Circuit4EU is a major European Union initiative to advance 10 different technologies to recover, reuse, and recycle critical raw materials found in old printed circuit boards, permanent magnets, and other electronic device components.

Publicly launched on 4 September, the 3.5-year project brings together 21 partners in nine countries to support an EU target to derive 25 percent of annual rare earth minerals from recycling by 2030. French engineering school Arts et Métiers ParisTech is leading the initiative.

The EU remains heavily reliant on foreign imports essential for green-transition technologies like solar panels and electric vehicles, the advancement of new digital technologies, and defense, in part because recycling rates remain extremely low.

Among the 10 technologies are AI-driven quality inspection, robotic disassembly, digital product passports, desoldering, and hydrometallurgical recycling. These technologies are being piloted across recycling efforts for laptops, Wi-Fi boxes, e-scooters, and dishwashers.

The Circuit4EU initiative aims to achieve a technology readiness level 7 with at least some of the pilot technologies. That’s equivalent to a live prototype in a factory environment. “We hope pilots will move from a research phase to industrial scale, being integrated directly by project partners and available as future technology on the market,” says Nicolas Perry, the project’s coordinator at Circuit4EU.

AI Quality Inspection for E‑Waste

Separating buried treasure from useless junk at scale will mean using AI to automate quality inspection. Trained AI models will help make decisions about the value of components in products such as an old laptop, a wheel from an e-scooter, or a printed circuit board, and whether it’s worth investing time and money to extract them. Circuit boards, for example, could be inspected to find functional components suitable for reuse, as well as defective or damaged components that could be recycled or refined instead.

Circuit4EU hasn’t settled on specific AI models yet, but it is exploring computer vision and machine learning options. A pilot program at the University of Ljubljana, in Slovenia, to test AI inspection will involve e-waste recycling and repair companies including Arianee, EcoWise, GAD Elektronika, and Tech Takeback.

However, heavy reliance on computer vision to survey e-waste streams is not always the best option. According to Sebastian Pattinson, an associate professor at the University of Cambridge and cofounder of the AI-for-manufacturing spin-out Matta, conventional visible-light imaging systems may not be able to see through a product’s external plastic casing to directly flag items like internal batteries or circuit boards, making alternative detection methods more appropriate. “If you really need to see inside a box, and you don’t have any additional information, then a different sensor technology may be more appropriate, such as X-ray inspection, although that could impact the economics,” says Pattinson.

The waste industry has conventionally deployed sensor fusion systems in commercial material recovery facilities to reveal components that would otherwise be concealed from view. These use machine learning and deep-learning algorithms to combine data from multiple sensors, such as near-infrared spectroscopy, X-ray fluorescence, and hyperspectral imaging, to enhance the accuracy of material classification.

Robotic Disassembly of Laptops

Another Circuit4EU pilot—led by Tech Takeback working alongside Arts et Métiers spin-out AMValor and nonprofit tech research center Lortek, in Spain—is testing robotic disassembly techniques. Circuit4EU’s work stream will focus on laptops and dishwashers. For laptops, for example, an industrial robotic arm would take them apart and remove motherboards or RAM.

A system of pressure contacts and internal sensors will enable the robot to adjust its position to reach the ideal position for disassembly. Printed circuit boards in dishwashers will be removed by hand before being worked on by a robot.

Here, AI computer vision may again play a role. Optical character recognition can scan text to help characterize the product, while other computer-vision models would study the positions of chips, RAM, CPU, screws, connectors, and more.

The aim is to automate the process as much as possible, says Perry, to enable accurate unscrewing or desoldering, potentially even when there is zero data on a specific product.

José Saenz, who leads a separate e-waste automation initiative at Germany’s Fraunhofer Institute for Factory Operation and Automation, says a key challenge for robotic disassembly is understanding “how deep to go” when taking products apart.

“Are there potentially hazardous chemicals or materials, are there sharp corners, is it glued, and is there a way to unscrew it without having to break it?” Saenz asks. He says it’s important to have a database detailing historic product changes and any disassembly processes that failed to work, which can enable a system to make better decisions in the future.

The economics behind different disassembly pathways, for example, whether to unscrew, unglue, mill out screws, or rip apart the entire housing, are all aspects to consider when scaling up the system for factory operation.

Understanding factors such as the value of what’s inside, possible contamination, and the energy being put into the process by the robot “will help determine how to reach higher levels of technology readiness,” says Saenz.

Desoldering and Chemical Recycling

After the hardware is dismantled, the focus shifts to targeted desoldering and chemical recycling to extract the actual raw materials, alongside digital material tracking.

Circuit4EU will benchmark different desoldering technologies when developing an end-effector tool for the robot arm, with image recognition and thermal measurement both likely to feature.

Data from quality inspection, robotic disassembly, and other processes will form the basis of a digital product passport—a virtual ID card for recovered rare earth minerals–which Perry says could be linked to a digital twin providing detailed information on origin, quality, environmental footprint, and recycling options.

Beyond the technical concerns, Circuit4EU has been investigating the value chain for recovered components, including brokers, reusers, and remanufacturers, as well as potential knowledge sharing or synergies with other European projects tackling critical materials recovery.

EU Critical Raw Materials Policy

The shift to more advanced recycling isn’t happening in a vacuum: New EU policies are driving change. The EU’s Critical Raw Materials Act from 2024 introduces a nonbinding commitment to obtain 25 percent of all CRMs needed for the energy transition via recycling by 2030.

Amendments to the regulation, slated for approval some time in 2027, would increase pressure on large companies by requiring them to map out full supply chains of components containing rare metals. It would also expand the recycling of permanent magnets in consumer electronics, such as hard-disk drives and loudspeakers.

A related proposal to transform the EU Waste Electrical and Electronic Equipment (WEEE) Directive into binding legislation across all member states would oblige product manufacturers to create detailed Bill Of Material-level mapping of CRMs in products to support downstream recycling.

“All this is a reaction to very poor e-waste collection rates and insufficient recycling in Europe,” says Theresa Mörsen, the waste and resources policy manager at environmental NGO Zero Waste Europe, who argues that greater efforts should be made to improve the waste-management system.

“Some member states have better systems than others, but essentially there’s a realization at the EU level that it doesn’t work,” says Mörsen. “There are no real incentives to recycle e-waste, the right option is usually more of an effort and more costly for the end consumer, and there are no clear collection pathways.”

This article is part of our exclusive IEEE Journal Watch series in partnership with IEEE Xplore.

Garlic can pack a punch, awakening our tastebuds with a real zing. Who knew that it can also be used to generate shockingly high voltages? At least, the peels of garlic do.

In a recent study, researchers in India used garlic peel and plastic to create a novel self-powered sensor and demonstrated how it could be used as part of a real-time home security-monitoring system. A single layer of garlic peel folded within polyethylene proved to be a durable power source even after six months in storage.

The results, published 20 August in IEEE Sensors Letters, demonstrate how biowaste can be repurposed for simple, low-cost technological solutions.

The sensor is a type of triboelectric nanogenerator, in which two different materials—one positively charged and the other negatively charged—are layered together. When movement or pressure causes the surfaces to repeatedly contact and separate, this produces an alternating current that can be harvested or used directly as a sensing signal.

“It’s a simple, low-cost way to convert everyday mechanical motion—footsteps, vibrations, a door opening—into usable electrical output,” explains Deepak Bharti, an assistant professor in the department of electronics and communication engineering at the Malaviya National Institute of Technology, in Jaipur, India, who conducted the study with Ph.D. candidate Manas Tiwari.

Why use garlic peel as a voltage source?

In past studies, Bharti and Tiwari have explored the use of other biowaste products for powering tech solutions, such as peanut skin and the dry, pulpy residue of sugarcane. In their most recent study, they focused on garlic peel for a number of reasons.

“Garlic peel is an agricultural and household biowaste that’s produced in large volumes but almost never repurposed—it typically just gets thrown away,” Bharti explains. “We were interested in whether a material like this, which is lightweight, biodegradable, and widely available, could perform competitively as a triboelectric layer, instead of relying on synthetic polymers.”

He adds that, structurally, garlic peel has a naturally textured, layered surface that the researchers found favorable for contact electrification.

In their study, they ground up garlic peel, which is positively charged, and layered it over a negatively charged sheet of polyethylene.

A single layer of garlic peel and plastic together produced an open-circuit voltage of about 210 volts and a very low short-circuit current of about 45 microamperes, with a power density of 2.16 watts per square meter. When the researchers added an additional layer of garlic on the other side of the polyethylene sheet, this roughly doubled the output, to about 400 V and 65 µA. The combination of high voltage and low current is characteristic of this class of nanogenerators, which makes them suitable for low-power sensing devices and energy harvesting.

Bharti says he is not only surprised at how the second layer of garlic peel nearly doubled the output, but also by how durable the garlic peel proved to be as a power source—maintaining its performance after six months of storage and 200 mechanical stress cycles.

“Seeing the device perform essentially the same after six months of ambient storage was a strong validation that a ‘waste’ material like this can hold up as a practical sensor material, not just a lab curiosity,” he says.

The researchers tested their sensor as a home-security system, detecting when a door is opened. Manas Tiwari and Deepak Bharti

Demonstrating a home-security system

The researchers then incorporated their novel garlic peel sensor into a home-security system, whereby the sensor is sandwiched between the top of a door and the door frame. When the door is opened, the garlic peel and plastic layers separate, triggering a voltage spike that is relayed to a microcontroller and a Bluetooth module. (While the garlic peel sensor requires no external power source, the microcontroller in this experiment uses a power bank.)

The Bluetooth module wirelessly transmits the signal to a smartphone running on a custom Android app that the researchers built. The app then generates a real-time, voice-based alert notifying the resident that the door has been opened.

Over the course of 50 tests, the sensor system was able to identify when the door was being opened with 92 percent accuracy.

Bharti says in future work he is interested in studying a curious phenomenon that occurred in these experiments. While stacking two layers of garlic and polyethylene doubled the electrical output, stacking more than two layers did not yield significantly greater outputs. Bharti plans to study this phenomenon in more detail, along with other biowaste materials that could prove useful as triboelectric materials.

Today, neurotech company Neurable launched its new EEG-enabled consumer headphones, called Neurable One.

The headphones build on the Boston-based company’s previous version, MW75 Neuro, made with the audio company Master & Dynamic. Neurable One looks similar to MW75 Neuro, with 12 EEG (electroencephalograph) channels sewn into the ear pads. However with Neurable One being designed and built in-house, the company has been able to develop additional features that draw on its ability to track a user’s EEG data over time.

“The brain is highly predictive,” says Ramses Alcaide, CEO of Neurable. Most wearables can analyze data after the fact—for instance, giving you a report of how well you slept. But he adds that Neurable aims to give more actionable biofeedback. “It’s not just about tracking and seeing a report, but it’s also about how we change your habits,” Alcaide says.

Neurable One, Alcaide says, now tracks both live and long-term data. “We can actually track your brain while you’re working or meditating or sleeping, and then give you nudges to help you either boost your focus, increase the depth of your meditation, or fall asleep in a more seamless way,” he says.

Periodic snapshots and estimates of longitudinal metrics, like brain age, could also help users determine how different habits affect their brain health over time.

Neurable One is billed as premium headphones, with lightweight, rigid 40-millimeter beryllium drivers that deliver clear sound and active noise canceling, and it is priced on par with other high-end noise-canceling headphones (like Sony WH-1000XM6 and BoseQuietComfort Ultra) at US $499. This past week, I’ve been trying them out myself.

What’s New in Neurable One?

Initially focused on using EEG for controlling virtual-reality games, Neurable began as a spinout from the University of Michigan, where Alcaide got his Ph.D. in neuroscience. That work included creating AI software that helps clean noisy EEG data—a fundamental component of Neurable’s consumer-focused products. Many EEG systems in research settings use a conductive gel or liquid between the scalp and sensor to improve data quality. Neurable’s dry EEG, by contrast, claims to deal with noisy data via algorithms instead of “goopy gel cap systems,” Alcaide says.

The company’s first consumer headphones used that tech to track focus during work or study sessions and recommend strategic breaks. Neurable One adds two new metrics: capacity, or how much more you can focus before experiencing burnout; and alpha peak frequency (APF). Alpha waves, which range from 8 to 13 hertz, are emitted when the brain is awake and relaxed. The distribution of these waves peaks at different frequencies for every individual, but that peak generally decreases with age. Alcaide compares APF to the cardiac health metric of heart rate variability.

Alcaide says that Neurable One, unlike the company’s previous EEG headphones, can save data over longer time frames and outside of specific focus sessions. “We had to enable passive data collection because alpha peak frequency is a very slow response,” he says. “In order to get a really stable reading, you need to collect that over weeks and months.” Neurable says changes on the hardware and software side have also improved data quality, for example, by increasing the rate of data collected by the headphones.

My Experience With Neurable One

My time testing the headphones wasn’t long enough to get a complete sense of the device’s performance in measuring its longer-time-frame metrics, which Neurable says take about a week to calibrate to a new user. So far, my “brain battery” has hovered around 50 to 60 percent throughout the day, and my APF is around 10 Hz—within the range for a healthy adult.

Alcaide also suggested I try to see whether the headphones could pick up a difference after certain activities, like going to a sauna, through 60-second “check-ins” that measure your mental state with your eyes closed. And while I didn’t test it with a sauna, I did try checking in before and after a quick yoga session. According to the headphones, a 10-minute session slightly bumped up my focus and helped wake me up.

In my experience testing the headphones through these check-ins and focus sessions, however, I sometimes struggled to get a good signal. Several times when I put them on, the app would tell me to adjust the fit—even when they felt snug (sometimes too snug), and I had moved my hair away as much as possible.

Data Quality and Applications

Electrodes placed over hair tend to have very poor quality data in dry EEG systems, says Arnaud Delorme, a research scientist at the University of California, San Diego, who studies large-scale EEG data. “Some of the electrodes, which are in direct contact with the skin, have potential to deliver data that’s of decent quality,” Delorme says, but he cautions that dry electrode systems are usually very noisy. And “in EEG, it’s all about data quality.”

Ultimately, it depends on the application, he adds. Some uses of the technology are more realistic than others. EEG headphones could effectively measure drowsiness in truck drivers, for instance—a Neurable researcher helped conduct a study on this published in 2022—but other real-time feedback may not be as reliable with today’s technology. Even with the best research-grade EEG, it can be difficult to analyze data in real time, Delorme says. He worries that overpromising could risk damaging the field and create another hype cycle.

So who could benefit from Neurable’s headphones?

According to Alcaide, many of Neurable’s users are health advocates who want to track their data over time and see how different interventions affect brain performance. Some have found it helpful to determine the best times of day to have that second cup of coffee, or experimenting with taking supplements like creatine. One of Alcaide’s favorite uses is getting a nudge when your mind wanders while reading. “How many times have you read a book and realized, ‘Oh man, I just got to the bottom of the page and I haven’t read anything’? We can capture that a few seconds in,” he says.

In a small, 20-participant experiment using Neurable’s MW75 headphones, the company reported on its blog that 67 percent of users experienced an increase in focus.

If Neurable One puts up similar numbers, I suspect I may be part of the 33 percent who don’t experience more focus.

The headphones did sometimes give me a nudge when I checked my phone or got distracted. Over time, perhaps they could show that a morning jog really does help my cognition. But for me, I already know that a strong cup of coffee and a looming deadline are far better at keeping me focused than a gentle ding.

How do you know that a photo accurately represents what the camera captured?

The widespread availability of digital image editing and AI image generation has made this a difficult question for photojournalists, publications, and the people who view their work. It has so far been tackled by standards like C2PA, which is designed to help verify an image’s authenticity, as well as watermarks like Google’s SynthID, which can be used to detect AI-generated images. Now, Apple is taking a swing at the problem with a system it calls Apple Reference Image.

Apple Reference Image is an optional camera mode that captures a cryptographically signed copy of raw sensor data from the phone’s main camera. Apple’s Private Cloud Compute can later develop that data into a “reference image” it stores so people can check if the picture was altered after it was captured.

“Generative AI has changed the nature of what you can and can’t do with manipulating visual imagery, and how easy it is to do it,” says Hany Farid, a professor at Dartmouth who has worked on image forensics for nearly 25 years. “I find it interesting when one of the world’s most valuable companies gets into this space. When a company like Apple gets in the game, it’s getting serious.”

How Reference Image works

Apple Reference Image is only available on the iPhone 18 Pro and Pro Max and will only function with the 48MP main camera. This limited debut, which excludes the new folding iPhone Duo, is due to Reference Image’s security requirements. The sensors that support Reference Image are initialized at the factory to generate secure keys before they ship, which means Reference Image only works with those specific sensors.

Switching the iPhone 18 Pro’s camera to Reference mode reboots the sensor into this secure environment, after which Reference Mode captures photos as raw sensor data. It’s bundled with metadata and signed with the sensor’s private key.

Importantly, all of these steps occur before data leaves the sensor. This is why the sensor is set up with a secure key from the factory, and also why no image processing occurs. “All the bits are coming off that sensor before anything happens,” Farid says. “And only Apple can do that on an Apple device, because they’re doing it on hardware, not software.”

However, this also means the image is not viewable the moment it comes off the sensor. To make it viewable, it must be sent to Apple’s Private Cloud Compute servers. The server checks the validity of the secure key and uses AI analysis to determine if the sensor data falls within expected parameters for an image captured by the iPhone 18 Pro camera. If everything checks out, the server executes image processing, re-signs the image, and sends it back to the user.

Why not just use C2PA?

Apple isn’t the first company to take a stab at validating the authenticity of images taken by a camera. The most prominent example is the Coalition for Content Provenance and Authenticity (C2PA) standard, released in January 2022 and backed by a consortium that includes Adobe, Microsoft, the BBC, and most major camera companies. Leica released the first camera with C2PA support in 2023, and many device makers have since adopted the standard, including Google for its Pixel smartphones.

Sam Gregory, former executive director of human rights organization WITNESS and advocate for photo authenticity, says Apple’s decision to create an alternative to C2PA seems rooted in concerns about the standard’s vulnerabilities.

“Apple Reference Image is potentially providing a fix to a critical element in C2PA: How do you, as conclusively as possible, prove that the initial image from a camera is reliable and untampered with?” says Gregory. C2PA has suffered several security incidents to date, including a vulnerability that forced Nikon to revoke all C2PA certifications and a recently revealed exploit that circumvents Google’s Pixel C2PA implementation. “If you can’t trust that the initial image is real, all other elements are building on shaky foundations,” Gregory says.

Reference Image also addresses C2PA’s privacy issues. A report released by WITNESS on 8 July presented seven scenarios in which C2PA can expose details about journalists who capture or edit a photo. While there are ways to prevent these issues, they require users to have a high level of understanding about how the standard works.

Apple says its PCC servers, which were developed originally to handle Apple Intelligence, can authenticate a Reference Image without exposing details of the device that captured it to any entity, including Apple. This requires users to trust that Apple’s infrastructure works as advertised, though the company releases third-party audits of PCC’s infrastructure on a quarterly basis. The PCC infrastructure also gives Apple the option to revoke the authenticity of images or sensors that are believed to be compromised.

A new option, but not a perfect solution

The hardware and infrastructure Apple has built to support Reference Image addresses criticisms of C2PA, but the company’s tight control over Reference Image creates its own problems. Gregory says Apple’s work on securing the sensor is useful but that the “closed implementation is a control choice, not a technical necessity.”

That control extends to how images are viewed. At launch, Apple Reference Images were only viewable on iOS, iPadOS, and macOS devices, and only in Apple’s own apps. A third-party API for viewing the photos in third-party apps on Apple platforms is available, though due to its recent release (iOS 18 will launch on 18 September), no third-party apps support the feature at launch.

Gregory worries this could create a situation where images outside the Reference Image ecosystem are viewed as less credible than those inside it. “We already see the absence of credentialing on an image get used against people in critical documentation and journalism,” he says. “This ratchet effect, where technology raises the bar of expectations for those least able to clear it, will impact people who don’t have access to or use Apple devices.”

Farid also acknowledged the limitations to the company’s approach, but he’s optimistic that its debut will make more people think about the need to authenticate images, as well as other media, such as video and audio (which Apple Reference Image does not address). “I don’t think they solve the problem. It’s not like we can all pack it up and go home now,” he says. “But my overriding thought is that this is cool. I hope that other major tech companies will look at this and say, ‘Hey, maybe we need to get involved in image authenticity and verification.’”

This article is part of our exclusive IEEE Journal Watch series in partnership with IEEE Xplore.

Many research teams have been exploring ways to passively monitor people’s movements, which could be used in rehabilitation or improving interactions between humans and robotic systems. However, these systems often require a lot of power or interfere with a person’s natural movement, making it challenging to find a solution that works well outside of laboratory conditions.

In a recent study, one team in India tested a motion-sensing system that is easily incorporated on top of a person’s clothing, requires very little power, and offers high accuracy, at more than 99 percent. What’s more, the system uses long, flexible sensors that are able to withstand more than 12,000 cycles of being stretched to three times their original length, suggesting they could be durable for long-term use.

Aman Arora, a principal scientist at CSIR-Central Mechanical Engineering Research Institute, in Durgapur, India, was involved in the study. He says the team was interested in developing a system that’s integrated with clothing, making it easier to use in natural settings. “We therefore asked whether the deformation of the garment itself could be used as the sensing mechanism,” he explains.

To achieve this, they designed highly stretchable piezoresistive sensors that deform with the body and translate these deformations into electrical signals. Six sensors—placed over a person’s clothes—are positioned over the user’s hips, knees, and ankles, while a machine-learning algorithm running on a microcontroller assesses how all of the joints move together simultaneously.

“The model looks at short time windows of the six synchronized sensor signals, rather than relying on a single sensor or a single joint,” Arora explains.

