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 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 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.

One of the singular features of Baotou, an industrial city of 2 million people in Inner Mongolia, is a 10-square-kilometer (4-square-mile) lake that is among the most polluted in the world. The artificial lake, referred to by the name of the dam associated with it, Weikuang Dam (or Baogang Tailings Dam), has for decades collected the effluent from minerals processing for China’s globally dominant iron, steel, and rare-earths industries.

Built in the late 1950s, the lake contains a highly toxic stew of industrial chemicals with substantial amounts of radioactive thorium, the latter resulting primarily from the processing of monazite ores for rare-earths. China has since greatly tightened its environmental regulations and restrictions on mining and minerals processing. For example, largely because of the environmental disaster that is Weikuang Dam, the country doesn’t permit the development or operation of stand-alone monazite mines—that is, mines where nothing is mined except monazite ore.

Now, with industrialized nations in North America, Asia, and Europe spending billions to shore up their rare-earths supply chains, and counting on monazite to help them do it, China’s history and experiences with monazite offer important lessons.

This article is a collaboration between Foreign Policy, the global magazine of politics and ideas, and IEEE Spectrum, the flagship magazine of the IEEE.

Monazite ore is a pillar of China’s rare-earths industry because it is a prime source of some less common but extremely important rare-earth elements, notably dysprosium, terbium, yttrium, and erbium. The country mines and processes vast quantities of monazite ore; it also imports huge amounts of monazite ore for processing.

Most of the domestically mined monazite ore comes from a single mine: the mammoth Bayan Obo complex, 150 kilometers north of Baotou, in Inner Mongolia. Bayan Obo, which is primarily an iron-ore mine, also happens to contain the largest known rare-earth deposit in the world. It’s the source of nearly half of all the rare-earth ore mined on the planet in a typical year. Most of the rare-earth ore mined at Bayan Obo is a type called bastnaesite; it’s the mining of that bastnaesite that makes it legal to mine monazite ore at Bayan Obo as well.

Related: Can a Radioactive Ore Break China’s Grip on Rare Earths?

All told, ore from Bayan Obo was used to produce 170,000 tonnes of rare-earth metal last year, according to David Merriman, research director at the minerals intelligence firm Project Blue. Roughly 30 percent to 40 percent of the rare-earth ore mined at Bayan Obo is monazite. That amount alone would be enough to make China the world’s largest processor of monazite for rare-earths.

China’s 2025 Monazite Imports Topped 70,000 Tonnes

But a more detailed picture of China’s interest in monazite emerges when you consider the country’s imports of the ore. Last year, they amounted to more than 70,000 tonnes, says Thomas Kruemmer, a Singapore-based consultant on minerals and metals. That figure becomes more impressive when you consider the unique challenges of shipping monazite by sea. Because of its radioactivity, concentrated monazite ore is classified as Class 7 dangerous cargo under the International Maritime Dangerous Goods Code. That means, Kruemmer notes, that the ore can be imported through only four ports in China.

In addition, importation of some kinds of monazite into China is supposed to go through a joint venture consisting of the China National Nuclear Corp. and Shenghe Resources, which are at least partly state-owned. But lots of imported monazite doesn’t go through those organizations, notes Merriman. That’s because heavy mineral sands, which may contain roughly 2 percent or 3 percent monazite by weight, are not subject to the same import controls as concentrated monazite ore. So companies import the heavy mineral sands; separate out various profitable concentrates, including the monazite mineral concentrate; and then process the monazite concentrate to recover the rare-earths.

China has lately stepped up its efforts to secure rare-earth ore around the world. One reason, says Kruemmer, is simply to make up for recently lost supplies. In 2025, the U.S. government invested US $400 million in MP Materials, which operates the Mountain Pass mine and processing facility in California. The deal put an end to long-standing contracts under which MP Materials had been selling tens of thousands of tonnes of rare-earth concentrate annually to a trading affiliate of Shenghe Resources.

But other Westerners speculate that China has strategic reasons for stepping up its efforts to acquire heavy mineral sands or other rare-earth assets abroad. By locking down rights to deposits or projects abroad, Chinese organizations are making them unavailable to non-Chinese developers, according to this reasoning.

Merriman points out that your view of these activities largely depends on where you sit. The U.S. government recently made a conditional loan commitment of up to $725 million to U.S. monazite processor Energy Fuels, which has been aggressively securing substantial monazite supplies in Australia, Brazil, and Madagascar. To a Chinese observer, those moves probably look rather strategic.

China, meanwhile, is doing what it has been doing for decades. “China’s playing its game,” says Merriman.

This article appears in the November 2026 print issue as “Radioactive Realpolitik.”

The global campaigns to break China’s dominance of vital rare-earth markets are shaping up into one of the signature geostrategic struggles of the early 21st century. Hundreds of mining companies, government agencies, and industrial facilities scattered across North America, Australasia, and Europe are striving to wrest share from Chinese firms, which control roughly 90 percent of the markets for rare-earth oxides and rare-earth metals.

Rare earths are the ingredients on which vast and crucial manufacturing categories depend: semiconductors, LEDs, lasers, displays, metal alloys, optical lenses, and catalysts for petroleum and chemical processing. Most important, though, are permanent magnets. Essentially all powerful electric motors are based on rare-earth permanent magnets. These motors are the heart of electric vehicles, heating and cooling systems, robots, appliances, and tools. Rare-earth magnets are also in the motors for military systems ranging from tactical drones to ballistic-missile submarines.

This article is a collaboration between Foreign Policy, the global magazine of politics and ideas, and IEEE Spectrum, the flagship magazine of the IEEE.

Succinctly put, any industrial power that wants to remain competitive in vital, high-tech industries must have access to a steady supply of rare-earth elements or products based on them. The traditional supply chain for these elements, originating in China, is increasingly fraught—so much so that some analysts are beginning to use the term “mineral war.” The countries that are now scrambling to piece together alternative supply chains, whether by themselves or in cooperation with each other, face daunting odds. Never mind the sheer momentum of the Chinese rare earths juggernaut. Few industries can match the rare earths business for instability, risk, intrigue, and sheer frustration.

And yet, many rare-earth experts see a way forward, and it depends on a specific kind of ore, called monazite. The ore contains huge amounts of valuable rare earths, including neodymium, praseodymium, and dysprosium.

“The future is monazite,” declares Jack Lifton, an analyst, consultant, and author. Lifton, who has been studying the rare earths industry since 1962, has seen countless trends, companies, and technologies come and go, and he isn’t given to sweeping statements. But on the subject of monazites, he’s bullish. “Right now, in my opinion, monazite is the most important rare-earth ore,” says Lifton, who is also cochairman of the Critical Minerals Institute, an organization that’s promoting the creation of a North American supply chain.

Related: Chasing China’s Monazite Monopoly

That path forward isn’t going to be smooth or straight, Lifton and others are quick to acknowledge. Monazite typically contains radioactive thorium and uranium, which means its handling, storage, and processing are tightly regulated by government agencies. It can be processed only with special techniques and procedures, and the mildly radioactive wastes generated by processing must be stored in isolation indefinitely.

