Extracting a meaningful signal from a noisy real-world environment is fundamentally an analog problem.”

Being face-to-face with customers at the conference was new to her.

“What struck me most was seeing how people react to the technology,” she says. “When you’re designing a block, you’re focused on specifications and margins. At CES, after explaining what the product does, I had a better understanding of why it matters to them.

“That connection between the circuit work and the final product is something you don’t always see from the design bench.”

Since the CES, Ashwath has presented other technologies she’s worked on. At the annual National Association of Music Merchants show in Anaheim, Calif., she demonstrated the A2B Automotive Audio Bus to the audio industry. The bus uses a low-cost unshielded twisted-pair cable to carry multichannel digital audio, control data, synchronized clock signals, and power in fewer than 50 microseconds, according to Analog.

“IEEE connects you with academia, which you don’t get much exposure to once you’re in industry. That connection between industry and academic work is valuable in both directions.”

At the company’s Global Technology Conference—an internal event held in Boston that draws engineers from the company’s more than 50 sites around the world—Ashwath led a demonstration of its solutions for augmented reality smart glasses.

“Each of these events has a different audience,” she says. “Learning to explain my technical work to different people has become a skill I’ve acquired.”

The most rewarding part of her job, she says, is seeing the end result of her work.

“You spend months on a design, and there’s a long wait between finishing it and finding out whether your instincts were right,” she says. “When the measurements match what you simulated, that’s the most satisfying moment in this job.

“The other part is knowing my work ends up in a real product. The circuits I design go into devices people actually use every day.”

Giving back to the community

Like many engineers, Ashwath joined IEEE to access research published in the IEEE Xplore Digital Library. It keeps her current in a field that keeps moving, she says.

During her first year of membership, in 2021, she discovered other perks including networking and educational opportunities.

“IEEE connects you to a community outside your own company,” she says. “Attending lectures and technical talks exposes you to approaches and problems you wouldn’t encounter in your own work.”

She eventually began seeking out volunteer opportunities, specifically with IEEE conferences and journals, because they are “how you stay current with what’s happening in circuits and systems beyond your own day-to-day work,” she says.

“As an analog designer,” she says, “being connected to that broader community matters. The field moves, and you need a way to keep learning from it.”

She serves as a peer reviewer of papers for several IEEE conferences including the IEEE International Symposium on Circuits and Systems and the IEEE International Midwest Symposium on Circuits and Systems. ISCAS is the flagship conference of the IEEE Circuits and Systems Society. MWSCAS is the oldest U.S. conference on the technologies.

“Someone has to assess whether submitted work is technically sound and worth presenting,” she says, “and that only happens if practicing engineers are willing to put in the time.”

The combination of learning from the community and contributing to it is what makes her membership worthwhile, she says.

“Analog design is a field where you’re always learning. Just when you feel you understand one architecture, there’s another one to learn,” she says. “Being an active learner isn’t optional in this field; it’s the requirement, and being part of the IEEE community helps with that.

“IEEE also connects you with academia, which you don’t get much exposure to once you’re in industry.

“Reviewing papers means reading what researchers are working on, which is often quite different from how the same problems get approached in a product environment. That connection between industry and academic work is valuable in both directions.”

This article is crossposted from IEEE Spectrum’s careers newsletter. Sign up now to get insider tips, expert advice, and practical strategies, written in partnership with tech career development company Parsity and delivered to your inbox for free!

If you’re an engineer, you may feel like your skills aren’t commanding the premium they once did. Many engineers have faced layoffs or restructuring. But even for those of us who still have jobs, something corrosive is happening. You may have heard some version of this:

“ChatGPT answered that in five seconds. Why can’t you?”

Or, “I made this app in a weekend. Why is our team taking so long?”

I worked on a team where leadership said things like this directly to engineers. This was in 2024 and 2025, when the AI craze was in full swing and Claude and ChatGPT adoption was rising.

Demo-Driven Development

Generative AI (and the marketing behind it) gives people the illusion that the work is simpler than it is. A working demo hides complexity, and once someone has a website that works, they may decide they no longer need engineers, or start to distrust the ones they do have.

Some CEOs want to believe they can replace an engineering team with a few sexy prompts. Many companies also want them to believe this.

That website works great in the demo—under specific conditions and without users with thumbs for fingers. Then it has to plug into a system with millions of lines of old code, satisfy legal requirements, meet accessibility standards, and be delivered in a complex deployment process.

Thinking that writing code was the only bottleneck for deploying new software is like the Titanic seeing the tip of the iceberg and assuming there was nothing underneath it.

What you can do about it

Engineers have a unique position in the AI space, because you have the capacity to understand how these models work much better than the average person. Use this to your advantage.

Learn how AI models work. If you can’t already explain that LLMs are probabilistic, what tokens are, or why training data matters, spend a weekend fixing that. You don’t need to become a full-blown AI researcher. A few fundamentals will put you ahead of most people making the AI decisions.

Teach upward. Turn that knowledge into teaching moments for the people who lead you. Explain the limits of these tools, and why a demo and a production system are different animals. It keeps people from truly believing that a notebook full of prompts can replace an engineering team.

Consider the criticism of today’s processes. Some of what we do has been treated as ceremony for too long. Be willing to say, “you know what, maybe you’re right, this process is too slow for the speed we can work at now.” Conceding that earns you the credibility to defend the steps that matter.

I still believe that speed at the cost of everything else is harmful—but that’s a topic for another time.

Find your voice. Engineers at the forefront of AI adoption need to push back and explain why we do things the way we do. Nobody else in the room is going to do it for us.

Don’t chase (too many) shiny objects

Your goal, in my opinion, is to save your organization from buying into the hype so much that they chase the wrong thing. Our job, before and after AI, is building safe and useful products for human beings, and engineers are still best suited for that.

—Brian

We Need to Embed Ethics Into Our Engineering Courses

The need to educate engineering students about ethics and social responsibility is widely recognized, but today’s curricula often miss the mark. Three professors propose a new approach: adding ethics training directly to core technical coursework.

Read more here.

The U.S. Just Bet $1 Billion on Quantum Chip Manufacturing

Looking for a career in quantum computing hardware? An IBM spin-off called Anderon could receive up to $1 billion in funding from the U.S. CHIPS and Science Act for its quantum foundry, established earlier this year. Could Anderon do for quantum hardware what TSMC did for conventional chips?

Read more here.

Meet the Anthropic Engineer Keeping Its AI Agent Trustworthy

When Gurmeet Saran joined Anthropic’s technical staff in 2025, he entered a role focused on a key question: How do you know whether you can trust a chatbot like Claude? Learn more about Saran’s path and his work at the AI company.

Read more here.

When San Antonio–based software engineer Nikhil Jathar first went to Washington, D.C., to advocate for federal support for technology education and research funding, he was in awe of the opportunity. It was spring of 2025, not long after Jathar settled in the United States from India, when he got an invitation to join the annual Congressional Visits Day organized by IEEE-USA.

“My first reaction was, ‘Wow, I get to visit Congress,’ ” says Jathar, CTO of AvanSaber, a business software company he cofounded in 2014. At the time, he did not have a specific lobbying position on behalf of the company. But he was intrigued by the invitation—which goes out to all U.S. IEEE members—and as he read more about congressional visits, his motivation deepened. He realized: “I need to try this. I want to see where it goes.”