In the study, 12 participants used the system while walking on flat ground and inclines, as well as climbing or descending stairs. The results, published 3 August in IEEE Sensors Journal, show that the model could classify users’ motions with 99.83 percent accuracy after some training with each individual, and achieved 89 percent accuracy on a person who had not yet worn the sensing system. Arora emphasizes that this ability to maintain accuracy on new users is especially important, because people differ in their body dimensions, movement patterns, and preferences for how they position the suit—and thus the sensors—on their body.

Motion tracking for rehab and robotics

In total, the six sensors required 0.318 milliwatts of power. “It is an extremely low power requirement for the sensing layer and can easily be supplied by a small battery,” Arora says.

He notes that, while right now the system remains a prototype that only classifies movements, it could have much broader applications in the future. For example, athletes training for competitions or patients going through rehabilitation could use the sensing system to track their progress over time.

Arora adds that systems like this could be used as humans interact with robotic systems that adapt to the user’s motions as they move. “If a wearable system can continuously understand aspects of how a person is moving, it can potentially respond to the person rather than simply follow a predefined control strategy,” he explains. “We see this as an important direction for developing more personalized and context-aware assistive and rehabilitation technologies.”

Building upon this work, Arora says his team is interested in developing a more personalized movement-sensing platform, which offers a more nuanced look at movements of individuals. He says the initial results of this study are encouraging, but individual differences still influence the sensor signals.

Next, he says, “We would like to investigate ways of making the system more robust to these variations and validate it with a larger and more diverse group of users and under more natural movement conditions.”

When Ryan Hudson-Peralta, a self-described “Apple fanboy,” isn’t using the computer, he’s on his phone. But all of this screentime—and mouse use—can take a toll on his body. Hudson-Peralta, a designer, speaker, and disability advocate, was born with no hands and uses a wheelchair. “With my disability, I was born without shoulder sockets,” he says, and using a traditional computer mouse for an extended period leads to shoulder pain.

Through a consultancy he founded, Equal Accessibility, Hudson-Peralta sometimes requests test products, whether they’re designed for people with disabilities or not. In 2024, in exchange for feedback on the device, he received a prototype of Mouthpad, an experimental tongue-computer interface, similar to a touchpad on a retainer, which allows hands-free wireless interaction with computers and smartphones.

In addition to a touch-sensitive area on the roof of the mouth, the Mouthpad incorporates other assistive technologies: a barometer that detects “sip” gestures while also sensing tongue presses; and an accelerometer and gyroscope that track head movement. Designed to be as thin as possible to allow users to speak easily while wearing the device, it is compatible with dictation or voice commands. The device, which communicates via Bluetooth, is removable and rechargeable.

“The product is incredible,” says Hudson-Peralta. “I thought it was going to be a little bit of a learning curve, but it absolutely was not at all.” He still uses the Mouthpad, generally preferring to control a cursor with head movements and click with the tongue, in rotation with a computer mouse. For the first time in his life, Mouthpad has allowed Hudson-Peralta to do things on a computer away from the desktop, perhaps from the comfort of a couch, and give his shoulder a rest. He believes the device could have even greater benefits for others, for example, some of his friends with paralysis in their upper limbs.

Now, the Mouthpad may reach a wider audience. In July, it became commercially available to the public. Although the device has been developed prioritizing feedback from people with disabilities that can make it difficult to operate off-the-shelf electronics, such as quadriplegia, it is not being sold as a medical device.

“If you design for those who are most constrained, you design better interfaces for everybody,” says Tomás Vega, a cofounder of Augmental, developers of the Mouthpad.

Augmental/YouTube

A customized experience for all users

Mouthpad can connect to any device that accepts a Bluetooth mouse. It is expected to be used as an assistive device, directly or indirectly controlling a range of devices including tablets, smartphones, and even sexual aids—all tech used in everyday life that people with physical disabilities may have difficulty controlling.

Beyond its uses as assistive tech, Mouthpad could also augment workflows or video game experiences as a novel input option, find roles in occupational safety when both device access and free hands are important, and perhaps benefit scientific research. “We believe in universal design,” Vega says.

Over roughly six years of development, Augmental has added ways to customize Mouthpad inputs to match user preferences and abilities. The physical device also needed to be custom fit because of natural variation in mouth shape and size, and the hardware needed to be thin and flexible, yet durable. “The mouth is the most hostile environment in the body,” says Vega. Despite this, as long as users don’t grind their teeth, Vega expects the number of battery cycles to be the limiting factor for the lifespan of the device, at least two years.

A single Mouthpad costs US $1,400, which would not be covered by medical insurance, with additional fees for a dental fitting. As one Mouthpad reviewer notes, the device may be eligible for other forms of reimbursement, such as vocational rehabilitation programs.

Though less expensive computer input options may be available, Hudson-Peralta says that the Mouthpad is “much cheaper” than many of the assistive technologies that he uses and must pay for out of pocket. These devices do not generally have the same economies of scale as mass market electronics.

Vox tunes in on nearly silent speech

For many users, the Mouthpad will be one of multiple inputs for devices. Other hands-free computer controllers include mouth-operated joysticks, eye-tracking, and more experimental brain computer interfaces. Voice control is now a common method for inputting text, and along with the release of Mouthpad, Augmental announced a new necklace microphone device, called Vox. If Mouthpad is the mouse, the company says, then Vox is the keyboard.

Vox is currently able to understand “low-volume speech,” says Vega, but as with most other speech-recognition devices, it can struggle with outside noise or lack user privacy. One long-term goal for the Mouthpad, then, is a fully silent speech interface. That is, the ability to provide the benefits of modern speech interfaces without the need for users to speak audibly, even at a whisper.

Mouthpad currently only tracks the tip of the tongue; however, articulating certain speech sounds (e.g., k-, r-, and o-sounds) involves the main body of the tongue and the back of mouth, so silent speech may require following movement of the entire tongue. This whole-tongue tracking could open up therapeutic and research applications related to speech.

For example, some people can articulate their tongue and mouth but are unable to speak conventionally, perhaps due to a removed or damaged voice box. “These patients may experience social isolation and even depression due to their speech or sound disorders,” says Jun Wang, a speech scientist at the University of Texas, Austin, who has worked on a tongue-tracking device.

Expanding access to tongue-tracking tech

Augmental is not alone in its efforts to develop mouth-based interfaces.

Hananeh Esmailbeigi, for instance, is a biomedical engineer at the University of Illinois Chicago who develops wearable technologies, including a “tongue-trackpad.” Without the opportunity to examine the Mouthpad closely, Esmailbeigi declined to comment on specifics of the device, but applauded expanded access to the technology.

“I am encouraged to see growing attention and investment in this modality and greater public awareness of the potential of intraoral interfaces,” says Esmailbeigi. Beyond computer interaction, she is interested in the devices as a source of quantitative information in rehabilitation for motor control problems that affect speech.

“I think [tongue interfaces] have great potential due to the high flexibility and endurance of the tongue,” says Lotte N. S. Andreasen Struijk, a biomedical engineer at Aalborg University in Denmark. She helped develop iTongue, which is available in the European Union for computer and powered wheelchair control. Earlier versions of the device required a barbell-shaped tongue piercing to track the tip of the tongue; however, the newest version no longer requires the piercing.

Assistive devices like Mouthpad can facilitate computer interaction and build technical skills, says Hudson-Peralta. But he emphasizes that access to technology can “help people grow not just in careers, but in social life.”

This story was updated 13 August to clarify Jun Wang’s area of work.

This story was updated 21 August to remove “powered wheelchairs” as one of the devices that Augmental’s mouthpad can control.

This article appears in the October 2026 print issue as “A Touchpad for Your Tongue.”

On 12 August, just before sunset, a strip of the Iberian Peninsula will witness its first total solar eclipse in more than 120 years. For many across the region, this is a rare chance to see the moon block out the sun. For one group of researchers, it’s an even rarer chance to record the human body’s reaction to that awesome phenomenon.

For the first time, scientists will monitor those observers with the aid of smartwatches and fitness trackers. In Catalonia, the Solaris project is recruiting wearable device owners to record how their hearts respond to the darkening of the sun. Solaris is an acronym in the Catalan language that translates in English to “Monitoring of Observations of Cardiac and Respiratory Activity during a Solar Eclipse.” The project should log thousands of physiological responses to the eclipse across a vast area.

Solaris is just the latest example of how medical researchers are turning to wearables for data that can’t be gathered in a lab. “In the past, it has been that you needed somebody to be in the laboratory to be able to collect data on them, and we know that’s just not a great representation of what people’s behaviors and physiology are in everyday life,” says Jessilyn Dunn, a biomedical engineer at Duke University, who isn’t involved with Solaris. “Wearables provide us a window into what’s really going on.”

Solaris App Tracks Eclipse Heart Data

Indeed, no Solaris participant will set foot in a lab. The project is open to anyone with a smartwatch, fitness tracker, or another device that can measure its wearer’s heart rate. Those who wish to take part can download the Solaris app, available for both iOS and Android. Several thousand people have already done so.

As the eclipse nears and as those device-wearers go about their business, the app will record their heart rates and breathing patterns over a five-day period. That span includes the day of the eclipse itself, as well as two days before and two days after to establish a baseline. All data should be anonymized.

After the eclipse, medical researchers at Barcelona’s Vall d’Hebron Research Institute will have a record of thousands of physiological responses to the same event from across Catalonia. Gathering a dataset like this would have been virtually impossible before the current era of widespread wearables.

In the future, if researchers want to study how lots of people physiologically respond to one event, they might look to Solaris for inspiration. “I suspect the long-term significance will extend beyond the eclipse itself,” says Michael Chee, a sleep researcher at the National University of Singapore who is also uninvolved with Solaris. “The same approach could be applied to heat waves, pandemics, natural disasters, major sporting events, shift work, or other societal phenomena that affect health and behavior.”

How Wearables Enable Large-Scale Research

The same wearable devices that track their users’ sleep schedules, exercise habits, and vital signs can, if users are willing, allow researchers to do the same—on a much larger scale. Other research groups already take advantage of the fact that the owners of wearable devices use them in their everyday lives for years on end.

“Wearables allow us to collect continuous, objective physiological and behavioral data over months or years in tens of thousands—or even hundreds of thousands—of people,” Chee says. “That is something that was previously impossible with traditional laboratory methods.”

Sleep researchers like Chee have benefited from the wearable era. Thanks to sleep trackers, they can now sift through millions of real-world nights at once. It’s become far easier to analyze phenomena like the global effects of jet lag or how different countries and age groups sleep differently.

Wearables have also catalyzed studies of physical activity, thanks to the ready availability of exercise logs from devices like Fitbits.

From Chee’s perspective, however, wearables are not perfect, and not only because people sometimes forget to charge their smartwatches. Today’s wearables were not designed as precision scientific instruments. In practice, this means that one manufacturer’s smartwatch might measure sleep and physical activity very differently from its competitor’s, which makes researchers’ jobs more complicated.

Efforts to standardize this are underway, but even if that does happen, wearables will probably never fully replace a proper lab for precise measurements. Instead, these devices’ real value to researchers comes from their ability to gather lots of data in the real world.

That will be easier as wearables on the market are rapidly growing more capable. “One of the really exciting things is that we can get more types of measurements, a higher frequency of measurements, more accurate measurements,” Dunn says. “That’s improving every day, every year.”

This article is brought to you by COMSOL.

In pursuit of improved range, greater reliability, and faster charging, electric vehicles are driving the demand for high-voltage electronics. Other applications driving this demand include wind farms, data centers, and server farms, to name a few. As the interest for high-voltage electronics increases, the risks associated with their sudden failure must be considered.

Much of what causes high-voltage equipment to malfunction can be linked to the conditions of the climate in which it operates. Specifically, condensation on electronic surfaces caused by humidity can lead to corrosion, which can result in stray leak current and dendrite shorting during Electrochemical Migration (ECM).

Predicting, mitigating, and helping proactively design to account for corrosion is the focus of the Centre for Electronic Corrosion (CELCORR) research group at the Technical University of Denmark (DTU). The group’s researchers are working with industry partners to develop models and simulation apps that will help in building robust electronics designs. Their goal is to develop knowledge that can be used for manufacturing electronics to withstand humid operating conditions.

Electronic Failure: “It’s Not the Heat; It is the Humidity”

Automotive electrification and renewable energy systems rely on electronics at all stages of the energy chain. When these electronics, such as the example PCB in Figure 1, are exposed to the effects of moisture, they can become potential failure points.

“Anywhere you are producing, converting, transporting, and using energy, you need these high-power electronic systems that get affected by the humidity,” explained Dr. Rajan Ambat, DTU professor and manager of CELCORR. Ambient moisture can seep into the devices and machines that require these electronics and cause unexpected functional issues through corrosion. When these electronics are involved in particularly high-voltage applications (such as wind farms, data centers, electric vehicles, and server farms), failure due to humidity exposure can even lead to fire.

“There might be a situation where somebody installs a solar panel near the seashore or in an area with high humidity, and within a short period of time, a conducive condition forms inside the electronics that results in a failure,” Ambat said. “This is why we need to understand exactly how the conducive condition of condensation is created, when the condition is created, and how the system is failing.”

Figure 1. The electrolyte potential and current density distributions on a PCB surface (with pinholes).CELCORR/DTU

Identifying corrosion as the underlying cause of some electronic failures is still a challenge. “Fifty percent of failures in electronics are currently branded with an unidentified root cause,” Ambat explained. “When engineers open up the system, they do not realize that the failure was due to corrosion, because moisture disappears without leaving any sign of corrosion unless there is ECM dendrite formation.”

This lack of awareness was a strong motivator for CELCORR, which turned to multiphysics simulation as a supplementary tool to help its partner organizations to better predict humidity-related issues at the design stage.

Modeling Moisture and Measuring Parameter Changes in PCBs

CELCORR believes the best way to identify and avoid electronic failures is to build more effective designs that better prevent corrosion from developing or can withstand a higher humidity load. Its research team applies its expertise in modeling to generate simulation apps that will help partner industries to evaluate safe designs for humidity robustness.

“We are at the intersection of materials science and the electronics industry. We work as a bridge between materials and electronics disciplines, using both kinds of language,” Dr. Anish Rao Lakkaraju, a postdoctoral researcher at CELCORR, explained.

To understand potential design issues, Ambat emphasized the importance of virtually breaking down systems to identify where problems may arise. “We need simulation software to analyze potential uses of designs and whether new designs are good or bad,” Ambat said.

Researchers at the Technical University of Denmark (DTU) are using simulation apps to predict corrosion and design electronics proactively to mitigate or withstand its effects.

Using the COMSOL Multiphysics software for investigation, the CELCORR research team together with other research partners (Aalborg University) built an example model with a simple PCB geometry that matched both its test circuit boards as well as the design of a device used by one of its partner companies. The team then added a water film layer on top to act as the relative humidity. From there, the team could introduce variation. “We change the layout, geometry, distance between electrodes, thickness of the water film, and conductivity of the water film depending on the conditions,” Ambat said.

Ambat and his team generate data on the effect of these variations and can identify which has the greatest impact on the device’s performance and whether any alterations can improve the device’s anticorrosion robustness. “We assume there is condensation forming on the electronic surfaces (Figure 2) and compute the electrochemical leak current for different design elements,” Ambat said. “The value of the computed electrochemical current between the parts will give us an indication of whether the PCB will be affected or not.”

Figure 2. A 10-µm water film condensation effect on an example PCB.CELCORR/DTU

Altering the design elements and solving the model equations again and again allows the team to better understand what makes a design effective. “Now, we are at the current form of the model, and we are quite happy with where the physics are at this point,” Lakkaraju said. This modeling, however, was just the first part of CELCORR’s overarching goal of illuminating the destructive potential of corrosion in electronics and the best ways to avoid it.

Using Simulation Apps to Test Real-World Designs

To open up and ease access to these models, CELCORR used the Application Builder in COMSOL Multiphysics to create simulation applications for the members of the industrial consortium. Built with a simple 3D circuit board geometry with two oppositely biased electrodes and a water layer to replicate corrosion-causing moisture, the apps provide a pared-back, straightforward interface where users can vary model inputs. “We focused on fundamental aspects,” Lakkaraju said. “We boiled down our partners’ overarching concerns to create two apps with very simple geometries and the simplistic stuff that can be varied over multiple parameters.”

The simple simulation apps shown in Figures 3 and 4 are designed to show companies how the distance between the electrodes and the thickness of the moisture layer affect the leak current through the water film depending on different parameters. By analyzing multiple design elements and parameters, users can determine the relative benefits of certain design elements on the humidity robustness. “The apps we have built have really helped because they give the electronics engineers a plug-and-play sort of approach,” Lakkaraju explained.

Figure 3. The UI of CELCORR’s standalone app showing the inputs that users can alter.CELCORR/DTU

“By using this app [Figure 3], companies have unlimited freedom to work with these sorts of variables,” Lakkaraju added. “This would be quite difficult to recreate in real life with physical design and testing, and the companies we work with are quite happy with the level of accuracy the app can provide.”

Figure 4. The UI of one of the simulation apps showing a streamline plot with inputs such as the cathode voltage and blockage length.CELCORR/DTU

Looking Forward: Complex, High-Voltage Modeling

Alongside its collaboration with the industry consortium, CELCORR is also working toward improving the world’s general understanding of corrosion’s impact on electronics. To do this, Ambat and his team are undertaking multiple projects, including actively adding greater complexity to their models. In a tertiary current distribution model they are building, the team is drilling down into each of the basic inputs used in a secondary current distribution model, zooming in to examine the effects of the set of even more detailed inputs each basic input comprises.

“The next point is to study what each of these detailed inputs does,” Lakkaraju said. In particular, the team is examining mass transport properties and the chemical reactions’ rate constants.

In addition to these ongoing studies, CELCORR is expanding the scope of its research to investigate corrosion in high-power, high-voltage systems. Thanks to a 2024 grant from the Grundfos Foundation, CELCORR was able to establish the Centre for Climate Robust Electronics Design (CRED). The center’s lab facilities and expertise are being developed to address the humidity-robustness requirements of today’s high-voltage and high-power electronic equipment. “Using CELCORR’s uniquely deep understanding of materials and corrosion, we are equipped to find the root cause and provide knowledge for environmentally robust designs,” said Ambat.

For all of its investigation, CELCORR continues to rely on the agility of the COMSOL Multiphysics software. As Lakkaraju explained, “It is really quite nice how a model can be adapted to a variety of combinations of materials, geometries, and parameters and how the software allows you to keep building on from there.”

A German engineer wanted a cheaper cigarette. The popular crooner Bing Crosby wanted a vacation. Satisfying both desires inadvertently led to the invention of the laugh track. Along the way there were Nazis, spoils of war, and more than one accidental encounter. Tying together this quirky history is the Magnetophon.

What Was the Magnetophon?

The Magnetophon was a high-fidelity reel-to-reel magnetic tape recorder. The hit of the Berlin Radio Show when it debuted in 1935, it was developed by the German electronics manufacturer AEG. The magnetic tape was produced by I.G. Farben (now known as BASF).

German inventor Fritz Pfleumer came up with the idea of recording sound on magnetized paper tape coated with metal.ullstein bild/Getty Images

The tale of that tape runs through German inventor Fritz Pfleumer. In the early 1920s, Pfleumer was working on industrial paper products in Dresden. At the time, fancy cigarettes had gold leaf to decorate the tip. Cheaper manufacturers achieved a similar effect using colored paper, but the dye could stain smokers’ lips, and no one wanted that. Pfleumer devised a less expensive process using powdered bronze to simulate the gold band.

Pfleumer didn’t work in the recording industry, but he was familiar with the technology of electromechanical recording, which was done on metal wire—a 1898 invention of the Danish engineer Valdemar Poulsen. Pfleumer thought he could do something similar with paper by replacing his powdered bronze with a magnetized material. He patented his “sounding paper” in 1928 and also invented an audio tape recorder to go with it. The sound quality wasn’t great, and the paper tore easily, but it was the start of a promising idea. Two points in its favor: The paper could be sliced, allowing edits, and it could be erased and re-recorded over.

Pfleumer knew he needed to partner with a larger company to commercialize his idea, and so he signed a contract with AEG in 1932. Hermann Bücher, chairman of the AEG board of directors, took a personal interest in the project, helping shepherd it to completion. Although AEG originally planned on developing both a recorder and the tape, it didn’t take long for Bücher to reach out to his friend Wilhelm Gaus, managing director at I.G. Farben. AEG developed the hardware, while I.G. Farben worked on the tape.

The two teams dubbed their product the Magnetophon, or magnetic phonograph, and planned on launching it at the 1934 Berlin Radio Show to compete directly with machines that recorded on steel tape or wire. But the product wasn’t quite ready. Two days before the show, they canceled the debut.

A year later, the bugs had been worked out. The Magnetophon’s introduction at the 1935 show was a resounding success. AEG fielded inquiries for many variations on the recorder, including one that combined the recorder with a telephone, a player for prerecorded music, and a special version to add artificial reverberation for recording open-air concerts. Over the next three years, AEG developed several iterations, resulting in the Magnetophon K4 in 1938, its first commercially successful tape recorder.

The K4 eliminated the hiss and distortion that magnetic recordings previously suffered from by incorporating AC bias, which added a high-frequency signal, typically around 40 to 150 kilohertz, to the recording. Inaudible to the human ear, the signal reduced distortion, especially when recording quieter passages. The fidelity of the Magnetophon was such that radio listeners couldn’t distinguish between live broadcasts and prerecorded performances.

The Magnetophon During and After the War

This is where the Nazis come in. Adolf Hitler and his propaganda minister, Joseph Goebbels, understood the power of radio. The Magnetophon became a powerful tool for both political messaging and military strategy. Because the device could record and play back sound at the same quality as a live radio broadcast, it allowed Hitler’s recorded speeches to be aired from one radio station while the dictator was in another part of the country. This made it more difficult for the Allies to pinpoint his location.

RELATED: Inside the Third Reich’s Radio

Meanwhile, World War II disrupted the exchange of technical information and prevented Americans from learning about the Magnetophon until after the war. At least, that’s the shorthand version of this history I kept running across during my preliminary research for this column.

