Skip to main content

Fast Company Technology

Anthropic’s IPO asks investors to bet that a product can build a company, not the other way around

1 hour 5 minutes ago

When Anthropic launches its initial public offering this Thursday, it will be another test of how much further the AI industry’s seemingly unchecked growth can go. The company is seeking a valuation of $1.4 trillion, asking investors to buy into the idea that artificial intelligence will become a major contributor to the economy.

But the IPO also asks investors to make a more unusual bet: on a company whose own product could eventually do much of the work of building the business itself.

In its amended S-1 filing with the Securities and Exchange Commission, released last week, Anthropic says that “each new version of a Claude model directly accelerates development of the next.” Claude already helps run experiments, generate training data, and test future models, the company says. As evidence, Anthropic notes that Claude wrote more than 90% of contributions to its code base in September 2026. In the three-month period through September, the daily volume of code contributed per engineer was more than 45 times its 2024 average.

Traditionally, companies go public asking investors to bankroll employees who will build and sell better products. Anthropic is pitching something different: a business whose products increasingly help develop their own successors—and, in the process, help build the company itself.

There is some precedent for a company using its own technology to improve it. When the software company GitLab went public in 2021, its prospectus described how employees used GitLab’s development platform to identify problems and determine what to build next, a practice known as “dogfooding.” But that was software helping human developers do their jobs, rather than AI increasingly performing the work itself.

“It may be the first IPO in which this type of thing is happening,” says Sinan Aral, a professor at MIT Sloan School of Management.

Still, Aral cautions against confusing AI’s growing role in developing products with the ability to decide which products should exist in the first place. Claude may be writing code, but humans remain responsible for the more consequential research decisions. “What to do, what goals to have and what to optimize, is generally being decided by human beings,” he says.

Even so, experts say Anthropic’s pitch represents a striking departure from the conventional IPO. Until now, discussions of recursive self-improvement—the idea that AI systems could improve their own capabilities—have largely centered on the risks of systems becoming misaligned with human goals, says Jay Ritter, director of the IPO Initiative at the University of Florida’s Warrington College of Business.

But Ritter sees a potentially transformative business opportunity, too. Unlike conventional processes of innovation, which often yield diminishing returns, AI-driven self-improvement could theoretically produce exponential gains. That would fundamentally change not only how products are developed, but how the companies behind them operate.

“Anthropic and its competitors are companies that are increasingly capable of building themselves,” Ritter says. “Investors are essentially being asked to pay for the electricity to do this.”

The prospect also raises questions about how much autonomy Anthropic is willing to grant Claude—and whether a system capable of directing its own development would make decisions aligned with human interests. For now, MIT’s Aral says, the company appears reluctant to surrender that control. Its S-1 contains “a risk section like we’ve never read before,” he says, reflecting the possibility that its technology could eventually begin making consequential decisions on its own.

Not everyone is convinced that such a future is imminent, or even plausible. Kirill Skrygan, CEO of JetBrains, a company known for its software development tools, is skeptical that Anthropic has created an AI system capable of improving itself without human intervention. He also disputes the notion that Anthropic represents an entirely new kind of business, one whose development could eventually escape human control.

“I think it’s more about hype before the IPO,” he says.

Indeed, a closer reading of Anthropic’s filing reveals a more qualified vision than its headline figures might suggest. The company acknowledges that errors in existing models could carry over into future ones, and says it does not expect to let Claude develop successive frontier models without meaningful human supervision—at least not yet.

Whether Anthropic can deliver on its vision of AI increasingly building the business behind it remains an open question. Investors who buy the stock this week will be betting that it can, even as the company itself acknowledges the risks and limitations of that ambition.

Jeff Bezos’s $41 billion AI startup Prometheus is shrouded in secrecy. We found out who’s building it

2 hours 5 minutes ago

For many in the tech industry, physical AI is the next world to conquer. While developers of large language models like OpenAI and Anthropic seek to push reasoning models distilled on billions of streams of text to their limits, an emerging crop of companies are focusing on an even more ambitious goal: developing AI programs designed to speak the language of physics. Should they succeed, these firms have a stake in everything from building models that might power self-driving manufacturing or enable drug discovery to those that can take on plane design and weapons development. 

The promise is enormous. Whoever cracks physical AI could control a slice of the software stack that becomes critical to a host of industries—and worth many, many billions. Already, a growing list of companies are racing to stake their claim. Among the buzziest is Prometheus, a startup backed by Jeff Bezos that’s already flush with cash.

Bezos isn’t just bankrolling Prometheus. He’s also coleading the company, marking his first executive role since stepping down as Amazon’s CEO in 2021. His cofounder is Vik Bajaj, a Stanford professor who previously led science operations at Verily, Alphabet’s health technology venture. Prometheus was incubated in 2024 at Foresite Labs, which Bajaj leads, with support from the venture capital firm Arch Venture Partners. A number of Prometheus employees also appear to have come from firms associated with Bajaj.

By combing through LinkedIn and researchers’ personal websites, Fast Company identified 130 people who say they’re currently working for Prometheus, and assembled what appears to be the first public database of its kind. Prometheus employs approximately 250 people, including numerous accomplished AI researchers, some of whose names have not been publicly disclosed. The company is recruiting not only AI model developers but also experts in fields like aerospace and chemistry, a strong hint that it may one day target manufacturing and healthcare customers. Employees bring expertise in ocean engineering, energy engineering, mathematics, and even high-performance computing.

In this Subscriber Exclusive, you’ll discover:

  • An exclusive database of the more than 130 people who have publicly joined Prometheus, Bezos’s buzzy and secretive AI startup
  • What those employees’ expertise tells us about where the company is headed, and direct insights from people within Prometheus about what they’re actually building
  • A list of rival companies competing in the physical AI space—and their investors
  • Deep insight into what physical AI actually is, and why the market could be worth hundreds of billions
  • Why Bezos’s other, more renowned companies, Amazon and Blue Origin, stand to benefit most if Prometheus succeeds

At a Semafor event on November 17, Bezos and Bajaj said they’ve spent much of their funding on compute and have built one single multi-physics model. Thus far, Prometheus has built a neural surrogate for numerical simulation and has used the technology to design a range of hardware, including an experiment that involved having the company’s models redesign the Mars helicopter.

This next AI revolution needs AI tools trained not on text but on data from the physical world, as well as subject-matter experts who can ensure that models are distilling a common core of engineering principles that can extend across myriad domains. This, in turn, involves recruiting not just AI coders but plenty of people with expertise in the physical realm, whether that’s manufacturing automotive parts or running biology labs.

“Most fundamentally, the biggest challenge is just the sheer breadth of what we’re doing. Engineering is vast. There’s no such thing as an engineer,” one leading AI expert and early-stage Prometheus employee, speaking on the condition of anonymity, tells Fast Company. “This idea of building something that’s general enough to cover all of that space, that’s been a really, really huge challenge—and a really interesting one.”

What Prometheus, and physical AI, aim to achieve

News that Prometheus (formerly called the Prometheus Project) was under development went public only last year, and its leadership has been tight-lipped about what it’s actually trying to build. We do know that the company was incubated at Foresite Labs, which is housed within the health and life sciences investment firm Foresite Capital. (In addition to Prometheus, Foresite Labs has backed companies like AI drug discovery startup Xaira and disease therapy outfit Alumis.)

Still, Prometheus’s bare-bones website includes only links to its LinkedIn and X pages, and government documents show that Prometheus used a code name (“FL2024-008 INC”) for a small amount of lobbying and when applying for H1-B visas for some employees.

We also know that Prometheus has attracted significant capital from investors interested in healthcare, AI, and tech manufacturing. The company originally raised about $6 billion in funding, and then another $12 billion—for a $41 billion valuation—earlier this year. Its backers, Semafor reported in June, include JPMorgan; BlackRock; and Arch Venture Partners cofounder Robert Nelsen, a major name in biotech. Other backers include Powerlaw, a closed-end fund that has also invested in high-tech companies like SpaceX, Colossal, and Waymo, as well as RoboStrategy, a venture capital fund that says it owns shares in the company.

Bezos has said that the company isn’t focused on robotics, though that’s what many people first associate with the term physical AI. Instead, Prometheus aims to develop what it calls an artificial general engineer, which Bezos likens to a much more advanced version of computer-aided design, or CAD, powered by artificial intelligence.

“My through line for the last few years has been AI,” Bezos said in an interview with Squawk Box anchor Andrew Ross Sorkin earlier this year. “My time at Amazon is spent on AI. My time at Prometheus is spent on AI—Prometheus is really an AI company. And my time at Blue [Origin] is largely spent on AI. And [at] Prometheus . . . the tools that we’re building are going to help companies like Blue immensely.”

A trademark application makes clear that Prometheus’s aspirations extend far beyond healthcare, spanning much of the engineering and manufacturing pipeline. The document references physical simulations and digital twin technology, along with structural, thermal, fluid, acoustic, and electromagnetic analysis, and even “acoustic and vibration analysis, noise prediction, and acoustic design optimization.” The application also discusses 3D printing, semiconductor manufacturing, and developing factory layouts.

It’s not hard to see why this interdisciplinary approach might one day prove lucrative. Companies spend an extraordinary amount of time researching and developing new components and hardware, a process that can involve the arduous work of finding and vetting suppliers. There’s also the expensive, resource-intensive process of simulating and testing how new systems perform once they’re built and under different conditions. Another challenge: developing production lines capable of manufacturing that hardware at scale.

Consider the development of a jet engine, a process that can require thousands of workers and more than a decade of design, prototyping, testing, and manufacturing. Prometheus envisions AI tools that could dramatically compress that timeline, from, say, a decade to just a few years. The company sees similar opportunities across everything from household products to large-scale industrial systems. 

The aspiration for the physical AI industry is to automate as much of the process as possible and make real-world testing and evaluation far more efficient. The idea is that this could speed up hardware development across a wide range of industries. Indeed, Prometheus reportedly already has its eyes on a network of spin-offs in which it might one day own stakes.