Nevertheless, major investment is now flowing to monazite mines and processing technologies. In the United States, Energy Fuels, which operates the country’s only conventional uranium mill, has developed its own extraction process and is producing mixed oxides of neodymium and praseodymium—and also pilot-scale quantities of dysprosium and terbium oxides, key heavy rare earths that are otherwise produced almost exclusively in China.

Energy Fuels is processing monazite into rare-earth oxides at its White Mesa Mill in southeastern Utah. The body of water at left in the image is a tailings impoundment, where radioactive waste from processing is stored.Energy Fuels

The company has spent more than US $200 million to secure monazite assets in Brazil and Madagascar, and plans to scale its rare-earth production to higher commercial levels, at an estimated cost of $410 million. Energy Fuels’ rare-earth ambitions received a significant boost in June, when the Trump administration announced that it had signed a $725 million conditional loan commitment with the company.

Meanwhile, though, a government-backed project in Saskatchewan, Canada, stands as a cautionary tale for monazite backers. After six years of construction and expenditures of more than CA $100 million (US $72 million), a processing plant there has produced only minuscule quantities of rare earths. The processing plant, built by the Saskatchewan Research Council, was supposed to be producing 40 tonnes of rare earths per month by the end of 2024. That didn’t happen. “Full operation,” with production of 400 tonnes of metal per year was planned for “early 2025.” That didn’t happen, either. The organization is now targeting late 2027—by which time the project will have cost more than CA $216 million, according to the Saskatchewan Research Council’s CEO, Mike Crabtree.

Outside of North America, major monazite projects are also planned by Astron Limited, Hastings Technology Metals, Iluka Resources, Mkango Resources, Sovereign Metals, Steenkampskraal Holdings, and Tronox, among others. Perth, Australia–based Lynas Rare Earths, the largest rare-earths company outside of China, is expanding its monazite mining and processing operations. Most notable is the Lynas mine at Mount Weld, Australia, one of the world’s top monazite resources; the ore is processed at a sprawling facility in Malaysia.

Of course, for countries outside of China, all this surging activity doesn’t guarantee a route to rare earths independence. It’s not even close. So the vital questions now are, What else has to happen to broaden and diversify the rare-earths supply chains? And can all of this be done safely?

What Are the Advantages of Monazite Ore?

Planet Earth has enormous quantities of monazite ore, in two different forms. In Australia, Canada, China, and South Africa, massive deposits of monazite take the form of veins of hard rock that run through mountains and calderas. Much of the new interest in monazite, though, is focused on a second type, called heavy mineral sands. The sands are the result of tens of millions of years of exposure to wind, water, and gravity crumbling mineral-rich rocks into grains, which then accumulated in coastal areas or alongside riverbeds.

A particularly desirable form of monazite ore is found in heavy mineral sands, such as these on a beach in Brazil. Travel Pix/Alamy

Within those grains, the monazite is relatively hard and chemically inert. The monazite might be just a few percent by weight, but it typically contains between 45 and 70 percent rare earths by weight, including, in many cases, all of the necessary rare earths for magnets: neodymium, praseodymium, dysprosium, and terbium.

“The best distribution of rare-earth elements that you can find anywhere is in monazites,” says Constantine Karayannopoulos, former CEO of Neo Performance Materials and now a director of or adviser to several rare-earths firms. “The real money is in those four elements.” Dysprosium and terbium, both examples of heavy rare earths, are particularly sought after and are usually not present in the most commonly mined ore for rare earths, called bastnaesite.

There are vast deposits of monazite-bearing heavy mineral sands in Australia, Brazil, India, Madagascar, Mozambique, and South Africa, with smaller ones in China, Idaho, Indonesia, Malaysia, Sri Lanka, Thailand, the southeastern U.S., and Vietnam.

A geologist points out dark veins of hard-rock monazite ore in the Steenkampskraal mine in South Africa [top]. Ore samples from the mine are collected for analysis [above].Rodger Bosch/AFP/Getty Images

What’s more, collecting the ore simply involves dredging or strip-mining the sands, as opposed to the blasting, drilling, and heavy earth moving required to get at hard-rock ores. Then, too, the sands are already in granular form, so they don’t need to be crushed or ground, like rock ores do. Separating out the heavy minerals from the sands can usually be done at the mining site, using relatively cheap magnetic or gravity-based techniques, leaving a relatively high grade of concentrate for further processing.

And now the bad news about heavy mineral sands. “The problem is they tend to have a lot of thorium,” which is weakly radioactive, says James Kennedy, president of ThREE consulting, a rare-earths advisory firm based in Missouri. “The thorium is the single biggest problem” with using the sands, he says. “If you were mining and processing monazites for rare earths, you would end up with hundreds and eventually thousands of tons of thorium-containing radioactive residues” per processing facility. The only markets for thorium are a couple of tiny ones for welding rods and some optics, he says, so “you have to store it.” And under the rules for nuclear material set by the U.S. Nuclear Regulatory Commission and the International Atomic Energy Agency, any stored radioactive material that doesn’t have a planned use or disposal pathway is considered nuclear waste. “So the [radioactive] liability suddenly far exceeds the value of anything you’ve mined and sold,” Kennedy says.

It wasn’t always this way. Up until the 1950s, monazite was the main ore mined for rare earths. Companies processed heavy mineral sands from Australia, Brazil, and India, and hard rock from South Africa, to produce rare earths for tracer ammunition, steel alloys, cigarette-lighter flints, and, later, color-TV phosphors.

The Mountain Pass mine and processing plants, in California, are the only large-scale rare-earth production facilities in the United States.David McNew/Getty Images

Several developments combined to shift rare-earths production from monazite to bastnaesite. In 1949, a vast deposit of bastnaesite ore was discovered at Mountain Pass in California. By the early 1960s, bastnaesite had eclipsed monazite as the primary rare-earth ore. In the 1990s, with rare-earth production shifting to China, the move toward bastnaesite continued with the exploitation of massive deposits in that country, such as those at Maoniuping, that were either entirely or mostly bastnaesite.

Now, though, any resurgence of rare-earth production outside of China may hinge on a monazite renaissance, and some observers are taking a hard look at the strict regulations that control the handling of thorium and other by-products of monazite processing. Those regulations, Kennedy insists, will throttle the rare-earth ambitions of the world outside China, unless governments take action now.

“You rationalize the regulations around thorium and you will end the rare-earth crisis in the U.S.,” he declares. “You’ll turn rare earths from something that’s always scarce to something that is common and just naturally flows out of other mining.”

How to Scale Rare-Earth Processing Using “Waste”

A prime example of Kennedy’s vision already exists at Energy Fuels, based in Lakewood, Colo. It’s the only company in the U.S. licensed to process, handle, and store compounds of uranium, and among the very few licensed to store radioactive materials of any kind. Those permits give Energy Fuels the distinction of being the only U.S. entity allowed to process monazite ore for its rare earths.