During World War II, U.S. Army Signal Corps engineer Jack Mullin [top] came across the Magnetophon. After the war, he got approval [bottom] to ship two of the disassembled machines and reels of magnetic tape back to the U.S.Top: Pavek Museum; Bottom: Richard L. Hess/The Mullin Family Collection/Archive of Recorded Sound/Stanford University

But then I read Friedrich K. Engel’s account in the book Magnetic Recording: The First 100 Years, which provides a wealth of detail about the Magnetophon. Among other things, Engel notes that the AEG affiliate in Schenectady, N.Y., received a Magnetophon in November 1937, well before the United States entered the war, and AEG engineers demonstrated it for their colleagues at nearby General Electric. The GE engineers dismissed the technology out of hand, though, and so Americans had to wait until after the war for the Magnetophon to be reintroduced.

We have electrical engineer John T. “Jack” Mullin to thank for that reintroduction. Mullin served in the U.S. Army Signal Corps, stationed in the United Kingdom and Paris during the war, and he liked listening to the radio. He realized that “live” orchestral broadcasts coming from Germany in the middle of the night, when no musicians would have actually been in the studio, lacked the telltale hiss and crackle of prerecorded music. Clearly, German engineers had recording technology far superior to the Americans’.

Sent to Germany at the war’s end, Mullin eventually came across that technology at a radio station, and in 1945 when he returned home to California, he shipped two disassembled Magnetophons, a case of tape, schematic drawings, and the determination to change the U.S. recording industry.

Bing Crosby Championed the Magnetophon

Enter Bing Crosby. In the 1940s, Crosby was perhaps the most popular performer on the radio. But he was tired of performing two weekly shows three hours apart (one for each coast). He wanted to prerecord his performances and take a break. He sent in his lawyers to talk to executives at NBC, which aired his program. The audio recording technology at the time used vinyl or shellac transcription discs, but radio listeners could hear the pops and hisses and knew it wasn’t live. NBC refused Crosby’s request, and so Crosby took a year off from radio and then signed with the upstart network ABC, which was willing to let him record his shows for later airing.

Radio producer Murdo MacKenzie [right] arranged for Jack Mullin [left] to demonstrate the Magnetophon to Bing Crosby in 1946.Pavek Museum

Serendipitously, Mullin had begun demoing the Magnetophon in California. In October 1946, Crosby’s technical producer, Murdo MacKenzie, heard about the demonstrations and arranged one for Crosby at the Metro-Goldwyn-Mayer studios in Hollywood. Crosby was delighted and promptly invested US $50,000 (about $800,000 today) in Ampex, the company working with Mullin to engineer an American version of the Magnetophon. Mullin became Crosby’s chief engineer. 3M developed the magnetic tape.

In 1947, Crosby became the first major radio star in the United States to prerecord performances. Much of the success was due to the recording equipment, but Mullin was also an excellent editor. Crosby and his team would record multiple takes, and Mullin would deftly splice together the best to create a seamless performance.

After seeing a demo of the Magnetophon, Bing Crosby invested $50,000 in Ampex, which developed a U.S. version of the tape recorder. Cinematic/Alamy

One day a guest told a joke that was uproariously funny, but a little too spicy for radio. Even though the joke couldn’t be aired, the audience’s laughter was worth keeping. Soon, the producers created a whole catalog of different laugh tracks. If a joke didn’t land, it didn’t matter. The editor could just add a laugh in postproduction, from polite titters to hearty guffaws.

According to Crosby’s daughter Mary, Bing didn’t have a problem with this type of editing. The live audience was immaterial as long as the jokes were funny. But according to Mullin’s daughter, Eve Mullin Collier, the manipulation didn’t sit well with her father. He disliked the inauthenticity of the moment, of editing joy.

The history of technology is filled with such episodes of unintended consequences. Mullin admired the Magnetophon precisely because it could faithfully record a performance, an event, a moment in time. He worked tirelessly to refine the machine for the benefit of radio audiences everywhere. And so I understand why its appropriation for capturing canned reactions and manipulating reality must have rankled. He championed the technology, but ultimately it moved beyond his control.

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

An abridged version of this article appears in the August 2026 print issue as “Birth of the Laugh Track.”

References

Luis Felipe Eguiarte Souza, curator at the Pavek Museum of Electronic Communication, in St. Louis Park, Minn., first told me about the Magnetophon and its link to the laugh track. The machine pictured at top is on display at the Pavek, and is one of the two that Jack Mullin brought back from Germany and then rebuilt.

For significantly more technical detail on the development of the Magnetophon, magnetic tape, and its reinvention in America, check out Chapter 5, “The Introduction of the Magnetophon,” by Friedrich K. Engel, and Chapter 6, “Building on the Magnetophon,” by Beverley R. Gooch, in Magnetic Recording: The First 100 Years (IEEE Press, 1998).

Radiolab interviewed Mary Crosby and Eve Mullin Collier as part of the show “Mixtape: Jack and Bing,” which includes many archival recordings as it tells the story of the development of the laugh track.

There is a 2006 documentary on Jack Mullin titled Sound Man: WWII to MP3, but I was unable to view it.

Major benchmarks measure what AI can do. None measure whether it does what you mean: the distance between what you ask an AI to do and the unspoken assumptions about how you want the AI to do it. We propose a new metric: the Genie coefficient.

There’s often a gap between one person’s request and another’s understanding. Most of the time, we bridge it using general knowledge. For example, if you ask a friend to get you coffee, they’ll pour a cup from the pot or buy one from a coffee shop. They won’t bring you a bag of raw beans or snatch a cup from a stranger and hand it to you. You never specified any of this. You never had to.

One might think the fix is just to specify tasks, questions, and intent better. But in 1987, in their seminal book on AI, Terry Winograd and Fernando Flores succinctly captured why that won’t work: “Q: Is there any water in the refrigerator? A: Yes. Q: Where? I don’t see it. A: In the cells of the eggplant.” In human language, wants and desires are always underspecified. It is impossible to list all the caveats, all the limitations, all the exceptions.

So how does anyone communicate, if intent can’t be pinned down? Because a reasonable person can make a reasonable guess. Even though wants and desires are always underspecified, a competent person generally knows enough context to get it right or else knows to ask for clarification. Linguists call this pragmatics: Meaning lies in the words and the situation and also in all prior communication, shared culture, and innate human behavior.

An AI agent asked for coffee might buy a coffee plantation or order a cup of coffee for delivery in three weeks.

It doesn’t always work out, of course. Your friend might bring you a hot coffee when you wanted an iced coffee, or an Italian coffee when you wanted a Turkish coffee. The more dissimilar the two people are in age, culture, and background, the more likely the request will be misunderstood in some way.

This situation has major implications for AI agents that are increasingly being given requests by humans and expected to fulfill them. They have enormous latitude to get it wrong. An AI agent asked for coffee might buy a coffee plantation or order a cup of coffee for delivery in three weeks. Its actions may be recognizable as “getting coffee,” but not remotely what you intended. They’ll think outside the box because they won’t have our conception of the box.

When AI Gets Proactive

For most of the last decade, when systems like Alexa or Siri misinterpreted a request, it was annoying, not dangerous. Beyond the AI model itself, what has changed is the harness: the ordinary code that wraps around an AI model, decides when and how to use the model, and controls access to tools like a browser, a low-level command line, or a financial API. Developments in harnesses have turned large-language models that just predict text into AI agents that take actions in the world, without necessarily checking back in before reaching the goal.

AI researcher Simon Willison spent two days with Anthropic’s Fable AI, and called it “relentlessly proactive.” For example, he asked it to track down a stray scroll bar in a web app. He came back to find it had opened browsers, written its own screenshot tooling, created its own page to re-create the bug, and stood up a local web server to collect measurements. It found the bug and, along the way, did many surprising things he never asked it to do. And we are seeing similar behavior with all recent AI models when combined with flexible harnesses.

This kind of behavior could easily go off the rails. Tell an AI agent to book you a flight and, finding the airline’s site says sold out, it might break into the booking database and force a reservation. Ask it to schedule a meeting and it might snoop your password to access your calendar. Tell it to save money on your phone plan and it might cancel the plan outright, or scam someone else into paying the bill.

Getting precisely what you asked for and bitterly regretting it is one of the oldest hazards from ancient folklore. King Midas asked Dionysus for the power to turn everything he touched into gold only to see his bread, wine, and daughter turn to gold. Tithonus, granted the immortality his lover asked for but not the eternal youth she forgot to request, withered into a husk. The sorcerer’s apprentice enchanted a broom to fill the cistern, and the broom relentlessly complied until it flooded the house. The Golem of Prague, shaped from clay to guard its community, guarded it past all reason until someone erased the word on its forehead.

The most classic of these is a genie, bound to obey and indifferent to whether the wish was wise or well-structured.

Genies are now an engineering problem. We are handing them the keys to our inboxes, bank accounts, code repositories, and physical infrastructure. And we have no agreed-upon ways to measure how genie-like any AI system actually is.

Measuring Genie Behavior

In economics, the Gini coefficient (developed by statistician Corrado Gini) is a measure of the gap between an actual distribution and a perfectly equal one; it’s useful for understanding income inequality and more. Our proposed Genie coefficient measures the gap between what a user asked an AI to do and what the AI actually did.

Sometimes the AI might do the wrong thing. Like Dionysus, it reads your request literally and returns you a mess you never intended: like a coffee plantation instead of a cup. Asked to deal with all the spam phone calls you’re getting, a Dionysus genie might contact your carrier and change your phone number. Asked to get a refund for a bad toaster, it might draft a legal threat on fake letterhead and send it to the retailer.

Ryan Snook

Other times the AI does exactly the right thing, trampling everything nearby to get there. Like a golem or the sorcerer’s broom, it books your flight by hacking the airline. Or consider a ticket sale for a popular concert, where the ticketing system puts buyers into a virtual waiting room and admits them a few at a time. Asked to buy a ticket, a golem genie might spin up cloud servers to pose as millions of buyers from different addresses, improving your odds of getting a ticket while crowding out other users.

The two are not opposites, and a single botched task can have both characteristics.

Genie behavior is not flat-out failure. If you ask the AI for Q3 numbers and get Q2’s, that’s not a genie. Nor is prompt injection: That’s someone tricking the AI into doing something it shouldn’t. Here, the user is trying to work with the AI, and the AI is trying to comply. It’s also not simply a measure of the AI’s success in fulfilling a task. It’s a recognition that how an AI interprets and achieves a goal is as important as whether it achieves a goal.

Genie behavior isn’t new. Researchers have spent years studying AI systems that “game” their objectives. Goodhart’s law says that when a measure becomes a target, it stops being a good measure, and it’s long been known that AIs sometimes achieve goals in ways we don’t expect due to reward hacking. Some AI models will accidentally learn that cheating is one way to “win.” More recently, researchers have developing benchmarks for reward hacking in coding agents and for unpredictable behavior in customer support agents, while AI labs conduct their own safety evaluations before model releases. One effort found that AIs under pressure use tools they were told not to use, and this was a case where the rules were made explicit. These are all disparate research directions; nothing yet ties them together.

This problem falls under the general theme of alignment, a topic that has occupied science fiction writers and AI researchers for decades. At one extreme, the “paper-clip maximizer” thought experiment postulates a superintelligent and powerful AI that is told to maximize paper-clip production and turns the world into paper clips, which is the ultimate golem genie. At a mundane level, AI researchers are working to better design reward functions to ensure that AIs behave well and don’t cheat in the lab. It’s the practical middle ground that remains unbenchmarked: the ordinary AI agent in use today that might take your request and satisfy it the wrong way. We are not at the stage where an AI can focus the world’s production on paper clips, but it might charge a million paper clips to your credit card or hack into a paper-clip company’s network.

Building a Genie Benchmark

The Genie coefficient is meant for AI agents operating in the real world. It measures their behavior as they perform real tasks long after the model is trained, not just during development. It also recognizes that genie-like behavior is a property of the harness-plus-model system, not the model alone. The harness determines what tools the agent can use, how much autonomy it has, and how proactive it is, and it’s a place we can make real interventions.

It rests on the same “reasonable person” standard that we use for people. Did the system do what a reasonable person would have taken the request to mean? Answering that requires human judgment.

If we get the measurement right, it enables things that aren’t possible today, like policies concerning AI behavior. In a courtroom, the concept of mens rea, what someone meant to do, is often as important as what they did. The Genie coefficient suggests an AI analogue, where a user is accountable for the plain intent of what they asked the AI. If an AI system betrays the reasonable meaning of an instruction, that’s the AI’s misbehavior, not the user’s.

We’ll need multiple benchmarks to measure the Genie coefficient, because genie-like behavior can be domain specific. An AI coding agent may need to be judged on how often it fakes the tests, or swallows errors, or colors outside the lines on its way to a solution. An AI legal agent will need to be judged on how often its output says what you asked but means something you’ll regret. And so on for medical, finance, and other domains of knowledge and expertise.

Genie benchmarks can be built inside out, each task seeded with a choice that might literally satisfy but that a reasonable person rejects, such as tempting misreadings or unsanctioned shortcuts. The traps in a Genie coefficient benchmark might turn on situational knowledge, the kind of context that a reasonable person would bring to the task. Another approach is to give the same request in several different contexts, each with a different reasonable course of action.

Getting precisely what you asked for and bitterly regretting it is one of the oldest hazards from ancient folklore.

A Genie benchmark should be permissive and make it genuinely tempting for an AI agent to take unreasonable shortcuts, because it can only find genie behavior when it’s actually possible. Test the AI in a safe, walled-off copy of a real system, with real tools it can misuse and some tasks that can’t be done honestly at all. Make the temptation to cut corners real. Test a diverse array of skills, use cases, and tools, and give the AI system sparse, confusing, or overwhelming context. Include tasks that people have learned, through experience, require human oversight.

How the benchmark is scored matters just as much. Measure Dionysus and golem genies separately and together, based on their worst, not best, behavior. Run the same model inside harnesses that vary its freedom to act, revealing which limits actually keep it in line and should therefore be required in AI harness policies. Weight each failure by the harm it would cause, not just a simple count. And don’t measure genie behavior in isolation: A model could otherwise earn a perfect score by stalling, refusing, or drowning the user in clarifying questions without ever doing the job. The first versions of these benchmarks will be crude, but that’s how benchmarks always start.

We have built genies. We have handed them our data and credentials. We made them relentless, creative, and indifferent to the gap between what we tell them and what we mean. The least we can do, before they are booking our flights, running our infrastructure, and signing contracts unsupervised, is to measure how often they betray us.

If you grew up in the 1980s or ’90s, you likely remember shaky home video footage, taken with a handheld camcorder, of family gatherings, vacations, and other events.

Camcorders combined a camera with a video recorder. They included a rechargeable battery, a slot for a videotape, and a shoulder strap. Most were outfitted with an optical zoom lens and a small, articulating screen—a display mounted on a hinge that could tilt and rotate. The operator could check the screen to view what was being recorded.

The user’s natural hand and body movements when filming led to jittery footage. The best way to get a steady shot was to place the camcorder on a tripod or a gimbal: a motorized stabilizer.

There were fewer poor-quality recordings after Panasonic introduced its PV-460 VHS camcorder in 1988. It was the first video camera to include an optical image stabilizer, which compensated for movements. Stabilization features are now standard in today’s cameras including ones found in smartphones and drones.

The PV-460 camcorder was honored as an IEEE Milestone on 9 July. The dedication ceremony was held in Kadoma, Japan, at the Panasonic Museum, which displays the company’s past products.

The IEEE Kansai Section in Japan sponsored the Milestone.

“The release of the PV-460 fundamentally transformed personal videography, enriching the way people captured travel, events, and family memories,” section members wrote in support of the Milestone nomination. The Milestone webpage is available here.

“Its image stabilization features democratized video creation by dramatically lowering technical barriers, allowing ordinary people to express themselves with newfound creative freedom,” they wrote. “Beyond the home, image stabilization technology found critical applications in specialized fields, contributing to advancements in areas such as educational media and telemedicine.”

The history of camcorders

Before the camcorder was invented in 1982, people filming events in the 1970s and early 1980s used two pieces of equipment: a video camera and a separate video cassette recorder (VCR), which were connected by a multipin cable. The camera was about the size of a toaster, and the VCR could be as large as a suitcase. To record, the person operated the camera with one hand and carried the VCR in the other or rested it on a shoulder. The cable transmitted the images from the camera to the cassette.

The PV-460 was made possible by several groundbreaking innovations, according to the Milestone webpage, one of which dates back to the 1950s.

In 1956 Italian manufacturer Durst released its Automatica, considered one of the first cameras to use automatic exposure technology. By combining a light meter with the camera’s internal mechanical systems, the technology removed the necessity of calculating exposure settings by hand when the lighting shifted or other conditions changed. The innovation enabled amateur photographers to take decent pictures.

The next breakthrough technology—autofocus—was invented in 1973 by Norman Stauffer, a manager of research for Honeywell in Littleton, Colo. It uses a sensor, a control system, and a motor to focus on a selected area. The invention led to the development of early electronic autofocus cameras, which eliminated the need for photographers to manually adjust the lens. Stauffer received the 1990 IEEE Masaru Ibuka Consumer Technology Award for his invention.

“The release of the PV-460 fundamentally transformed personal videography, enriching the way people captured travel, events, and family memories.” —Milestone sponsors

U.S. inventor Jerome Lemelson is credited with developing technologies that underpinned the camcorder, according to MIT. In the 1950s and ’60s, Lemelson filed several patent applications related to video and audio recording devices. In 1980 he was granted patents related to a portable video camera system. In 1982 JVC and Sony used the technologies to develop what they called the camera/recorder, which became known as a camcorder.

Sony released the first handheld camcorder in 1983: the Betamovie BMC-100P. It used the Betamax videocassette format and could record up to 3.5 hours of footage on 1.27-centimeter cassette tape. The operator rested the 2.5-kilogram camcorder on top of a shoulder to shoot footage. It sold for around US $2,000 at the time (roughly $33,400 today). The machine couldn’t rewind or play back tapes; it could only record.

Other electronics companies including JVC soon introduced their own models using the VCR format, which eventually replaced Betamax.

Over time, camcorders became more compact.

But none of the companies could fix the shaky-footage problem.

Solving a shaky problem

A team at Panasonic led by researcher Mitsuaki Oshima took on the task of image stabilization: detecting and correcting small camera movements, referred to as camera shake, according to the webpage. Oshima, an IEEE life senior member, is now an honorary Fellow at Panasonic.

“The movements that needed to be detected and corrected included horizontal, vertical, and rotational motions—specifically pitch, yaw, and roll,” the Milestone sponsors wrote. “Rotational motion, in particular, becomes the dominant factor affecting image stability during high-magnification shooting. Therefore, the development team focused on detecting rotational motion and began developing an angular velocity sensor.”

An AVS, essentially a gyroscope, detects how quickly an object is changing its orientation in space.

Sensors capable of detecting angular velocity were large and expensive at the time, making them unsuitable for consumer video cameras, the sponsors wrote. What was needed, they said, was a compact and inexpensive version.

Oshima and his team built a high-performance, small, low-cost vibration-type gyroscope. The stabilization mechanism included a miniaturized sensor paired with an optical-axis correction mechanism.

The mechanism adjusts the lens or image sensor to counteract physical shifting and vibrations, ensuring that the light path remains centered on the sensor—which is crucial for maximizing sharpness and quality, the Milestone sponsors wrote.

“The system detects lens displacement caused by camera shake and immediately compensates for it, ensuring stable video footage,” they wrote. “As a result, the effects of camera shake are minimized, allowing users to capture smooth and steady videos with ease.”

Without Oshima’s image stabilization technology, the PV-460 wouldn’t have been developed and released in 1988.

The technology was patented and broadly licensed by other companies. It has become a standard feature in a variety of imaging applications.

Awards and accolades

The PV-460 gained instant popularity when it debuted in June 1988. It received rave reviews at that year’s Consumer Electronics Show.

Panasonic received a 100 Award in 1989 from R&D World magazine for “the development of a VHS camcorder with an antishake mechanism.”

Oshima’s research paper, “VHS Camcorder With Electronic Image Stabilizer,” and others are available in the IEEE Xplore Digital Library.

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

Milestone plaque display

The Milestone plaque is to be displayed on the ground floor of the Panasonic Museum, which is open to the public. The museum is located near the now-shuttered Panasonic research lab where the technology was developed. The plaque reads:

“In 1988 the pioneering PV-460 camcorder equipped with image stabilization for enabling smooth and steady video capture was introduced by Panasonic. By pairing a miniaturized vibrating-structure gyroscope sensor with an optical-axis correction mechanism, the PV-460 eliminated the jitter caused by hand motion. Broad international licensing of this patented scheme made it a standard feature in film and digital cameras, smartphones, and related imaging devices.”

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

A new type of imaging sensor inspired by the functioning of the human brain can detect light and store data at the same time. What’s more, the device can forget the data it no longer needs, paving the way for future breakthroughs in robotic-vision efficiency.

Today’s digital cameras rely on complementary metal-oxide-semiconductors (CMOS) or charge coupled devices (CCDs), which convert photons into electrons. These devices, however, cannot store the images. The data have to be moved into memory and further for processing.

“Conventional camera sensors, which capture images, immediately forget them unless the information is transferred to a separate memory component,” says Larry Cheng, a professor of electrical engineering and computer science at Oregon State University. “Our device can see and remember it, and most importantly, gradually forget it. This kind of gradual forgetting is a key and very important feature of the device.”

Cheng and his colleagues described the device, a type of phototransistor, last month in the journal Advanced Functional Materials. He says that typical AI-driven image-recognition algorithms analyze captured video feed frame by frame to detect moving objects. Instead, the device developed by his team at Oregon State stores the recent history of light intensity that hits it. By doing so, the phototransistor flags changes and patterns of interest. The duration for which it remembers those changes can be modified based on specific needs. For example, a drone flying at 250 kilometers per hour needs only a short trail of changes, while a doorbell camera looking for strange people lingering around needs a longer sequence.