“Today it takes days to weeks for engineers to set up one simulation for one design. There are a lot of repetitive, manual, low-added-value tasks that they have to go through while using this software,” explains Thomas von Tschammer, cofounder of Neural Concept, a physical AI startup and Prometheus competitor. “The opportunity with AI is first of all to automate and compress a lot of this low-added-value work from those engineers so that they can actually spend time engineering.”

Inside Prometheus

One early Prometheus employee, an expert in neural networks who was among the company’s first hires, was excited to “get back to basics” and pursue fundamental research. At the Bezos-backed firm, the employee says he’s working on building an “end-to-end” system at the intersection of physics and AI, developing “accurate, high-fidelity simulations” that could help engineers with initial designs.

The goal, the employee says, is to create an agentic artificial engineer that could one day drive “simulations and use them to make design decisions at the highest level.” Such an AI system would ideally be generalizable across engineering disciplines, much as current LLMs can be applied to a wide range of text-based tasks and types of content.

“If we can get a tool that across many different domains is able to simulate faster and more accurately, you can see how this generalizes to fields that wouldn’t otherwise have had contact very much at all,” the Prometheus employee says.

A good number of staffers come from other major AI companies, including xAI, and many have particular experience in AI-based materials science. Notably, a number hail from companies where Bajaj was formerly affiliated, including Google, Foresite Labs, and Foresite-backed Xaira.

Before leaving Foresite, Bajaj helped incubate a range of biotech firms and filed a patent for AI-based nucleic acid sequencing. Several employees have also come to Prometheus from General Agents, the applied AI research lab that Prometheus quietly acquired last year. The lab released “Ace,” a real-time computer autopilot, last year, and its former CEO, Sherjil Ozair, is now on Prometheus’s technical staff.

Other staffers have far more eclectic backgrounds. In its quest to build powerful simulations, Prometheus has hired at least one former employee of video game developer Epic Games, and a former AI researcher from Meta’s Reality Labs. The company has also recruited employees from Waymo and Red Bull Advanced Technologies, which helps translate Formula 1 innovations into applications for the automotive and aerospace industries.

Prometheus’s hires also suggest that it may eventually want to sell to the government. The company’s team includes a technical adviser who also serves as a chief scientist for the U.S. Army’s ground systems research. Another Prometheus adviser is Chase Johnson, who previously served in the Marine Corps’ innovation unit and worked as an attorney for defense tech company Anduril. Johnson, in his LinkedIn bio, suggests Prometheus could one day transform the “defense industrial base.”

Another employee who works at Prometheus government relations tells Fast Company that what brought them to the Bezos-backed firm was the desire to work on “cutting-edge” problems. In particular, they’re interested in how an AI engineering agent could aid industries with critical supply chains, including the automotive and semiconductor sectors.

“We are entirely focused on building models and products for the physical economy, and to accelerate the dream to build loop,” the employee says. “We are exclusively focused on design, engineering, and manufacturing, and I think that is an advantage. Focus is always an advantage, and we know that we are purpose-built for design, engineering, and manufacturing.” 

window.addEventListener("message",function(a){if(void 0!==a.data["datawrapper-height"]){var e=document.querySelectorAll("iframe");for(var t in a.data["datawrapper-height"])for(var r,i=0;r=e[i];i++)if(r.contentWindow===a.source){var d=a.data["datawrapper-height"][t]+"px";r.style.height=d}}}); The quest for physical AI

Bezos may already have a ready-made customer for Prometheus: himself. His space company, Blue Origin, is developing reusable rockets to launch satellites and humans into orbit, with ambitions that extend to orbital data centers. (At that Semafor event, Bezos said Blue Origin was already using some Prometheus tech.) All of that requires extensive manufacturing and engineering work, which could benefit from the kind of AI-powered software Prometheus is developing. Bezos has acknowledged the potential connection, though it remains unclear whether—or how—Prometheus might ultimately integrate with his other businesses.

Then there’s Amazon, which does plenty of its own in-house manufacturing, maintains frontier AI and robotics research groups, and has invested heavily in warehouse robotics. Bezos has also invested in Physical Intelligence, a startup developing AI models for the physical world and robotics, though Prometheus says there is no formal relationship between the two companies. Taken together, these interests raise the possibility that Bezos is positioning himself to build a new, far-reaching empire centered on physical AI.

In the meantime, a range of other startups are competing in the physical AI race. Anima Anandkumar and Benedikt Jenik were once approached about leading Prometheus before launching their own startup, Accelerated Understanding. (That offer, Reuters previously reported, came from Bajaj and would have included a major stake in the company and a million-dollar salary.) Their focus is on building models capable of making predictions in four dimensions, at any resolution. To do that, they say they need a team that isn’t constrained by traditional disciplinary silos or the limitations of conventional LLMs.

“Even if one were to use LLMs to automate execution, the real world stays the bottleneck,” Anandkumar tells Fast Company. “If we instead teach AI models to directly simulate and understand physics, that bottleneck goes away.”

Other firms envision a similar formula for building physical AI teams: a mix of machine learning generalists and specialists with engineering and manufacturing experience.

Paul Eremenko, of P-1 AI, which makes an agentic AI tool called Archie that works directly within engineering teams, argues that it’s often much easier to train someone with an aerospace background in the latest AI research than to teach an AI specialist the nuances of aerospace engineering.

“We don’t compete with OpenAI or Anthropic for talent,” Eremenko says. “The prototypical AI researcher or AI engineer for us is somebody who comes from a physical world background—a mechanical engineer, aerospace engineer, electrical engineer—who became very interested in AI.”

Back at Prometheus, the competition for talent is only expected to intensify. An employee working in government relations tells Fast Company that the Prometheus expects to expand its workforce exponentially, even as hiring remains highly selective. For the early employee specializing in AI, the recruiting pitch comes down to the opportunity to tackle problems that have long eluded conventional engineering tools. “In terms of wooing them, I think I’m often selling people on the novelty and the excitement.”

The smartest AI might be the one that leaves us alone

2 hours 35 minutes ago

One of the ideas I’m always trying to tell my students is that their brains are something, an asset, if you like, they need to protect. The human brain is an amazing optimizer: Provide it with an effortless way to achieve something, and it will quickly start to use it. We used to memorize a ton of phone numbers a few decades ago. Now we can hardly memorize our own, simply because our brain knows the number is a few clicks away. 

A clear example is why I never use AI to write my articles: simply because, for me, in my personal methodology as a professor of innovation, writing is my way of making concepts stick in my brain, concepts I will need later to teach and discuss. If I ask ChatGPT, Claude, Gemini, or any other chatbot to write my articles for me, I would lose much of the learning that writing gives me.

This doesn’t mean I don’t use AI: I can use it to translate certain parts, to search for relevant articles to back up an assertion, or to assess whether the concepts I try to explain come across clearly. But I don’t want to use it for writing, because writing is a skill I want to protect. And like a muscle, if you don’t exercise it, it quickly atrophies.

For a manager, the skill worth protecting might be decision-making: An AI that makes every choice could undermine precisely the judgment that person needs to develop. The useful behavior depends on the objective, and sometimes that objective requires the system to hold back. 

What happens when we move from personal to corporate use? Can an AI learn when its intervention would make things worse? That capability might receive very little applause in a product demonstration, but it could be enormously valuable in practice.

The sales agent that lets us be 

Imagine a company using a sales agent. When it identifies a customer approaching renewal, it follows a standard procedure: It quickly reviews the account, writes a personalized email, offers a discount, and schedules a follow-up. All of it in mere seconds. And the audience claps enthusiastically. 

But what if our customer would have renewed anyway? Then, the end result is that we have wasted an interaction, risked alienating the customer, and surrendered margin unnecessarily. What if our customer is a happy one, and the right thing to do would have been just to wait for him or her to renew voluntarily? 

Even when all tasks are perfectly executed, the decision to perform them could have been perfectly wrong. 

The cost of being helpful

We know how this feels when we’re working with real people: Experienced employees often develop a sense of when to follow up, when to escalate an issue, and when making too many interventions will look messy, send the wrong signal, or make the customer upset. How do they know that? It comes with experience; it’s learning that compounds over time, with more and more iterations.

That judgment is something we pay to have. However, when we evaluate AI, we tend to be fascinated with visible activity—with messages generated, tasks completed, workflows executed, etc. The equivalent is if we were dealing with a hyperactive employee. And as we probably know, someone who sends twice the number of messages is probably not twice as valuable. 

In other words, just because an action can be automated and, therefore, performed easily, doesn’t mean that we should be performing it all the time. 

Look what happens with AI coding assistants: In “When to show a suggestion? Integrating human feedback in AI-assisted programming,” Microsoft researchers used data from 535 programmers to retrospectively evaluate a system that selectively withheld suggestions that were apparently likely to be rejected by the developers. They also found that optimizing for suggestion acceptance could produce lower-quality suggestions. Getting people to accept more assistance was an imperfect measure of helping them. 

Learning when to intervene

In the case of our renewal agent, a system that learns from a lot of comparable cases could perhaps investigate which interventions help with whom. For instance, some customers don’t like the feeling of being “directed” or “influenced,” and are better left to their own devices. Some like a particular type of interaction through a particular channel, but they hate others. Or they want to talk to the same customer representatives every time because they remember them. 

The function to protect would be the ability to spot these cases. Do not automate; optimize. But optimization is not necessarily something that today’s corporate AI does well. Today’s corporate AIs can generate splendid emails, but following the before and after and deciding to do nothing is usually not an alternative. 

It’s the experience from all previous cases, processed along with the specific characteristics of every customer, that leads to the lessons learned, to the continuous improvement, to the process being optimized. And that ability to improve decisions through accumulated experience is precisely what you need. 

Give waiting a fair chance

When buying or deploying AI, managers should, therefore, find out what outcome the system is theoretically trying to improve, and design ways to know whether its actions helped or not. We need to be honest here: A renewal following an email does not necessarily establish that the email was the reason for the renewal. Customers differ, circumstances change, and useful comparisons may require carefully designed experiments. Simply storing more interaction histories will not settle the question. 

This includes maintaining the necessary human oversight, the now-famous “human-in-the-loop”: A missed warning or an unanswered complaint can be costly. People must define the obligations and limits within which the system can learn. 