“Supersacks” containing a tonne of neodymium-praseodymium oxides were loaded on pallets at Energy Fuels’ White Mesa Mill.Energy Fuels

The company is moving aggressively to capitalize on that distinction. Energy Fuels’ main business is still uranium—last year it produced nearly 500 tonnes of it. But it has already started selling rare-earth oxides, following a modest $22 million investment in a small rare-earth processing facility at White Mesa Mill in Utah, where it also processes uranium. In September 2025, the company announced that it had sold 1.2 tonnes of neodymium-praseodymium oxide to South Korea–based POSCO International Corp., a leading maker of permanent-magnet assemblies for EV motors. Around the same time, Energy Fuels disclosed that it had raised a further $700 million.

In December 2025, the company announced it had produced small quantities of dysprosium oxide. Tiny amounts of dysprosium are needed for motor magnets to ensure they keep functioning at high temperatures. Today, essentially all of the world’s dysprosium comes from China.

Energy Fuels says it can produce 1,000 tonnes of rare-earth oxide a year at White Mesa Mill. But its rare-earth production there currently shares facilities with the uranium operations, so they cannot both be produced at the same time. Its monazite comes from the chemical company Chemours, which mines heavy mineral sands in Florida and Georgia to produce titanium and zirconium. The monazite ore sold to Energy Fuels is what remains after Chemours strips out the minerals it needs, namely ilmenite and zircon.

The chemical company Chemours mines heavy mineral sands at sites in the U.S. states of Georgia [seen here] and Florida. Monazite extracted from the sands is sold to Lakewood, Colo.-based Energy Fuels, which processes the ore into rare-earth oxides.Chemours

“Effectively, monazites are waste,” Karayannopoulos points out. “So you don’t have to dig another mine” to produce them.

Chemours and Energy Fuels have agreements under which Chemours ships 500 tonnes of monazite annually to White Mesa Mill. Energy Fuels is also negotiating or already has deals that would give it access to enormous reserves of monazite ore in Australia, Brazil, and Madagascar. These will be essential for its ambitions to become a world-class supplier of rare-earth oxides.

Mark Chalmers, former CEO of Energy Fuels, led the uranium company’s entrance into the rare-earths business.Jesse Rieser/The New York Times/Redux

“We’re playing a long game,” says Mark Chalmers, who recently retired as Energy Fuels’ CEO and is now a consultant to the company. “We’re doing the engineering right now on a larger facility that will have the capacity to produce up to 6,000 tonnes per year” of rare-earth oxide, he said in an interview. For comparison, Lynas Rare Earths produced around 12,000 tonnes of rare-earth oxide in its most recent fiscal year.

In announcing the $725 million conditional loan commitment with Energy Fuels, the U.S. Defense Department’s Office of Strategic Capital said the deal would help Energy Fuels scale up its output of rare-earth oxides and also help realize the company’s plans to turn the oxides into metals, a key step preceding the fabrication of magnets.

What Are the Risks in the Rare-Earth Supply Chain?

To see what can go wrong with a monazite project, consider the project in Saskatoon, Canada, where the Saskatchewan Research Council (SRC) has built and demonstrated a pilot plant to produce rare-earth oxides from monazite ore. It’s the only rare-earth extraction and separation plant in the country. In an email, a spokesperson said the plant “is scheduled to be commissioned by year’s end and after a production ramp-up phase, operational within the calendar year 2027.” When in “full operation,” at a time not specified by the spokesperson, the plant will have a capacity somewhere between 400 and 600 tonnes of rare-earth metal per year.

But even that relatively modest output will depend on a reliable supply or “feedstock” of monazite. In an industry small enough that many of the top people are on a first-name basis, SRC’s inability to secure such a supply is an open secret. “Their problem is they haven’t got enough feedstock to run on a continuous basis, so they’re making sample runs,” a well-connected industry observer told IEEE Spectrum. (SRC declined to make an official available for this story.)

In April 2024, SRC announced it had reached agreements with a Vietnamese organization, the Hung Thinh Group, for 3,000 tonnes of rare-earth carbonate per year for five years. It’s unclear how much of that feedstock has made it to the SRC facility. Vietnamese government restrictions on the export of monazite and bastnaesite have been in effect since 2022 and became law this past January.

But the real twist here is that one of the largest, highest-grade monazite deposits in the world is just 800 kilometers from Saskatoon, near Alces Lake in Saskatchewan. It’s likely that the SRC facility was conceived and built with that monazite deposit in mind. After all, as a government-owned entity, SRC says its mission is “to help the people of Saskatchewan strengthen the economy with quality jobs and a secure environment.”

The Alces Lake deposit is owned by Appia Rare Earths & Uranium Corp. In an interview, Appia’s CEO, Tom Drivas, talked up the possibility of a link between SRC and Appia. “SRC would be our logical option because, you know, it’s in the same area, same province. They need the monazite. And if we can supply the monazite, that’s sort of a natural, right?”

It probably won’t happen, though. The logic of rare-earth mining is utterly different from the logic of mining a commodity like, say, copper. Multiple experts contacted by Spectrum suggested that the cost of developing a monazite mine at Alces Lake, including such necessary infrastructure as roadways, electricity, and water, would be in the hundreds of millions of dollars. Such costs, and the mercurial exigencies of the rare-earths market, make it unlikely that the government of Saskatchewan would commit that kind of money, even with help from the Canadian federal government. These factors help explain why Canada, which has some of the most attractive deposits of rare-earth ores on the planet, along with impressive mining companies and resources, has no operating mines for rare earths.

Meanwhile, SRC is pursuing other options. Under a deal announced in May with REalloys, in Euclid, Ohio, REalloys will spend $20.6 million for upgrades to SRC’s Saskatoon facility, while SRC develops technology to produce rare-earth metals from rare-earth oxides. The deal also gives REalloys the right to purchase up to 80 percent of the SRC plant’s output. Where will the feedstock come from? Hard to say, but REalloys also struck a deal with U.S. Critical Materials Corp. for 10 percent of the hypothetical output of a proposed mine at Sheep Creek, in Montana. Sheep Creek reportedly has substantial amounts of hard-rock monazite that’s unusually low in thorium. But it’s far from certain that those deposits will ever be mined. The proposal for developing a mine there, in a breathtakingly scenic area near the headwaters of the Bitterroot River, has attracted fierce opposition from environmental groups, outdoor enthusiasts, and local officials.

Taken together, SRC’s recent moves have left industry veterans nonplussed. “I’m not sure what their business model is,” said one. “They say, ‘we’re a research institute. Our mandate is to help industry.’ The reality is, they’re competing with industry.”