What’s Your Ask?

Congressional staffers expect clarity from advocates. Be clear on what you plan to ask for when you arrive. Here are some common options:

• Cosponsor (or oppose) an existing bill or amendment.
• Preserve, remove, or modify legislative text in a bill under consideration.
• Add language to an appropriations bill or committee report.
• Ask a federal agency a specific oversight question.
• Request a briefing, hearing, Government Accountability Office review, or stakeholder roundtable.
• Encourage an agency to clarify a technical rule.

Congressional staffers say they welcome the knowledge and perspective that engineers like Jathar and other visitors from beyond Washington, D.C., provide. It’s also important for those visitors to be clear on what they are advocating for. One staffer for the House Committee on Science, Space, and Technology, who agreed to speak anonymously due to the office’s media policy, says, “Ideas for new policies come from multiple places. It could begin at the [legislator] or staff level, an ask from a constituent, or needs from industry. It’s up to us to figure out what would do the most good,” the staffer said. “For instance, if you leave it to the lobbyists, they’re understandably going to do what’s in their clients’ interests.”

For engineers in the United States, speaking with elected representatives is an effective way to shape key policies, but it takes some preparation and the right approach, as Jathar learned over the course of two visits.

Mapping Your Way Into Washington

Congress’s job is to act on behalf of the people, and it has rules for how people can lobby it. It distinguishes between broader advocacy—conversations between constituents and U.S. representatives—and professional lobbying, which involves paid professionals and is governed by the Lobbying Disclosure Act. Many professional groups like IEEE also organize congressional visits for their members. Those visits can be subject to lobbying rules, so the organizers told Jathar that: “You are representing the organization, not your own views.” He would need to stick to the talking points IEEE-USA had prepared for all its participants, regardless of their specialty, to promote across several congressional offices.

On Jathar’s first visit, those organizational talking points included advocating for funding smart grids, authorizing Department of Energy research, and investing in STEM education. That represents one of the trade-offs of access in Washington. IEEE-USA’s priorities didn’t all overlap directly with Jathar’s work, but they were broad enablers of science and technology and, he says, “STEM education funding is close to my heart.”

Anyone can approach their representatives in Washington with their personal priorities, Jathar notes, but given how pressed for time congressional staff are, they are more likely to prioritize a meeting that represents the views of as many of their constituents as possible. “I could have gone outside the IEEE structure,” Jathar says, but congressional staff “are more likely to listen if you are part of a much bigger organization.”

For Jathar’s first congressional visit, organizers assigned him to call on the offices of several representatives from both political parties based on where he lives: Representatives Greg Casar (D-Texas), Lloyd Doggett (D-Texas), and Barry Moore (R-Ala.). “Meeting a staffer or a senior member from their office is good. They tend to relate to these issues more than a representative, because a representative has 10 different high-level things they have to worry about whereas the staffer can be more focused on the policy details,” Jathar says.

A typical visit like this may last anywhere from 15 minutes to an hour, so Jathar kept his talking points simple and short. “First and foremost you’re gonna get a very small window,” he says, “Make sure you’ve got everything ready to present in bullet points and not a lot of jargon. You can always bring that out later if they ask more questions.” Simple visual aids, in the form of one-page handouts, may be helpful for communicating numeric or technical information, too.

“If you leave it to the lobbyists, they’re understandably going to do what’s in their clients’ interests.”–Staffer for the House Committee on Science, Space, and Technology

The House Science Committee staffer agrees: “I prefer meetings where people know what they want and know what we’re doing in Congress.” In addition to having a clear request, the staffer says engineers and other advocates should keep track of where Congress is in its cycle: Is it appropriations season, and when is the next budget being set? How far along is a given piece of legislation in the process, and what is realistic to change at that point?

During Jathar’s 2025 visit to one Capitol Hill office, he learned firsthand the importance of preparation. For one of the three representatives he visited, staff politely pointed out that the legislator they worked for was not on the committee with jurisdiction over the IEEE-USA talking point they were discussing. “They explained that there are only certain things [the legislator] can do on that issue and that there are other people we need to talk to, to achieve more,” Jathar recalls.

Shortcuts to Influence

Earlier this year, Jathar decided to go back to Capitol Hill with his newfound knowledge, again within the framework of an IEEE-USA visit. This time he was able to provide something else legislative staff want from engineers: authoritative technical knowledge.

“When I introduced myself as a software developer to one staffer, his first question was, ‘Do you think AI will replace coders?’ ” Jathar says. The conversation went back and forth, with the staffer asking Jathar and his other engineering companions questions. Jathar went on to explain that his company developed an open-source business-management software it calls ERPClaw, designed to be used with AI agents such as OpenClaw. That led naturally to discussing the role of open source in AI, which plays a role in one of IEEE-USA’s proposals for democratizing AI. This time Jathar was able to align his organization’s advocacy with his expertise and the staffer’s areas of interest.

It’s too soon to know what legislative impact that advocacy might have, but it’s the sort of informed, relevant proposal that legislative staff seek out. “The people who build trust have more cachet with congressional staff because they’ve asked for something that’s reasonably thought out,” the House Science Committee staffer says. In turn, congressional staffers can advocate better within their offices for policies brought to them by people they trust.

And talking about the issues you care the most about can deepen that trust. The House Science Committee staffer once visited Congress as an advocate before switching to the other side of the table. On that visit, as on Jathar’s, the professional society organizing the visit provided a pitch. “I was making a pitch that was given to me without a lot of personal buy-in,” the staffer says. Instead, “you should feel emotionally connected to any issue you’re advocating for.”

Jathar’s support for the open-source movement is like that. Open source gives smaller companies a foundation they can adapt and build upon, while allowing other developers to extend that work. “We can progress a lot faster,” Jathar says. “I’m not doing it only because it’s good for my business, but because I feel it’s good for any technological advancement.”

IEEE Life Fellow Manfred “Fred” J. Schindler has been elected as the 2027 IEEE president-elect. He will begin serving as president on 1 January 2028.

Schindler, who was nominated by the IEEE Board of Directors, received 16,421 votes in the election. Senior Member David Alan Koehler received 13,874 votes.

Schindler’s Top 3 Goals as President-Elect

1. Strengthen IEEE’s connection with industry by coordinating events across geography, industry, and technology to revitalize professional activities, especially ones for young professionals.
2. Champion the improvement of IEEE’s digital infrastructure to eliminate manual processes and modernize interfaces.
3. Prioritize the advancement of technology to benefit future generations by expanding and supporting IEEE initiatives and programs.

At press time, the results were unofficial until the IEEE Board of Directors accepts the IEEE Teller’s Committee report in November.

Schindler, an expert in microwave semiconductor technology, is an independent consultant supporting clients with technical expertise, due diligence, and project management.

Throughout his career, he led the development of microwave integrated-circuit technology, from lab demonstrations to high-volume commercial products. He has numerous technical publications and holds 11 patents.