I haven’t seen an AI system that reviews an account, checks all the available evidence from the corporate ERP or CRM, and decides that contacting the customer would probably make things worse. So the final outcome would be that nothing gets sent, no discount is offered, and we just let the customer evolve naturally. No hyperactivity. Why? Because our corporate learning optimized it, and the accumulated learning behind that decision should remain under our company’s control, available even if we replace the model or the vendor powering our agents.

It might be a disappointing demo. But it could be excellent business. 

Apple is turning your iPhone into a car key for 3 popular brands. Unlocking is only the beginning

3 hours 5 minutes ago

Cadillac, Chevrolet, and GMC drivers will now be able to unlock their vehicles with their phones.

As first reported by MacRumors, General Motors announced that it has begun rolling out the Apple Wallet feature to select models. More vehicles are expected to be added over time. Eligible customers will receive notifications in the “What’s New” section of their vehicle’s app when digital-key support is available. 

Users will be able to use their smartphones or Apple Watches much like a physical key to unlock their vehicles. GM said supported vehicles will automatically unlock as a driver approaches and lock as the driver walks away. According to the company, the vehicle’s trunk and climate control system will be accessible with the new Apple Wallet feature. 

“The digital key experience will continue to evolve beyond simply unlocking and starting a vehicle. By leveraging ultra-wideband technology, GM is laying the groundwork for more deeply integrated and personalized experiences like automatically adjusting vehicle settings as drivers approach and other intelligent features that make ownership more convenient,” said the company in a press release.

Apple first announced the concept at its Worldwide Developers Conference in June 2020. It later launched exclusively for select 2021 BMW 5 Series models. Since then, the tech company has added other major automakers, such as Mercedes-Benz, Hyundai, Kia, Genesis, and Audi, to the offering.

The digital key may work even when a phone’s battery is dead. The company said that depending on a user’s iPhone model, keys could still be available even when a device needs to be charged. However, it warned that they will be available only for up to five hours.

The evolution of digital keys

Samsung rolled out its own version of the digital keys in 2021. In October, Samsung announced it would support digital keys for select GM vehicles. 

The digital key is the latest evolution in car-key technology. In the 1980s, automakers began using radio frequencies to create remote keyless entry fobs. 

In 2008, Nissan and tech company Sharp launched a mobile phone with a built-in intelligent car-key chip in Japan. Then, in 2017, Tesla advanced the concept by making its smartphone app the primary, default key for the Model 3, replacing the standard physical key fob. 

Apple and GM did not immediately respond to requests for comment. 

Get Inc.’s free CPG newsletter, Packaged Goods, for the latest on the founders, investors, and trends reshaping consumer brands.

—By Moses Jeanfrancois, News Writer

This article originally appeared on Fast Company’s sister website, Inc.com. 

Inc. is the voice of the American entrepreneur. We inspire, inform, and document the most fascinating people in business: the risk-takers, the innovators, and the ultra-driven go-getters that represent the most dynamic force in the American economy.

Nvidia’s first earnings call without Jensen Huang had one message: It’s business as usual

12 hours 38 minutes ago

Nvidia on Tuesday held its first earnings call in its 33-year history without cofounder Jensen Huang, who resigned five days ago after a scandal involving an extramarital affair with a suspected Chinese spy. After reporting strong third-quarter numbers, interim CEO Colette Kress (who has been Nvidia’s CFO since 2013) kept the call focused on Nvidia’s immediate business concerns. 

Analysts wanted to know what the company would say about the succession plan after Huang’s departure, and about the future of Nvidia’s pivotal role as the company powering the AI race. 

Most of the salient information about Huang was given at the front of the call by Harvey Jones, Nvidia’s chair of the board. Jones said Nvidia doesn’t intend to reinstate Jensen as CEO, providing a clapback to the growing contingent on social media lobbying for the ousted CEO’s return. Jones said the executive search firm Heidrick & Struggles is handling the search for a new CEO, but that Nvidia is looking at both internal and external candidates for the job. 

“The board is acting deliberately and in the best interest of the company’s employees and investors,” Jones said. “With that we will take no further questions about the circumstances of Jensen’s resignation.”

The numbers

Kress then took over the call and kept things focused on Nvidia’s numbers. Kress has run Nvidia’s earnings calls for years, but always alongside Huang, who set the company’s vision and did much of its dealmaking with customers and partners. Today her main job was to report the quarterly numbers, and convince investors that the world’s most important AI chip company can run without its founder.

The numbers, at least, are good. Nvidia reported $115.8 billion in revenue for its fiscal third quarter, well above Wall Street’s roughly $109 billion forecast, and projected $130 billion for the current quarter. Investors expected a beat. 

Kress said Nvidia has seen “explosive” adoption of its latest Vera Rubin computing platform, which includes a Rubin GPU and a Vera CPU. “Our momentum is real,” she said. “Our ambition will not change, our passion for innovation will not change.” 

Revenue for the quarter came in at $115.8 billion, about $6.5 billion higher than the roughly $109 billion analysts expected. Adjusted earnings were $2.68 per share, 21 cents above the $2.47 consensus. 

But all that happened during a third quarter that ended October 25—that is, before the Huang scandal broke. That’s why analysts were more concerned about the guidance numbers for a future without Huang. Some analysts had even questioned whether Nvidia would venture to give a full year 2028 guidance. 

But it did. For the short term, Nvidia forecast fourth-quarter revenue of $130 billion. The company had previously said it expects to see 70% growth in 2028, but Kress said its moving that guidance up to 75% growth, citing continued strong demand for AI infrastructure, including Vera Rubin, and despite the chance that chip exports to China will constrict somewhat in the wake of the Huang scandal.

That may have settled some analysts’ nerves about the AI sell-off that started in October, which was driven by the fear that Big Tech companies would slow their spending on AI data centers. 

Why this call mattered

Huang, who cofounded Nvidia in 1993, resigned on November 12 amid two scandals. The first involved an alleged affair with a woman U.S. intelligence believe may have been spying for China. Beijing has strongly denied the allegations. Nvidia’s board has opened an internal investigation into whether company or national security secrets were compromised. Huang stepped down shortly after reports of a second alleged affair, this one with Kimberly Powell, the company’s vice president of healthcare (and the CEO’s direct report).

The board made Kress interim CEO. She also stays on as CFO until a permanent CEO is named. Longtime director Harvey Jones became Nvidia’s first board chair. 

The stock was already falling before the scandal. Nvidia closed at a record $239.24 on October 6. Shares then fell on fears that AI spending would slow, and fell further as the Huang story unfolded. By Tuesday the stock was below $154, its lowest level since mid-2025. That’s a loss of about $2 trillion in market value. That’s why Wall Street was listening to Kress so closely. 

Before the call, analyst Dan Ives of Yorkville Ives predicted Kress would talk very little about Huang, that her main job would be to “restore order and give the Street confidence in the Nvidia story going forward,” he told Fox Business. 

And that’s exactly what happened. After the brief section at the front, the call had a business-as-usual tone, complete with a blizzard of numbers showing that Nvidia—as of the third quarter, at least—continues to sit at the center of the AI world. 

Inside Hank Green’s plan to protect kids (and their parents) from YouTube

18 hours 42 minutes ago

Pity the parent trying to pull the plug on YouTube, Netflix, or Disney Plus mid-video. The tears! The world-ending wails! 

Hank Green, host of popular YouTube channels Crash Course and SciShow, has been that parent. A longtime science educator and the co-creator of VidCon, Green has fought his fair share of “turn it off” battles with his son. 

“These systems are designed to keep kids watching,” he says of advertising-based platforms like YouTube and TikTok. “They don’t have any incentives to even consider designing a ‘stop’ experience.”

But parents, on the other hand, are highly incentivized to seek out an alternative to the endless scroll of sometimes questionable content that streams to family devices. Green hopes they will give his new venture, ClockwiseTV, a chance. 

Launching publicly today, ClockwiseTV is a subscription-based app featuring curated creator videos for kids ages 8-12. Parents establish a session length when opening the app; once kids start watching, ClockwiseTV helps them pick a video sequence that concludes when their session concludes. 

The app’s constrained choice philosophy extends to the video selection experience. Instead of infinite options, ClockwiseTV presents kids with a menu of just a dozen videos at a time, based on their interests (animals, space, history, etc). Each video has been vetted by a human and scrubbed of any inappropriate language. Creators in the initial library include Animal Wonders Montana, NerdForge, Crash Course Kids, and MinuteEarth. 

At launch, a subscription costs $6.99 a month or $69.99 a year. There are no ads, and the app is compliant with the Children’s Online Privacy Protection Act, or COPPA. 

[Image: ClockwiseTV] No autoplay, no rabbit holes

ClockwiseTV joins a growing array of products designed to keep kids entertained without algorithms or ads. For screen-free audio content, there are Yoto players and Tonie boxes. For kids who want to gab with their besties, there is the Tin Can phone, which mimics the experience of a landline for an approved set of friends and family. There are even sticker printers that use AI to turn kids’ ideas into black-and-white stickers they can color and decorate. 

Hank Green [Photo: ClockwiseTV]

Green says he’s been working on ClockwiseTV for about a year. The decision to focus on tweens stems from his own experience as a parent. “My son is this age, so that’s part of it,” he says. “There’s a kind of no-man’s-land of content where it’s like, you’re done with Mr. Rogers, here’s all of YouTube. I think that’s probably not the right way to do it.” 

Over time, he adds, ClockwiseTV may pursue “other slices of the pie” and expand its target audience. 

Leslie Morgan [Photo: ClockwiseTV]

More than half of children own a smartphone by age 11, according to Common Sense Media’s 2026 census. The average tween spends five hours a day watching entertainment on a screen, the census found. 

Leslie Morgan, who has worked in digital media with creators for more than a decade, joined ClockwiseTV as CEO in January. Like Green, she hopes to help parents avoid the internet’s darker corners. “There are no rabbit holes, there’s no autoplay,” Morgan says of the app. 