SRC has also drawn scrutiny for some questionable business practices. In 2022, SRC board member Dennis Fitzpatrick was fired after raising concerns about the qualifications of an engineer hired to manage and direct operations at SRC’s Saskatoon plant. The president and CEO of SRC at the time, Mike Crabtree, was responsible for the hire. A subsequent investigation of Fitzpatrick’s claims by the Canadian Broadcasting Corp. detailed a troubling web of relationships and practices surrounding the SRC’s rare-earths initiative. (Fitzpatrick did not respond to multiple emails.) For example, past business associates of Crabtree’s were reportedly appointed to positions of authority without any other candidates being considered and without their connections to Crabtree being made public.

In one case, in 2021, Crabtree disclosed that he was offering a venture capital firm, Westbridge Capital, an opportunity to invest in the rare-earths plant. But according to Fitzpatrick, Crabtree never mentioned that the VC firm was run by an SRC board member who also happened to be one of Crabtree’s business associates—or that Crabtree himself was a trustee of Westbridge and an investor in the company, as well as a paid advisor.

Crabtree remains as SRC’s CEO.

Managing Toxic By-products From Rare Earths

Putting aside SRC’s missteps, the organization does have one advantage: It holds multiple licenses from the Canadian Nuclear Safety Commission to handle, process, and manage radioactive materials. And that’s huge.

Most of the radioactive burden associated with processing monazite ore comes from substances produced during processing. In one of the early steps, pulverized, powdery ore is dried and then leached in strong acids—a process called cracking—to produce a rare-earth concentrate. Waste products of this step include a slurry containing thorium hydroxide and much smaller amounts of uranium hydroxide. The rare-earth concentrate then undergoes an extraction process to isolate specific rare earths, such as neodymium or dysprosium. More thorium and also radium are produced during these steps.

Waste barrels were photographed in 2024 at Energy Fuels’ White Mesa Mill, where ores containing uranium and rare-earths are processed and refined.Jesse Rieser/The New York Times/Redux

The thorium hydroxide slurry, meanwhile, is typically deposited in heavily regulated waste cells, called tailings impoundments or tailings ponds. The thorium hydroxide, either in the form of a slurry or a powder, must be stored in isolation from the environment indefinitely. Energy Fuels stores waste from its longstanding uranium production operations, as well as its new rare-earth operations, in tailings impoundments at its White Mesa Mill facility. The impoundments are just 10 kilometers from the community center of the Ute Mountain Ute Tribe, a group of Indigenous people, which has vigorously opposed operations at White Mesa Mill for more than 20 years.

Virtually all of the thorium is the isotope thorium-232, which is only weakly radioactive. Its radiation is alpha particles, which cannot penetrate human skin. Nevertheless, improper storage can have serious consequences, particularly if the thorium is inhaled or ingested.

“Have a look at the gravitational separation of HMS [heavy mineral sands] in Nigeria,” Thomas Kruemmer, a rare-earth consultant and analyst, wrote in an email. “There is lots of dust in the Nigerian processing sites, which the workers inevitably breathe in. And inside your respiratory system there isn’t any protective skin, so alpha radiation develops its full potential. This is something many people do not consider.”

In 2012, protestors in Malaysia [top] rallied against the Lynas rare earths processing facility [above], then under construction near Kuantan. Left: AFP/Getty Images; Right: Goh Seng Chong/Bloomberg/Getty Images

Moreover, adds Kruemmer, “it is not thorium itself. It is the decay isotopes of thorium that are highly radioactive.” One decay product of thorium is polonium, he points out, “the preferred poison for [killing] Russian dissidents.”

In 1982, a company called Asia Rare Earth, a subsidiary of Mitsubishi Chemical, began extracting yttrium from monazite ore in Perak, Malaysia. Local residents soon began complaining of various illnesses. With the consent of the Malaysian government, the company had built a waste channel to a nearby disposal site, where the waste was stored in open steel barrels and in an unlined waste pond. The waste overflowed from the pond whenever it rained heavily. In 1984, a Japanese professor measured radiation levels at and near the pond and found them to be, in the most extreme cases, hundreds of times higher than internationally accepted maximums. A subsequent investigation found extremely high levels of birth defects and eight cases of leukemia in the small community near the waste dump.

In 2011, at age 69, Lai Kwan prepared to bathe her 29-year-old son, Cheah Kok Leong [top]. Cheah Kok was born with severe mental disabilities that Lai Kwan believed were caused by exposure to radiation that she received while working at a rare earths processing plant in Bukit Merah, Malaysia [bottom, photographed during its disassembly].Rahman Roslan/The New York Times/Redux

A protracted legal fight among local residents and Asia Rare Earth and the Malaysian government culminated in the closure of the plant in 1992. By that time, so much radioactive material had been carelessly dumped that cleanup took more than two decades and cost Mitsubishi an estimated $100 million. The episode still haunts the rare-earths industry in Malaysia, where Lynas now processes monazite concentrate at the largest rare earths processing facility in the world. That plant, in Kuantan in the state of Pahang, has been the subject of organized and active local opposition that began even before its commissioning in 2012. The Malaysian government has sporadically tightened its regulations on the plant, stipulating that, for example, radioactive wastes be shipped out of Malaysia or that the processing stages that create radioactive residues be accomplished outside of Malaysia.

Kennedy notes that if monazite ore is to become a major source of rare earths, much larger quantities of thorium will need to be disposed, probably more than could be economically accommodated in lined and isolated tailings ponds.

One of the very few options for using such large amounts of thorium involves nuclear power. Years ago, countries including the U.S., Canada, Norway, and Germany had thorium-reactor research programs, and India has had a slow-moving program for decades. But today, the only operating thorium power reactor is in China. The Shanghai Institute of Applied Physics of the Chinese Academy of Sciences has been operating, since 2023, a 2-megawatt molten-salt-cooled reactor that is thorium-fueled.

The world’s only operating thorium reactor is a 2-megawatt unit at the Shanghai Institute of Applied Physics of the Chinese Academy of Sciences. In October 2025, technicians used mechanical arms to analyze samples of the reactor’s molten-salt fuel.Jin Liwang/Xinhua/Getty Images

A couple of Western startups are keeping the thorium reactor dream alive. A Franco-Dutch company called Thorizon says it has designed a reactor that would be fueled by a mixture of thorium and spent nuclear fuel. Startup Copenhagen Atomics is working on a 1-MW reactor that’s a prototype for a 100-MW, mass-manufacturable, molten-salt reactor.

Processing of monazite ore could provide fuel for an enormous number of thorium reactors far into the future. Consider a single large facility producing 5,000 tonnes of rare-earth oxide a year. If the monazite contains 7 percent thorium, the plant would also be producing roughly 700 tonnes of thorium a year. According to Rusty Towell, a professor at Abilene Christian University in Texas who directs an advanced-reactor project there, a tonne of thorium would be sufficient to generate about 350 MW (electric) for a year. So the large processing facility alone would be throwing off enough thorium annually to keep 2,450 100-MW reactors operating for a year.

That amount of power—245 gigawatts—is about half of all the generating capacity growth forecast for the United States over the next decade, a period during which analysts expect accelerating electricity growth due to increasing numbers of data centers and electric vehicles.