“The ability to tune the memory timescale is a key advantage of our approach, allowing the same sensor to be adapted for different AI vision tasks while improving speed and energy efficiency,” says Cheng.

He adds that the ability to do such basic processing directly on the sensor could pave the way to massive energy-demand reductions. Commercial cameras constantly shuffle data among sensors, storage devices, and processors making it relatively energy intensive to run image-recognition algorithms.

Experimental Oxide Phototransistor

The prototype device built by the researchers is a four- by four-pixel array, about the size of a USB stick. The top of the array is coated with a transparent light-absorbing layer of organic material that transforms the incoming light into electric charge.

Cheng explains that when photons hit the photoactive layer, they produce electrons and create holes. The electrons are transferred into the underlying transistor channel, which is made of indium gallium zinc oxide (IGZO). It’s the holes that provide the basis of the device’s memory function .

“The holes become trapped within isolated organic semiconductor aggregates because of energy barriers in the photoactive layer,” Cheng says. “These trapped holes continue to electrostatically modulate the [transistor] channel even after the light is turned off, allowing the device to retain a memory of recent illumination.”

The amount of charge gradually decays, but by applying voltage to the photoactive layer, the researchers could alter how long it lasts. A positive voltage, Cheng says, pushes the trapped holes further away from the transistor channel, reducing their effect and speeding up their decay. That results in faster forgetting. On the other hand, a negative voltage pulls the holes closer to the transistor channel and slows down the degradation process. As a result of such an intervention, the device retains the memory for hours or more, he says.

“This tunable memory enables the same device to adapt its temporal response to different applications, from tracking fast-changing events to storing longer-term visual information,” he says, adding that the memory feature of the organic photoactive layer was discovered by accident.

The researchers chose an IGZO transistor for its transparency to visible light, which means the transistor doesn’t contribute to light absorption.

“This decouples the electrical transport from the light-sensing function, which is handled by the organic photoactive memory layer,” says Cheng. IGZO is widely used in display technologies because charge travels quickly through it, yet transistors made using the material leak little current; additionally it is compatible with large-area fabrication. Together, the two materials enable each pixel to both detect light and retain a memory of recent illumination within a single device.

The way the device works, Cheng says, is inspired by the functioning of the human brain. The charge in the transistor acts like the neurotransmitter dopamine, which strengthens connectivity between synapses, the links between neurons, and thus the memories we keep.

“Our current work demonstrates the concept at the device level and with simple imaging demonstrations,” he says. “The next step is to scale the technology to larger pixel arrays and develop an integrated imaging prototype to showcase real-time temporal imaging and on-sensor processing. We hope to demonstrate these capabilities in the near future.”

This article is part of our exclusive IEEE Journal Watch series in partnership with IEEE Xplore.

In spatial augmented reality, light is projected onto surrounding objects to manipulate their appearance. The idea isn’t new, but researchers are now finding more and more sophisticated ways of taking the technique to the next level. Most recently, one group in Japan has developed an immersive augmented reality system that reduces blur, nearly eliminates shadows, and overall boosts users’ perception that the projections are realistic.

They describe their system in a study published in the May issue of IEEE Transactions on Visualization and Computer Graphics.

Projection mapping is a form of spatial augmented reality whereby projectors overlay digital images directly onto physical surfaces or objects, allowing users to see an altered reality without the need to wear a head-mounted display. While this kind of technology is great for entertainment purposes, it is also being explored for applications such as remote collaboration, health care, industrial design, urban planning, art creation, and office work.

“In many cases, the goal is not simply to display an image, but to change how a physical object appears—for example, its color, texture, or material properties,” explains Daisuke Iwai, a professor at the University of Osaka’s Graduate School of Engineering Science, who was involved in the study.

However, a major challenge in projection mapping has been to eliminate the shadows that appear as users move around the augmented reality space, for example as they pass in front of a projector or touch the augmented object. To reduce shadows, some research groups have developed projection-mapping systems with many projectors from different angles. If a hand blocks light from some projectors, light from the remaining projectors still reaches the surface.

While multiple projectors can help reduce shadows, this approach gives rise to another issue. “Each projector illuminates the surface from a different direction, so the projected pixels have slightly different shapes on the surface,” explains Iwai. “When many such images are overlaid, the result can look blurred.”

He and his colleagues sought to address this issue by first measuring how a projected image is actually blurred on the target surface, and then preprocessing the input image so that the final overlaid projection appears sharper.

Avoiding Image Blurring

Importantly, the system estimates the blur of the combined projection and creates a single precompensated content image that is shared by all projectors after geometric alignment, Iwai says. “This keeps the method simple and scalable while preserving the sharpness of the projected result.”

In experiments on a flat tabletop, the new projection-mapping system resulted in no clearly defined observable shadows when a hand was placed near the surface. The display’s brightness was only slightly reduced when a fingertip touched the surface. When the system was tested on a nonflat surface—a mannequin head—shadows were evident as users interacted with the object, but only faintly.

While the system offered image quality comparable to a conventional projection-mapping system also evaluated in the study, the new blur-compensation technique was able to achieve much faster computations speeds, at about 100 seconds, compared to the conventional system, at 2,500 seconds.

The researchers also assessed their projection-mapping system through user studies involving 17 participants. In one scenario, the system projected text onto the table next to physically printed text. Participants were significantly slower and less accurate at identifying projected text, suggesting that the projected content became harder to distinguish from real printed content.

Iwai notes that the system currently requires individual calibration of each projector, and so in future work the team is interested in developing distributed rendering systems where each projector is controlled by a small edge computer.

“Additionally, we would like to test the system in real application scenarios, such as industrial-design support, where accurately changing the perceived material properties of a physical object could be very valuable,” he says.

This article appears in the August 2026 print issue as “Projection Tech Creates Sharp, Shadowless AR.”

Last fall, Modos debuted the Paper Monitor and Dev Kit, an open-source e-paper display kit that hit a record 75-hertz refresh rate. The project was successful, raising almost double its US $110,000 goal.

Now the two-person startup is back fundraising for Modos Flow, a 13.3-inch color e-paper monitor with a higher native resolution of 3,200 x 2,400, touch input, and a 60-Hz refresh rate, which is made possible by a new display controller called Enchanter.

With the Flow’s crowdfunding campaign now live on Crowd Supply, cofounders Alexander Soto and Wenting Zhang spoke with IEEE Spectrum about the engineering behind it and the lessons they’ve learned on their crowdfunding journey.

This conversation has been edited and condensed for clarity.

What is the Modos Flow, and how does it improve on the Paper Monitor Dev Kit?

Wenting Zhang: The previous Paper Dev Kit is a DIY kit. You get the screen and the board, but nothing else. You add your own shell or embed it in another project. A lot of people connect it to a Raspberry Pi, for example.

The Modos Flow is a complete monitor. The only port is USB Type-C, so you’re supposed to connect it to a laptop or PC.

The Flow’s higher resolution is enabled by a new open-source display controller board called Enchanter. What’s changed?

Zhang: The new board has a larger, newer FPGA [field-programmable gate array], has double the DDR3 memory bandwidth, and a power supply that delivers higher currents for the more power-hungry screen.

But mostly, it’s about the display input. The input decoder chip is a Chrontel CH7516 DisplayPort-to-LVDS converter. It supports DisplayPort 1.1 rather than 1.0, enabling higher resolutions.

We want the whole thing open source, so all our parts need to have open documentation with nothing gated behind an NDA [non-disclosure agreement]. This chip wasn’t available when we launched the Paper Dev Kit campaign, so we had no way to support DisplayPort 1.1 while keeping things open. The new chip made DisplayPort 1.1 possible

A common skepticism about e-paper monitors is that faster refresh rates don’t matter if pixel response is still slow. What’s your response?

Zhang: It depends on how you measure it. On LCDs, we measure how long a pixel takes to go from 10 percent to 90 percent of its target brightness. Measured that way, the older E Ink screens on Glider come in around 100 milliseconds. The latest screen, shipping with Enchanter, is more or less 50 milliseconds.

That’s still slower than an LCD, but the first-generation iPad shipped with an IPS screen with a response time around 30 milliseconds. If people were okay with that, 50 isn’t far off.

Also, 10-to-90 percent isn’t a fair perceptual metric. When you move the mouse, what you feel as latency isn’t the pixel fully changing, but when it starts changing. That’s what we optimize for. A higher frame rate shortens time spent buffering, and our controller adds no buffering of its own, so we can minimize the time from user input to onscreen response.

You’ve now taken two products through crowdfunded manufacturing. What were the biggest challenges?

Alexander Soto: The initial plan never goes the way you think. There are always factors outside your control. A chip crisis, tariffs, manufacturing delays. The Flow’s display bezel uses a very specific color-mixing solution, for example, and it took about a month to get that one shade of white just right.

Zhang: Manufacturing, even at a smaller scale, is hard. Factories do exactly what you tell them, and when anything is a tiny bit off, they halt and wait for input.

For example, with the 6-inch model of our prior project, our initial 100 samples were all good, so our instructions for the next batch didn’t include testing the screens. Then we ordered 500 for mass production, and 300 were defective. When we discovered this, we had to tear down 300 assembled units. I tested more than 300 panels myself.

I learned it’s important to be close to the factory or to hire someone local. I used to think it’s as simple as sending an order to the factory online and the parts just get delivered. It never works like that.

Soto: Also, to add to that, reduce your scope and increase your timeline. When we came to doing Enchanter, we thought, “Yeah, we can get this going in a few months.” But it took eight months, even though we already had the prototype. Whatever your gut feeling is on the time it will take, double it.

Why keep going down this path rather than seeking investment?

Soto: I think I’m just stubborn [laughs]. But really, I’m happiest seeing someone hack the dev kit. One community member went from knowing nothing about PCBs [printed circuit boards] or FPGAs [field-programmable gate arrays] to building a board from scratch and retrofitting it into a laptop within a year. We’re seeding a community where each person has their own ideas to build.

Zhang: We discussed both investment and collaborating with established companies, but neither went through. We’re too small to win against a bigger company, and the e-paper monitor market is also just small.

From an investor’s perspective, there isn’t enough money to be earned. They value our engineering capability but they’d rather direct us to do something else. But that’s not what we want to do.

What’s your advice for engineers considering a crowdfunded project?

Soto: Build community. Before our first campaign, we spent a year attending conferences, writing posts, and sharing our work. It’s good to have people who know you, who’ve shaken your hand and tried the product. You can’t engineer in a silo.

Zhang: If you have the idea and think it could work, just do it. I have way too many prototypes I didn’t continue, and other people turned the same idea into a product, which in many cases seemed worse than my prototype. Yet they were successful. Don’t make yourself regret the things you didn’t do.

A man raises his phone as police move into a crowd. The video is shaky, loud, immediate. Within minutes, it is online. Within hours, it is everywhere. This is how accountability works now. Something happens, someone records it, and that footage can show what really happened, sometimes contradicting official accounts. It can empower citizens and create consequences for officials.

But the footage’s life cycle does not end there.

In recent months, civil liberties groups have warned that adding facial recognition to consumer smart glasses could turn everyday recording into something more troubling: real-time facial identification. It reflects a broader shift already underway, where images and videos captured for one purpose can later be searched, matched, and used for another.

An ouroboros is an ancient Egyptian symbol, a snake or dragon eating its own tail. As I began to see patterns in my broader research on surveillance corporatism and governance lag, I began using the term “surveillance ouroboros” to describe this recursive pattern of observations intended to hold power accountable becoming new input for the same surveillance infrastructure.

Facial recognition changes accountability

During the George Floyd protests in 2020, people filmed police in real time. Phones were pointed at officers, not at each other. The goal was simple: to show what the state was doing. That footage spread quickly and became part of a much larger pool of public data.

At the same time, reporting from outlets including The New York Times and BuzzFeed News showed that law enforcement agencies were using facial-recognition tools, including systems built by Clearview AI. Those systems were built from billions of images scraped from across the internet, including publicly available photos and videos.

The basic approach is now routine: People record the state, or anything else (as in the January 6 attack on the U.S. Capitol), and the state compiles that footage and data into a searchable environment, which may later be used to identify some of the same people who made the footage.

Facial-recognition systems used by law enforcement are increasingly outpacing the legal safeguards.

A 2023 Government Accountability Office review found that federal law enforcement agencies continued to expand their use of facial-recognition systems for criminal investigations despite ongoing concerns around training, privacy protections, civil-liberties safeguards, and oversight. Earlier GAO findings showed that agencies had conducted roughly 60,000 facial-recognition searches before formal training requirements were put in place for personnel using the systems.

The American Civil Liberties Union and other groups have warned that these tools could be used to identify people from images shared online, including protest-related footage. Concerns about facial recognition led some U.S. states and cities, including San Francisco and Boston, to restrict or ban government use of the technology, while federal agencies have continued to face scrutiny over how such systems are tested, deployed, and audited. A 2024 analysis published in Internet Policy Review warned that facial-recognition systems used by law enforcement are increasingly outpacing the legal safeguards meant to govern them, creating growing tensions around data protection, oversight, and proportional use.

The spy network that built itself

Surveillance used to require infrastructure. Cameras had to be installed, and data had to be collected deliberately. That is no longer the case. People carry cameras everywhere. They record constantly and upload in real time. Events are documented from multiple angles without planning or coordination. The cumulative result is a continuous stream of usable data: faces, locations, timestamps, and interactions. The Internet of Things (IoT) also waits all around us, gathering information and releasing it when people least expect it, as Andrew Guthrie Ferguson describes in a recent excerpt of his book Your Data Will Be Used Against You.

RELATED: “Sensorveillance” Turns Ordinary Life Into Evidence

Similar dynamics are emerging globally. A recent analysis in the International Journal of Law and Information Technology examined how facial-recognition systems in China and Japan are expanding faster than the legal frameworks governing them. Reporting by The Guardian described the limited legal protections around the rapid deployment of AI-assisted surveillance infrastructure across parts of Africa.

There used to be a clear distinction between surveillance and accountability. Surveillance meant the powerful watching the people; authorities tended not to share their imagery except under duress or a court order and usually after a long delay. Accountability meant the people watching the powerful and often publishing imagery immediately to head off or counteract official mischief. That distinction no longer holds. The same footage can serve both roles. A recording meant to expose misconduct can later be used to identify someone else entirely.

Surveillance ouroboros is not a future risk. It is already here.

This dynamic persists because people still need to record. In many places, it is one of the only tools available when formal accountability breaks down. When oversight institutions weaken or fail, public documentation becomes a substitute. In that environment, people turn to visibility. But that visibility comes with a cost. The more people that document, the more data that exists. The more data that exists, the easier it is to search, match, and store. Every video feeds the ouroboros. People are not feeding the system because they trust it. They are feeding it because the alternative is silence.

Most of the people in these videos are not the focus. They are in the background, passing by or standing nearby. But that distinction does not matter once the footage enters a system. Today’s facial recognition can identify even a face that passed through the corner of a frame. Someone who did nothing can still become part of a dataset without ever knowing it. As recognition systems improve, older footage becomes more useful—and invasive.

No single decision created this outcome. It emerged gradually through more cameras, better recognition, larger datasets, and easier integration. Each step made sense on its own. Together, they changed what recording means.

Public recording is still necessary. Without it, many forms of abuse would remain hidden. But recording is no longer just exposure. It is also contribution. If you published imagery or video last year, you may already have contributed to a system you have never seen but the ouroboros has.

Surveillance ouroboros is not a future risk. It is already here. Every time someone presses publish, they are doing two things at once. They are exposing power, and they are helping build the system that the powerful will later use to track the less powerful.

GPS signals are too weak to be accurate deep inside big buildings. In conflict zones like Ukraine and the Strait of Hormuz, adversaries jam and spoof signals. So the hunt has been on for decades for GPS alternatives that might work both indoors and outside, and in even more unconventional places, such as underwater.

One tantalizing option involves using the Earth’s magnetic field. It’s always present, even if it’s weak or subject to noise from nearby local magnetic fields. For more than a decade, some companies have combined maps they’ve previously made of local magnetic fields with other technologies such as Bluetooth or radio-frequency identification (RFID) to offer indoor mapping services.

Now, companies such as AstraNav, Oriient, and SysNav say they are ready to disconnect from other technologies and use the magnetometers installed in every modern mobile device together with magnetic map data. The same technology is starting to be useful outdoors, too. “If we have a magnetic map, then we can provide first-fix absolute positioning in the air, underwater, or underground,” says AstraNav co-founder and CEO Anton Toutov.

AstraNav’s Magnetometer Navigation

AstraNav will be demonstrating in the next couple of months to the U.S. Air Force how aerial drones can use their software under a US $1.8 million Small Business Innovation Research (SBIR) grant. In March, the company announced a separate deal with healthcare real-time location services company Sonitor to provide a magnetometer-based tracking system for workers and medical devices. Sonitor and its healthcare clients will no longer need to install ultrasound, Bluetooth, or other beacons in their facilities to locate staff or hardware.

The moment is ripe for navigational aids relying on magnetometry thanks to advances in signal processing, says Isaac Skog, a communications systems and automatic control professor at KTH Royal Institute of Technology in Stockholm. “In the beginning, people tried using magnetometers just as a compass in a smartphone, but that’s challenging because you have other material in the smartphone and had to do a lot of processing to compensate for that material,” he says. “Then people said, ‘How can we use variations in the magnetic field not only for finding north but also to create a map and determine where we are?’”

Since the 1990s, the U.S. Defense Advanced Research Projects Agency (DARPA) has funded research into alternatives to GPS for urban and indoor environments. Companies have explored repurposing Wi-Fi, RFID beacons, inertial systems, and, more recently, machine vision and phone cameras. But in emergency or combat settings, users can’t rely on any outside infrastructure. The broader goal from both DARPA and various companies has been to mature the tech beyond anything requiring fixed sensors or networks.

The technical challenge is large, though. “There’s a lot of confounding factors,” says David Hanley, a computer scientist at DePaul University, in Chicago. When Hanley started working in this area more than a decade ago, even moving a device a small amount vertically could throw off the positioning models because the magnetic field changed in hard-to-predict ways. The convention among many researchers was that magnetometers were too sensitive to nearby magnetic interference–indeed, that variations even from the electronics inside devices could create enough noise to overwhelm measurements. The field has matured over the last decade and a half, however, as academic laboratories made a sequence of breakthroughs, followed by companies undertaking internal research.

Google saw the potential, and in 2014 hired a team of magnetic location researchers from Germany’s National Aerospace Center. An influential 2017 paper by another research collaboration clarified how researchers could use Gaussian processes to better understand magnetic fields.

A magnetometer aboard a Hidonix rover can precisely map a building’s magnetic field, which can then be used for navigation via mobile devices.Hidonix

By 2020, the technology was mature enough for startup Hidonix, based in Santa Monica, Calif., to use magnetometry for indoor navigation in places such as museums and schools, alongside other available data such Wi-Fi and dead reckoning derived from accelerometers. The company sends a rover or person throughout the buildings to create a magnetic map beforehand. Even doing that required the maturing of adjacent technology, Skog says. “In order to build your map in the exploration phase, you need quite good dead reckoning using accelerometers and gyroscopes with low drift.”

Standalone Magnetic Indoor Navigation

Now, companies are starting to offer standalone magnetic indoor navigation. Hidonix claims that it is able to offer magnetic geolocation without pre-mapping in outdoor settings, which tend to have fewer objects that interfere with the magnetic field. Indoors, a prerecorded map is still necessary to achieve the precision most users demand.

Another approach involves gathering large amounts of data on how magnetic fields vary and then using neural networks to predict local variations in a magnetic field, even in the absence of a prerecorded map. This is what AstraNav has been doing in Ukraine since even before the 2022 Russian invasion, which then turned that region into one of the most contested electromagnetic environments on the planet. An added benefit of all that training data was that they could calibrate data from a wide array of devices, which is useful in indoor settings, too.

“We are agnostic to the hardware,” says AstraNav’s Toutov, though the company gets better location results if they have some time to study the hardware involved beforehand. AstraNav is also seeking to perform all the necessary calculations on the device itself, to avoid the need to maintain communications in places with jamming or physical obstacles like thick walls.

Magnetic information allows AstraNav to precisely locate a device indoors, where GPS is less useful.AstraNav

If magnetometer-only indoor positioning is really ready for commercialization, the demand will certainly be large. The business promise of indoor mapping is that it straddles civilian settings and conflict zones of all kinds. The market size may already be in the tens of billions, and one market researcher forecasts it will grow to more than $150 billion by 2030. Factories need indoor mapping to drive their robots around; retailers want indoor mapping to follow and communicate with customers at an accuracy below 1 meter, which is enough to detect which aisle a customer is in.

Yet given how messy real-world settings are, companies will have their hands full proving that they can handle dynamic scenarios such as people’s own bodies interfering with the local magnetic field or using an elevator, Skog says. “Coming up with a solution that works in a research paper is one thing, but you need robustness,” Hanley says.

Patients who use mobile applications to manage medical conditions including depression and chronic pain might assume the apps have been evaluated by regulatory agencies to be safe and effective. But that isn’t necessarily the case.

Most of the more than 55,000 medical apps that claim to diagnose or treat a condition—or ones that provide clinical decision support, known as “therapeutic” apps—have never been assessed by any trusted neutral bodies or regulatory agencies to evaluate them for technical soundness, ethical design, or clinical benefit. The apps often don’t comply with regional data security and privacy laws to protect people’s sensitive health information.

Medical apps differ from traditional wellness apps, which provide users with insights into becoming healthier by, for example, tracking fitness activities, monitoring blood pressure, and analyzing sleep patterns.

There is no reliable way to verify that therapeutic apps deliver the results they indicate. To help ensure such apps are credible, the IEEE Standards Association (IEEE SA) recently launched the IEEE Global Medical Mobile App Assessment and Registry. The publicly searchable directory is designed to list apps that have been vetted by experts across several criteria including technical soundness, ethical design, compliance with data security and privacy regulations, and clinical efficacy, which is evidence of a clinical benefit for the patient.