Incentivizing kid-friendly creators

ClockwiseTV is paying creators to license their content. “A lot of these creators would love to get in front of a younger audience, and it’s not financially viable,” says Morgan, explaining that labeling content as “made for kids” on YouTube typically lowers creators’ AdSense revenue. “For them, it doesn’t make economic sense.” With ClockwiseTV, she argues, creators can access new audiences and supplement their income from other platforms. 

When I watch YouTube with my kids, it’s often a stressful experience. Typically they’ve requested to revisit a favorite video that, in their view, is very specific (“the tap dancing man in the hat!”). It’s up to me to weed through a confounding list of thumbnails and keyword-optimized titles, searching for an elusive winner, while they squirm impatiently. 

[Image: ClockwiseTV]

ClockwiseTV, in contrast, is built to reward parental trust. I told the app that my son was interested in history videos, and had just 15 minutes to watch. The app suggested a short video about fossils and a slightly longer video about hunting for 50 million-year-old shark teeth. Both were educational and engaging. 

“That’s it for now,” the app announced when the allotted time was up. “Go do something else that’s awesome!” 

I found myself wishing that Instagram would tell me the same thing. 

People’s AI secrets raise questions about the use of chatbots in personal life

18 hours 50 minutes ago

A purple box set up on a folding table in San Francisco’s Alamo Square could have been mistaken for a tissue box. Instead, it held dozens of secrets.

Olivia Tai was nearby, watching as passersby puzzled over a sign that read, “Confession box: Everyone has an AI secret.”

Among confessions about farming out too much work or using artificial intelligence to suggest a first-date spot were several admissions about using it to write deeply personal messages: “I had AI write my breakup text.” “I had ChatGPT date my boyfriend. Every text.” “I used AI to write my girlfriend’s grandmother’s 80th birthday letter.”

“I used AI to win an argument. Lol, not proud, but I won.”

Tai, an artist who has repeated the project more than a dozen times in the Bay Area, New York and Denver, said calling it a “confession” suggested people were admitting that using AI made them feel at least a little guilty. But they used it anyway.

As more people turn to a chatbot to write emails or handle mundane tasks, it raises the question: Does using it for personal messages cross a line? Is it ever OK to use AI to write a holiday card or birthday message? What about a love letter or wedding vows?

Some use AI to write personal messages, even if they feel guilty about it

As part of research for a Ph.D. in consumer behavior, Danielle Hass conducted vignette-based experiments to determine how people felt about using AI to write personal messages. She found that people who had used AI before felt just as guilty as those who hadn’t.

“Everyone felt guilty,” said Hass, a professor at Kennesaw State University in Georgia. “The driver of this relationship was dishonesty, that people felt like they were manipulating.”

Even if people are honest about their AI use, others still view them negatively, said Scott Claessens, a senior researcher at the University of Bristol in the UK.

Claessens says his research suggests that people who use AI are perceived as lazier, less competent and less trustworthy.

He noted that people often already supplement social tasks, such as using pre-written birthday cards or poems in love notes. But something about the effortless way that AI works turns people off, he said.

“What we really want in our social relationships is for people to put effort in, because it shows that as a signal that you’re invested in the relationship,” he said.

How to know when it’s not OK to use AI

Sarah Newman, who collaborated with Tai as director of art and education at Harvard University’s metaLAB, said AI adoption is happening so quickly that norms will likely emerge about when it’s appropriate and when it’s not.

She suggested using conflicted feelings as a guide. If you imagine using AI to write your wedding vows, would you be comfortable telling your partner you used it? If not, then don’t.

“Personal messages don’t need to be perfect,” Newman said. “Part of what makes these messages authentic and endearing is that they are imperfect.”

Tai and Claessens said some situations could warrant using AI in personal situations, such as people in relationships who use it to communicate better with a partner who doesn’t share a native language. People with learning disabilities, such as dyslexia, could also use it for help.

Hass believes people are quickly accepting that AI can make them more efficient in some work contexts, but she said she’s glad people feel guilty using it for personal messages.

“Guilt’s purpose in this context is to show us our interpersonal wrongdoings,” Hass said. “It’s showing us, ‘hey, this is where you’re supposed to be human.'”

Albert Stumm writes about travel, food and wellness. Find his work at https://www.albertstumm.com

—Albert Stumm, Associated Press

Are school phone bans working so far? Here’s what early studies show

21 hours 20 minutes ago

The argument for school cellphone bans is simple: Phones offer endless distraction. Take them away and students can focus on classwork, engage with peers and, ultimately, learn more.

But the reality is more complicated, as emerging research shows. Schools can lock phones away in pouches or banish them to backpacks, but they can’t shut out the digital world.

In the past few years, more than 40 U.S. states have passed laws that restrict cellphones at schools, according to the Phone Free Schools State Report Card, which tracks state phone legislation. But those laws vary widely and often let schools design their own policies, resulting in a patchwork of rules and enforcement across the country.

A large-scale study published Tuesday found strict cellphone restrictions — ones that require students to put phones in locked pouches for the entire school day — result in small academic gains. But, the researchers said, the gains can add up if schools continue banning phones.

Phone bans appear to help, to a point

The study, led by Harvard’s Center for Education Policy Research, found the schools with strict phone restrictions saw “modest” improvements in academic growth, compared with students in similar schools without those restrictions. The study, which analyzed test scores of 3.3 million students at 4,600 schools nationally, estimates the academic gains were equivalent to about an additional week or two of learning over a school year.

Student achievement has been falling globally since 2013, which is roughly when social media-enabled smartphones took off. The researchers sought to determine if the decline is being reversed now that schools have started taking phones away, said Harvard professor Thomas Kane, one of the study’s authors.

“Our answer is yes, phone bans do have a positive effect on growth. But it’s not a panacea,” Kane said. Learning loss was gradual over a decade, he notes, and it’s not surprising that gains would also be gradual. The study used math and reading test data for grades four through 10 from the past five academic years.

“While schools should not expect large boosts in achievement immediately, storing students’ phones seems to be a cost-effective way to add 1-2 weeks of instruction each year, which could add up over time, just as the initial declines did,” the study says.

The boost is equal to about half the annual decline in test scores since smartphones became mainstream, the study said.

Academic gains are overall ‘disappointing,’ one researcher says

Still, the connection between phones and learning is not clear-cut.

A different study published in May concluded strict school phone bans had “close to zero” effect on test scores. It compared schools with bans with tens of thousands of other schools.

The National Bureau of Economic Research study looked at a different set of standardized test scores than Harvard’s study, as well as GPS phone activity, teacher surveys and data from Yondr, the leading provider of lockable phone pouches, at 4,600 schools that use the pouches.

The study had found “modest positive effects” in high school test scores over three years, particularly in math. But the report’s authors, researchers at Stanford, the University of Michigan, Duke and the University of Pennsylvania, were hesitant to conclude overall learning benefits from the bans.

Both studies point to a similar conclusion, said Stanford education economist Thomas Dee, an author of the NBER study. “The early results from school phone bans are largely disappointing and raise unresolved questions,” Dee said.

Noting that the Harvard study showed learning gains may accumulate over time, Dee said: “I hope they’re right. However, I’m not comfortable making that extrapolation from estimates that are small, sensitive and based on just a few years of phone bans.”

The reasons, Dee said, could be that students are still distracted on other devices at school. Or, he said, “it may be that we see these disappointing results because out-of-school phone use is harming child development.”

Phones aren’t the only tech distracting teens

Just because schools or states have phone bans does not mean they’re being enforced, said Jean Twenge, a psychologist at San Diego State University and author of “10 Rules for Raising Kids in a High-Tech World.”

Ask any teacher — or teenager — and they’ll tell you that some phone bans are largely cosmetic. Teens come up with creative workarounds, including stowing burner phones or even phone-shaped rocks in their locked pouches. At schools that rely on students to keep phones in backpacks, adherence to the ban can be spotty.

“It happens pretty often that there’s a policy of no phones during the school day, bell-to-bell, and you go into the school and phones are everywhere,” Twenge said.

Even without phones, “Do they have laptops at school? Are they just doing social media on laptops instead?” Twenge said.

Phone use outside of school affects academics too, Twenge said. “If they have that phone available when they’re studying, how much are they really taking in?”

Kane, of the Harvard study, agrees the effect of cellphone use outside of school is “the big question.” But researchers have not been able to investigate parents and educators’ suspicions that late-night binges on YouTube and other platforms are costly for students.

“I find it astonishing that social media companies know how much time kids spend on their platforms, but that data has never been linked to student outcomes,” Kane said. “We can’t provide answers until the social media companies share data.”

The Associated Press’ education coverage receives financial support from multiple private foundations. AP is solely responsible for all content. Find AP’s standards for working with philanthropies, a list of supporters and funded coverage areas at AP.org.

—Jocelyn Gecker, AP Education Writer

Lab-grown chicken is finally becoming affordable. Will consumers bite?

22 hours 5 minutes ago

There’s something new on the menu at Epic Steak, an upscale restaurant across the street from Google in San Francisco: “Cultivated” chicken, made partly from chicken cells grown in bioreactors, is now offered to diners right alongside staples like porterhouse and Wagyu.

Epic is one of eight California restaurants that serves the food, along with some corporate and college campuses, including Stanford University. The Bay Area startup behind the cultivated chicken, Upside Foods, says it has reached a new milestone in scaling up its production—and that has helped bring the price down in the range of premium farm-raised chicken. Its new product includes a blend of cultivated chicken and plant-based ingredients like soy and olive oil.

[Photo: Upside Foods]

At Epic, “Southern Fried Upside Chicken” with kimchi slaw and hot honey aioli sells for $12. Waterbar, a neighboring seafood spot with the same chef as owner, sells a five-spice version of Upside chicken for $15. At Chimmelier and Jilli, popular Korean fried chicken restaurants in Los Angeles and San Francisco, K-Pop Chicken made from the new meat sells for $19.

“We can put it at a price point on the menu where it can be available for everybody,” says Parke Ulrich, executive chef and owner at Epic and Waterbar. Two of his other Bay Area restaurants, Mersea and Hurrica, will also serve dishes made with Upside chicken.