Schindler served as CTO of Anlotek, and director of Qorvo and RFMD’s Boston design center. He was applications manager at IBM, engineering manager at ATN Microwave, and a lab manager at Raytheon.

An IEEE volunteer for more than 30 years, Schindler served as the 2024 vice president of IEEE Technical Activities and the 2021–2022 Division IV director. He was chair of the IEEE Conferences Committee from 2015 to 2018 and president of the IEEE Microwave Theory and Technology Society (MTTS) in 2003. He received the society’s 2018 Distinguished Service Award and its 2015 N. Walter Cox Award. His award-winning micro-business column has appeared in IEEE Microwave Magazine since 2011.

He also led the 2025 One IEEE to Enable Strategic Investments in Innovations and Public Imperative Activities adhoc committee.

Schindler is an IEEE–Eta Kappa Nu member and a MTTS honorary life member.

Ellia Lebleu is a founding member of the WikiProject AI Cleanup, a volunteer effort launched in 2023 to find and remove AI-generated material from Wikipedia that breaks the website’s guidelines. When the cleanup began, AI writing was often betrayed by conspicuous clichés. Today, the project’s volunteers are developing tools to detect subtler forms of machine-generated content and debating AI’s role in the world’s most widely used encyclopedia.

When did you realize Wikipedia had a problem with AI-generated material?

Ellia Lebleu: It was around December 2023. A few other editors and I were chatting on Wikipedia’s Discord server, and we had all seen very strange articles with the same odd characteristics. They repeated phrases like “stands as a testament” or “rich cultural heritage.” They sometimes had conclusion sections, or sections called “Why is this important?” That was very typical of ChatGPT in 2023 and early 2024 but very odd for Wikipedia articles—and very easy to spot. Then I decided, let’s create an official project on Wikipedia so more people can deal with this.

Ellia Lebleu

Ellia Lebleu is a founding member of the WikiProject AI Cleanup. Lebleu is pursuing a Ph.D. in AI at the Inria Center of the University of Rennes, part of France’s national research institute for digital science and technology.

How do editors spot AI-generated writing today?

Lebleu: By 2025, the signs of AI writing were becoming subtler and more structural. For example, AI loves the “rule of three,” negative parallelisms, and other structures of speech that aren’t conclusive by themselves but become meaningful as they accumulate. We also look beyond the article. If someone writes five articles of thousands of words each in 5 minutes, there is reason to suspect AI.

Does the AI problem just appear in the main text, or also in citations?

Lebleu: Fake or misleading citations have become a major worry. Sometimes an AI hallucinates a citation entirely, although that’s much rarer today because of retrieval-augmented generation that allows the AI to find sources. More often, it finds a real source but gets its content completely wrong. AI citations that don’t actually support what is being written are extremely commonplace.

English Wikipedia recently prohibited editors from using LLMs to generate articles. How much AI-generated content is still getting through?

Lebleu: I’ve been working with [AI-detection company] Pangram to estimate the proportion of AI-generated content. I tested it on Wikipedia content from before ChatGPT and found no false positives, so its results provide a very conservative lower bound. In late 2025, it flagged perhaps 20 percent of new articles and drafts as AI-generated. In early 2026, that went down to 15 or 16 percent. But we don’t yet know whether that’s because our recent policies have had an effect or because models are getting more powerful and Pangram hasn’t had time to adapt.

Is Wikipedia using AI to combat AI?

Lebleu: Another Wikipedia editor has made an AI source verifier, which goes through the sources on a page and checks whether they match the claims. Since it is itself an AI model, there is no guarantee that it will be correct every time. But if you want something that works at scale and filters out obvious cases, it’s a good first step. The Wikimedia Foundation is also working on a tool called Paste Check, which can flag text pasted directly from some AI chatbots.

And we’ve seen Claude sometimes say, “I can’t write a Wikipedia article; it would go against Wikipedia’s policies.” If all major AI companies implemented that kind of prompt refusal, it wouldn’t solve the problem entirely, because there are still open-weight models that can be tweaked to avoid restrictions. But it would help with people who use commercial models and aren’t especially familiar with the technology.

Can you tell if the rule prohibiting AI writing is working?

Lebleu: It’s still a little early to tell. We see possible effects, such as prompt refusal and the apparent slight drop in AI-generated content. But again, we don’t yet know if that is because models are getting better or because beginners are learning the rules. The vast majority of AI-generated content is submitted by beginner editors, so we’ve made many efforts to make the rule very visible in welcome messages, editing help, and advice.

What if AI becomes capable of writing truly reliable Wikipedia articles?

Lebleu: That’s one of the big questions we’re facing. When the guideline was passed, it was for very practical reasons: We were flooded with AI content, we didn’t have good tools, and we needed to put a stop to it. But since then, it has taken on a more philosophical dimension. Many people have called Wikipedia “the last good place on the internet.” Being a place that was, in theory, free from AI slop has become part of Wikipedia’s identity.

A lot of Wikipedia’s communication is now centered on the idea that we’re a human-written encyclopedia. Even AI companies find that interesting. It can be valuable to have a purely human-written corpus on which to [train] AI systems. The fact that Wikipedia is still that, and will remain that for at least the near future, makes it a precious resource that isn’t replicated anywhere else.

But will the people who read and contribute to Wikipedia want it to become partially AI-generated if that makes it more comprehensive? We could preserve Wikipedia’s identity as a human-written corpus while also offering an AI-assisted option.

Has Wikipedia reckoned with the idea of “Google Zero”—a possible future in which websites no longer get traffic from Google search?

Lebleu: If everything goes through AI search and websites plummet to zero visibility, how do we expose Wikipedia and show that it can be edited and that people can still contribute? Wikipedia is frequently cited in [Google’s] AI Overviews, but users may only see an AI answer citing Wikipedia. They may not realize that Wikipedia is a living project they can edit and contribute to.

The Wikimedia Foundation is working on an attribution framework that could give AI users more information about the article [being cited]: when it was last edited, how comprehensive it is, whether editors have flagged problems, and perhaps even a call to action. In the future, readers might even contribute through AI models: “The AI cited Wikipedia, but there seems to be an error. Can I make an edit request and add this other source I found?”

I can see a very strong symbiotic relationship between Wikipedia and AI models in the future. I don’t see Wikipedia disappearing, but I can see it taking a very different shape. Yesterday, AI was something to clean up. Today, it’s a tool to help us with the cleanup. Tomorrow, it will be the air we breathe.

A condensed version of this article appears in the November 2026 print issue as “Ellia Lebleu.”

Advancing technology for the benefit of humanity isn’t just the tagline for IEEE’s mission as a public charity. It’s something that the organization’s more than 500,000 members strive to do all the time. IEEE’s recently released 2025 Social Impact Report showcases a variety of programs volunteers worked on to increase public awareness of technology’s impact on society, many of which The Institute covered.