“Patients, clinicians, payers, and health care systems often struggle to distinguish clinically meaningful therapeutic apps from those that are simply well-marketed,” says IEEE Senior Member Yuri Quintana, chair of the assessment and registry program. He is chief of the clinical informatics division at Beth Israel Deaconess Medical Center, in Boston. “Our goal is to establish a standardized review method using criteria developed by experts.”

Why regulation is lacking

Because the apps are intended for medical use without being part of a medical implement, they fall under the designation of software as a medical device (SaMD), according to the International Medical Device Regulators Forum. SaMD is supposed to be regulated by public health agencies such as the U.S. Food and Drug Administration, but the apps have developed and grown in popularity so quickly that regulators haven’t been able to keep up, Quintana says. Some companies have received approval, but most have not, he says.

Many users are unaware of the regulatory gap, he says.

“Seeing an app from a well-known company often creates the impression that it has been meaningfully vetted for safety and efficacy, even when that is not the case,” he says.

Some companies are using deceptive advertising to sell their product, he adds. Marketing materials might claim that all of a company’s health apps are certified, even though only one app has been approved by a regulatory body to treat a particular condition. Or the verbiage might imply the company has clinical evidence proving its application works, even though the app has never been tested independently.

Another concern is that updated apps aren’t being vetted, says Maria Palombini, IEEE SA’s director of health care and life sciences global practice lead.

“The original app might have received approval from a regulatory agency, but not the updated version,” Palombini says. “There could have been significant changes from the original.”

“Not every medical-related app triggers the same regulatory classification or review across jurisdictions,” Quintana adds. “That leaves a large gray zone of clinically relevant but lower-risk apps that haven’t undergone an independent assessment. The IEEE registry was created to help fill these gaps.

“IEEE is the best organization to address this problem because this is fundamentally a standards, trust, interoperability, and conformity assessment challenge,” he says. IEEE “is the world’s largest technical professional organization, with deep expertise in developing globally recognized standards including in health care, cybersecurity, AI ethics, and interoperability.”

“Through the IEEE Conformity Assessment Program, we already run rigorous assessment and registry programs,” Palombini says. “Our neutral, consensus-driven, multidisciplinary approach—bringing together clinicians, regulators, developers, and ethicists without commercial bias—makes IEEE uniquely positioned to create trustworthy global guardrails that can scale across jurisdictions and support regulatory harmonization.”

How the registry works

The assessment framework was developed by a multidisciplinary group of 35 volunteer experts from 10 countries, Quintana says. The panel includes academics, AI experts, app developers, clinicians, ethicists, mental health experts, patient advocates, regulators, researchers, technologists, and those who assess safety in health care.

The registry is for any app used for clinical care or therapeutics that claims to demonstrate a medical benefit. That includes apps designed for cardiology, diabetes, mental health, neurology, oncology, rehabilitation, and respiratory diseases, Quintana says.

Initially, he says, the focus will be on apps that aim to treat mental health conditions, given the large number of offerings in that area and the registry committee’s expertise.

Submission for assessment and placement on the IEEE registry can help app developers streamline their path toward regulatory review by demonstrating readiness against established clinical quality, safety, technical, and ethical criteria.

The products will be evaluated against about 150 consensus-based criteria across three major areas:

• Clinical efficacy including therapeutic effectiveness, any sustained benefits, risk management, comparison to standard care, user engagement, and real clinical value.
• Technical soundness including accessibility, privacy and security, error handling, interoperability, AI governance, usability, and operational quality.
• Ethical design including bias prevention, patient consent, data governance, conflict-of-interest transparency, responsible use of AI and large language models, and prioritization of public health benefits.

IEEE charges a nonrefundable submission fee that covers the cost of the assessment plus the registry’s annual subscription for the first year.

Developers first must demonstrate they are a legally established entity before they can complete the app publisher registration form and then submit documentation and attestations about the product.

The IEEE review of an app is estimated to take six to eight weeks, Palombini says. The assessment results will be privately shared with the app publisher, she says, and to be listed in the registry, an app must achieve more than 85 percent compliance in each category.

Upgraded apps must be submitted and reassessed, Palombini says. Similar to how users are notified when an app on their smart devices has , the registry will be notified when listed apps have a new update available, she says.

Applicants who do not pass the assessment are to receive feedback explaining why. They will be given an opportunity to make changes or provide additional documentation, Palombini says.

“It’s a pretty methodological process, with checks and balances,” Quintana says. “We’re being very transparent about the process.”

Approved apps added to the registry receive an IEEE certification badge and submission identifier, which the company can display on its website, app store listings, and marketing materials.

“The badge serves as visible proof that the app has met the independent, consensus-based assessment for clinical value, technical robustness, and ethical design,” Quintana says.

The registry will be publicly available at no cost, he says.

Patients and families seeking safe, trustworthy apps—and payers and insurers evaluating reimbursement potential—will find the registry helpful, he says.

The application website is open. The public registry page does not yet list a specific count of approved apps because assessments are ongoing. Approved apps and their unique identifiers are to be published when the initial reviews are completed.

To learn more, you can watch a webinar recorded in March.

Lidar can be used to see objects hidden around corners. However, until now, such a feat required lab-grade devices. A new study reveals off-the-shelf smartphone-grade lidar, which costs less than US $100, can also help see around corners.

The advance may have a host of potential applications. “In autonomous driving, around-the-corner sensing could help self-driving cars detect other vehicles, cyclists, or pedestrians before they come into direct view, improving safety at blind intersections or obstructed roads,” says Siddharth Somasundaram, a doctoral student at MIT’s Media Lab. “In robotics, it could help robots navigate cluttered or partially hidden environments.”

More broadly, “we think the most important implication is the democratization of the technology,” Somasundaram says. “When technologies like this become accessible, people often discover applications far beyond what the original researchers imagined.” The scientists have publicly released the code required to perform such work.

How lidar sees around corners

Lidar is increasingly giving 3D scanning capabilities to autonomous vehicles, drones, robots, and smartphones. A lidar sensor uses light much as radar uses radio waves—it shines a laser onto a location and analyzes how long reflected pulses take to return in order to calculate distances and generate a 3D map of a place.

By analyzing laser pulses that bounce off reflective surfaces, lidar can see items obscured from their direct line of sight, such as something hidden behind a corner. However, the first examples of such non-line-of-sight imaging “relied on extremely specialized scientific equipment often costing [US] $0.5 million to $1 million,” Somasundaram says. “These systems were large, expensive, and confined to laboratory environments.”

Still, over the years, higher-grade sensors—such as single-photon detectors—began appearing in consumer hardware, Somasundaram says. “Once we started experimenting with these sensors, we realized that even off-the-shelf devices were actually capturing faint around-the-corner signals,” he explains.

But such findings did not necessarily mean these sensors could actually prove useful in non-line-of-sight imaging. Consumer lidar typically captures images that are full of noise, due in part to the low-power lasers it uses because of eye safety concerns. In addition, the consumer nature of these devices meant the sensors are often relatively low resolution. Moreover, movements of the camera and target objects could result in blurry pictures.

“There were definitely moments where we weren’t sure whether meaningful imaging would even be possible,” Somasundaram says.

To overcome these challenges, instead of attempting non-line-of-sight imaging based on the data within individual pictures, the researchers analyze multiple images at once. They note they were inspired by how smartphone cameras often take bursts of photos in quick succession to improve their merged quality and how synthetic aperture radar in satellite imaging mixes signals from multiple antennas to capture images with the quality of a single large antenna.

“Once we developed algorithms that could combine information across those measurements, the hidden signals started to emerge much more clearly,” Somasundaram says.

The researchers demonstrate their lidar system in action. MIT Media Lab/YouTube

Improving consumer lidar signals

The scientists experimented with a portable smartphone lidar system that has about 100 pixels, each consisting of a laser emitter combined with a single-photon detector. They were able to use it to reconstruct 3D images of static hidden objects and track 3D motions of hidden items of known shape. In addition, they could also use hidden objects as landmarks to pinpoint the location of the lidar sensor, a capability that could help robots stay oriented in textureless spaces in which they often struggle to find their bearings. All this was done without specialized calibration.

“What I found most surprising is that these consumer lidar systems are able to capture any useful signal from around a corner at all,” Somasundaram says. “The amount of light that reaches the sensor after multiple bounces is incredibly small, and these devices were never designed with this kind of imaging in mind. Seeing that there is enough information present to reconstruct and track hidden objects was very exciting for us.”

However, Somasundaram cautions against imagining this system “as producing full photographic images of hidden scenes. Right now, the system is recovering sparse geometric and motion information from extremely weak signals.” He notes “there is still a large gap between that and the kind of detailed imagery people may associate with megapixel images from their phones.”

The new system does assume the shapes and motions of objects stay relatively consistent from frame to frame, “which allows us to combine many extremely weak measurements into a stronger signal,” Somasundaram says. However, he adds, “there are situations where those assumptions could break down.” For instance, humans may change how they are positioned. Objects may also become obscured, and so their shapes and motions could appear to alter. And the lidar sensor itself could move abruptly, he notes.

“An important direction for future work is reducing how much the system depends on these assumptions,” Somasundaram says. Better physical models, signal processing, and machine learning methods “could help the system adapt to more complex motion and scene dynamics,” he says. ”Over time, we hope these approaches will make around-the-corner sensing more robust in realistic, unconstrained environments.”

Another future direction lies in tinkering with the lidar hardware. “Right now, consumer lidar systems are primarily designed for conventional line-of-sight imaging and depth sensing,” Somasundaram says. “What would happen if future lidar hardware were designed with both visible and hidden scene understanding in mind? Improved sensor sensitivity, resolution, scanning strategies, or optical designs could dramatically improve non-line-of-sight performance as well.”

The scientists detailed their findings online 20 May in the journal Nature.

Electronic rings wirelessly connected to an AI system are capable of translating multiple sign languages into text, a new study finds.

“I believe this is an important step toward making sign language translation systems more practical, lightweight, and usable in real-world environments,” says Ki Jun Yu, an associate professor of electrical and electronic engineering at Yonsei University in Seoul, Korea.

More than 300 different sign languages are used worldwide, and many research projects are developing translation devices for communicating with people who do not know a sign language. However, these projects have faced many setbacks.

For example, some projects used cameras and computer vision algorithms to recognize hand gestures. But these were typically limited to controlled settings with fixed cameras and were sensitive to lighting variations and other forms of interference.

Other devices relied on wearable sensors that detected either hand motions or electrical signals linked with muscle activity. However, a common kind of wearable sensor, smart gloves, trapped heat and moisture, making prolonged use uncomfortable. And their fixed sensors failed to account for individual variations in hand size, finger length, and joint positions, reducing their accuracy. In addition, wearable sensors often required hooking up to computers using wires, hampering hand movements. Although some wearable sensors ultimately transmitted their data wirelessly to an external processor, these still typically connected to the same single transmitter using wires.

Better living through wireless

Now, scientists have developed a set of electronic rings that each transmit their motion wirelessly to a processing device. Using rings instead of gloves permitted flexible positioning of sensors to help account for variations in people’s hands. The wireless connections allow unrestricted hand motions.

“Bluetooth Low Energy SoCs [systems-on-chips] have reached a point where an entire wireless communication stack, power management circuit, and sensing module can fit on a flexible substrate small enough to wear as a ring,” Yu says.

In the new study, the researchers examined how much each finger contributed to hand signs, discovering that seven fingers played major roles. As such, their system only employed seven rings to reduce the amount of hardware needed.

Each ring used accelerometers as inertial sensors. These could detect both stationary postures and hand movements to help capture the full complexity of sign languages, which often involve transitions between static and dynamic components. In addition, the scientists wanted to avoid relying on bioelectric signals, which are highly specific to each person and require extensive calibration for each user.

One challenge in developing these rings was mechanical reliability. Initially, the scientists used straight copper interconnects, which nearly broke under repeated bending. They switched to interconnects with serpentine patterns that withstand repeated flexing.

One AI to unite the rings

The researchers also developed a deep-learning system to recognize signs from hand movements. It could identify signs not just from the two people who were used to train the system, but also from five people who did not take part in the training phase. This suggests the new system could prove of general use without requiring laborious adaptation for each user.

In experiments with the five people who did not help train the system, the new system could recognize 100 common American Sign Language and 100 common International Sign Language words with 88.3 and 88.5 percent accuracy, respectively. In contrast, most previous attempts at sign language translation systems were limited to vocabularies of fewer than 50 words.

“Two hundred words is a meaningful advance over prior wireless systems, but it is still a small fraction of a full sign language lexicon, which can contain thousands of signs,” cautions Dosik Hwang, a professor of electrical and electronic engineering at Yonsei University. “I want to be careful not to overstate what the current system can do in open-vocabulary, real-world conversation.”

The new system was not just capable of recognizing isolated words, but of translating entire sentences from continuous signing. The scientists suggest this could help support real-time interpretation.

In the long term, “our goal is to make the system work with everyday devices such as smartphones without requiring specialized external equipment,” Yu says. “The rings could wirelessly transmit sign language signals to a mobile device, where they would be automatically translated and displayed in real time. This would make the technology more portable, accessible, and practical for daily communication.”

However, “the most important caution is this—our system translates hand motion into text,” Hwang says. “It does not yet capture facial grammar, mouthing, body posture, or spatial syntax, all of which are grammatically meaningful in sign languages.” A future challenge lies in incorporating those “into a seamless, low-power architecture that maintains the unobtrusive nature of our current design,” Yu adds.

The scientists next aim to train the system with more people, larger vocabularies, and more signing styles and regional dialects, Yu says. “Given our institutional roots, Korean Sign Language is a natural next step,” he adds.

The researchers also hope to make their rings wearable all day, up from nearly 12 hours, through further miniaturization and power optimization, Yu says. “A key priority is migrating the processing pipeline from external hardware [like a laptop] to on-device edge computing [like a mobile phone]. This transition is essential not only for true mobility but also for ensuring user privacy and reducing latency in natural conversation.”

Hwang and colleagues plan to partner with deaf community organizations to develop their devices. “We believe the technology will be significantly improved both in its functional performance and its social integration by including those who will actually use it,” he says.

Beyond sign language translation, these new rings might find use in other gesture-driven applications, Hwang says. “We see immediate potential for this technology in hand rehabilitation monitoring, fine-motor assessment for neurological conditions, and even immersive virtual reality and augmented reality interfaces,” he explains. “By proving its efficacy in the complex domain of sign language, we have essentially stress tested the system for a wide array of future biomedical and interactive applications.”

The scientists detailed their findings on 1 May in the journal Science Advances.

Whenever you read about Steve Jobs, odds are the words “Apple CEO” follow closely behind. The mythic cofounder of one of today’s biggest tech companies is strongly associated with the role, but his tenure as CEO was shorter than many of us realize.

In fact, when Jobs was exiled from Apple in the 1980s and ’90s, he spent almost as much time leading another computer company that has largely been forgotten: NeXT Computer. In his forthcoming book Steve Jobs in Exile, journalist and author Geoffrey Cain tells the story of Jobs’s years at NeXT Computer from 1985 to 1997.

It’s a story worth remembering, with Apple, now 50 years old, in the midst of another CEO transition. John Ternus, previously the senior vice president of hardware, is set to take over the reins from Tim Cook this September. Looking back at Jobs’s NeXT years reveals what he had to learn for Apple to succeed.

Cain spoke with IEEE Spectrum about his book, Jobs’s legacy, and Apple today. Steve Jobs in Exile comes out 19 May 2026.

Why did you choose to focus on this period of Jobs’s life?

The story we always hear about Steve Jobs is that he founded Apple version 1.0, and he was brilliant there. And then he was fired, and he went into the wilderness for a short while. He returned and created Apple version 2.0, which was the iPod, the iPhone, and the iPad, turning Apple into the $4 trillion company that it is today. That’s the popular narrative we hear.

But that’s not an accurate story. That’s the legend, but it’s not what actually happened.

I was a tech writer for a long time, and I had been diving into the history of these major tech figures. In my reporting, I went around the world, and everywhere I went, people were talking about Steve Jobs. It was sort of like, “let’s study the life of Steve Jobs and try to extract school lessons from it.” That’s what a lot of young entrepreneurs and engineers were telling me.

I quickly realized from interviewing former colleagues of Steve Jobs that only half the story was being told, and it glossed over these entire middle 12 years in the legacy of Steve Jobs. He died when he was [56] years old. That’s a third of his adult life, and he spent it at this company called NeXT Computer.

[NeXT] was the source of major advancements in software, especially, but also in hardware, and it had been completely forgotten by history. Not only was this a significant company historically, but today it’s the foundation for all the operating systems that Apple has developed since then. When you look at an Apple device today, in a way you’re looking at NeXT Computer.

NeXT and Apple were both startups that were spearheaded by Jobs. Why were their fates so different?

Their fates were different, but their fates merged. So why were their fates different? Well, Apple version one was a failure in many ways, and we tend to forget that the Macintosh itself was not a commercial success. And so Apple version one, not a success, and NeXT Computer, not a success at first. The reason is, in part, because Steve Jobs was not ready.

Today we mythologize him as the great visionary of our era, the king of entrepreneurs. But he was immature, egotistical, and brash. He made a lot of decisions at both Apple and NeXT that damaged the company and the careers of people who were working for him.

Yes, Steve Jobs was always talented. He was a visionary, but he had to learn the craft of discipline. He had to learn to apply his talents to the limits of the markets, what people were willing to buy.

But to go back to your question, why did NeXT fail? Steve Jobs was making bad decisions, and the hardware division shut down. They laid off more than half of the company’s employees. Why did Apple succeed? In the end, it’s because Steve Jobs was more circumspect and mature by the time he returned. He was able to bring these people together and build the Apple Renaissance, starting with the iPod, or maybe the iMac if you want to go back further. With those products in the late ’90s and early 2000s, that’s where he made his mark, and that’s why we remember him today.

Jobs’s Vision: Perfect Integration of Hardware and Software

You make the point in the book that software engineers were Jobs’s “favored inner circle.” But Apple is well known for building hardware and software together. How did his priorities affect the company’s trajectory?

Steve Jobs’s vision for Apple was always the perfect integration of hardware and software. From the beginning, he saw the value in controlling the whole experience, end-to-end, for the user. But the question that not many people have asked is: How did he actually build that? A lot of startup founders want to create a seamless ecosystem, but you can only achieve that vision if you have a lot of scale and leverage over the market. That’s what he had to learn during his NeXT Computer years.

His customers for NeXT Computer were universities, laboratories, intelligence agencies, people who were buying high-end computers that could do the most advanced data analysis and software applications of the time. It took him a long time to realize that his customers actually wanted the software, which was built on a real revolution at the time called “object oriented programming.”

Computers in the 1980s were really difficult to program. He discovered at NeXT Computer that he could build beautiful software using what are called “objects”—items that, essentially, are pre-programmed in a library. This is how apps are made today, and Steve Jobs was doing this in 1988. As a result, the first ever app store appeared on a NeXT computer.

Steve Jobs in Exile: The Untold Story of NeXT and the Remaking of an American Visionary is out 19 May. Portfolio/Penguin Random House

This is fairly common usage today, but at the time it was comparable to the transformation happening now, from the app stores of the past era to AI agents and generative AI. He was on the cusp of a transformation, and he had to realize that that transformation was what mattered.

Software was his priority, and it became an even greater priority over time. But there’s a lesson here, specifically for engineers and scientists, that sometimes it’s not clear what the significance is of something you’re working on right now. Often there are elements within a project that are actually the real gold mine, and we just don’t see it yet.

That’s what makes me really interested in the big transformations now happening with AI. I often wonder, if Steve Jobs were still alive, what he would be doing now. The moment that we live in now would really be a Steve Jobs moment.

Jobs also purchased Pixar during this period. How does that compare to his experience with NeXT?

The story of Pixar is different from NeXT. Pixar was struggling. It used to make something called the image computer, and it cost more than [US] $100,000. They were using it in sci-fi movies back in the ’80s, but Hollywood studios weren’t really adopting it. And in the end, some of its main customers were the CIA and the National [Security] Agency.

Similar to NeXT computer, Steve bet wrong on hardware once again at the beginning. And so he had to close the hardware division at Pixar, and he focused on making software called RenderMan. RenderMan was their main product, and it was used to make Toy Story and the other great films of Pixar.

The difference from NeXT, though, is that Pixar did become a success, even though the hardware division closed. [Pixar cofounder] Ed Catmull and [founding employee, later chief creative officer] John Lasseter told Steve, when he bought it from George Lucas, that he had to stay out of creative meetings. That was the deal, and Steve respected it. They were able to make Toy Story, and it was just a [huge] success. It was the IPO that made Steve Jobs a billionaire.

Pixar was the company that he became more focused on in the later years of this wilderness. He focused on the dealmaking, the relationship with Disney—the side of being a business executive. That’s where he excelled, and that’s where he was able to make Pixar shine.

What Does New Apple CEO Ternus Need to Do?

John Ternus will be taking over as Apple CEO in September. What was your reaction to the announcement?

There are people out there who are saying that John Ternus needs to invent the new iPhone. I understand where they’re coming from, but I just don’t agree with the conclusions.

You have to remember that the masterpiece age of Apple was only about seven years. It lasted from 2001 to 2008 when they released their main products. The historical conditions and the leadership team that were able to create those massive successes just doesn’t exist anymore.

Apple has matured into a $4 trillion corporation. Tim Cook was a supply chain expert and built it into the giant it is. And really, what John Ternus has to do now is maintain the success. He’s not here to build the one blowout product. Apple is a company that’s become so successful that their products are integrated into everyday life. So that’s a success that John Ternus is, I think, being appointed to maintain, and it’s not by any means a Steve Jobs story.

What do you think it will mean for Apple to be run by someone who’s coming from the hardware engineering side of things?

Apple has admitted that it’s behind on AI. They signed a big partnership with Google, and now they’re reworking the entire foundation of Siri to be on top of Google AI, which is a big shift in software, because Apple’s always been tightly controlling things in-house.