Upside, which launched in 2015 and has raised around $600 million, has spent years on R&D for its cultivated meat—meat produced without raising animals or using slaughterhouses, using real animal cells that grow in tanks the company calls cultivators.

The company was originally called Memphis Meats, but rebranded in 2021 with a name that it saw as more consumer-friendly. It’s one of more than 140 companies working in the space of cultivated meat. A handful are selling products on the market, including Vow, which sells cultivated quail and fois gras in Australia, and Wildtype, which sells cultivated salmon at a few U.S. restaurants.

In theory, producing meat this way could help reduce the enormous environmental footprint of animal agriculture, though that depends on how it’s made, including how much energy is used.

The space has been slow to take off, and some other companies have faced difficulties. Good Meat, another Bay Area-based cultivated meat brand, stopped commercial production in the U.S. in 2024. Two other startups, Meatable and Believer Foods, folded late last year.

Over the last several years, “the biggest challenge was the external skepticism that this would have worked,” says Dr. Uma Valeti, cofounder and CEO of Upside Foods. The company first unveiled a commercial facility in 2021, though it didn’t initially use its large-scale bioreactors. (Wired reported that it struggled to use the equipment; the company says that its early R&D challenges have been fully addressed.)

In 2023, when the company got regulatory approval of an earlier version of its chicken, it was produced at a small scale. Customers had to make reservations for the dish, available only in limited quantities, at a restaurant that sold it temporarily, or find it at one-off events.

[Photo: Upside Foods]

Now the company says large-scale production is underway. Upside has built a much larger production line, with 20,000-liter tanks in addition to its earlier equipment. Valeti says the two-story-tall tanks “are the largest of their kind in the world,” including anything ever used in the pharmaceutical industry. (The company had previously planned to build a factory in Illinois, but retrofitted its Emeryville headquarters with the giant equipment instead.)

In a year, it claims that it has the capacity to produce more than a million pounds of Upside Chicken on less than half an acre of land. One cultivator produces 1,500 to 2,000 pounds of meat a day; the next “harvest” will begin later this week. Inspectors from the U.S. Department of Agriculture work on-site, using a new regulatory process that was developed for this type of food.

[Photo: Upside Foods]

Scale helps bring down cost. The company makes a key ingredient for its cell feed, something that was expensive to source previously. The chicken itself has also been redeveloped and now includes some plant-based ingredients. That helps improve the texture and nutrition of the food, Valeti says, but also reduces cost. The company declined to comment on the percentage of the food that’s made from cultivated chicken cells versus plant-based ingredients, saying only that cultivated chicken is the second ingredient after water.

[Photo: Upside Foods]

For restaurants and cafeterias, the chicken comes in whole cuts, cubed, or shredded varieties. Ulrich, who has been developing recipes for his restaurants with Upside Foods over the last several months—experimenting with everything from soup to a barbecue chicken sandwich—says the chicken flavor is subtle, since the product is a blend. But the texture is “fantastic,” he says. It “peels apart like a chicken breast would. It has those strands. And it’s just very silky in texture.”

[Photo: Upside Foods]

Upside wants to show demand at the restaurants and cafeterias it works with now, and then will scale up production more and potentially begin retail sales. With more scale, and production facilities built outside the expensive Bay Area, the cost will come down even further.

The big question: How many consumers will be willing to eat it? The market for alternative meat has been struggling for a few years. Plant-based meat sales in the U.S. are roughly half of what they were in 2021. As consumers increasingly look for “clean” ingredients—simple, recognizable, and natural—cultivated meat’s production methods are a harder sell. Even as the taste gets closer to conventional chicken, the food is still likely to be perceived as different.

Texture is key, chef Ulrich says: “My focus has been trying to get the texture there, because I think people eat with their eyes, and it’s almost more important sometimes than how they taste it. Because if it doesn’t look like chicken, if the look isn’t there, their taste buds automatically go somewhere else.”

Correction: An earlier version of this article misstated Upside Foods’ approval timeline and federal inspection procedures.

FEMA has a cord-cutting problem

1 day 2 hours ago

On Tuesday, countless TV viewers will find their favorite shows interrupted for a one-minute test of the Federal Emergency Management Agency’s Emergency Alert System (EAS). The test will cut into live broadcasts and cable reruns alike, and also be heard on radio stations across the country. However, anyone getting their TV fix from Netflix, Disney+, or any other streaming service likely won’t notice a thing.

That’s bad news, especially when the next real emergency happens. According to recent Nielsen data, U.S. consumers are already spending half their TV viewing time with digital video platforms like YouTube and Amazon Prime. In other words: Half of America’s TV viewers are missing out on emergency alerts because they’re watching online instead.

“It is vital that emergency alert systems adapt to changes in consumer technologies,” says University of Colorado Denver communications professor Hamilton Bean, an expert in public alert and warning systems. “Whether through streaming services, mobile devices, or IoT [Internet of Things], emergency alerts work best when they reach people where they are—at home, at work, or in public settings.”

The federal government first began testing emergency alerts in the 1950s to warn citizens of a possible Soviet nuclear attack. Federal alert systems have been modernized multiple times since those early days, including an expansion to mobile phones in 2012. But in recent years, regulators and emergency management officials have struggled to keep up with rapidly changing media consumption habits and shifts in the TV business.

This has led to confusing gaps in the Emergency Alert System. One example: Consumers who have switched their TV subscriptions from a cable company to an internet-based TV service like YouTube TV, Fubo, or Hulu + Live TV may also miss out on alerts. These so-called “virtual” pay TV operators don’t participate in the EAS, which means they don’t interrupt their programming to carry such warnings.

But if a YouTube TV subscriber happens to be tuned into a local broadcast network like ABC or Fox, they may see the alert after all. Those channels are frequently ingested from local broadcast TV affiliates, which in turn do work with FEMA. In other words: YouTube TV subscribers who watch Dancing With the Stars on ABC during an emergency may get a warning, while those tuned into Bravo’s Southern Charm may be blissfully unaware of any impending danger.

“Regardless of the other aspects of regulation, when it comes to public safety issues, [traditional and internet-based pay TV services] should be on equal footing,” says John Bergmayer, legal director of the nonprofit technology advocacy group Public Knowledge. Virtual pay TV services already need to know a subscriber’s location to serve up local broadcast channels, which should help them target consumers with the right alerts. “And like cable, they have the technical capacity to both interrupt a program or to add some sort of crawl along the bottom of the screen,” Bergmayer says.

Streaming services and their media company owners have generally resisted the idea of being forced to carry emergency alerts. In a 2021 regulatory filing, the Motion Picture Association, which counts Netflix, Disney, and Skydance among its members, declared that it would be “enormously technically difficult, if not entirely impracticable” to expand emergency alerts to streaming.

One problem the industry association pointed to is the federated nature of America’s Emergency Alert System. While Tuesday’s test alert goes out nationwide, real warnings about wildfires or flooding are necessarily local. Requiring a streaming service like Netflix to relay those alerts would mean building out extensive monitoring infrastructure in every state.

Some hope that the new Advanced Television Systems Committee (ATSC) 3.0 broadcast standard, which brings modern technologies to antenna TV, will come to the rescue. In theory, modern TVs equipped with ATSC 3.0 tuners could interrupt a Netflix stream as soon as they receive an emergency alert via antenna. In practice, there’s no regulation in place to require this yet. Critics also fear that ATSC 3.0’s content protection technologies could hamper emergency warnings.

In the end, the lesson from Tuesday’s test may be that emergency alerts can’t rely on just one technology. “There is no silver bullet,” says Art Botterell, author of To Warn Is Human: A Cultural History of Public Warning. Instead of trying to find a tech fix, he’d prefer that we invest in civic infrastructure and strengthen the bonds between neighbors so they can alert and help one another. “We need to facilitate that person-to-person sharing,” Botterell says.

That, in turn, may also help people who happen to tune into the wrong TV show at the wrong moment.

Holiday gift guide 2026: 20 thoughtfully curated tech gifts that will charm and delight everyone on your list

1 day 3 hours ago

As the world obsesses over AI and conventional tech products balloon in price, the thought of buying gadgets for anyone this holiday season might feel a little intimidating. But that’s not to say that there aren’t options out there. I often find that the best gifts are charming, focused devices with a touch of style. 

My selection process for this gift guide was entirely subjective. Almost everything on it is a gadget or design object with an element of tactile satisfaction for how it handles an everyday task. You won’t find a boring black rectangle anywhere on the list, and you shouldn’t have to think too hard about the soaring cost of memory.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/12\/multicore.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/12\/multicore-mobile.png","eyebrow":"","headline":"\u003Cstrong\u003E\u003Cstrong\u003ESubscribe to \u003Cem\u003EMulticore\u003C\/em\u003E\u003C\/strong\u003E\u003C\/strong\u003E","dek":"\u003Cem\u003EMulticore\u003C\/em\u003E is about technology hardware and design. It’s written from Tokyo by Sam Byford. To learn more, visit \u003Ca href=\u0022https:\/\/www.multicore.blog\/\u0022\u003Emulticore.blog\u003C\/a\u003E","subhed":"","description":"","ctaText":"SIGN UP","ctaUrl":"https:\/\/www.multicore.blog\/","theme":{"bg":"#f5f5f5","text":"#000000","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#000000","buttonHoverBg":"#3b3f46","buttonText":"#ffffff"},"imageDesktopId":91454027,"imageMobileId":91454030,"shareable":false,"slug":"","wpCssClasses":""}} Apple: The First 50 Years by David Pogue

Published to mark Apple’s half-century this year, veteran tech journalist David Pogue has marked the occasion with a hefty tome that serves as a biography of the company itself. Pogue conducted 150 interviews for the book, including figures like Jony Ive and Steve Wozniak, and the presentation benefits from hundreds of photos throughout. This is a book that’s worth owning in print.

Twelve South Airfly SE

Some airplane seatback screens support Bluetooth, but it’s not something you can always rely on when boarding a long-haul flight, as many screens still won’t connect to noise-canceling headphones. Twelve South’s AirFly SE is an inexpensive dongle that plugs into your seat’s headphone jack and transmits the audio to your cans. You’ll never need to use airline earbuds again.