Here is a selection of the programs from the report, with links to our related articles:

• Shaping trustworthy digital experiences for children online. An IEEE collaboration with Indonesian policymakers led to a new regulation to protect children, a first of its kind in the region. The advanced frameworks used for the design of Internet platforms and systems are based on the IEEE 2089-2021 and 2089.1-2024 standards. The IEEE 2089-2021 Standard for an Age-Appropriate Digital Services Framework Based on the 5Rights Principles for Children provides practical steps to qualify online products and services. The follow-up, the IEEE 2089.1-2024 Standard for Online Age Verification, provides a framework for designing, specifying, evaluating, and deploying age-verification systems. You can learn more about the two standards in “IEEE Makes Strides to Improve Online Safety for Kids.”
• Catalyzing humanitarian outreach. Last year IEEE launched its Mobile Outreach Vehicle (MOVE) program in India. The vehicle brings educational resources to communities and can provide communication and technology support to relief workers and survivors during emergencies. In 2024, for example, MOVE helped survivors of Hurricanes Helene and Milton in the Southeastern United States. When the vehicles are not deployed, IEEE MOVE remains engaged in outreach efforts, participating in 25 STEM educational events such as one in San Diego in 2023.
• Preserving Tech History. The IEEE Global Museum, developed by the IEEE History Center, presents traveling exhibits to educate the public about the history of technology and its impact on society. In “IEEE’s Global Museum Brings Engineering History to You,” you can learn about current exhibits “Unseen Signals: E. Howard Armstrong’s Radio Revolution” and “Microchips That Shook the World.”
• EPICS in IEEE. This program provides nonprofit organizations with technology solutions while broadening undergraduate students’ hands-on experiences. Last year US $290,000 was awarded to a record-breaking number of service projects around the world. Among the 48 projects was one in which a student team in Macedonia created IoT-driven sensing and display devices to help measure and promote awareness of pollution and air-quality issues in the capital city of Skopje.
• IEEE TryEngineering. This initiative inspires preuniversity students to pursue technology and engineering careers. It awarded US $70,379 in grants to 58 programs around the world. One is its TryEngineering Summer Institute, a nine-day summer sleepaway camp for students ages 13 to 17. More than 310 students participated last year, learning about aerospace and artificial intelligence.
• IEEE–Eta Kappa Nu. IEEE-HKN has more than 260 collegiate chapters worldwide and more than 200,000 members. The honor society encourages and recognizes excellence in education, professional practice, and meritorious work within IEEE’s designated fields of interest. IEEE-HKN members last year contributed 110,442 volunteer hours to universities and communities, and they held 827 service and outreach events and 137 STEM outreach activities. Last year IEEE and HKN celebrated the 15th anniversary of their merger.
• IEEE Connecting the Unconnected Challenge. This annual global competition solicits innovative solutions from startups, grassroots organizations, universities, and others working to bridge the digital divide. The Institute wrote about last year’s winners. They included a solar-powered community broadband network in Tanzania, a low-cost method for accessing the Internet that uses FM radio and a short message service, and a strategy for utilizing India’s rural broadband infrastructure to deliver medical services to people living in isolated, tribal, and other underserved regions.
• Supporting Sustainable Climate Efforts. IEEE helps foster the exchange of technical knowledge and information to develop technologies for a sustainable climate. “IEEE Plays a Pivotal Role in Climate Mitigation Talks” reported how the organization enhanced its standing as a trusted, neutral authority on the role of technology in climate change mitigation and adaptation. Leaders, including 2023 IEEE President Saifur Rahman, showcased tech solutions at COP30 and the International Symposium on Achieving a Sustainable Climate. Also, “IEEE Young Professionals Talked Sustainability Tech at Climate Week NYC” reported on the group’s activities during the 2024 Climate Week NY, the largest annual climate event, featuring more than 600 activities throughout New York City.

Engineering: A force for good

“The IEEE global community of more than 500,000 members and volunteers demonstrates what is possible when expertise is mobilized for the public good,” Sophia Muirhead, IEEE executive director and chief operating officer, said in a news release about the Social Impact Report. “Technology has the power to expand opportunity, strengthen communities and improve lives, and IEEE’s mission is to ensure that this power is used responsibly, ethically, and in service of humanity. The 2025 report shows how IEEE members and volunteers are advancing solutions that make a real difference in people’s lives around the world.”

Engagement research shows that members increasingly cited “giving back to my profession and the world community” as a reason for joining IEEE and renewing their membership, IEEE President Mary Ellen Randall said in her September column about the report.

“Humanitarian programs are not only meaningful,” she wrote. “They also are magnetic. They attract younger engineers, keep them engaged, and help them build a professional identity rooted in purpose and impact.

“IEEE has the opportunity to become the world’s leading convening force for developing interdisciplinary technology solutions to solve humanity’s most important challenges. Our future relevance will be defined not only by the technologies we advance but also by the problems we choose to help solve.”

Agility Robotics, the company I cofounded out of my robotics lab at Oregon State University in 2015, has just launched Digit 5. It’s designed to be the first humanoid robot that can safely and economically do useful work at scale, allowing us to bring tens of thousands of humanoid workers into the real world. Engineered to lift 23-kilogram bins onto high shelves and with a battery that can run for 90 minutes on a 9-minute charge, Digit 5’s goal is to handle heavy warehouse tasks shift after shift, filling the need for labor in jobs that are too physically stressful or too dull for humans to want to do.

Digit 5 is what Agility has been working toward for over a decade. But the robot is heavy and bulky, lacking the graceful natural motion of its predecessors, and it’s farther than I thought possible from the robot I dreamed of building when I started this company. To deliver a useful commercial product, Agility has taken a path that is very different from what I first imagined.

From ATRIAS to Digit

My dream has always been to build robots that combine locomotion and manipulation to move through the world and interact with it as people do. I seek a mastery of motion, achieved through devices designed to navigate the constraints of physics with elegance. When I started the Dynamic Robotics Laboratory at Oregon State in 2008, we began with a theory derived from decades of work in biomechanics: that a simple “spring-mass” model (a body supported by legs that act like springs) should be able to reproduce every known walking and running gait found in animals.

We put theory into practice by building the bipedal robot ATRIAS—short for “Assume The Robot Is A Sphere,” a joke among physicists about reducing a complex system to its simplest possible form. ATRIAS looked like a large microwave on stilts because it was only intended to demonstrate the physics of human walking and running. This was the first glimmer of achieving my dream; even though ATRIAS wasn’t designed to look like us, it still seemed natural, and even somehow alive, when it settled into a walking gait.

With the success of ATRIAS, the barriers to progress shifted from scientific to engineering and organizational, and a startup company seemed like the way to bridge the gap from a one-off research concept to an actual robotic product. I felt confident of two things: that I would be able to build robots that could go where people go, and that this capability would be incredibly useful. This combination of ambition and naivete is necessary for anyone to found a new company—how hard could it be?

Good advice for starting a company is to understand a customer problem, and then find a way to solve it using your unique capabilities. My cofounders, Damion Shelton and Mikhail Jones, and I did not do this. We relied on our belief that this technology would enable new possibilities and trusted ourselves to figure out the business as we went. So we simply focused on building our first product: the bipedal robot Cassie.

Cassie was the first dynamic bipedal robot that researchers could buy, able to maintain its balance through active, fluid motion rather than careful, rigid steps. It had ostrichlike legs, not because we were imitating ground-running birds, but as a result of engineering strictly for performance. It was not intended to be commercially useful, and had no torso, arms, or head. We sold Cassie robots to Harvard, Caltech, the University of Michigan, the University of California, Berkeley, and a handful of other research institutions, ultimately saturating the market for US $250,000 bipedal research robots after having sold a total of 10 of them.