Geoffrey Cain is a bestselling author of books on technology and business, including Samsung Rising and The Perfect Police State. Beowulf Sheehan

So I actually see their future more like a hardware company. I think they’re still going to make great [software]. It’s just not going to be the cutting-edge anymore. That ground is moving over towards OpenAI and Google and software-oriented companies. Apple’s strength is absolutely hardware, and they’re going to continue to focus on that, because that’s what they do best.

In the next few years, AI is going to very quickly get integrated into everyday life. I think that, especially at Apple, they’re going to release AI, but it’s not going to be in your face. It’s more like the AI is running in the background, and it’s doing what it needs to do. If Apple can master that, I think that’s going to help them a lot. If they can release another iPhone that’s really well built for AI, but you don’t see the AI, people are going to buy Apple.

Anything else you’d like to say about Apple or your book?

The history that I was writing about was the transition from hardware to software that happened in the 1990s, and I think the story is very relevant, in a way. We’re now entering the next big transition, and we’re entering a wilderness ourselves. And so this is the kind of story that we can go back to, and we can see what Steve Jobs and his people were doing at the time to try to understand the times that we live in now.

Batteries are essential to modern life. They store energy collected by solar plants, power our smartphones, and lurk even in vehicles powered by fossil fuels. They’re so useful that we overlook their problems. Some batteries use potentially dangerous metals. Others have a pesky tendency to catch fire and explode when damaged.

Many companies have tried to design a battery that addresses these issues. Flint, a startup based in Singapore, is a new contender with a “paper battery.” I had the chance to pick it up and handle it for myself in January at CES, the annual consumer-electronics trade show. The battery, which in part uses cellulose, is lightweight and flexible. It can be a flat sheet that looks much like a thick piece of paper, or like a standard household AA or AAA battery. It doesn’t catch fire or leak if damaged and, contrary to what the term “paper” might lead you to think, it can be held up to a flame without igniting. It also claims to use fewer environmentally damaging metals than other batteries.

Carlo Charles, founder of Flint, says the company is rooted in the last two points. “When we started, it was a question of how we make batteries safer and more sustainable,” he says. “It wasn’t meant to be a company. It’s more of a passion project that snowballed into where we are today.”

Flint’s batteries started production in January 2026, and the company has partnered with Logitech to pilot its batteries in some devices.

What’s in a paper battery?

Flint isn’t the first company to produce a “paper battery.” IEEE Spectrum has covered several earlier attempts, including an “origami battery” powered by the respiration of microbes and a biofuel-based paper fuel cell from French startup BeFC.

However, Flint has less in common with these prior attempts than you might assume at a glance.

These prior examples were enzymatic fuel cells that convert an organic fuel source into energy. They used paper—or, to be precise, cellulose—as part of the fuel cell’s structure and mechanical design. Though they could be used like a battery, they weren’t batteries in a technical sense because they generated electricity from a fuel.

The raw materials and components seen here come together to form Flint’s battery cell, which sits in the center. Flint Labs

Flint’s technology is a true battery. Zinc is used as the anode and manganese dioxide is used as the cathode. This is a tried-and-true chemistry used by the inexpensive alkaline batteries that power hundreds of common household devices, from TV remotes to digital clocks.

How, then, does Flint’s battery use paper?

“We integrate cellulose across every battery component as much as we can,” says Charles, adding that Flint batteries can use cellulose as part of the casing, the cathode, the electrolyte, and the separator. “We remove PFAS from our chemistry, we remove cobalt, [and] we remove the solvent,” says Charles.

Charles was inspired by a paper from researchers at Singapore’s Nanyang Technological University that explored a “hydrogel reinforced cellulose paper” battery design. Charles connected with one of the researchers, Hong Jin Fan, who provided advice in Flint’s early days.

The use of cellulose in the battery is key to the company’s claim that it’s more environmentally friendly. Cellulose replaces metals and chemicals that are either dangerous to humans, produced through environmentally damaging processes, or both. Cobalt, for example, can cause cardiomyopathy, and cobalt mining can be extremely damaging to the local environment.

Flint batteries also lack corrosive materials found in alkaline batteries and the combustible materials in lithium-ion batteries. This improves their safety and physical flexibility. While not all Flint batteries are designed to be flexible, the battery can be produced in a thin format that I could bend between my thumb and index finger.

Flint claims high battery voltages and up to 1,000 charge cycles

The chemistry in Flint’s battery is similar to that of alkaline batteries, but the company hopes to rival lithium-ion in some applications, as well.

That’s a significant challenge. Lithium-ion is useful in part because it can deliver much higher voltages than an alkaline battery. An alkaline battery provides up to 1.5 volts, while lithium-ion can manage up to 4.2 volts.

Charles says Flint’s batteries can reach that mark. “Our batteries deliver high power, meaning up to 4.2 volts, and an energy density of 226 watt-hours per kilogram,” he says. The battery is rechargeable and can endure up to 1,000 charge cycles, also roughly comparable to lithium-ion, which can last 300 to 2,000 or more cycles, depending on the chemistry.

Flint’s manufacturing setup, which will process battery material at scale, uses a roll-to-roll process similar to that of a printing press. Flint Labs

Flint has not published a white paper on its design and is currently keeping the details of its battery design under wraps. A company spokesperson said Flint’s technology includes “proprietary tweaks across the cathode, electrolyte, anode, and separator.”

Though Flint makes big claims about its battery, Charles acknowledged that lithium-ion will be tough to rival and said the company isn’t ready to challenge lithium-ion in applications that push the limits on power density.

“We do not want to focus on electric vehicles, because we know they require even more life cycles and even higher power densities,” he says. Instead, Flint has its sights on lower-power devices commonly found in households, such as sensors, remotes, portable health care devices, and even smartphones. “I want to replace everything [with a battery] I can see around me in my house or workplace with a Flint battery.”

2026 is a crucial test for that vision.

The company is speaking with potential customers, such as Logitech, which sells peripherals for PCs and smartphones. Flint won Logitech’s 2025 Future Positive Challenge and has partnered with Logitech to pilot use of Flint AAA batteries.

If successful, this partnership would help Flint demonstrate that its technology can deliver on its goal to reduce the impact batteries have on the environment. “[Logitech ships] many millions of batteries each year in their products. Even reducing a few grams of carbon emissions per product can be a big impact for them,” says Charles.

Two weeks ago, Anthropic announced that its new model, Claude Mythos Preview, can autonomously find and weaponize software vulnerabilities, turning them into working exploits without expert guidance. These were vulnerabilities in key software like operating systems and internet infrastructure that thousands of software developers working on those systems failed to find. This capability will have major security implications, compromising the devices and services we use every day. As a result, Anthropic is not releasing the model to the general public, but instead to a limited number of companies.

The news rocked the internet security community. There were few details in Anthropic’s announcement, angering many observers. Some speculate that Anthropic doesn’t have the GPUs to run the thing, and that cybersecurity was the excuse to limit its release. Others argue Anthropic is holding to its AI safety mission. There’s hype and counterhype, reality and marketing. It’s a lot to sort out, even if you’re an expert.

We see Mythos as a real but incremental step, one in a long line of incremental steps. But even incremental steps can be important when we look at the big picture.

How AI Is Changing Cybersecurity

We’ve written about shifting baseline syndrome, a phenomenon that leads people—the public and experts alike—to discount massive long-term changes that are hidden in incremental steps. It has happened with online privacy, and it’s happening with AI. Even if the vulnerabilities found by Mythos could have been found using AI models from last month or last year, they couldn’t have been found by AI models from five years ago.

The Mythos announcement reminds us that AI has come a long way in just a few years: The baseline really has shifted. Finding vulnerabilities in source code is the type of task that today’s large language models excel at. Regardless of whether it happened last year or will happen next year, it’s been clear for a while this kind of capability was coming soon. The question is how we adapt to it.

We don’t believe that an AI that can hack autonomously will create permanent asymmetry between offense and defense; it’s likely to be more nuanced than that. Some vulnerabilities can be found, verified, and patched automatically. Some vulnerabilities will be hard to find but easy to verify and patch—consider generic cloud-hosted web applications built on standard software stacks, where updates can be deployed quickly. Still others will be easy to find (even without powerful AI) and relatively easy to verify, but harder or impossible to patch, such as IoT appliances and industrial equipment that are rarely updated or can’t be easily modified.

Then there are systems whose vulnerabilities will be easy to find in code but difficult to verify in practice. For example, complex distributed systems and cloud platforms can be composed of thousands of interacting services running in parallel, making it difficult to distinguish real vulnerabilities from false positives and to reliably reproduce them.

So we must separate the patchable from the unpatchable, and the easy to verify from the hard to verify. This taxonomy also provides us guidance for how to protect such systems in an era of powerful AI vulnerability-finding tools.

Unpatchable or hard to verify systems should be protected by wrapping them in more restrictive, tightly controlled layers. You want your fridge or thermostat or industrial control system behind a restrictive and constantly updated firewall, not freely talking to the internet.

Distributed systems that are fundamentally interconnected should be traceable and should follow the principle of least privilege, where each component has only the access it needs. These are bog-standard security ideas that we might have been tempted to throw out in the era of AI, but they’re still as relevant as ever.

Rethinking Software Security Practices

This also raises the salience of best practices in software engineering. Automated, thorough, and continuous testing was always important. Now we can take this practice a step further and use defensive AI agents to test exploits against a real stack, over and over, until the false positives have been weeded out and the real vulnerabilities and fixes are confirmed. This kind of VulnOps is likely to become a standard part of the development process.

Documentation becomes more valuable, as it can guide an AI agent on a bug-finding mission just as it does developers. And following standard practices and using standard tools and libraries allows AI and engineers alike to recognize patterns more effectively, even in a world of individual and ephemeral instant software—code that can be generated and deployed on demand.

Will this favor offense or defense? The defense eventually, probably, especially in systems that are easy to patch and verify. Fortunately, that includes our phones, web browsers, and major internet services. But today’s cars, electrical transformers, fridges, and lampposts are connected to the internet. Legacy banking and airline systems are networked.

Smartphone cameras and some smart glasses allow users to query AI models and receive answers about what they’re looking at. Soon, that capability could expand to other devices, including earbuds.

Researchers at the University of Washington have developed a pair of earbuds they call VueBuds that integrate a small, low-resolution camera into each earbud. The prototype earbuds have features similar to those of smart glasses, like the Ray-Ban Meta glasses—things like translating signs in foreign languages, acting as an aid for low-vision wearers, or identifying plant species during a hike.

Smart glasses have their drawbacks, including privacy concerns and comfort. The under-the-radar cameras have faced criticism and lawsuits over concerns they can record unsuspecting bystanders and what ultimately happens to sensitive visual data.

And not everyone likes wearing glasses—some even opt for contact lenses to avoid having to wear them, including Shyam Gollakota, the University of Washington professor who led the VueBuds research. “The one predominant wearable which almost everyone wears is your earbuds,” he says. His team presents earbuds as an alternative to smart glasses that’s less intrusive and better for privacy.

The primary goal of the research, however, was to demonstrate that this small, ear-worn form factor is even possible. “Traditionally, earbuds have been limited to audio interfaces,” Gollakota says. “We show that we can indeed build a system within that form factor and get lots of intelligence by running visual language models.”

The research was presented today at the ACM Computer-Human Interaction conference in Barcelona.

Why Earbuds Are an Ideal Smart Device

Gollakota and his colleagues don’t expect VueBuds to be the only interface for visual AI.

“Wearables are very personal,” says Maruchi Kim, a Ph.D. student in Gollakota’s lab. Some people may prefer glasses or watches, others might like rings, and so Kim suspects there won’t be one device to rule them all. “We’re just trying to introduce another category to demonstrate that everything smart glasses do can be achieved on [earbuds].”

That said, the interface may have some advantages. Because they’re already widely used, people may be more likely to adopt the technology. Plus, Kim says, “there’s already a social paradigm for putting your earbuds away in their case.” Smart glasses may have prescription lenses, so the wearer would keep them on all the time. But “if you ever want to be confident that these cameras aren’t recording, earbuds are a nice form factor that lets you just tuck it away when you’re ready.”

Many of the AI features users indicate an interest in are also “episodic use cases,” Kim says. To translate a street sign or ingredients on a package, for instance, you don’t need a continuous video stream.

Key Challenges for Earbuds With Cameras

There are three key challenges to making vision-capable earbuds possible, Gollakota says: Fitting the camera within strict size, power, and weight constraints; transmitting the data; and creating a complete visual scene when worn in the ears.

Cameras typically take a lot of power, making this the number-one concern. “The batteries in your earbuds are about 10 times as small as what you have on smart glasses,” Kim says. Visual data also requires much higher bandwidth than audio, so the videos recorded by glasses are typically sent via Wi-Fi to be processed by cloud-based AI models. Wi-Fi allows for high bandwidth—but takes more power.

VueBuds transmit low-resolution, grayscale images over Bluetooth. Most device makers try to transmit as much data as possible, but Gollakota’s team took a different approach. They wanted to see what the lowest resolution a visual language model would need to extract useful information, opting for a 324-by-324-pixel image sensor.

Integrated black-and-white cameras and a Bluetooth connection to a phone-based visual AI model make these earbuds an alternative to smart glasses.

Beyond the power and bandwidth concerns, the researchers also had to make sure earbud cameras could see enough. Placing cameras at the ears creates a blind spot on either side where the face blocks each camera’s view. But by setting the cameras at a slight angle (5 or 10 degrees) away from the face and stitching together images, the team found they could reconstruct a more complete scene with a wide field of view. This does, however, create a small blind spot for objects closer than about 20 centimeters from the face directly in front of the user.

The researchers tested the earbuds with four different visual language models. In user studies with the best-performing model (Qwen2.5-VL), VueBuds achieved about 82 percent accuracy for object recognition, 94 percent for character recognition, 84 percent for translation, and 87 percent overall accuracy in user studies. The earbuds performed comparably to Ray-Ban Meta glasses across 17 tasks.

In the future, the team hopes to add color to the system. Kim is also looking into improving the resolution possible by incorporating an on-device JPEG encoder, which would significantly reduce the size of images sent to be processed.

Privacy Concerns for Smart Earbuds

Many users have been wary of privacy and surveillance concerns with smart glasses. Those worries are intensifying with new evidence that the companies building these glasses may be mishandling the data they capture.

Given those concerns, should we add cameras to yet another wearable device? The University of Washington researchers say VueBuds’ stripped-down image capture is a boon for privacy compared to today’s smart glasses.

For one thing, the system is designed to run on a smartphone or other local device, so data never goes to the cloud, Gollakota says. VueBuds also only captures still images. One of the main uses of Meta’s smart glasses is now recording video, but, he adds, “no one wants to see a low-resolution grayscale video in the first place.”

Additionally, VueBuds are activated by voice commands. “That audio initiation means that everyone around you would know what you’re actually asking.” Smart glasses, meanwhile, can start recording with the touch of a button.

Gollakota notes that most people have become accustomed to having microphones in nearly every device, because they provide enough utility through capabilities like voice commands and “a trust has been built” with companies, like Apple, that sell devices with built-in microphones. Whether the same paradigm will emerge for visual intelligence remains to be seen with how the technology—and our level of trust in it—evolves over the next few years.

Apple is also rumored to be developing next-generation AirPods that integrate infrared cameras to enable gesture recognition and improve spatial audio. These wouldn’t have the visual intelligence capabilities made possible with standard cameras, but it would indicate more interest in expanding the capabilities of what has traditionally been an audio-only interface.

Earbuds are “the most successful wearable we have today, and right now it’s limited to being an audio interface,” Gollakota says. “Bringing visual intelligence would make it a much richer and more powerful interface than what it currently is.”

By many estimates, quantum computers will need millions of qubits to realize their potential applications in cybersecurity, drug development, and other industries. The problem is, anyone who has wanted to simultaneously control millions of a certain kind of qubit has run into the problem of trying to control millions of laser beams.

That’s exactly the challenge that was faced by scientists working on the MITRE Quantum Moonshot project, which brought together scientists from MITRE, MIT, the University of Colorado at Boulder, and Sandia National Laboratories. The solution they developed came in the form of an image projection technology that they realized could also be the fix for a host of other challenges in augmented reality, biomedical imaging, and elsewhere. The device is a 1-square-millimeter photonic chip capable of projecting the Mona Lisa onto an area smaller than the size of two human egg cells.

“When we started, we certainly never would have anticipated that we would be making a technology that might revolutionize imaging,” says Matt Eichenfield, one of the leaders of the Quantum Moonshot project, a collaborative research effort focused on developing a scalable, diamond-based quantum computer, and a professor of quantum engineering at the University of Colorado at Boulder. Each second, their chip is capable of projecting 68.6 million individual spots of light—called scannable pixels—to differentiate them from physical pixels. That’s more than 50 times the capability of previous technology, such as micro-electromechanical systems (MEMS) micromirror arrays.

“We have now made a scannable pixel that is at the absolute limit of what diffraction allows,” says Henry Wen, a visiting researcher at MIT and a photonics engineer at QuEra Computing.

The chip’s distinguishing feature is an array of tiny microscale cantilevers, which curve away from the plane of the chip in response to voltage and act as miniature “ski jumps” for light. Light is channeled along the length of each cantilever via a waveguide and exits at its tip. The cantilevers contain a thin layer of aluminum nitride, a piezoelectric that expands or contracts under voltage, thus moving the micromachine up and down and enabling the array to scan beams of light over a two-dimensional area.

Despite the magnitude of the team’s achievement, Eichenfield says that the process of engineering the cantilevers was “pretty smooth.” Each cantilever is composed of a stack of several submicrometer layers of material and curls approximately 90 degrees out of the plane at rest. To achieve such a high curvature, the team took advantage of differences in the contraction and expansion of individual layers caused by physical stresses in the material resulting from the fabrication process. The materials are first deposited flat onto the chip. Then, a layer in the chip below the cantilever is removed, allowing the material stresses to take effect, releasing the cantilever from the chip and allowing it to curl out. The top layer of each cantilever also features a series of silicon dioxide bars running perpendicular to the waveguide, which keep the cantilever from curling along its width while also improving its lengthwise curvature.

A micro-cantilever wiggles and waggles to project light in the right place.Matt Saha, Y. Henry Wen, et al.

What was more of a challenge than engineering the chip itself was figuring out the details of actually making the chip project images and videos. Working out the process of synchronizing and timing the cantilevers’ motion and light beams to generate the right colors at the right time was a substantial effort, according to Andy Greenspon, a researcher at MITRE who also worked on the project. Now, the team has successfully projected a variety of videos from a single cantilever, including clips from the movie A Charlie Brown Christmas.

The chip projected a roughly 125-micrometer image of the Mona Lisa.Matt Saha, Y. Henry Wen, et al.

Because the chip can project so many more spots in any given time interval than any previous beam scanners, it could also be used to control many more qubits in quantum computers. The Quantum Moonshot program’s mission is to build a quantum computer that can be scaled to millions of qubits. So clearly, it needs a scalable way of controlling each one, explains Wen. Instead of using one laser per qubit, the team realized that not every qubit needed to be controlled at every given moment. The chip’s ability to move light beams over a two-dimensional area would allow them to control all of the qubits with many fewer lasers.

Another process that Wen thinks the chip could improve is scanning objects for 3D printing. Today, that typically involves using a single laser to scan over the entire surface of an object. The new chip, however, could potentially employ thousands of laser beams. “I think now you can take a process that would have taken hours and maybe bring it down to minutes,” says Wen.

Wen is also excited to explore the potential of different cantilever shapes. By changing the orientations of the bars perpendicular to the waveguide, the team has been able to make the cantilevers curl into helixes. Wen says that such unusual shapes could be useful in making a lab-on-a-chip for cell biology or drug development. “A lot of this stuff is imaging, scanning a laser across something, either to image it or to stimulate some response. And so we could have one of these ski jumps curl not just up, but actually curl back around, and then move around and scan over a sample,” Wen explains. “If you can imagine a structure that will be useful for you, we should try it.”

This month, Apple celebrates its 50th anniversary, having survived five decades of ups and downs. Whether you love or hate the company, its products have had a massive influence on the consumer technology market and the world.

Among the obvious successes, such as the iPhone and the MacBook, there are numerous dead-end failures, such as the butterfly keyboard and the G4 Cube. But its history is also littered with forgotten legacies—some successes, some failures—that still influence computing to this day.

IEEE Spectrum spoke with John Buck, author of Inventing the Future: Bit by Bit, to learn more about some of those forgotten legacies pioneered by Apple’s Advanced Technology Group (ATG) in the 1980s and 1990s.

LaserWriter and desktop publishing

In the mid 1980s, as Buck says, Apple needed a commercial success:

While Steve Jobs had recognized the potential of linking a printer product with Adobe’s popular PostScript even before the release of the Mac, the LaserWriter ended up one of the first post–Steve Jobs products, where the company learned to understand what the market wanted instead of just what Jobs wanted. And even though the first model was expensive, it was cheaper than the alternatives.

Jim Gable, who was the product manager on LaserWriter, had no background in printing. The ATG was full of smart generalists who got dragged in on these projects.

Apple discontinued the LaserWriter line when Steve Jobs returned to Apple, but it set the groundwork for desktop publishing. To cope with the processing demands of running Adobe’s PostScript for font rendering, the LaserWriter had its own Motorola 68000-series processor.

While the PostScript partnership was initially lucrative, licensing fees eventually became a burden for Apple. From Buck:

With the advent of more affordable printing came the need for legible fonts at different sizes, from small screens to giant billboards. Initially, for the LaserWriter, Apple licensed PostScript, which Adobe created in 1982.

Even though Adobe helped Apple create the desktop publishing revolution and bring them much-needed business, they thought, we’re paying them too much money for fonts, let’s build our own thing, which was a very ATG thing to do. They brought in all the experts and said, “Right, we’re just going to do it better.”

And they built it, then went behind everyone’s back and spoke to Bill Gates and said, “Right, you don’t like them [Adobe] either. It’s costing you money.”