Sharge Flow Mini Go

Sharge’s whole thing is designing utilitarian gadgets with a stylish twist, and the Flow Mini Go is one I can recommend to anyone who ever finds themselves running low on a charge. It has a 5,800-milliampere-hour (mAh) battery, 22.5-watt charging and a built-in cable, as well as an option to attach it directly to your phone’s USB-C port. Just as importantly, it’s only about as big as a lipstick tube. 

Nothing Ear (a)

At $79, Nothing’s Ear (a) model is my pick for affordable wireless earbuds. They sound great and look even better, especially if you opt for the construction-equipment-yellow pair. Even the charging case is cool, with a design inspired by medicinal blister packs.

Logitech MX Master 4 Developer Edition

If you know someone who uses a mouse all day but has never really thought about the actual mouse they’re using, the MX Master 4 might just change their life. The MX Master series has long been considered the leading productivity mouse due to its unique free-spinning scroll wheel, and the latest model adds a customizable haptic thumb panel that buzzes to confirm a shortcut. The new Developer Edition adds a “Hello, World” prompt and a reference to the Linux kernel’s release date.

BOOX Palma 3

The Boox Palma 3 is an E Ink reader the size and shape of a smartphone, offering a convenient antidote to doomscrolling that you can take anywhere. Because it runs on Android, you can easily add your Kindle and Kobo libraries straight from the Google Play store. This latest version features a streamlined aluminum design and stylus support.

8BitDo Retro Mechanical Keyboard

8BitDo is best known for its excellent range of gaming controllers, and its first mechanical keyboard turned out to be just as haptically satisfying. Launched in versions that evoked the Nintendo Entertainment System and the Famicom, the lineup now includes models inspired by the Commodore 64 and the IBM Model M.. Each comes with enormous programmable buttons you can place next to the keyboard, making copying and pasting feel like a session at the arcade.

Balmuda The Lantern

Balmuda is a hip Japanese appliance maker best known for its minimalist steam toaster, but its equally obsessive take on the camping lantern may be a more practical gift. It has a single rotary dial that takes you from a warm, flickering candle-style effect at the low end to a brighter reading light when you turn it up. The light is diffused through a pipe for a soft glow, and its rechargeable battery means you can use it anywhere from the bedroom to the balcony.

Analogue Pocket

The Analogue Pocket feels like it came from an alternate universe where Apple designed the Game Boy. It has a gorgeous stripped-down design, an amazing high-resolution screen, and programmable hardware that has been tuned to provide the most authentic reproduction possible of Game Boy and Game Boy Advance games. The newer Analogue 3D does much the same thing for the Nintendo 64, if your gift recipient is more of a couch potato.

Lego Game Boy

Lego’s Game Boy is another meticulous reproduction of the Game Boy, but from a totally different angle. This kit is incredibly fun to build, making you build up the hardware from the circuit board level and work on the mechanisms of the satisfyingly squishy buttons and D-pad. As a display piece, you can swap out cartridges and lenticular screen inserts for Super Mario Land and The Legend of Zelda: Link’s Awakening.

Backbone One

The BackBone One has been around for a while, but together with its Pro version, it still remains the top name in smartphone game controllers. And the proposition is more compelling than it used to be, with modern phones now powerful enough to run solid versions of high-end console games like Assassin’s Creed Mirage and Death Stranding. Backbone offers models with both PlayStation and Xbox controller layouts, depending on your preference.

Aura Carver Mat

Digital photo frames have been a clichéd throwaway gift for a long time, which is why Aura’s products stand out. The Carver Mat is a landscape, color-calibrated matte screen with a paper-textured border that automatically dims its lighting to match the ambient surroundings, making it feel much more natural than other digital frames. It can display two portrait photos side-by-side and even supports Live Photos from an iPhone. 

Kodak Charmera

The Kodak Charmera hits on two recent trends: Labubu-style blind-box collectibles and the throwback use of lo-fi digicams. Each $35 box contains one of seven retro designs, and the tiny camera itself has several filters and frames. The image quality is not great, of course, but the Charmera is well-designed and well-named—it’s the perfect stocking stuffer.

Fujifilm Instax Pal 2

If you like the idea of a tiny pocket camera but want the results to be a little more usable, Fujifilm has you covered with the $169 Instax Pal 2. It’s shaped like a vintage range finder, like many of Fujifilm’s more advanced mirrorless cameras—but even though it fits in the palm of your hand, it makes room for an optical viewfinder, a 1-inch display, a command dial for color effects, and a lever to swap between aspect ratios. Photos can be transferred directly to your phone or to an Instax Link printer.

Peak Design Mobile Tripod

The biggest problem with tripods is finding a way to take them with you. But here’s one you won’t even notice in your jeans pocket. Peak Design’s take on the mobile tripod folds three legs and a full-on ball head into a package that’s about 5 millimeters thick. It snaps onto your MagSafe-compatible phone with minimal obtrusion, and unfolds into a stand you can angle for group shots, video calls, or hyper-lapses.

Native Union Pop cables and phone

I’ve long been a fan of Paris-based Native Union, which puts out some of the most thoughtfully designed tech accessories around. Its coiled Pop Cables are particularly great, avoiding tangles and saving desk space—and now the company has revived its original Pop Phone, giving you an old-school handset to conduct your Zoom calls like you’re in an episode of Mad Men. There’s even a hang-up button.

Xiaomi Portable Electric Air Compressor

If you know anyone who rides a bicycle and plays basketball, this gadget might just change their life. For everyone else, it’s still worth keeping around—I’ve recommended it to countless friends. This USB-C-powered air compressor has modes tailored to all sorts of devices and displays the real-time pressure rating on its LCD screen.

Nintendo Sound Clock: Alarmo

Trust Nintendo to come up with a bizarrely useful take on something as mundane as the alarm clock. Anyone who shares a bed with a partner may consider it a hard sell to start each morning with sounds from Mario Kart or The Legend of Zelda. But the motion-sensing feature alone is a game changer for the groggy moments before getting up; the screen stays switched off until you turn your head to check the time.

Teenage Engineering Pocket Operator

Swedish design company Teenage Engineering first made its name through its iconoclastic synthesizers, and it sounds like the company might be moving on from that heritage. If that’s the case, now could be a good time to stock up on its $99 Pocket Operator series. It’s a range of exposed circuit boards with clicky switches, rotary dials, and monochrome LCD panels, all amounting to remarkably capable instruments.

Yukai Engineering Nekojita FuFu

In Japan, if you’re sensitive to hot food or liquids, you’re said to have a nekojita—in other words, a cat’s tongue. Yukai Engineering is attempting to solve this condition with a tiny, rechargeable cat that attaches to the rim of your coffee cup or ramen bowl. Its internal fan drops the liquid’s temperature to a drinkable level in just a few minutes, and the cat looks cute in the process.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/12\/multicore.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/12\/multicore-mobile.png","eyebrow":"","headline":"\u003Cstrong\u003E\u003Cstrong\u003ESubscribe to \u003Cem\u003EMulticore\u003C\/em\u003E\u003C\/strong\u003E\u003C\/strong\u003E","dek":"\u003Cem\u003EMulticore\u003C\/em\u003E is about technology hardware and design. It’s written from Tokyo by Sam Byford. To learn more, visit \u003Ca href=\u0022https:\/\/www.multicore.blog\/\u0022\u003Emulticore.blog\u003C\/a\u003E","subhed":"","description":"","ctaText":"SIGN UP","ctaUrl":"https:\/\/www.multicore.blog\/","theme":{"bg":"#f5f5f5","text":"#000000","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#000000","buttonHoverBg":"#3b3f46","buttonText":"#ffffff"},"imageDesktopId":91454027,"imageMobileId":91454030,"shareable":false,"slug":"","wpCssClasses":""}}

The FAA’s ground stop has a simple culprit. Its causes may not

1 day 14 hours ago

On Friday afternoon, L3Harris Technologies was doing something it has done at Federal Aviation Administration facilities month after month: moving a communications line off decades-old copper and onto fiber, work that federal auditors count as part of a $12.5 billion overhaul of the country’s air traffic control system.

Shortly after a technician installed a new router at an FAA radio site in El Dorado, Arkansas, the device began pumping out data—so much that the flood disrupted communications at 17 of the 22 so-called en route centers that handle high-altitude traffic.

At 3:36 p.m. Eastern, with about 3,500 planes in the air, the FAA held all departures nationwide. The stop lasted more than four hours, until 8:15 p.m. Airlines canceled more than 5,000 flights by Saturday and about 1,500 more on Sunday as they repositioned crews and aircraft. It was the first nationwide ground stop since January 2023, when the system that sends safety notices to pilots failed.

“We haven’t determined what the fault was,” Transportation Secretary Sean Duffy said at a Saturday news conference. Investigators are trying to establish whether L3Harris installed or configured the router incorrectly, or whether the equipment itself was at fault. Adam Bauer, a spokesperson for Hewlett Packard Enterprise, which owns router maker Juniper Networks, told Fast Company its devices appeared to have worked as intended. “Based on what we know and have been told by others, our equipment was working as configured.”

But the failure exposed weaknesses the FAA has been warned about for years: a contractor that has run the FAA’s network since 2002 and has a record of outages, an integrator the agency’s own staff rated a financial risk, a technician workforce the FAA admits is understaffed, and a $12.5 billion overhaul that auditors say is racing ahead without a master schedule. There is no evidence yet that any of them caused Friday’s failure.

Still, some critics see Friday’s chaos as years of warnings come true. “The sad fact is this has happened before,” Mary Schiavo, a former Transportation Department inspector general, told NPR’s Weekend Edition.

The sprawling network of copper and fiber that carries the country’s air traffic communications across roughly 4,400 FAA sites still runs largely on technology that dates back decades. Some older equipment, including radar, requires replacement parts that can only be obtained through online auction websites or made with 3D printers. In 2023, an FAA-commissioned panel of outside experts warned that the condition of the agency’s facilities and equipment was “elevating system risk to unsustainable levels.” The agency under President Joe Biden had commissioned another group of aviation experts to examine air traffic control in October 2024, but weeks after Trump took office, the group learned it was being disbanded, The Washington Post reported. 