ATRIAS, Cassie, and Digit 4 EVT [left to right] are on display at Agility Robotics’ RoboFab in Salem, Ore. The 70,000-square-foot facility has the capacity to produce more than 10,000 Digit 5 robots per year.Agility Robotics

Because Cassie’s hardware enabled research into the physics of walking and running, the legged-robot research frontier shifted to control software. Traditionally, engineers controlled legged robots with detailed mathematical models that determined how the machines should move. With Cassie, researchers pioneered the new approach of reinforcement learning for bipedal legged locomotion: A simulated robot learns through trial and error, and the resulting neural network controls the physical machine. This approach is what enables virtually all of the dynamic behaviors that you see on legged robots today, and Cassie looked alive as it recovered from stumbles while climbing stairs, ran 5 kilometers across the Oregon State campus, and broke the robot world record for a 100-meter dash.

Cassie got us closer to being able to move through the world with confidence and grace, but to make it useful, we needed manipulation. After exhaustively exploring different configurations, we found that an upright torso helped the robot balance, while two arms mounted near the top of the torso allowed it to pick up large objects and to catch itself when it fell. Digit 1, which added a torso and arms to the Cassie legs, was our first mobile manipulation robot, and another step on the path toward a product that could do useful work for commercial customers.

Why Digit Began With Warehouse Bins

As we explored the first applications for Digit, potential customers kept asking us: Why legs? And why a humanoid? These were tough questions for me to answer, because they exposed a core business reality: It didn’t matter to Agility Robotics if we solved a cool science problem in a beautiful way. What mattered to the business was solving real problems for customers.

After years of engagement with industry partners, we found one answer in warehouses and manufacturing facilities: moving bins. Bin handling is a clearly defined task where two arms are needed to lift large bins up high and legs are the best way to balance and move in narrow spaces. Because the task can vary widely from one deployment to another, it hasn’t been broadly addressed by single-purpose automation. Yet it was constrained enough to give Digit an economically viable first job while it developed a broader set of skills.

In 2019, Digit began demonstrating its capabilities to do real work. With Ford Motor Company, we showed how Digit 1 could carry packages from an autonomous vehicle to a front porch. We developed Digit 2 and 3 while working with a certain global package delivery company on moving and sorting cardboard boxes. Amazon tested Digit 4 on moving and stacking bins, and it became the world’s first commercially deployed humanoid at the logistics provider GXO. In 2024, Digit appeared on the cover of Time magazine.

Despite Digit’s apparent success, we began to run into conflicts between my dream and the needs of actual customers. We were no longer working with relatively patient university research labs. Agility has to sell robots to survive, and customers will only purchase robots that provide value. If bins weren’t moving, our customers were losing money. We spent a huge amount of time and effort keeping our first deployments operational, which arguably eliminated any profit we made. Yet this also kept our company learning and alive the way no lab demo ever could.

I had no idea how hard it would be, or that running a company would involve so many things that have nothing to do with solving technical problems: deployments across the country; shipping, service, and customer support; fundraising and investor relations; more fundraising; human resources and IT; developing fleet-management software; managing a supply chain; and even more fundraising. I had no experience with any of these things. My career had focused on solving the technical challenges of robots interacting with the world, and it turned out that was only a tiny piece of building a successful robotics company. But Agility was the only path I saw to continuing the technical work and having an impact.

Turning Digit Into a Product

I knew that Agility needed to focus more on robots as a product, but I didn’t know how to make that happen. Hiring the people we needed to run the company meant giving up an immense amount of control, but we had no other choice. We recruited a new executive team with the experience required to commercialize this new technology, and they quickly made changes that were both painful and necessary.

Digit versions 1, 2, and 3 [left to right] show the evolution of the robot from a dynamic walking platform into a useful mobile manipulator. Agility

We had intended the next revision of Digit, Digit 4 EVT (Engineering Validation and Test), to be a relatively minor upgrade. But Agility’s new leadership insisted that the robot was missing key features that were essential if Digit was going to become a real product. So Digit 4 EVT instead incorporated a series of major changes in a highly compressed timeline, such as larger batteries and system-wide upgrades for safety and regulatory compliance, which added enough mass to dramatically alter the robot’s appearance and capabilities.

While early Digits were headless, Agility Robotics was able to justify a head for Digit 4 to optimize antenna and camera placement, and it turned out to be surprisingly useful for human-robot interaction.Christie Hemm Klok

I could only watch as others brought their own talents to bear to make Digit into a successful product, in gut-wrenching ways that didn’t always align with my vision. There were some especially difficult times when I felt like there were compromises made that didn’t seem necessary, but making those decisions was no longer my sole responsibility. I had to accept that Digit had grown far beyond me, and look for ways of influencing decisions rather than directing them. In the moment, it was incredibly frustrating, as Digit 4 EVT took a few steps backward from my vision of how humanoids should look and move.

Agility is in our name because agility is core to the future of the company. I want Digit to be lithe and graceful. When people see Digit, I want them to like the robot, to understand what it can do at a glance, and to feel that it’s there to help. But with Digit 4 EVT, all of that was sacrificed to functionality, because of course it had to be—the customer just needs the robot to do the work. The design process became a ruthless triage, with aesthetics and elegance of movement being delayed over and over again in favor of brutal functionality. For an engineer whose mind lives more in the future than in the present, it has been hard for me to accept this trajectory. And yet, through Digit 4 EVT’s commercial deployments, we began to really understand specific applications and what they required. We mapped the first 10 use cases Digit would perform, and then began designing Digit 5 from scratch based on everything we had learned.

Digit 5

Digit 5 will be Agility’s first real step toward a general-purpose robot. It marks our transition from imagining what a future humanoid robot should be to designing a product around the realities of customer deployments. These realities include things like European CE regulatory compliance, North American UL electrical-safety certifications, and Federal Communications Commission limits for electromagnetic radiation—all of which felt tangential to my dream for Digit, but were absolute requirements.

Digit 5 is designed to be the first humanoid robot that can operate safely next to human coworkers while working 20 hours a day and lifting more than a human can.Christie Hemm Klok

Customer requirements helped us realize that safety in particular was one of the most critical barriers to scaling. In Digit 4’s commercial deployments, the robots had to work behind physical barriers, separated from humans at all times, to meet insurance and regulatory demands. This seemed silly to me, because I’ve never felt that our robots were dangerous, but we have to prove that they’re safe. And since keeping the robots in safety cages severely limited their usefulness, enabling Digit 5 to operate safely around people became a primary design goal. We spent two years on this challenge.

Digit 5 is designed to be the first humanoid robot that can operate safely next to human coworkers. It will rely on an array of sensors to locate nearby people in a manner that’s certifiably safe, and then take actions based on that information. As a person approaches, Digit 5 is programmed to stop what it’s doing, set down its payload, sit on the floor, and cut power to its motors before the person gets close enough to touch the robot. This conservative approach reduces the risk to nearby people in a verifiable way, and we believe it’s the only approach that makes sense as a first step toward working alongside humans.