What Could Strangle Rare-Earth Supply Chains?

Virtually all rare-earths experts foresee a growing role for monazite. But given the problems with thorium, local opposition to some projects, and political instabilities connected to others, some have muted expectations. “The sole and single monazite hope rests with Iluka’s expensive project in Australia, IMHO,” writes Kruemmer. Iluka Resources, based in Perth, is now building a massive rare-earths refining facility at Eneabba, 275 km north of Perth. It will process a million-tonne monazite stockpile that accumulated over decades from the company’s other mining operations. The cost of the project is estimated at AU $1.8 billion (about US $1.3 billion).

In a US $1.3 billion project, Iluka Resources is building a massive facility in Eneabba, Australia, to extract and separate rare earths from a million-tonne stockpile of monazite, visible in the background of this photo.Iluka Resources

For other experts, though, the biggest challenges facing monazite projects isn’t thorium, or radiation, or regulations. It’s the blend of quirks, misconceptions, and self-dealing that dog the entire rare-earths industry.

“The rare-earth mining business is the Afghanistan of mining,” quips Kennedy. It’s a place where people and companies go with grand ideas, only to become ensnared in quagmires.

One fundamental problem is the highly speculative nature of almost every aspect of the rare-earths business, which lends itself to backroom deals. In particular, the industry is plagued by speculators who buy stakes in a project with no intention of staying with the project through completion; they cash out as soon as they can profit from any future favorable news—a technique known as “pump and dump.”

Consider the Trump administration’s decision to take a 10 percent stake in USA Rare Earth in exchange for $277 million and to provide a $1.3 billion secured loan to the company. USA Rare Earth has the rights to a deposit called Round Top near Sierra Blanca, Texas. The company hired the Wall Street firm Cantor Fitzgerald to raise $1 billion in equity financing for the deal. Cantor Fitzgerald, as it happens, is run by the sons of Howard Lutnick, Donald Trump’s Commerce Secretary.

Meanwhile, there is no rare-earths mine at Round Top, and it’s a distinct possibility that there never will be.

“The current administration consistently fails to seek expert advice and is recklessly pouring billions of taxpayer dollars into speculative [rare-earth element] processes and projects, like Round Top, that have little chance of commercial success,” says Terry McNulty, a mining consultant based in Tucson.

Geoff Atkins, vice president of business development of the Canadian rare-earths firm Ucore Rare Metals, had advice for his colleagues at a convention of the Critical Minerals Institute in 2022. “Instead of focusing on your speculative value,” he said, “instead of being able to say we’ve got the biggest project in the world, instead of, ‘we’ve got a [net present value] of $3 billion or $4 billion,’ how about you say, ‘we’ve got a project which is going to get into operation, and this operation is going to provide the greatest value to our shareholders in the long term, not just in the short term when we can pump this up and, you know, the [early] investors can sell out.

“Because unless there is that shift,” he warned, “we’re not going to have any hope of getting the volumes we need to give the certainty to the downstream supply chain.”

In short, the rare-earths industry outside of China needs to grow up. And sooner rather than later.

This article appears in the November 2026 print issue as “Can a Radioactive Ore Break China’s Grip on Rare Earths?”

This White Paper shows RF and microwave engineers and technicians how compact analog signal generators support core amplifier and receiver test tasks, from gain and distortion measurements to blocker rejection, power calibration, and local oscillator substitution.

What you will learn about:

• How to measure amplifier gain and gain compression, including the P1dB point that marks the transition from linear operation into saturation.
• How two-tone intermodulation distortion testing reveals unwanted signal products created by amplifier nonlinearities, and why third-order intercept (IP3) is used to compare amplifier linearity.
• How to test receiver performance under blocker and interferer conditions, using high output power and fast switching speed to simulate real-world interference.
• How power calibration and RSSI testing depend on accurate, repeatable signal sources across multiple frequency channels and bands.
• How signal generators can substitute for local oscillator and clock subsystems during early-stage receiver development, before final circuit prototypes exist.

Progress continues in restructuring IEEE’s geographic regions, as outlined in The Institute’s February 2023 article “IEEE Is Realigning Its Geographic Regions.”

• The “Region 1” designation will no longer be used.
• The new Region 2 will merge what is currently Region 1 (Northeastern U.S.) with the current Region 2 (Eastern U.S.). It will be called Region 2 (Northeastern and Eastern U.S.).
• The current Region 10 (Asia Pacific) will be divided into two separate regions. The new Region 10 will encompass North Asia, and the new Region 11 will cover South Asia and the Pacific.

The restructuring is defined in the image of the redrawn map in this article.

What’s underway

Although the new regional alignment will not officially take place until 1 January 2028, several activities have begun within IEEE Member and Geographic Activities (MGA) and across other IEEE organizational units. Within MGA, volunteers and staff have been actively addressing system enhancements, director/director-elect elections, governance updates, communications, and regional finances and offerings.

No changes to membership

Work will continue from now until the effective date of realignment in 2028. The change will not affect your IEEE membership, nor will it change either your local IEEE section or chapter.

If you have any questions regarding the realignment, contact Reg-Realignment@ieee.org.

If you’ve ever wondered how Claude and similar AI agents can answer complex questions in seconds and where they get the information, you’re not alone.

Gurmeet Saran works to find that answer every day as an engineer at Anthropic in San Francisco, the AI research company behind Claude. The IEEE senior member is developing an infrastructure to understand the mechanics behind the chatbot’s behavior and to ensure its underlying architecture produces reliable outputs.

Gurmeet Saran

Employer

Anthropic in San Francisco

Title

Member of technical staff

Member grade

IEEE senior member

Alma maters

University of Kalyani, in India; Birla Institute of Technology and Science, in Pilani, India

“My job is to make sure the data is accurate, structured, and trustworthy, and that private data stays private,” he says. “As I learned earlier in my career, privacy can’t be a layer you bolt on—it has to be part of the architecture from the start.” If you can’t trust the logic, he adds, you can’t make reliable claims about the behavior, and that has real safety implications.

Safety is a priority for Anthropic, he says. According to Claude’s mission statement, the company wants the AI agent to be “genuinely helpful to the people it works with or on behalf of, as well as to society, while avoiding actions that are unsafe, unethical, or deceptive.” Crucial to the mission is ensuring that researchers can observe and understand how AI models behave, a challenge that is especially acute for large language models, Saran says.

“We have decent tools to monitor databases and web services,” he says, “but for LLMs, we’re still figuring it out.”

Solving data quality problems at Yahoo, Intuit, and Meta

Saran’s interest in computing systems dates back to his childhood in Kolkata, where he logged as much time as he could in the school computer lab.

“I remember the first time I used one of the computers,” he says. “What struck me was that you could write instructions, and the machine would follow them exactly. No ambiguity.

“That felt magical to me.”