Last May, Duffy said the FAA was speeding up its overhaul of the system under a program called BNATCS, or the Brand New Air Traffic Control System. Congress approved $12.5 billion for the effort, which aims to modernize the telecom network by the end of 2028, before the end of Trump’s term, and 21 months ahead of an earlier September 2030 target.

In September, the Government Accountability Office released its latest report on the overhaul. The FAA was moving at an unprecedented speed, auditors said, but without a reliable lifecycle cost estimate or an integrated master schedule. Instead, auditors counted 11,389 separate project schedules that had not been coordinated, so different programs were booked to visit the same sites repeatedly or in ways that risk “excessive disruptions to air traffic control operations” and could jeopardize the FAA’s “aggressive deployment schedule.” Delays, in turn, could leave controllers reliant on outdated systems that are prone to failure, and “lead to more system outages, ground stops, flight delays, and safety incidents.”

The GAO auditors also found the agency had exempted the effort from nearly all of its standard acquisition rules. Schedule pressure on federal programs can lead teams to skip or compress testing, they warned. “In essence, the FAA is buying delivery speed at the risk of cost uncertainty and potential disruptions,” the GAO wrote.

The incident has also drawn attention to Peraton, the Herndon, Va.-based defense contractor hired last December to manage the broader overhaul, under a contract with a $1.5 billion ceiling. Politico reported in June that one lobbyist—a former Duffy spokesperson who sits on the board of a nonprofit linked to the Transportation Secretary—lobbied for Peraton in the period when it won its FAA contract. (The FAA has said that Duffy played no role in the company’s selection.)

By August, FAA officials were frustrated with Peraton’s performance, according to a September Politico report, citing people familiar with the matter who described slipping schedules and a contractor that was not managing the project. Peraton was “not sophisticated with how the FAA works” and was being asked to sprint from the start, one person told Politico. The same person said the FAA was partly to blame, accusing the agency of blocking Peraton’s efforts to work with its own staff.

As the FAA began its investigation into the cause of the router issues on Saturday, Schiavo told NPR that uncoordinated schedules and a rush to patch the system had made such a failure predictable. “What did the GAO conclude? That it will result in air traffic outages and disruptions. And there we are.”

The FAA and L3Harris take different approaches to scheduling 

The trouble first showed up as interruptions at the FAA’s Indianapolis center, as a surge of data flooded the network’s older routers. A flurry of erratic radar feeds and communication drop-offs with aircraft led officials to order a ground stop. Soon troubleshooters traced the issue to the El Dorado router and to a second one at a radar site in Amarillo, Texas; when both were cut off from the network, the problems stopped. Officials have zeroed in on a feature on the new Juniper routers that broadcast data more widely than necessary. The FAA has since switched the feature off on all of its Juniper routers while the two suspect routers are being tested at L3 Harris’s offices in Florida, with help from Juniper and Cisco engineers and the FBI. 

“Was it a faulty installation? Was it a faulty configuration?” FAA Administrator Bryan Bedford asked on Saturday, adding that the agency would have to rethink how it carries out upgrades.

Asked why the work was done on a busy Friday afternoon, Duffy cited the urgency of the upgrades and said the site was far from major flight corridors. But he added: “there will be a reevaluation of when we do this work.” 

The timing contrasts with how L3Harris describes its own method. In an interview published in March, Chris Collings, the L3Harris vice president who runs the business, said the company never switches network segments off. At critical sites with two paths, he said, crews upgrade one path, validate it, then upgrade the other, during low-traffic windows. The network is designed to contain a local fault, so why one remote radio site could disrupt most of the country’s en route centers remains unexplained.

Duffy repeatedly blamed the contractor. “We’re going to hold people accountable,” he said, adding that many contractors “take our money and provide really poor service.” He also told reporters he was “stuck with L3Harris.” 

In a statement, L3Harris said it was supporting the agency’s investigation, and that it “takes the nationwide ground stop and its impact on travelers, airlines and air traffic operations seriously.”

The FAA said that one of its representatives helped an L3Harris technician gain access to an agency site on Friday. The union for FAA technicians, however, said its members had no part in the work. Historically, it said in a statement Saturday, contractors have relied on FAA employees to assess risk and coordinate maintenance, but its “technicians were not involved in the upgrades that led to the ground stop.” 

The FAA has acknowledged that its technician workforce is understaffed and released a workforce plan in May 2026 calling for an increase of at least 750 more certified technicians by 2030. “However, little progress has been made on hiring those new technicians,” the union said. The FAA and L3Harris did not respond to a series of emailed questions for this story.

A push to shift from L3Harris to AT&T

L3Harris, based in Melbourne, Fla., specializes in defense, but it has run the FAA Telecommunications Infrastructure network, known as FTI, since its predecessor Harris Corp. won the contract in 2002. The company says it has already converted 70% of the copper-to-fiber conversion. 

The agency has had warnings about the network for years, including a failure strikingly like Friday’s. On Nov. 19, 2009, an FTI outage hit 129 air traffic facilities. The inspector general concluded that configuration and procedural errors by Harris caused the outage and delayed restoring service. A Harris engineer had put a new router into service in Los Angeles that carried a configuration error. Even after fixes, the inspector general noted, FAA and Harris officials acknowledged a continuing risk of critical outages.

The FAA has since faulted the company publicly. After a September 2025 telecom outage at the Dallas approach-control facility halted flights at both of the city’s major airports for hours, the agency blamed failures in service from local carrier Frontier Communications. It added: “Oversight by L3Harris, an FAA contractor, failed to ensure that redundancies in the system functioned properly.” L3Harris responded in a statement at the time that it and the carriers had warned the FAA for years that the circuits needed replacing, but the money never came. “Now, under the leadership of Secretary of Transportation Sean Duffy, L3Harris is on contract to replace these legacy analog circuits (which total over 5,000) with new modern circuits.”

This September, a primary circuit serving the Philadelphia approach-control facility failed, and after L3Harris advised switching to a backup fiber line, which turned out to have been cut by a construction crew in New Jersey. About 9,500 U.S. flights were delayed or canceled. “When it fails, we just don’t have the redundancy” of a modern architecture, Bedford said at the time.

The FAA is already moving to replace FTI. In August, using waiver authority tied to the modernization program, it made a single, undefinitized award to AT&T to lead the FAA Enterprise Network Services program, FTI’s successor. The award started with a $74.3 million task order for what the FAA expects to be a multibillion-dollar contract lasting up to 15 years. In September, Bedford called for an additional $10.5 billion from Congress for the project. On Saturday, Duffy said that Verizon and Lumen were also “ready to go” on the upgrade. Until that transition is done, L3Harris remains the operator, and the contractor converting the FAA’s copper lines.

The FAA network is a small part of a company that has become more central to the administration’s priorities. L3Harris, which reported $21.9 billion in revenue last year, converted a Boeing 747 donated by Qatar into an interim Air Force One this year. It worked against a July 4 deadline that, Democratic senators say, forced it to scale back the usual upgrade package. (The Air Force has said the work would cost less than $400 million; the contract is classified.) 

In January, the Pentagon agreed to invest $1 billion in L3Harris’s missile business, which the company plans to spin off—a rare direct federal stake in a major defense supplier. And in July, the Space Force awarded L3Harris up to $955 million to build 18 missile-defense satellites for the Golden Dome shield, bringing its orders for U.S. missile-tracking and defense satellites to at least 70.

What the GAO found about the FAA upgrade

The GAO’s September report examined the FAA’s upgrade. Of the $12.5 billion Congress allocated, $9 billion is for systems modernization. Telecommunications is one of 13 programs in the overhaul’s first phase: replacing 5,170 copper lines with fiber and moving nearly 30,000 services to Internet Protocol. GAO figures show the program’s estimated cost rising from about $4.75 billion to about $5.91 billion. 

The auditors’ main findings:

  • There’s no reliable cost estimate. In June the FAA put phase one’s systems work at $10.6 billion, but officials could not produce the analysis behind the figure, telling auditors it would take a tremendous effort to find. The estimate omits government costs, most operating and maintenance costs and all of phase two. The FAA says it needs another $10.8 billion.
  • There’s no master schedule. Peraton, the prime systems integrator, delivered a preliminary integrated master schedule in April. As of May, though, the 11,389 project schedules had not been integrated, which risked inefficiencies and conflicts with airport operations. Officials said loading them all into their scheduling tool “would not be useful or helpful.” The FAA planned to use an artificial intelligence tool to flag conflicts, but as of May it was still being tested and rules for using it had not been set.
  • Rules were waived. In August 2025, Bedford issued a blanket waiver exempting the program from nearly all of the FAA’s acquisition management requirements, including the one that would require a lifecycle cost estimate. The GAO called that a key reason no such estimate exists.

The Transportation Department partly agreed with the GAO’s two recommendations: build a reliable lifecycle cost estimate and a detailed integrated master schedule. In its reply, the department said 12 of 13 workstreams, covering over 95 percent of project tasks, were on plan as of July. According to the FAA’s public Modern Skies dashboard, 11 out of 12 programs are at or ahead of plan. The exception is the replacement of analog protocol converters, devices that turn analog signals into digital ones, which stands at 8 percent complete, with just 45 of 69 planned sites completed in the most recent month recorded. 

The FAA has said it will answer the GAO’s September recommendations in detail within 180 days of the report, a deadline that falls in March.

Whose upgrade was it?

Bedford tried to separate Friday’s failure from the modernization program. “We’re conflating modernization with this upgrade at this particular facility. We spend $1 billion a year maintaining and sustaining the existing 1980s [legacy] telco network,” Bedford said. “This is not part of the Peraton-FAA $12.5 billion modernization effort.”

That contrasts with the GAO’s account, which lists the copper-to-fiber program as part of phase one of the modernization. And according to the GAO, existing program contracts were modified to give Peraton authority to manage the other contractors on the FAA’s behalf. “Peraton, think of them as an extension of me,” Bedford said in April. 