Yet this strategy also feels like a massive slowdown on the path toward robots becoming part of everyday life. Enabling coexistence with humans has influenced almost every aspect of Digit 5’s design: the shape of the legs and the body, the safety-rated sensors within protective housings, the isolated computers dedicated to safety, and more. All of these features increase the robot’s complexity, power consumption, mass, and physical size.

Digit 5 is Agility’s largest, heaviest, strongest, and most capable humanoid robot worker.

The robot’s design was driven by safety, with integrated protection for the robot and for the humans it will work around.Agility Robotics

Weighing over 120 kg, Digit 5 is a huge robot. It needed to achieve the lifting limits recommended under the U.S. National Institute for Occupational Safety and Health guidelines for human workers, which include repeatedly lifting 23-kg bins and reaching shelves 1.75 meters high. Digit is designed to perform tasks like these for 20 hours a day—something no human can do. This requires motors with enough power and torque to lift the load, but also enough precision to delicately control the force Digit applies when it makes contact with the world. To power those motors, we needed a large battery and a dock for rapid autonomous charging. The dock will deliver 240 amperes at 48 volts, which is enough current to make its thick charging cables visibly wiggle under the induced magnetic forces, while powerful fans blow air through cooling channels in the robot’s torso to keep the battery from overheating. The battery and cooling system takes up about 75 percent of Digit’s torso, but it will allow the robot to operate for 90 minutes and fully recharge during a 9-minute break, enabling near-continuous work.

All of this extra bulk means that Digit 5 can no longer run or jump the way previous generations could, and the svelte form I imagined has become physically impossible. In some ways, Digit 5 feels more like a forklift than the graceful humanoid I set out to build. But looking at the situation rationally, I believe this trade-off is necessary for any company that claims its humanoid can do this kind of work today: If you see a robot that looks like a human working in a factory or warehouse, it either can’t lift the required weight or can’t make it through the whole shift. And if it’s not covered with sensors, it’s likely not safe to stand next to.

The most visible change with Digit 5 is that humanlike legs replace the ostrichlike legs of our earlier versions. While the “backward” leg configuration was optimized for walking and running, we discovered that it created problems with squatting because the ankle will strike the ground. Digit 5 will spend a lot of time lifting heavy objects from near the floor, so we used a humanoid leg configuration with knee and hip motors that both contribute torque to help the robot stand up again.

When Digit detects an approaching human, the robot will lower itself to the floor and cut power to its motors. Agility believes this is the only near-term approach to verifiably reduce risk when the robot is working near people.Agility Robotics

Other companies building humanoid robots have copied the human form without ever understanding its purpose. It’s very frustrating that, to an outside observer, it now looks like we discarded Agility’s distinctive leg design and simply did what everyone else was doing. Yet I know that we arrived at this shape based purely on functional engineering choices. Making a robot look like a human is the easy part—without the function to match, and the understanding of the physics underlying that function, the whole point of the human form may be lost.

With the power, battery life, and safety features that we are designing into Digit 5, we will have a truly useful robot. But it also compromises on some things that are core to my vision and mission. I understand why we made the choices we did and can justify them all, but together they leave me unsatisfied. It’s the curse of the engineer.

What Comes After Digit 5

I believe Digit 5 will be the scaling moment for Agility Robotics, the inflection point where we can begin deploying robots in large numbers. While the robot will continue carrying bins, our road map includes many more tasks, from picking individual items and tending machines, to preparing parts for assembly and shipping, and eventually doing assembly work itself. Its interchangeable tool interfaces and modular design will enable us to gradually upgrade its subsystems. It will be a platform for improvement, a source of data for future behavior training, and a viable product all at the same time.

Hurst stands between the previous version of Digit, Digit 4 [left], and Digit 5.Christie Hemm Klok

But I cannot love Digit 5. It’s the largest, heaviest, most expensive robot Agility will ever build, full of design compromises that enable it to be exactly the robot it needs to be right now rather than the robot I dreamed of. In the beginning, I envisioned steady progress on agility and dexterity, with each version becoming more dynamically capable. Instead, developing Digit 5 has required large detours to incorporate requirements that seemed unrelated to my original vision. However, I now recognize that tackling these challenges was always necessary. Customers and product requirements didn’t ruin my vision; they revealed what my vision was missing. This is how robots will achieve commercial success and make a real-world impact, which will bring me closer to my scientific dream as well.

We won’t have a humanoid that matches the capabilities of humans until it achieves that mastery of motion that inspired me to found Agility over a decade ago. I saw glimmers of it with the spring-mass locomotion of ATRIAS, and more glimmers with Cassie’s running gait.

I know that such mastery is physically possible with the tools and hardware that we have right now, and despite all of its compromises, Digit 5 will give us a practical foundation for pursuing that dream. I want every future version of Digit to be lighter and more graceful, to interact more naturally with humans, and to be better at manipulating objects and moving through the world, enabling robots to enter aspects of our lives that they haven’t been able to help with before. By the time I’m old enough to need assistance at home, I expect Digit will be up to the task.

With Digit 5, Agility has built a robot that will be useful in our world. Now I’m getting back to work to make it beautiful, too.

This article appears in the November 2026 print issue.

This white paper explains how edge AI and physical AI help cities move beyond passive monitoring. It shows how urban infrastructure can sense, reason, and act on real-time conditions across public safety, mobility, and infrastructure management.

What you will learn about:

• How the shift from smart cities to cognitive cities depends on infrastructure that reasons and acts, rather than infrastructure that only monitors.
• Why edge AI supports low-latency decisions, keeps sensitive data within city boundaries, and reduces dependency on centralized cloud infrastructure.
• How computer vision, multimodal sensing, and generative and agentic systems work together to support public safety, mobility, and infrastructure applications.
• How a tiered edge architecture assigns monitoring, planning, and action tasks across the far edge, regional, and cloud layers, and what durability, security, and interoperability requirements this places on hardware and software.

This article was originally published by Canary Media.

America’s energy supply crunch is driving a surge of federal and private investment in new electricity generation—from polluting gas turbines to cleaner developments like solar farms and geothermal plants. But gigawatts of power are already available for the taking across the U.S. economy in the form of wasted energy from an unexpected source: motors.

Some 50 million electric motors drive the fans, pumps, and compressors found in virtually every factory and building. Together, they gobble enough electricity to equal nearly a third of total U.S. power demand. Often, the motors use far more energy than they actually need to keep America’s machines humming.

Companies can install devices known as variable frequency drives, or VFDs, to make motors run super efficiently. Yet adoption remains relatively low, in part because of the perceived hassle of retrofitting equipment and because they’re typically not required. On average, only 16 percent of motor systems at industrial facilities and 4 percent of commercial-sector motors use VFDs, according to a 2021 market report by Lawrence Berkeley National Laboratory.

However, there are now more reasons than ever to equip the nation’s motors with power-saving devices, energy experts and device manufacturers say. They’re urging federal agencies, utilities, and deep-pocketed tech giants to spur a wider rollout of VFDs through financial incentives and efficiency standards.