After graduating from high school in 1999, he stayed in India to pursue advanced degrees. He earned a bachelor’s degree in electronics and telecommunication engineering in 2004 at the University of Kalyani. In 2010, he earned a master’s degree in software engineering from the Birla Institute of Technology and Science in Pilani.

While a graduate student, he began his career in 2007 at Yahoo as a technical lead of the data team in Bengaluru, India. While there, he encountered the project that shaped his career trajectory: working on web analytics and business intelligence reporting solutions.

“We kept running into a frustrating problem,” he recalls. “Different teams would calculate the same website metric—something like daily active site users—using slightly different logic, and end up with conflicting answers. It sounds like a small issue, but when leadership is making decisions about resource allocation and product development using that data, not knowing which number is correct becomes a problem.”

The team built a central data repository with defined metrics, essentially serving as a single source of truth. Saran created a model that defined what “daily active site users” means, breaking down which events counted, how to handle edge cases (or when standard code fails), and more.

With the new model, data teams across the company could rely on a standard method to determine the number of daily active users for the projects and sites they were responsible for and obtain comparable results using the same logic.

From there, Saran and his team built Yahoo! Digits, an internal web analytics tool used for all the company’s products.

Saran left Yahoo in 2010 to join Intuit as a staff engineer in Mountain View, Calif. He helped build the analytics infrastructure that unified data across the company’s products, including TurboTax and QuickBooks accounting software. During his tenure, Intuit developed a feature that detected fraudulent tax filings in real time. It required Saran and other engineers to build a system to protect taxpayer data while adhering to strict U.S. Internal Revenue Service policies.

“IEEE operates in a lot of countries where engineers don’t have easy pathways into the global professional community. That kind of connection matters as AI development spreads beyond a handful of organizations.”

“This taught me that privacy and governance can’t be layers you bolt onto software after the fact,” he says. “They have to be part of the architecture from the start.”

Saran joined Meta (known as Facebook at the time) in 2016 as a lead software engineer in Menlo Park, Calif. The company was facing controversy over its video metrics: Meta publicly acknowledged an error that caused it to overstate the number of views for posted videos.

“This wasn’t a coding bug; it was an architectural problem,” Saran says. He built a system that worked similarly to what he had developed at Intuit, whereby one single truth was defined to govern all video view metrics across the site.

That system, Metric360, eventually became part of the site’s architecture. By the time he left Meta in 2021, he says, Metric360 was being used across the company to calculate video metrics.

He left the company to become a principal software engineer at Wealthfront, a financial services company in Palo Alto, Calif. He helped build the company’s data infrastructure.

Building trustworthy AI systems at Anthropic

Saran joined Anthropic in 2025 as a member of the technical staff.

To describe the implications of his work, he uses a medical analogy: “Imagine a doctor is monitoring a patient’s vital signs and making treatment decisions based on those numbers. What if one or a few of the medical devices used were slightly miscalibrated? Even if most readings are fine, it’s a serious problem because the difficult decisions happen in the margins.

“It’s the same with AI systems. My job is essentially to make sure the instruments are giving accurate readings—that the data means what it says it means.”

His job comes with high stakes and complex problems to solve.

“AI agents are especially tricky to observe,” he says. “When a system is taking sequences of actions [generating conversational responses with different users], you need a way to capture what it did, what information it was working from, and whether outcomes matched what the end user expected. And you have to do all of that without exposing anyone’s private information—the systems are designed to recognize sensitive data and keep it protected as it flows through. It’s a much more demanding observability problem than anything I’ve worked on before.”

Solving for system reliability while AI is still in its early days of adoption is both Saran’s passion and the problem that keeps him up at night, he says.

“I think making AI systems genuinely auditable and transparent is one of the more important engineering problems of the next decade,” he says, “and it doesn’t get talked about as much as model capabilities do.”

IEEE can play a role in shaping AI development

Saran joined IEEE in March 2025.

“After more than 20 years of building data infrastructure across organizations like Yahoo, Intuit, Meta, and Wealthfront,” he says, “I wanted to contribute back to the technical community through peer review, publishing, and engaging with the broader engineering research ecosystem.”

He says he believes IEEE can play an essential role in shaping the future of AI through its standards work.

“We have standards for everything, from building codes to software testing,” he says. “AI infrastructure needs something similar, and IEEE has the credibility and the reach to do that.”

Another problem that IEEE can solve for, he says, is access: “IEEE operates in a lot of countries where engineers don’t have easy pathways into the global professional community. That kind of connection matters, especially as AI development spreads beyond a handful of organizations in a few cities.”

This article was originally published by Canary Media.

The geothermal startup Fervo Energy is now selling power to the grid from its Cape Station project in Utah—marking the first time a utility-scale, next-generation geothermal facility has achieved such a feat worldwide.

The Houston-based firm turned on an initial 33-megawatt unit late last week and has been testing and ramping up output since. On Thursday, Fervo will officially begin commercial operations, supplying carbon-free electricity to Shell Energy, a retail power provider in California.

Cape Station, a $2 billion project that will eventually be 500 MW, is unique among the world’s commercial geothermal plants. While other facilities draw on naturally occurring hot water and steam, Fervo has engineered its own geothermal resource deep underground. The startup uses drilling techniques from the oil and gas industry to create reservoirs in hot, dry rocks that generate steam to produce electricity.

“Traditional geothermal would have never been done in southern Utah,” says Dawn Owens, Fervo’s senior vice president and head of development and commercial markets.

With the first unit online, Fervo says it is working to commission two more 33-MW power blocks by 1 January, 2027. The project’s remaining 400 MW—which is supported by a new agreement with Google—is under construction and slated to come online in 2028.

Fervo’s Next-Generation Geothermal Journey

Fervo already operates a 3.5-MW commercial pilot plant in Nevada, which went online in late 2023. Owens says the pilot, called Project Red, was important for demonstrating the technology itself. But firing up Cape Station proves that Fervo’s approach can achieve utility scale and export power to wider markets.

“This is massive for both Fervo and the geothermal industry at large,” she says.

Fervo, which went public in May, is leading the fast-growing movement to develop clean, always-on geothermal power to meet soaring demand from data centers and balance growing supplies of intermittent solar and wind energy. Next-generation geothermal firms have raised billions of dollars in private capital in recent years to develop and deploy technologies that can harness energy in harsh, deep environments.

Drilling continues at Fervo Energy’s Cape Station geothermal plant in Beaver County, Utah, where the company plans to ramp up production to 500 MW. Fervo Energy

Today, geothermal represents just 0.4 percent of total U.S. electricity generation every year. Expanding geothermal’s share of the mix will require not just building more projects like Cape Station but also tackling high upfront drilling costs and other economic and regulatory barriers—all while working to limit potential impacts on communities and the environment, experts say.

“Fervo’s first power achievement shows investors the size and scale of the opportunity to fuel many more such milestones across developers and sites in the future,” says Mitchell Smith, who serves as grid lead for the Mountain West Geothermal Consortium. The bipartisan initiative is calling for increased public and private financing “to help bring more projects online responsibly and at the speed of demand,” he added.