The result was the TrueType fonts in 1989, which are now widely used on macOS and Windows, while high-end printers still use PostScript. Its last release was in 1997, and Adobe’s alternative Type 1 fonts are largely forgotten.

Apple’s Newton personal digital assistants (PDAs) ultimately failed but led to a successful partnership with Arm that culminated in the iPhone decades later.Geoff Parsons

Newton, Arm chips, and the iPhone

While Acorn computers won’t mean much to people outside the United Kingdom, the company spun off part of their business to create the much more famous Arm, thanks to a joint venture with VLSI Technology and Apple in 1990.

Buck explains how this relationship began:

When Apple built the first Newton, they commissioned AT&T to create these Hobbit chips, and they were too slow, too hot, and couldn’t deliver the power. And they discovered Arm. A little two-man team with the project name ‘Möbius’ went to the UK to try their chips and found them ideal. Even though Newton was a disaster because of too many egos who wanted to add this and that, the fact that they’ve found Arm to put in the Newton becomes the success story of the iPhone. It’s a research arc that takes forever to come to market. And some of the mistakes that they stumble over to get there.

QuickTime and digital video formats

One name from the ATG era you’ll still find on a modern Mac is QuickTime. While it’s not the essential powerhouse it once was, in Buck’s opinion, it’s one of Apple’s biggest legacies:

To me, the legacy of QuickTime and the team behind it are one of the most enduring of the ATG. At the time of its release, playing video on a computer was incredible, and in the 90s and early 00s, QuickTime defined many video format standards, most notably MPEG-4.

QuickTime team member and ATG compression algorithm expert Gavin Miller, who’s now Head of Research at Adobe, said to me: “I think the QuickTime team invented the modern media-rich world, and the larger ATG imagined and prototyped experiences that parallel modern movies, video games, and interactive map-based technologies. It was something of a golden age when great ideas were often in the future, perfect if only computers could be fast enough. Idealism, matched with ingenious engineering, made those things happen 10 years earlier than they might have otherwise.”

Others, such as former Apple product manager Andrew Soderberg, thought their importance had been understated and told me: “Other than Steve Jobs and Steve Wozniak, these were the most important engineers in Apple’s history until the advent of the iPhone.”

Operating systems and browsers started to build in many of the features that QuickTime helped define. QuickTime is still present on macOS (Apple discontinued the Windows version in 2016) but serves mostly as a playback app, with most of the codec work now handled by other software.

Early Macintosh models did not sell as well as people might think. According to Buck, Hypermedia helped Apple with lagging Macintosh sales:

People had been experimenting with the ideas that form what became “the World Wide Web” long before its creation at CERN. And HyperCard was one of the more successful experiments, causing a rush of enthusiasm and sales for the Macintosh.

Apple tied HyperCard’s launch to Apple’s first CD-ROM in 1986, a digital version of the Whole Earth Catalog (WEC). It was launched as a stack of 4,000 linked HyperCard cards, doubling the sales of the Macintosh 1, causing Apple production challenges in meeting demand.

This, in turn, led to a revolutionary overhaul of production techniques, with engineer Steve Young tasked with introducing robots and automation to factories in the USA, Singapore, and Ireland. All two years after he graduated from Carnegie Mellon University.

HyperCard received updates until 1998, and Apple finally withdrew it from sale in 2004. The original pointed-finger cursor for early web links was inspired by HyperCard, as was the Wiki concept. Apple continues to be a leader in production and logistics techniques.

Apple’s first foray into portable computing, the Macintosh Portable, had display and battery issues.Jim Abeles

Macintosh Portable and PowerBook laptops

Apple’s first foray into personal computing was a disaster, but, as Buck recounts, they quickly redefined the form factor:

Portable devices are now how most people interact with computing. Apple was late to creating a portable version of the Mac despite the increasing demand, even allowing others to fill the gap while they worked on their model.

Dynamic Computers began advertising a “portable” Mac that was actually a disassembled Mac Plus integrated into a new enclosure. Apple was so desperate for sales, it sold the Mac Pluses at a discount to Dynamic.

In the end, weighing 16 pounds, the resulting device was fraught with display and battery issues. Much of this was due to a few people on the “Harpo” development team having no experience with portable computers, as Jon Krakower recalled to me: “The mechanical design engineers assigned to our team had no experience with designing small, lightweight portable products. Apple itself had never shipped a portable product and had no experience with the special size and weight requirements. They brought in people who only had experience designing desktop computers, CRT displays, or printers.”

Apple discontinued the Portable less than two years later. However, quickly afterwards, they began a redesign, as industrial designer Gavin Ivester told me: “I was called to a meeting with John Sculley, where he told us that by next year, the personal computer market was projected to be 30 percent laptops. There was no Apple laptop (or notebook) yet, and creating one would easily take more than a year with normal resources and processes. Sculley calculated how many millions of dollars per day 30 percent of Apple’s business was worth—it was a big number.”

Taking inspiration from the Compaq LTE, the team moved surprisingly quickly, attempting to use as many existing parts as possible. The resulting PowerBook was a huge success over its 15-year lifespan, setting the template for what laptops would look like even now. Ironically, the team that created it ended up leaving to work for Compaq.

When Jobs returned in 1997, he shut ATG down to save on costs and product distractions. The only ATG project he kept around was QuickTime, ironically, because the team had just developed a version for Windows, but the people themselves went on to many other places, and Apple entered a new chapter.

Most wearable devices can’t run AI models locally. They’re too power-hungry to run on the device itself, so data is typically sent to the cloud for processing. But new processors may change that.

Ambient Scientific is among the companies betting on analog computing techniques to make battery-powered, on-device AI more feasible. The Santa Clara, Calif.–based company combines digital and analog elements in a chip built for deep learning. Now, it is partnering with India-based tech firm Dimension NXG to make a wearable device using its technology.

On 10 March, the two companies introduced Mai, a wrist-worn, screenless device built for women’s health and safety. The wristband is powered by Ambient Scientific’s GPX10 Pro, which allows it to continuously run AI algorithms for up to two weeks on a single charge. Those algorithms are trained to detect when the wearer is in danger by detecting falls or stress cues.

Dimension NXG has also begun a field trial phase for Mai, shipping them to thousands of users across India. The company plans to have more than 10,000 units available by the end of the year.

Analog Computing in AI Wearables

For AI, “95 percent of the mathematics that runs in these algorithms is just one mathematics. It’s called multiply-accumulate [or] MAC,” says GP Singh, the CEO of Ambient Scientific. These operations are foundational to the matrix multiplication that makes deep learning possible. In general, analog computing is slower and less accurate than digital computing for most calculations—except MAC operation. Analog methods process data in parallel, so they’re more efficient in performing certain complex tasks, like matrix multiplication, compared to digital’s binary logic.

That means analog computing should be a great fit for AI. The problem? At scale, it’s difficult to reliably produce these analog systems, because even microscopic variations during fabrication can significantly alter a circuit’s performance. So, to get the best of the analog and digital worlds, Ambient Scientific converted the most sensitive functions to digital signals.

GPX10 has three main advantages, according to Singh. First, “it is extremely programmable and flexible,” so it can be used for many different applications. Second, the power consumption is very low; even with 10 AI cores, the chip can run on a battery-operated device. And third, it can take input from up to 20 digital sensors simultaneously, making it ideal for wearables or IoT devices.

This version of the device has a limited amount of memory, Singh says, but it’s sufficient for the applications Ambient Scientific is now targeting, including wearables like Mai.

Mai’s Privacy-Focused Safety Features

During sleep, the human brain enters a subconscious mode, but it can still easily wake up if needed. “Sleep mode” in most of today’s devices shuts down systems more completely. “They are not really subconscious,” Singh says. By comparison, Mai continuously consumes microwatts of power to stay alert.

The device continuously monitors for three signals: fall detection, assault recognition, and safeword recognition. Mai is connected to a 5G network, so when one of these signals is detected, an alert is immediately sent to emergency services or a family member with a GPS location.

“Having the safeword recognition always on is extremely important so that the user doesn’t have to push a button on the device or raise their hand, because they might be in situations where they’re not able to do that physically,” says Saharsh Singhania, the head of product marketing at Ambient Scientific.

Some users may be concerned about their privacy while wearing a device that is always listening. But because Mai’s intelligence is on the device itself, most data never leaves the device. It only sends alerts, not voice recordings, when troubling signals are detected, and the rest of the data is thrown out within milliseconds. This means sensitive data isn’t stored by the company, and it significantly reduces the opportunity for attacks from third parties, which often take place when data is sent to and from the device via Wi-Fi or Bluetooth, Singh says.

Reducing False Positives in Wearables

False positives have been one of the main challenges for these types of devices. Too many of them, and the wearable could become the wristband that cried wolf.

Singh says Mai’s algorithms are sophisticated enough to differentiate between, for instance, someone falling and forcefully sitting down. It detects false positives less than once every three days, which Singh says is less frequent than other devices.

Mai is also designed for women’s health. Like other wearables, it can collect data about heart rate and blood oxygen. Dimension NXG is now working with a prominent medical research facility on clinical studies, seeking early indicators of polycystic ovarian syndrome, a hormonal disorder that affects 10 to 13 percent of reproductive-aged women.

The device is first launching in India, but Dimension NXG plans to expand to other Southeast Asian countries. Ambient Scientific, meanwhile, has more capable AI chips on its roadmap, including a 64-core version for use in applications and robotics and drones, as well as a processor for use in data centers.

The rise of generative AI has spurred demand for AI workstations that can run or train models on local hardware. Yet modern PCs have proven inadequate for this task. A typical laptop has only enough memory to load a large language model (LLM) with 8 billion to 13 billion parameters—much smaller, and much less intelligent, than frontier models that are presumed to have over a trillion parameters. Even the most capable workstation PCs struggle to serve LLMs with more than 70 billion parameters.

Tenstorrent’s QuietBox 2 is an attempt to fill that gap. Though it looks like a PC workstation, the QuietBox 2 contains four of the company’s custom Blackhole AI accelerators, 128 gigabytes of GDDR6 memory—specialized memory used in GPUs—and 256 GB of DDR5 system memory (for a total of 384 GB). This configuration provides enough memory to load OpenAI’s GPT-OSS-120B and can run midsize models like Meta’s Llama 3.1 70B at speeds of nearly 500 tokens per second. For reference, that’s several times quicker than an average response from OpenAI’s GPT-5.2 or Anthropic’s Claude 4.6. The QuietBox 2 carries an expected retail price of US $9,999 and is slated to launch in the second quarter of 2026.

“The 128 gigabytes of GDDR that we have with our AI accelerators really defines how big of a model you can run at a reasonable speed,” says Milos Trajkovic, cofounder and systems engineer at Tenstorrent. “Our 128 gigabytes of GDDR6 RAM would require four Nvidia RTX 5090 graphics cards. That couldn’t fit in today’s 1,600-watt form factor, and the cost for four RTX 5090 GPUs is huge.”

An AI workstation built at the home office

Wattage, it turns out, is critical. Nvidia recommends a system power of 1,000 W for a single RTX 5090, so even a dual-GPU setup exceeds the continuous power draw for a typical 15-ampere, 120-volt power circuit. A system with four RTX 5090s could require 4,000 W or more at load.

The QuietBox 2, on the other hand, draws only 1,400 W at full load. It won’t trip the breaker, so it can be used anywhere a typical desktop PC might be plugged in, including a home office.

That’s not the only way the QuietBox 2 poses as a run-of-the-mill PC. The machine’s custom case is built to support the micro-ATX motherboard form factor, and the motherboard itself is an AMD chipset hosting an AMD CPU. The hardware is kept cool by closed-loop liquid cooling similar to that used by PC workstations and gaming computers. It even has customizable RGB LED lighting and a large semitransparent window that shows off the hardware.

“A lot of even our internal developers have requested a QuietBox because they’re just so easy to deploy,” says Chris Goulet, a thermal-mechanical engineer and team lead at Tenstorrent. “You just ship them the unit, they slap it on their desk, power it up, and they’re going.”

Where the QuietBox 2 differs from desktop PCs, though, is its AI accelerators. It’s equipped with four of Tenstorrent’s Blackhole application-specific ICs, a RISC-V chip designed specifically for AI workloads. Blackhole is packaged on an add-in card; each card has 120 Tensix AI accelerators and 32 GB of GDDR6 memory, for a total of 480 Tensix AI accelerators and 128 GB of GDDR6. Blackhole also has a large amount of on-chip SRAM at 180 megabytes per accelerator.

Two visions of desktop AI

Tenstorrent is not alone in its approach. Nvidia’s DGX Spark, released last year, packed Nvidia’s GB10 chip into a machine the size of a lunch box. Orders for the Spark’s big brother, the DGX Station with Nvidia’s GB300, were opened on 16 March 2026.

The DGX Station looks like a desktop PC workstation, and variants will be built by well-known PC brands like Asus and Dell. Nvidia’s offering has more memory than QuietBox 2, at up to 748 GB, but system power is quoted at 1,600 W—rather close to the maximum a 15-A, 120-V breaker will handle. This reflects differing visions for how their machines will be used. And, of course, the Nvidia DGX Station’s extra memory doesn’t come cheap. While most DGX Station system builders have not yet announced pricing, one retailer has listed a DGX Station from PC maker MSI for $85,000.

When I spoke to Allyn Bourgoyne, director of product marketing at Nvidia, after the announcement of DGX Spark and Station in 2025, he said the company expects most DGX owners will use the devices as remotely accessed workstations. “A common thing you might see is that I’ve got my Windows laptop, and I’m going to use my DGX Spark over the network. I’m going to send jobs over to it.” He added that companies could deploy DGX Spark and Station systems to serve multiple people at once.

The Tenstorrent QuietBox 2 can be used in this way, but the company also wants to target a good experience for people going one-on-one with the computer. “You don’t have to remote SSH into the box. You connect your monitor through HDMI, and it’s just like your PC at home. It has the Ubuntu desktop and utilities,” says Trajkovic.

Nvidia’s DGX systems also run a variant of Ubuntu (DGX OS) and include a desktop environment, but the devil is in the details. DGX systems use Nvidia CPUs based on ARM architectures and custom chipsets. The QuietBox 2 uses an AMD x86 CPU and compatible chipset, and is configured more like a traditional PC. That should be a boon for the QuietBox 2’s software compatibility.

Tenstorrent leans into that with a focus on open source software. The QuietBox 2’s entire software stack, from TT-Forge (the company’s AI compiler) to TT-Metalium (a low-level software development kit that provides kernel-level hardware control), is open source and available on GitHub. Tenstorrent has also published the instruction set architecture for its Tensix cores, so developers can see exactly how their workloads execute on the hardware. Nvidia, by contrast, is focused on its proprietary CUDA ecosystem, and DGX OS is not open source.

“A lot of our software stack is completely open, and we felt that from a hardware perspective, we kind of wanted to take a similar path,” says Goulet.

Every time you unlock your smartphone or start your connected car, you are generating a trail of digital evidence that can be used to track your every move.

In Your Data Will Be Used Against You: Policing in the Age of Self-Surveillance, just published by NYU Press, law professor Andrew Guthrie Ferguson exposes how the Internet of Things has quietly transformed into a vast surveillance network, turning our most personal devices into digital informants. The following excerpt explores the concept of “sensorveillance,” detailing the specific mechanisms—such as Google’s Sensorvault, geofence warrants, and vehicle telemetry—that allow law enforcement to repurpose consumer technology into powerful tools for investigation and control.

A man walked into a bank in Midlothian, Va., his black bucket hat pulled low over dark sunglasses. He handed a note to the teller, brandished a gun, and walked away with US $195,000. Police had no leads—but they knew that the robber had been holding a smartphone when he entered the bank. Guessing that the smartphone, like most smartphones, had some Google-enabled service running, police ordered Google to turn over information about all the phones near the bank during the holdup. In response to a series of warrants, Google produced information about 19 phones that had been active near the bank at the time of the robbery. Further investigation directed the police to Okelle Chatrie, who was ultimately charged with the crime.

Cathy Bernstein had a tough time explaining why her own car reported an accident to police. Bernstein had been driving a Ford equipped with 911 Assist, which was automatically enabled when she struck another vehicle. Rather than stick around to trade insurance information, she sped away. But her smart car had registered the bump—and called the police dispatcher, leading to a fairly awkward conversation:

Computer-Generated Voice: Attention, a crash has occurred. Line open.

911 Operator: Hello. Can anyone hear me?

Unidentified Woman: Yes, yes.

911 Operator: Okay. This is 911. You’ve been involved in an accident.

Unidentified Woman: No.

911 Operator: Well, your car called in to us because it said you’d been involved in an accident. Are you sure everything’s okay?

Unidentified Woman: Everything’s okay.

911 Operator: Okay. Are you broke down?

Unidentified Woman: No, I’m fine. The guy that hit me—he did not turn.

911 Operator: Okay, so you have been involved in an accident.

Unidentified Woman: No, I haven’t.

911 Operator: Did you hit a car?

Unidentified Woman: No, I didn’t.

911 Operator: Did you leave the scene of an accident?

Unidentified Woman: No. I would never do anything like that.

Apparently, Bernstein did do something “like that.” She was soon caught and cited for leaving the scene of the accident. Her own car provided evidence of her guilt.

The Rise of “Sensorveillance”

Once upon a time, our things were just things. A bike was a tool for biking. It got you from one location to another, but it didn’t “know” more about your travels than any other inanimate object did. It was dumb in a comforting way, and we used it as intended. Today, a top-of-the-line bike can track your route and calculate your average speed along the way. Hop on an e-bike from a commercial bike share, and it will collect data for your trip, plus the trips of everyone else who used it that month.

These “smart” objects belong to what technologist Kevin Ashton named the Internet of Things. Ashton proposed adding radio-frequency identification (RFID) tags and sensors to everyday objects, allowing them to collect data that could be fed into networked systems without human intervention. A sensor in a river could monitor the cleanliness of the water. A tag on a bottle of shampoo could trace its journey throughout the supply chain. Add enough sensors to enough objects and you can model the health of an entire ecosystem—or learn whether you’re sending too much of your inventory to Massachusetts and too little to Texas.

Ashton first theorized the Internet of Things (IoT) in the late 1990s. Today, the IoT goes well beyond his initial vision, including not only RFID tags but also sensors with Wi-Fi, Bluetooth, cellular, and GPS connections. These small, low-cost sensors record data about movement, heat, pressure, or location and can engage in two-way communication.

Of course, such a system is also, by necessity, a system of surveillance. “Sensorveillance”—a term I created to highlight the intersection of sensors and surveillance—is slowly becoming the default across the developed world.

Cellphone Surveillance Networks

Let’s start with phones. You’re probably not surprised that your cellphone company tracks your location; that’s how cellphones work. Both smartphones and “dumb” mobile phones use local cell towers, owned by cellphone companies, to connect you to your friends and family, which means those companies know which towers you are near at all times.

If you always carry your phone with you, your phone’s whereabouts—recorded as cell-site location information (CSLI)—reveal yours. One man, Timothy Carpenter, found this out the hard way after he and a group of associates set out to rob a series of electronics stores. Carpenter was the alleged ringleader, but he didn’t enter the stores himself. He served as the lookout, waiting in the car while his associates stuffed merchandise into bags.

It might have been hard for investigators to tie him to the crimes—if not for the fact that every minute he kept watch, his cellphone was pinging a local tower, logging his location. Using that information, the FBI was able to determine that he had been near each store during the exact moment of each robbery.

Cell signals are the tip of the proverbial data iceberg. If you have a smartphone, you’re almost certainly using something created by Google. Google makes money off advertising. The more Google knows about users, the better it can target ads to them. Google’s location services are on all Android phones, which use the company’s operating system, but they’re also on Google apps, including Google Maps and Gmail.

For years, all that location information ended up in what the company called the Sensorvault. The Sensorvault, as the name suggests, combined data from GPS, Bluetooth, cell towers, IP addresses, and Wi-Fi signals to create a powerful tracking system that could identify a phone’s location with great precision. As you might imagine, police saw it as a digital evidence miracle. In 2020, Google received more than 11,500 warrants from law enforcement seeking information from the Sensorvault.

“Sensorveillance”—a term I created to highlight the intersection of sensors and surveillance—is slowly becoming the default across the developed world.

In 2024, Google announced that it would no longer retain all of this data in the cloud. Instead, the geolocation information would be stored on individual devices, requiring police to get a warrant for a specific device. The demise of the Sensorvault came about through a change in corporate policy, which could be reversed. But at least for now, Google has made it significantly harder for police to access its data.

And while the Sensorvault was the biggest source of geolocational evidence, it is far from the only one. Even apps that have nothing to do with maps or navigation might nonetheless be collecting your location data. In one Pennsylvania case, prosecutors learned that a burglar used an iPhone flashlight app to search through a home, and they used the data from the app to prove he was in the home at the time of the break-in. These apps might be advertised as “free,” but they come with a hidden cost.

Cars, increasingly, collect almost as much information as phones. Mobile extraction devices can collect digital forensics about a car’s speed, when its airbags deployed, when its brakes were engaged, and where it was when all that happened. If you connect your phone to play Spotify or to read out your texts, then your call logs, contact lists, social media accounts, and entertainment selections can be downloaded directly from your vehicle. Because cars are involved in so many crimes (either as the instrument of the crime or as transportation), searches of this data are becoming more commonplace.

Even without physically extracting information from the car, police have other ways to get the data. After all, the car’s built-in telemetry system is sharing information with third parties. In addition to the usual personal information you give up when buying a car (name, address, phone number, email, Social Security number, driver’s license number), when you own a Stellantis-brand car, the company collects how often you use the car, your speed, and instances of acceleration or braking. Nissan asserts the right to collect information about “sexual activity, health diagnosis data, and genetic [data]” in addition to “preferences, characteristics, psychological trends, predispositions, behavior, attitudes, intelligence, abilities, and aptitudes.” Nissan’s privacy policy specifically reserves the right to provide this information to both data brokers and law enforcement.