For one, slashing motors’ power use would free up grid capacity for large electricity users like data centers and electric vehicles, they argue. It would help curb costs for factories that are switching from fossil-fueled heating systems to electrified, motor-based solutions like heat pumps—all at a time when industrial power prices are climbing fast. Plus, VFD technology has become far more sophisticated and less expensive to install than earlier iterations.

“This is low-hanging fruit that gives a lot [of energy] back and is reasonably simple,” Nik Sawe, a senior policy analyst at the nonpartisan think tank Energy Innovation, says of the technology.

What are Variable Frequency Drives?

Energy-wise, the main problem with electric motors is that they’re always running at full capacity. Consider an office building, which empties out in the evening. The heating, ventilation, and air-conditioning system can wind down in those hours, with dampers and valves restricting the flow of cold or hot air. But the motors and fans that drive the whole setup are still working in the background, wasting energy.

VFD technology, such as the Danfoss iC7 series, has gained some momentum in recent years among factories and buildings trying to combat rising electricity costs. Danfoss Drives

VFDs are electronic controllers with temperature sensors that continually adjust the speed of motors, automatically ramping up or down to match demand. The technology, which is also called adjustable speed drives, has the potential to curb a motor’s energy use by around 40 to 60 percent, says Neal Elliott, director emeritus of the American Council for an Energy-Efficient Economy.

All those energy reductions add up. Toyota, for example, installed VFDs on over 200 motors inside the HVAC system at its auto manufacturing plant in Georgetown, Kentucky, as part of a US $1.25 million pilot project in 2015. The automaker saved more than $700,000 in annual electricity costs thanks to the VFDs, and it continues to use the technology at other manufacturing facilities across the U.S., the company said by email.

“We’ve known for many, many years that there’s a huge opportunity for energy savings by varying the speed of the motor,” says Elliott, who noted that he literally wrote the book on energy-efficient motor systems in 2002.

Electric Motors in Nearly Every Factory

Certain aspects of the technology have historically held the drives back since they debuted in the late 1960s. Measuring the flow of water, air, refrigerant—whatever is moving through the system that the motors are driving—was for decades an expensive engineering process. Accurately controlling motors to deliver the right amount of juice, at just the right time, faced similar issues.

But recent advances in sensors and machine-learning algorithms have significantly slashed the cost and complexity of using drives, Elliott says. Modern VFDs have also benefited from the increased production of advanced power chips that are used in high-voltage applications such as electric vehicles and solar power inverters.

Another long-standing challenge has been that commercial and industrial building developers often prioritize reducing upfront costs over netting long-term energy savings, and so they don’t always build to the highest efficiency levels, says Zack Valdez, senior director of regulatory and industry affairs at the National Electrical Manufacturers Association, whose members include drive and motor makers.

“If the facility that you’re putting it into doesn’t require it, then oftentimes people will overlook [VFDs], because they’re just trying to build to a minimum set of energy requirements,” he says. And if companies don’t install drives from the start, they’re likely less inclined to take time to retrofit an operating system.

The trade association recently launched a pilot program with manufacturers, utilities, and energy-efficiency groups to collect real-world data about how VFDs reduce motors’ electricity use and lower utility bills. The goal is to demonstrate how the technology can benefit not only individual facilities but also the wider grid—giving utilities a concrete reason to provide rebates for industrial customers.

VFDs Slash Motors’ Electricity Use

Proponents say they hope the results will help the U.S. Department of Energy shape future efficiency measures and incentives that encourage adoption of the drives, including in federally owned buildings. Despite the Trump administration’s open hostility toward home appliance and equipment mandates, agency staff have apparently been receptive to discussing solutions that could meet the soaring power needs of data centers.

“If you added a drive to all the motors that are out there, you could provide [power for] up to five years of data center growth without having to build any new generation,” says Kirk Anderson, director of government relations and public affairs for ABB. The global manufacturing giant makes VFDs domestically near Milwaukee.

Anderson floated the idea of having tech companies help pay for factory owners to install VFDs to free up room on the grid. After all, hyperscalers are already spending many billions of dollars to revive shuttered nuclear plants, support next-generation geothermal projects, deploy fleets of grid-scale batteries, and supercharge the build-out of gas turbines and power plants.

“Instead of ‘bring your own power,’ it should be ‘offset your own power,’” he says, referring to a growing effort by states to get data centers to secure their own power resources and prevent rising costs for regular customers.

Michael Reckamp, the director of enterprise strategy for Danfoss Drives, North America, another VFD manufacturer, said that despite the traditionally low adoption rates, the technology has gained some momentum in recent years among factories and buildings as rising electricity costs and other grid-wide challenges make it harder to justify wasting energy.

Still, “it hasn’t permeated enough to get everyone utilizing it,” he says. “In the current situation, where we’re having so many issues with grids, brownouts, and now data centers wanting to take the load, I think there should be a renewed focus on putting drives out there to really reduce the stress on the grid and improve energy savings.”

More and more city governments are using advanced technologies to improve livability for their residents. Municipalities use data from smartphones and sensors for traffic enforcement, traffic control, and waste management. They also have implemented artificial intelligence and deep learning programs.

Dubai, Paris, and Coral Gables, Fla., are among the cities that have invested heavily in programs to become smarter.

But along with the conveniences that come with a smart city transition, there have been challenges including safety and privacy issues and eroding relationships between residents and city officials.

Although many initiatives have delivered real innovations, not all have been successful. Inefficient global coordination has hindered projects with unclear standards, socioeconomic differences, and a lack of worldwide governance, leading to a disconnect between the technologies and the ability to adapt to them.

Citizens’ trust in the technologies suffers because of privacy breaches, personal data leaks, and other issues.

Finding solutions

Smart city projects are complex and expensive endeavors that involve politicians, citizens, technocrats, and engineers. Successful projects require citizens to be fully engaged and to have strong collaborations with the participants, especially engineers, as engineering is the cornerstone.

Collaboration with universities, research institutions, and technology experts can bridge knowledge gaps on the behavior of complex socio-technical systems. What’s missing, though, is a global platform where those involved can discuss their experiences.

To help solve those and other challenges, the IEEE Smart Cities Technical Community (formerly the IEEE Smart Cities Initiative) was launched in 2013 by IEEE Future Directions, which incubates emerging technologies and new applications of current ones. The technical community provides city leaders and others with credible, unbiased information and educational content developed by IEEE experts.

Since then, IEEE Smart Cities has launched services including its flagship International Smart Cities Conference, a media center with educational and technical resources, and a contest to recognize the best smart project completed by city officials. The platforms bring together practicing engineers, academic researchers, policymakers, and city leaders to collaborate and exchange ideas and experiences about moving toward a smart-city model.

Better serving humanity

In 2021 the IEEE Power & Energy Society began managing the Smart Cities program, which is supported by the IEEE Communications, Industry Applications, and Intelligent Transportation Systems societies and other IEEE entities.

Thanks to volunteers’ dedication and determination, IEEE Smart Cities launched several services. But the initiative’s structure limited the development of long-term projects and the creation of new activities.

To address the situation, the IEEE Technical Activities Board (TAB) created the second-generation technical community ad hoc committee, also in 2021. It was a pilot program whereby IEEE Smart Cities formed a board of governors to oversee the community’s operations and leadership, and created a policies and procedures manual.