Enhanced Geothermal Energy Gets Cheaper

Fervo has declined to disclose the terms of its power purchase agreements for in-progress and future development, which together represent about 1,000 MW, or enough to power roughly 1 million U.S. homes every year. Phase 1 of Cape Station is set to deliver power at $7,000 per kilowatt, while Phase 2 should come at $5,500 per kilowatt, the company previously told Canary Media. (By comparison, the average cost to build a natural gas–fired power plant was $2,157 per kilowatt in 2025.)

Along with slashing drilling costs, enhanced geothermal projects could improve their economics by delivering power when it’s most valuable to utilities, according to researchers. While geothermal plants traditionally provide around-the-clock, baseload power, next-generation facilities might operate flexibly during certain periods—like in the evening, when solar output declines—to reduce overall grid costs. Geothermal plants could also operate “behind the meter” to directly power data centers so that the costs of adding new electricity capacity aren’t passed on to utility customers, Owens says.

“We’ve been saying all along that our technology makes geothermal more scalable and affordable in more places,” she says. On Cape Station’s powering-up, she adds, “That’s just something no next-generation geothermal company has ever been able to do before.”

This White Paper provides roboticists and control engineers with a comprehensive overview of how motion planning and motion tracking can be decoupled to preserve professional athletic style in humanoid robots, and why adaptive, speed-aware control is essential for closing the gap between simulation and real-world deployment.

What you will learn about:

• Why combining a professional athlete’s motion style with reliable task performance remains an open challenge in humanoid ball-sports robotics, and how separating a high-level planner from a low-level tracker addresses this trade-off.
• How motion data extracted from broadcast video and motion-capture recordings can be reconstructed, retargeted, and physically corrected into reference motions a humanoid robot can safely track.
• Why the sim-to-real transition introduces compounding tracking errors, and how training a tracker on randomized execution speeds, paired with a learned motion-speed adapter in the planner, improves robustness on physical hardware.
• How this adaptive framework was validated across two humanoid platforms, including full-size, in-the-wild serving performed without any motion-capture system.

Today’s engineers face an unprecedented acceleration in AI hardware complexity, as explained in the recent research article “Revisiting Edge AI: Opportunities and Challenges.” The article examines the rapid growth of edge AI and the challenges it creates, including resource constraints, model architecture limitations, and network demands across edge-AI deployments.

The acceleration is driven by a fundamental shift in how modern AI models are built and scaled. As the models have become much larger and more complex, they are computationally more demanding because they contain more parameters and require more calculations.

To meet the demands of scaling deep neural networks, the industry is increasingly developing AI chips that are designed for specific tasks.

One major reason is that moving data between memory and the processor has become a major limitation on AI performance.

The movement to confront the hardware bottleneck—the AI memory wall—has altered the trajectory of semiconductor innovation, shifting architectural priorities toward domain-specific accelerator platforms.

No longer can engineers evaluate systems statically; they must master joint hardware design and network-algorithm co-optimization to navigate the critical trade‑offs between throughput, latency, and operational efficiency.

The challenges are addressed in the new AI Processor Architecture, Design Principles, and Performance program, developed by IEEE Educational Activities with support from the IEEE Computer Society.

Topics covered

The five-course program provides a structured exploration of AI processor technologies, from fundamental design principles to advanced architectures and real-world deployment.

The topics are:

• Fundamental principles of design and functionality.
• Practical insights into advanced architectures.
• Understanding neural processing units for industry deployment.
• Emerging trends and evolving architectures.
• Designing for edge, cloud, quantum, and the Internet of Things (IoT).

The program addresses the needs of professionals across the AI hardware ecosystem, including hardware architects, chip designers, embedded systems developers, data‑center hardware engineers, and innovators exploring next‑generation processor ecosystems.

It is also valuable for those transitioning into AI chip design or seeking to understand the architectural forces shaping modern machine learning acceleration. For many, it provides the bridge between theoretical knowledge and the collaborative, cross‑disciplinary reasoning required in engineering environments.

The program is designed to explore how modern AI processors are conceived, structured, and optimized. The architectural layers that define contemporary AI hardware will be covered, including compute units, memory hierarchies, dataflows, and the performance characteristics that emerge from design decisions. The curriculum bridges theory and application, enabling participants to interpret architectural foundations, analyze trade‑offs, and understand how hardware structures shape computational efficiency across diverse environments.

By the end of the courses, learners will be able to evaluate processor behavior with the analytical precision expected of professionals working at the frontier of AI hardware design.

AI-generated avatars explain concepts

The program also uses a dialogue‑driven learning approach. Learners view conversations between AI-generated avatars that engage in scenario‑based dialogues. The avatars represent engineers tackling the same problems from various perspectives based on their different roles. The simulated storytelling and dialogues make advanced engineering concepts approachable without sacrificing depth or interactivity, because the user is occasionally challenged to decide the correct answer that leads to the best course of action.

A hardware engineer might push back against a systems engineer’s demands, for example, revealing the friction between physical constraints and algorithmic ambition. A computational validation specialist might interrogate a chip performance engineer’s optimism, exposing the gap between theoretical throughput and real‑world behavior. A heterogeneous systems architect could debate a multiprocessor coordination specialist about synchronization overheads, while a standards development engineer discusses regulatory implications with a technology strategy and compliance architect.

In the final course, an IoT systems architect and an embedded AI optimization engineer dissect the realities of deploying AI in constrained environments.

By the end of the course, learners will be able to evaluate processor behavior with the analytical precision expected of professionals working at the frontier of AI hardware design.

The avatar-driven learning environment creates a psychologically safer environment for learners, who might feel intimidated by traditional expert‑led videos. Research published in 2024 in IEEE Transactions on Learning Technologies showed that avatar‑based instruction can increase a learner’s confidence by up to 25 percent and improve retention of complex technical material.

Each course concludes with a module in which the two experts from different disciplines debate, question, and challenge each other’s assumptions.

Modeling expert reasoning through dialogue

The cross‑disciplinary conversations can do more than explain concepts; they can model how experts think. They can reveal the negotiations behind architectural choices, the competing priorities that shape system design, and the analytical rigor required to balance performance, efficiency, scalability, and compliance.

Learners observe engineering discourse, gaining insight into the reasoning patterns that drive innovation in AI processor development.

The program goes further by interrupting the dialogue at key moments and inviting the learner to step in. Instead of passively absorbing information, the learner decides how to resolve a trade‑off, predict the outcome of a design choice, or select the most defensible engineering path. The experience can feel less like a course and more like an apprenticeship inside an engineering team.

A new standard for advanced engineering education

By merging rigorous technical content with an innovative, dialogue‑driven delivery model, the program sets a standard for teaching advanced engineering. It captures the complexity of real‑world problem‑solving, humanizes the learning experience, and challenges participants to think like the engineers who are defining the future of AI hardware. It is not simply a new course; it is a new way of learning.

For individual access, visit the IEEE Learning Network.