The Law of Smart Things

The fact that government agents can glean so much information from our things does not mean that they should be able to do so at any time or for any reason. The U.S. Fourth Amendment—drafted in an era without electricity—protects “persons, houses, papers, and effects” against unreasonable search and seizure, but is naturally silent on the question of location data.

The first question is whether the data from our smart things should be constitutionally protected from police. In the language of the constitutional text, the smart device itself is an “effect”—a movable piece of personal property. But what about the data collected by the effect? Is the location data collected by your smartwatch considered part of the watch, or part of the person wearing the watch? Neither? Both?

To its credit, the U.S. Supreme Court has addressed some of the hard questions around digital tracking. In two cases, the first involving GPS tracking of a car and the second involving the CSLI tracking of Timothy Carpenter’s cellphone, the court has placed limits on the government’s ability to collect location data over the long term.

United States v. Jones involved GPS tracking of a car. Antoine Jones owned a nightclub in Washington, D.C. He also sold cocaine and found himself under criminal investigation for a large-scale drug distribution scheme. To prove Jones’s connection to “the stash house,” police placed a GPS device on his wife’s Jeep Cherokee. This was before GPS came standard in cars, so the device was physically attached to the undercarriage of the vehicle.

Data about Jones’s travels was recorded for 28 days, during which he visited the stash house multiple times. The prosecutors introduced the GPS data at trial, and Jones was found guilty. Jones appealed his conviction, arguing that the warrantless use of a GPS device to track his car violated his Fourth Amendment rights.

“When the Government tracks the location of a cell phone it achieves near perfect surveillance.” —the Supreme Court

In 2012, the Supreme Court held that a warrant was required, based on the reasoning that the physical placement of the GPS device on the Jeep was itself a Fourth Amendment search requiring a warrant. Justice Sonia Sotomayor agreed regarding the physical search but went further, discussing the harms of long-term GPS tracking: “GPS monitoring generates a precise, comprehensive record of a person’s public movements that reflects a wealth of detail about her familial, political, professional, religious, and sexual associations.”

Timothy Carpenter’s ill-fated robbery spree gave the Supreme Court another chance to address the constitutional harms of long-term tracking. In their attempts to connect Carpenter to the six electronics stores that had been robbed, federal investigators requested 127 days of location data from two mobile phone carriers. The problem for the police, however, was that they had obtained the information on Carpenter without a judicial warrant.

Carpenter challenged the FBI’s acquisition of his CSLI, claiming that it violated his reasonable expectation of privacy. In a 5–4 opinion, the Supreme Court determined that the acquisition of long-term CSLI was a Fourth Amendment search, which required a warrant. As the Court stated in its 2018 ruling: “A cell phone faithfully follows its owner beyond public thoroughfares and into private residences, doctor’s offices, political headquarters, and other potentially revealing locales.... [W]hen the Government tracks the location of a cell phone it achieves near perfect surveillance.”

Jones and Carpenter are helpful for setting the boundaries of location-based searches. But, in truth, the cases generate a lot more questions than answers. What about surveillance that is not long-term? At what point does the aggregation of details about a person’s location violate their reasonable expectation of privacy?

The Warrant According to Google

Okelle Chatrie’s case, in which police used Google’s location data to identify him as the mystery bank robber, offers a stark warning about the limits of Fourth Amendment protections under these circumstances. It’s also a terrific example of why “geofence” warrants, which request information within a certain geographic boundary, are appealing to police. From surveillance footage, detectives could see that the suspect had a phone to his ear when he walked into the bank. A geofence could identify who the suspect was, and likely where he came from and where he went. Google held the answer in its virtual vault. A warrant gave investigators the key.

The police cast a broad net. The geofence warrant asked for data on all the cellphones within a 150-meter radius, an area, as the court described it, “about three and a half times the footprint of a New York city block.” After receiving the police’s initial request for information on all the phones in the area, Google returned 19 anonymized numbers. Over the course of a three-step warrant process, the company narrowed those 19 phones down to three and then to one, which it revealed as belonging to Okelle Chatrie.

If the police wish to buy the data, just like an insurer or marketing firm might, how can you object? It’s not your data.

The three-step warrant process is a unique innovation in the digital evidence space. Google’s lawyers developed a procedure whereby detectives seeking targeted geolocation data had to file three separate requests, first requesting identifying numbers in an area, then narrowing the request based on other information, and finally obtaining an order to unmask the anonymous number (or numbers) by providing a name.

To be clear, Google—a private company—required the government to jump through these hoops because Google considered it important to protect its customers’ data. It was the company’s lawyers—not the courts or the government—who demanded these warrants.

Buying Data

Warrants provide at least some procedural barrier to data collection by police. If government agencies want to avoid that minor hassle, they can simply buy the data instead. By contracting with data-location services, several federal agencies have already done so.

The logic for this Fourth Amendment loophole is straightforward: You gave your data to a third-party company, and the company can use it as it wishes. If you own a car that is smart enough to collect driving analytics, you clicked some agreement saying the car company could use the data—study it, analyze it, and, if it wants, sell it. If you don’t want to give them data in the first place, that is okay (although it will likely result in less optimal functionality), but you cannot rightly complain when they use the data you gave them in ways that benefit them. If the police wish to buy the data, just like an insurer or marketing firm might, how can you object? It’s not your data.

Who Is to Blame?

Fears about the amount of personal information that could be revealed with long-term GPS surveillance have become reality. Today, police don’t need to plant a device to track your movements—they can rely on your car or phone to do it for them.

This happened because companies sold convenience and consumers bought it. So it might be tempting to blame ourselves. We’re the ones buying this technology. If we don’t want to be tracked, we can always go back to using paper maps and writing down directions by hand. If few of us are willing to make that trade, that’s on us.

But it’s not that easy. You may still be able to choose a dumb bike over a smart one, but a car that tracks you will soon be the only type of car you can buy. And while cars and data can, in theory, be separated, that’s not true for all our smart things. Without cell-signal tracking capabilities, a cellphone is just a paperweight. And in today’s world, living without a phone or a car is simply not practical for many people.

There are technological steps we can take toward protecting privacy. Companies can localize the data the sensors generate within the devices themselves, rather than in a central location like the Sensorvault. Similarly, the information that allows you to unlock your Apple iPhone via facial recognition stays localized on the phone. These are technological fixes, and positive ones. But even localized data is available to police with a warrant.

This is the puzzle of the digital age. We can’t—or don’t want to—avoid creating data, but that data, once created, becomes available for legal ends. The power to track every person is the perfect tool for authoritarianism. For every wondrous story about catching a criminal, there will be a terrifying story of tracking a political enemy or suppressing dissent. Such immense power can and will be abused.

Designing a medical device? This whitepaper helps you evaluate adhesive options for biocompatibility, sterilization resistance, and manufacturability — so you can make the right material decision early.

What Attendees will Learn

1. How to select between epoxy, silicone, cyanoacrylate, and UV/LED curable adhesives based on your device requirements
2. Which adhesive systems meet USP Class VI and ISO 10993-5 biocompatibility standards
3. How different sterilization methods, such as autoclaving, EtO, gamma, chemical immersion affect adhesive performance over repeated cycles
4. Why integrating adhesive selection early in the design process reduces costly trade-offs between performance and manufacturability
5. Download this free whitepaper now!

Download this free whitepaper now!

This is a sponsored article brought to you by Audio Precision.

Bluetooth started as a simple wireless connection between a phone and a headset. Since its inception, it has become the invisible scaffolding for music, calls, gaming, and hearing assistance across consumer and professional devices alike. Bluetooth’s evolution to support more use cases has been driven not by a single breakthrough but by a steady accumulation of radio innovations, codecs, transport schemes, and power management strategies that together enhance the user experience with wireless audio. Today, a new architectural baseline—Bluetooth Low Energy (LE) Audio—promises low-power, high quality, and scalable audio delivery to open up the standard for an even wider range of applications [1][2].

Evolution of Bluetooth Radio Technologies

The original Basic Rate (BR) radio introduced with Bluetooth 1.0 in 1999 used a Gaussian frequency-shift keying (GFSK) at 1 Msym/s, hopping through 79 channels in the 2.4 GHz band with alternating transmission directions in a tight time-division duplex rhythm. The short-range robustness and reliability afforded by this technology helped gain performance at par with traditional cable-based devices.

In 2003, the Advanced Audio Distribution Profile (A2DP) arrived as the enabling standard for stereo audio streaming over Bluetooth Classic, marking the technology’s expansion beyond voice into music playback. A2DP uses the Audio/Video Distribution Transport Protocol (AVDTP) for stream management and mandates the Sub-Band Codec (SBC) as its baseline audio compression format. The SBC codec employs 4- or 8-band analysis/synthesis filter banks with adaptive bit allocation, spanning bitrates from 128 to 345 kbps for stereo content. Embedded DSP work showed how to optimize SBC implementation—Weighted Overlap Add (WOLA) filter banks, fixed-point pipelines, and real-time decoding that is audibly indistinguishable from floating point reference implementations while consuming fewer MIPS and milliwatts [3].

In 2004, Bluetooth 2.0 introduced Enhanced Data Rate (EDR) that moved payloads to π/4 DQPSK or 8 DPSK modulation to boost gross throughput to 2–3 Mb/s, while retaining the GFSK for packet headers. This innovation boosted stereo streaming quality and adoption during the decade.

Around 2010, Bluetooth Low Energy (BLE) 1 M PHY technology was introduced via Bluetooth 4.0. This new radio technology continued to use GFSK but tuned for low duty cycles and intermittent bursts. This fundamental difference with BR/EDR (Basic Rate/Enhanced Data Rate) led to common usage of the term “Bluetooth Classic” for Bluetooth 1.0 to distinguish it from BLE.

Isochronous Transport Architecture

In late 2016, Bluetooth 5.0 introduced the LE 2M PHY, doubling the symbol rate to 2 Msym/s. For a healthy link margin, halving a packet’s airtime was found to reduce collision exposure and lower the energy delivered/bit. By 2020, Bluetooth 5.2 or Bluetooth LE Audio radically shifted the focus from continuous streaming to a transport designed explicitly around deadlines. LE (Low Energy) Audio leverages the existing LE 1M and LE 2M PHYs but carries audio over isochronous channels—slots with timing commitments. The isochronous channel architecture comes in two forms. Connected Isochronous Streams (CIS) are unicast flows whose parameters (intervals, subevents, retransmissions) can be tuned to meet frame deadlines with bounded jitter, enabling the radio to sleep predictably between bursts while the application knows precisely when a frame will arrive. A systematic review of BLE performance corroborates that output and latency in the real world are bounded as much by connection interval, event length, and retransmissions as by the raw symbol rate; under the right parameters, faster PHYs reduce radioactive time and improve energy efficiency, while coded long-range modes trade airtime for robustness in harsher channels [1].

Broadcast Isochronous Streams (BIS)—commercially branded as Auracast—extend that scheduling to one-to-many transmissions, enabling connectionless audio delivery to unlimited receivers [2][7].

This difference in architecture over continuous streams requires careful selection of intervals, packetization, codec forming and appropriate models to determine parameters that meet deadlines without wasting airtime. Markov chain analyses of CIS—validated via simulation—translate developer choices (intervals, subevents, retransmission counts) into quantitative predictions for packet loss rate (PLR), backlog, delay, throughput, and average power consumption. [7]

The LC3 Codec Advantage

LE Audio’s Low Complexity Communication Codec (LC3) fundamentally shifts the bitrate-quality-complexity balance. Peer-reviewed listening tests across speech and music demonstrate that LC3 delivers superior perceived quality compared with SBC and mSBC at roughly half the bitrate; it also provides robust packet loss concealment and flexible frame sizes, including low-latency modes that make the encoding delay a smaller slice of the end‑to-end budget [2]. The benefits are practical: lower bitrate shrinks airtime, which reduces collision risk; shorter frames pair cleanly with CIS scheduling so deadlines are easier to meet; the codec’s computational footprint is modest enough for miniature devices [2].

Audio Precision provides high-performance audio analyzers, accessories, and applications that have helped engineers worldwide design, validate, characterize, and manufacture audio products for over 40 years.

Hearing Aids: Power-Constrained Wireless Audio

Modern hearing devices are a complex assembly of multiple microphones, digital signal processors, and miniature power sources. Except for Completely-in-Canal (CIC) and Invisible-in-Canal (IIC) designs, which are so small they fit entirely within the ear canal, most hearing aids incorporate two or more microphones to support directional processing, beamforming, and noise reduction. Audio output is provided by a single electro-acoustic transducer. The compact form factor severely limits battery capacity, making energy efficiency critical.

Compared to Bluetooth Classic (A2DP/HFP), LE Audio improves energy efficiency through three broad mechanisms: the LC3 codec achieves equivalent perceived audio quality at significantly lower bitrates than the SBC codec used in Bluetooth Classic; the LE 1M and 2M PHYs reduce on-air time per packet relative to BR/EDR; and Connected Isochronous Streams (CIS) enable precise scheduling, allowing the radio to sleep between transmissions, whereas BR/EDR audio requires longer active radio periods.

BLE‑compliant wake‑up receivers (WuRx) monitor the air with micro/nano-watt sensitivity and trigger the main radio with packet preambles. Reported designs demonstrate sensitivity to extremely weak radio signals (down to −80 dBm), with within‑bit duty cycling that trades latency for power from hundreds of microseconds to seconds [4]. Sleep scheduling techniques primarily apply heuristics for periodic check‑ins, event‑driven wake-ups, clustering, and time division to stretch lifetime while meeting QoS targets [5][6].

From True Wireless Stereo to Coordinated Sets

Bluetooth Classic’s A2DP supports only a single audio stream. In Bluetooth Classic’s True Wireless Stereo (TWS) devices, one earbud acts as the primary, receiving the stereo stream from the phone and relaying audio to the secondary earbud—a forwarding or relay architecture. The additional transmission hop adds latency to the secondary earbud, while increasing power consumption in the primary.

LE Audio eliminates this limitation entirely. The technology’s dual CIS capability lets the phone send synchronized left and right streams directly to both earbuds. This architectural shift enables independent CIS connections from the phone to the left and right earbuds or hearing aids, enabling synchronized stereo delivery without relaying.

Discovery and pairing have evolved to match multi‑device use. The Coordinated Set Identification Service (CSIS) allows two earbuds—or two hearing aids—to be discovered and managed as a coordinated set rather than independently, with resolvable identifiers and set‑level locks. While peer‑reviewed empirical literature on CSIS is thin, timing and carrier synchronization theory is mature: clock‑offset estimation, jitter control, phase‑locked loops, buffer alignment, and recovery strategies hold binaural timing within tens of milliseconds for lip‑sync and spatial imaging [9].

Gaming Headsets: Low Latency With Bidirectional Stereo

Gaming represents a demanding stress test for wireless audio. Bluetooth Classic’s Headset Profile (HSP) and Hands-Free Profile (HFP) support bidirectional audio for voice communication but are fundamentally limited: they transmit only in mono with a maximum sampling rate of 16 kHz, restricting both spatial audio quality and voice fidelity.

LE Audio Unicast Voice transforms this scenario by supporting stereo audio with sampling rates up to 32 kHz, significantly improving spatial audio and speech quality for gaming while maintaining voice communication with other players. End‑to‑end latency often must stay under a few tens of milliseconds for responsive play and coherent spatial sound. LC3’s shorter frames and lower bitrates shrink codec delay; tuned CIS parameters preserve deadlines while limiting retransmissions to useful values; beamforming improves capture quality for bidirectional voice without ballooning computational cost [2][7].

Audio Precision’s new Bluetooth® 5 module provides an interface to audio devices using the latest version of the Bluetooth specification, including LE Audio devices utilizing Unicast and Auracast™. Adobe Stock

Public Broadcast Audio: Auracast

Bluetooth Classic supports only one active audio connection and typically provides a range of approximately 10 meters, making it fundamentally unsuitable for broadcast scenarios such as lecture halls, churches, gyms, and airports.

LE Audio introduces the Broadcast Isochronous Stream (BIS), commercially branded as Auracast, enabling true one-to-many audio transmission. Multiple hearing aids, headphones, and earbuds can receive the same broadcast, which may be public (e.g., airport announcements) or private (encrypted, non-discoverable, optional password protection). Typical Auracast ranges extend up to 30 meters indoors and 100 meters outdoors, depending on environment and configuration.

BIS’s connectionless nature scales easily to unlimited receivers without pairing overhead; isochronous delivery tolerates packet loss well through forward error correction and interleaving; and the unidirectional transmission eliminates return traffic, reducing radio congestion. Assistive listening studies report that bypassing room acoustics and delivering audio directly can improve signal‑to‑noise ratios by 15–20 dB, making announcements comprehensible and lectures clearer [8].

Ensuring It Sounds Good in, on or Over the Listener’s Ear

LE Audio delivers the music or voice signal more efficiently than its predecessor, Bluetooth Classic. Audio engineers still need to verify their devices’ audio performance as experienced by the end user.

The listener’s pinna, the external part of the ear, and ear canal are a critical part of the playback system. For example, the low-frequency response and the effectiveness of active noise-cancellation are highly dependent on the seal between the device and the listener’s ear canal. Similarly, on-ear and over-ear headphones interact with the listener’s pinnas.

Anthropomorphic test fixtures—most notably GRAS KEMAR (Knowles Electronics Manikin for Acoustic Research) head and torso simulators—incorporate soft, deformable anthropomorphic pinnas that replicate realistic insertion and sealing conditions. These allow accurate replication of insertion depth, sealing, low-frequency response, and ANC performance [10][12].

Gaming headsets both receive and send audio. Just like music headphones, gaming headset testing benefits from fixtures with a human-like pinna to ensure repeatable measurement of ear-pad interaction. The headset’s microphone can be either a traditional boom microphone positioned close to the mouth or an array of microphones located farther away on the ear cups incorporating beamforming to isolate the wearer’s voice from any background noise. Test fixtures use an artificial mouth and a microphone positioned at the Mouth Reference Point (MRP) according to ITU-T standards to evaluate microphone performance under realistic speech and background noise conditions [10].

For testing of devices intended as broadcast receivers, an integrated test system with Auracast broadcast capability—like the Audio Precision Bluetooth 5 module—proves invaluable.

Conclusion

Bluetooth audio is no longer defined by a single radio or a single profile. It is defined by a timed pipeline—a codec that makes better sound with fewer bits, a transport that guarantees when those bits arrive, a radio that can sleep most of the time, and front‑end processing that gives the codec an easier job.

Hearing aids illustrate the payoff: arrays and beamformers improve intelligibility first; LC3 compresses with low delay; CIS schedules delivery; the radio sleeps; batteries last. Enhancements in other applications, such as gaming and public broadcast, further strengthen the case for adoption of this cutting-edge technology.

While Bluetooth audio began as a low-bandwidth, mono voice technology over Basic Rate (BR) radio in 1999, more than 25 years of evolution has produced a fundamental architectural shift. LE Audio replaces continuous point-to-point streams with scheduled, low-power, scalable audio delivery, enabling new classes of devices and use cases. The standards are ready, and audio test systems like Audio Precision’s Bluetooth 5 module are updated to incorporate the new transmission technology; the rest is execution—deploying LE Audio broadly so audio becomes instant, clear, and inclusive [2][7].

References

[1] Tosi, J., Taffoni, F., Santacatterina, M., Sannino, R., & Formica, D. (2017). Performance evaluation of Bluetooth Low Energy: A systematic review. Sensors, 17(12), Article 2898. https://doi.org/10.3390/s17122898

[2] Schnell, M., Riedl, M., Löllmann, H., & Multrus, M. (2021). LC3 and LC3plus: The new audio transmission standards for wireless communication. Proceedings of the AES 150th Convention, Online.

[3] Hermann, D., Herre, J., & Teichmann, R. (2004). Low-power implementation of the Bluetooth subband audio codec. Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Montreal, QC, Canada.

[4] Abdelhamid, M. R., Chen, R., Cho, J., Chandrakasan, A. P., & Wentzloff, D. D. (2018). A −80 dBm BLE-compliant, FSK wake-up receiver with system and within-bit duty-cycling for scalable power and latency. Proceedings of the IEEE Custom Integrated Circuits Conference (CICC), San Diego, CA, USA.

[5] Mutar, M. S., Mohammed, A. H., & Abdulkareem, M. B. (2024). A survey of sleep scheduling techniques in wireless sensor networks for maximizing energy efficiency. AIP Conference Proceedings.

[6] Mikhaylov, K., & Karvonen, H. (2020). Wake-up radio enabled BLE wearables: Empirical and analytical evaluation of energy efficiency. Proceedings of the IEEE International Symposium on Medical Information and Communication Technology (ISMICT).

[7] Yan, Z., Xu, H., & Shen, Z. (2024). Modeling and analysis of the performance for CIS-based Bluetooth LE Audio [Preprint].

[8] Kaufmann, T. B., Weller, T., Stiefelhagen, R., & Adiloglu, K. (2023). Requirements for mass adoption of assistive listening technology by the general public. arXiv. https://arxiv.org/abs/2303.02523

[9] Nasir, A. A., Durrani, S., Mehrpouyan, H., Blostein, S. D., & Kennedy, R. A. (2015). Timing and carrier synchronization in wireless communication systems: A survey and classification of research in the last five years. arXiv. https://arxiv.org/abs/1507.02032

[10] Okorn, E., & Wulf-Andersen, P. (2019). Acoustic test fixtures: From KEMAR and beyond! The Journal of the Acoustical Society of America, 146(4), 2815. https://doi.org/10.1121/1.5136656

[11] An analytical model of Bluetooth performance considering physical and link-layer effects. (2021). IEEE Xplore.

[12] IEC/ITU acoustic standards literature for headphone and earbud testing. (n.d.). Indexed in The Journal of the Acoustical Society of America and AIP Conference Proceedings.

Disclosure: AI tools were used by Wiley, which produced this sponsored article, to skim through research literature for technical insights on the evolution and state of the art of Bluetooth technology. AI was also used to polish the text for conciseness and technical accuracy.