In November 2025 IEEE Smart Cities became a TAB technical community.

IEEE Smart Cities now

Today more than 7,000 volunteers collaborate on various aspects of the community within at least one of the following technical tracks:

• Smart grid.
• Smart water management.
• Green electronics for smart cities.
• Smart governance and planning.
• Hydrogen for smart cities.
• Smart and sustainable campus.
• Security for smart cities.
• Smart lighting for cities.
• Smart mobility for cities.
• Artificial intelligence for cities.

New webinar series, workshops, conferences, educational resources, and other offerings are being created.

Two events were launched recently: the IEEE European Technology and Engineering Management Summit, held in May 2025 in Belgium, with the next one planned for 2027; and the IEEE Sustainable Smart Lighting World Conference, scheduled for 2 to 6 December in Athens.

The Smart Cities Ambassadors program aims to raise the technical community’s visibility outside IEEE by holding roundtables and other events at non-IEEE functions.

IEEE Smart Cities volunteers are working on launching journals. In addition, standards are being developed in collaboration with the IEEE Standards Association, societies, and councils.

The program is collaborating with the Open and Agile Smart Cities and Communities network, the Instituto de Competitividad e Innovación para el Desarrollo y Transferencia de Tecnología, and the International Project Management Association.

All of those actions have the ultimate objective of boosting IEEE Smart Cities’ growth and sustainability, as well as strengthening the leadership and educational offerings to technical community volunteers.

If you would like to volunteer, visit the Smart Cities website.

This article is brought to you by COMSOL.

These days, no one blinks an eye at two-day, one-day, and even same-day package deliveries. Despite the plethora of localized fulfillment and distribution centers that have cropped up to meet the demand, delivery trucks still need to make considerably long drives, often over highways. To reduce fuel and maintenance costs as well as carbon emissions, companies are investing in electric trucks for their delivery fleets. While turning to electric trucks is a promising solution, the battery packs incorporated into these automotive designs often fall short in the lifetime needed to make long-distance highway travel feasible.

To help overcome the issue of battery lifetime, a consortium of universities and industry partners incorporated multiscale simulation and a custom simulation app into the battery manufacturing process. The organizations involved in this project were Aurobay, Resolvent, the European Commission, the Technical University of Denmark (DTU), Vinnova (the Swedish Agency for Innovation Systems), Innovation Fund Denmark, and Volvo Group Trucks Operations.

The simulation app, built by Resolvent, a Denmark-based engineering consultancy that specializes in green energy, enables the R&D department at Volvo to easily test battery lifetime performance.

LaserWho?

The Multiscale Simulation of Laser Welding for Optimal Battery Pack Manufacturing project — referred to as “LaserBATMAN” — was initiated in May 2022 and wrapped up in May 2025 (Ref. 1). Although BATMAN here refers to battery manufacturing and not the famous masked superhero, the goals for the project also strove for greatness.

The LaserBATMAN project aimed to optimize two key performance metrics for battery packs: robustness and lifetime (Ref. 1). One element of the battery manufacturing process that affects both of these metrics is the laser welding used in component joining processes (for example, to connect battery cells in a pack via a busbar). If done correctly, laser welding can positively impact battery life, but if done incorrectly, laser welding can increase the risk of battery degradation. Computational modeling can be used to analyze and predict the effects of welding and joining processes at different stages of the manufacturing process.

Martin Refslund Nielsen, former chief commercial officer and partner at Resolvent, explained that for this type of multiphysics and multiscale problem, “COMSOL Multiphysics is just the right tool.”

R&D Workflow for Multiscale Model and App

The Resolvent team used the COMSOL Multiphysics software to study how laser welding affects the electrochemistry and degradation of battery cells. At least, that’s how the project started.

“Ambitions grew,” Nielsen said, explaining that the team ended up modeling the performance of the entire battery, from cell to pack level, and even “turned [the model] into a design tool in the end”, in reference to the app.

After training the deep neural network of the battery model, the team built a custom simulation app to efficiently optimize busbar tabs in the configuration of a battery pack.

First, the team built an electrochemical model of a battery cell by inputting the formulas for the relevant chemical reaction kinetics (Figure 2). The resulting cell-layer model was then fed input data, and the output data was used to train a deep neural network (DNN). With this reduced-order model, the team was able to quickly gain insight into cell behavior.

After training the DNN of the battery model, the team built a custom simulation app to efficiently optimize busbar tabs in the configuration of a battery pack.

Figure 2: The Resolvent battery model.COMSOL

A “Tool Within a Tool” for Busbar Weld Design

The different features of the simulation app, as described by Nielsen, include its options to input dimensions for the pack geometry (Figure 3), the number of cells, and operating conditions such as ambient temperature and charge–discharge current profiles. The app also makes it possible to apply a simple charging current and load customized cycles. Users are then able to see how the battery performs with each set of specific inputs.

Figure 3. The battery pack model developed by Resolvent.COMSOL

Another important aspect of Resolvent’s app is the Busbar Weld Design Tool button, which Nielsen described as “a tool within a tool.” This button can be used to efficiently calculate busbar weld performance after a number of cycles. After clicking the button, a window opens with options for setting different parameters, including the shape of the weld, busbar thickness, and penetration depth of the laser (Figure 4).

“[This button] is interesting from a robust design perspective,” said Nielsen. Users can apply a variant to the input parameters to see what happens in production if the weld does not reach far enough into the busbar.

Figure 4. Resolvent’s simulation app, showing how certain parameters can be customized. The model shows how the busbar weld is affected by a cooling airflow.COMSOL

“We certainly get deep technical insight here, and we can dimension the design in the sweet spot. Deep technical insight is also key to optimizing product performance,” Nielsen said. For example, if the results suggest better cooling, the team could potentially reduce the design by two battery cells and retain the same performance, easily quantifying the tradeoff between cooling and setup.

Predicting Pack Performance

The app’s user interface also includes a Compute Battery button, which can be used to see the temperature of the weld in a given state, the state of charge, and the solid–electrolyte interface (SEI) layer thickness (Figure 5). The results can help the engineering team further determine if there is room for optimization in the design. For instance, a user can apply a cooling flow in the app and see if turbulence or standing air occurs.

As explained by Nielsen, simulation apps offer more consistent analysis and reporting, and the intuitiveness of custom apps makes it easier to share knowledge. “We are in a much better place to optimize performance and increase robustness,” said Nielsen.

Figure 5. Temperature in the given state (left), state of charge (middle), and SEI layer thickness (right) in the battery pack.COMSOL

The custom simulation app developed by the LaserBATMAN consortium can be used to help produce more robust battery designs with ranges that can support highway travel over longer distances with fewer failures — all while promoting the use of green energy. A heroic project indeed.

References

1. “LaserBATMAN,” University of Skovde, accessed 08/21/2025, https://www.his.se/en/research/virtual-engineering/virtual-manufacturing-processes/LaserBatman/.
2. R. Nielsen, “Keynote: Optimizing Lifetime and Robustness of Electric Truck Batteries with Simulation Apps,” COMSOL Conference 2024 Florence, October 2024.

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.