For customized organizational options, contact a content specialist to discuss volume pricing.

In mid-September, Anthropic CEO Dario Amodei joined a growing group of researchers calling for the AI industry to slow down. While some have questioned these researchers’ motivation for voicing concerns about the rapid pace of development, the plea highlighted a key tension between safety and speed. The underlying problem is not simply a technical error, but a sociotechnical failure: Pressure to move quickly has made it hard for engineers to fully assess and manage new risks of increasingly powerful AI systems. It is a clear warning that responsible engineering requires more than technical expertise; engineers must also understand workplace pressures, moral and ethical rules, and the social consequences rooted in the systems they create.

Cases like this, alongside invasive surveillance technologies and biased AI tools, have exposed serious shortcomings in how engineers recognize and address the ethical and social implications of their work. While the need to educate graduates in ethics and social responsibility is widely recognized, a persistent gap remains—these topics are too often treated as afterthoughts, disconnected from the technical content of engineering coursework. This siloed approach leaves graduates ill-equipped to address real-world challenges; indeed, only about a third of engineering professionals in the United States report having received training on their public welfare responsibilities during their formal engineering education.

We are working to bridge this gap by designing two complementary instructional approaches, one that integrates social issues directly into foundation technical coursework and a second that prepares students to act as “public welfare watchdogs” through a standalone engineering course. By embedding ethics and social responsibility training directly into the core of undergraduate engineering coursework, we equip future professionals with the agency they will need to protect the public good.

The gap in instruction about ethics and social responsibility

Across the engineering profession, there is a persistent belief that technical skill-building is central to engineering, while ethical, social, and public welfare concerns are peripheral or distracting. This technical/social dualism is evident even in the evaluation of scholarly research. For instance, we recently submitted an academic position paper discussing this ethics education gap to a leading peer-reviewed engineering journal. It was dismissed by reviewers along the very same dualism it described.

One reviewer warned that it would be “highly irresponsible” to challenge technical skill-building as “the foundation of any kind of engineering” by adding ethics and social responsibility training into engineering classrooms. Another reviewer objected that integrating these considerations into foundational courses risks “distracting the students from fully absorbing the technology.” These comments miss the core argument of our work: that ethics and social responsibility are integral to responsible technical practice.

Today, accreditation agencies like ABET (formerly known as the Accreditation Board for Engineering and Technology) and professional licensing bodies like the National Society of Professional Engineers mandate student outcomes related to social and public welfare consequences. But these curriculum requirements tend to be unhelpfully generic and minimally enforced. As a result, engineering students often receive ethics instruction through introductory or non-engineering coursework, where the focus is on codes of conduct and academic integrity, rather than deeper topics of social responsibility. This instruction is almost always disconnected from the complex real-world dilemmas engineers face.

Indeed, we find that only ethics instruction in actual engineering courses (not in other parts of the curriculum) can effectively instill ethical awareness as professionals. Concerningly, only about 4 in 10 engineers in the United States recall receiving this kind of training, and 3 in 10 report never receiving training on their public welfare responsibilities at any point in their careers. This leaves most of the profession underprepared to handle the complex challenges of engineering practice.

A new approach: Integrating technical and social dimensions

To advance a new paradigm—one that embeds ethics and social responsibility issues within the undergraduate engineering curriculum—we’ve designed two innovative course-based approaches: a library of one-hour sociotechnical modules for foundational circuits courses (described here and here) and a new engineering course dedicated to public welfare responsibilities.

The first approach integrates discussions about relevant social issues directly into foundational engineering courses, where students begin to form their identities as engineers. “Introduction to Circuits” is typically the first course in the electrical engineering major and is a requirement for many other engineering disciplines, making it a powerful site for introducing social content and ethical reasoning. We designed several one-hour modules for this course, where social issues provide context for a technical topic, offering a manageable entry point for faculty who may feel ill-prepared to teach about ethics and social responsibility.

While many engineering faculty recognize the importance of teaching students about these topics, they struggle to integrate them into their courses amid heavy teaching and research demands. To overcome these challenges, each module provides ready-to-use materials, including lecture slides with scripts and homework and exam problems. Integrating these modules directly into the course signals to students that sociotechnical awareness is not an “add-on,” but a required professional competency.

Our module on hospital power prioritization, for instance, connects the technical topics of power and equivalent circuits to the ethical challenges of healthcare infrastructure. Students learn about the reality of “red outlets” in medical settings, debating the power needs of specific wards (e.g., the intensive care unit, operating rooms, and physiotherapy rooms) to decide which branches should receive emergency power during a hypothetical outage.

Our conflict minerals module connects the technical topic of capacitors to the global tantalum supply chain. Students calculate the amount of tantalum used in cell phones and identify mining locations in regions like the Democratic Republic of the Congo, where revenue often fuels armed conflict and human rights abuses. Students research the conflict-mineral policies of major companies such as Apple and Samsung and reflect on their future roles as ethical designers and consumers.

Student responses to these modules have been overwhelmingly positive. In interviews, students told us that the content helped them get away from the “stigma that engineers only worry about math” and made the work more meaningful. We plan to share the modules widely to help circuits instructors across the globe bridge the technical-social divide.

Beyond these integrated modules, our second approach is a standalone engineering course—Public Welfare Responsibilities in Engineering (PWRE)—that helps round out students’ sociotechnical education by preparing them to fulfill their professional obligations as engineers. While most dedicated ethics courses end by introducing engineers’ abstract responsibilities, PWRE helps students recognize specific public welfare concerns, enhances their motivation to act when they encounter them, and teaches concrete, practical intervention strategies.

In addition to teaching students about professional codes, such as the IEEE Code of Ethics, PWRE engages students in discussion of where these formal definitions may fall short, and how structural barriers such as workplace culture can hinder engineers from fulfilling these responsibilities. Then, the course equips students with intervention strategies. Students learn about their rights and responsibilities as whistleblowers, review tactics for seeking advice from professional societies, and practice writing op-eds to raise the alarm about ethical issues to the broader public. Our research shows that the course produces notable shifts in students’ ethical agency, strengthening their ability to serve as “public welfare watchdogs” in the modern workplace.

Ethical engineers for the future

Today’s engineering graduates face complex sociotechnical challenges beyond AI expansion, including climate change and global supply chains; they need tools to navigate ambiguity and act ethically in uncertain contexts. To meet these evolving demands, engineering education must move beyond outdated modes of instruction that focus only on building technical skills—some of which may soon be outsourced to AI. We must fundamentally reimagine the role of ethics and social responsibility, embedding them across the engineering curriculum and throughout the professional formation of engineers.

All engineers learn to solve problems, but ethical engineers can also discern which problems are worth solving and anticipate the consequences of innovation for society. By centering ethics and social responsibility as core priorities of the profession, we can prepare engineers to be stewards of technology and trusted voices in public discourse.

As one student reflected after participating in our modules, “We are a part of the issue if we don’t decide to fix it.”

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.