Peraton was seen as an underdog for the FAA job, in part because it had little prior experience working with the agency. The company itself grew out of the government IT services unit of Harris Corp., L3Harris’ predecessor, which private equity firm Veritas Capital bought and rebranded as Peraton in 2017. 

Before the FAA award, agency staff had rated the company a high risk for financial instability in an October 2025 presentation to Duffy, Bloomberg reported in September. When agency staff evaluated the bids last year, they graded Peraton and Parsons, which bid alongside IBM, on four technical criteria, according to the presentation. Parsons rated better on three and the same on the fourth. Staff wrote that Peraton’s answers to use cases and direct questions “often focused on matters irrelevant to the discussion topic.”

Peraton scored better on past performance, helped by a positive review from NASA. The FAA told Bloomberg that strong ratings from Peraton’s customers, including NASA, were among its reasons for the choice, along with what it called the company’s significant telecommunications experience. The presentation also showed that Peraton sought a higher award fee than Parsons; it is unclear whether that changed later in the process.

Politico reported in June that a former Duffy spokesperson, Mark Bednar, who now sits on the board of the nonprofit that paid for Duffy’s road-trip video series, Great American Road Trip Inc., had lobbied for Peraton in the period when it won its FAA contract. At a December 2025 congressional hearing, Bedford said career FAA staff ran the selection without interference from Duffy or Trump. The recommendation went from Bedford to Duffy and then to the president, he said, and Trump demanded a $200 million cut in the fee.

Concerns about Peraton’s work have since dogged the program. A few months ago, the FAA called Peraton in and told it to improve its performance, and the company blamed the agency in return, Bloomberg reported, citing a person familiar with the matter. The FAA has said it hasn’t penalized Peraton or issued the company a formal warning about poor performance.

A Peraton spokesperson told Bloomberg in August that the program is on track. But the company said the schedule is among the largest and most complex in government history, covering 330,000 independent projects and half a million dependencies. Peraton did not respond to an inquiry for this story. 

Duffy used the outage to renew and even scale up his call for Congress to fund the larger telecom modernization effort, arguing the $12.5 billion already allocated is not enough. For weeks, he has been calling for $30 billion in additional funds—with only about $10 billion of that request specifically for air traffic modernization—but during the outage he raised the demand to $50 billion. Bedford said the FAA would review its upgrade protocols. Neither said whether copper-to-fiber cutovers would be paused while investigators work out why and how one router could halt the country’s air traffic.

Representative James Walkinshaw (D-Va.), a member of the House Oversight Committee, said on Saturday that Duffy must be questioned by Congress, including about why the work was being conducted “on a busy travel day, as opposed to in the middle of the night where a disruption wouldn’t have been so damaging because they’re rushing.”

Ahead of its IPO filing, Anthropic pledges to prioritize humanity’s benefit. Here’s how

1 day 19 hours ago

When Anthropic goes public on Wall Street, the new pressure to deliver profits will test a commitment it has made since its founding: to prioritize humanity’s benefit as it develops artificial intelligence.

The San Francisco startup, which plans an initial public offering this month, has sketched out a governance structure that it says will protect the company’s broader mission against the pull of financial interests. But the mechanisms it describes give the company wide leeway to decide how to measure whether it’s benefiting society, experts say.

Like its rival OpenAI, Anthropic is a public benefit corporation, a legal designation used by companies that pledge to balance shareholder interests with the public good. How that will play out for companies worth potentially trillions of dollars is uncharted territory.

“I think then the real test emerges if and when the question of public benefit becomes costly” — or when the public benefit comes into conflict with making a profit, said Sarah Kreps, a professor and director of Cornell University’s Tech Policy Institute.

In its regulatory filings, Anthropic pledges to pursue long-term benefits of humanity through the responsible development of advanced AI. “We strive to help bring humanity safely into an era where transformative AI exerts tremendous and potentially decisive influence over a broad range of human undertakings,” the company said. Anthropic also has pledged at least 5% of its stock to a “public benefit” fund and its founders have committed to donate 80% of their wealth.

A pledge to do good is baked into corporate structure

Other tech companies have pledged to do good including Google, which prominently featured its then-motto, “Don’t Be Evil,” in its investor prospectus when it went public in 2004. But the words were not legally binding, and critics have accused the tech giant of leaving them behind.

Elon Musk’s takeover of Twitter in 2021 shows how the conflict between profit and public good might play out in real life, said Stavros Gadinis, a law professor at UC Berkeley School of Law.

“Let’s say that Twitter’s directors felt that Elon was the wrong guy to police the public square because they didn’t trust him, they didn’t like him, they didn’t think he’d be fair, whatever it was,” Gadinis said. “But he made them an offer that was so high compared to what their advisors told them the company was really worth that it would be against their fiduciary duties to not sell to Elon. This is the situation that the public benefit charter is trying to avoid.”

Had Twitter been a public benefit corporation, it could have been required to consider factors beyond maximizing shareholder value when selling the company.

There aren’t many public benefit corporations whose shares are publicly traded and they are all much smaller than Anthropic and OpenAI. They include the eyewear company Warby Parker, the insurance company Lemonade and software developer Veeva Systems, which had a conventional corporate structure when it went public in 2013 and converted to a public benefit corporation in 2021.

Anthropic’s idealistic ambitions for how it develops AI could run up against competitive pressure, such as from China, Kreps said.

Anthropic and OpenAI’s models have gone from answering almost any question, even if it was dangerous, to becoming much more circumspect. But as they face more competition, including from open-source Chinese models that likely will tell you whatever you want, she said, “will they start to rethink some of those guardrails? And what happens then?”

The setup generally is not preferred by shareholders because it gives more power to the company’s board than it does to shareholders.

Whatever a company decides as its public benefit mission, it will likely be much more difficult to measure than profits. That’s particularly true for AI companies, whose leaders have exalted AI as the solution to many of humanity’s problems — from curing cancer to creating economic prosperity.

“There are these echo chambers, the Silicon Valley bubble where they really think that what they’re doing is for the good of humanity,” Kreps said. And someone immersed in that environment, she added, might not even ask the right questions or understand how “public benefit” should be measured.

A trust aims to hold Anthropic accountable

A Long-Term Benefit Trust created by Anthropic is intended to operate as an independent body of trustees with expertise in AI safety, national security, public policy, and social enterprise.

Through a special Class T common stock, the trust “generally has the authority to elect and remove” Class T directors, who together will constitute a majority of Anthropic’s board of directors, according to the IPO filing. The Class T common stock is a single share that allows the five trustees to vote as a unit. It has no economic rights — meaning they get to vote but they don’t see any financial benefit or losses from the share and can’t sell it.

“The LTBT’s arrangements are designed to minimize the LTBT’s and the Trustees’ exposure to our financial performance and grant the Trustees sufficient independence to ensure we are striking this balance appropriately,” the filing said.

Three of the company’s board directors are elected by the trust directly, and another three are elected by other shareholders, which include the company founders with a 50.1% voting majority. A seventh director, with a potentially tie-breaking vote, is nominated by the trust and the holders of the Class T shares, but cannot be appointed or removed without consent from the board’s nominating committee, which is expected to be controlled by the founders.

Effectively, the trust can only hold Anthropic accountable on issues such as AI safety and the company’s broader mission up to a point, Gadinis said.

The trust “cannot determine the everyday decisions of these directors or the management of the company,” he said. “If they disagree, they will have to go through the effort of removing directors, and they may not be able to remove a majority.”

As a public company, Anthropic also will be required to report finances and generally be more transparent than a privately held one. Many public companies dealing with AI, including Google and Microsoft, regularly update the public on efforts to develop AI responsibly.

The incentive for transparency could be as powerful as other mechanisms set up in Anthropic’s corporate structure, Gadinis said.

“I do think that by being public, we are going to get a lot more information, because a lot of people are going to be asking questions, and companies will be responding to these questions. not only because they’re required by security laws, but also because they want to build this kind of trust with the investing community,” he said.

—Barbara Ortutay, AP Technology Writer

What is the FAA’s SMART AI system? Officials are denying a link to the nationwide ground stop

1 day 21 hours ago

The airline ground stop that wreaked havoc on travel since Friday was not related to the FAA’s new, $875 million AI-powered system the agency is rolling out to help predict problems in the sky, according to government officials and the artificial intelligence company behind the technology.

But the timing of the ground stop, so close to the national rollout of the Federal Aviation Administration’s SMART system, has raised questions about what the system does, its benefits and risks compared with the existing human system and what guardrails there are to address potential problems.

Officials say no connection to ground stop

The SMART system is based in a single location near Washington where a team of experts analyzes flight schedules, weather forecasts and hundreds of other streams of data. Then that office sends recommendations to the FAA’s command center, which relays them to air traffic controllers across the country for consideration.

Homeland Security Secretary Markwayne Mullin said Friday there’s no indication the communications issues that caused the ground stop are linked to the new computer system, and the top officials at the Transportation Department and FAA reiterated that Saturday.

The startup that designed the system, Air Space Intelligence, or ASI, said the ground stop wasn’t caused by “the SMART system, any product of ASI or our integration with FAA and ATC.”

But lawmakers and outside groups raised questions about the SMART system before Friday’s ground stop. Last month, House Democrats sent a letter to the U.S. Government Accountability Office requesting an audit of the new software, raising concerns about safety and guardrails.

“Concerningly, the nation’s front line aviation safety workforce apparently had little to no input into SMART’s development and have received limited training and information during the current deployment process,” the Oct. 13 letter said. “The use of artificial intelligence within SMART has rightfully raised serious questions about its reliability and safety benefits, particularly on its potential to interfere with decision-making by experienced air traffic controllers or to hallucinate safety-critical information.”

SMART, however, is not — at least not yet — set up for AI to be in charge of flight control. Rather, it is making recommendations to humans who do so.

What is the SMART system?

Transportation Secretary Sean Duffy and FAA Administrator Bryan Bedford unveiled the system in September, calling it a “cutting-edge tool that centralizes 200 data streams — including weather patterns, flight paths, traffic flow, and controller staffing metrics — into one platform.” SMART stands for Strategic Management of Airspace, Routes and Trajectories.

The system’s AI-supported engine synthesizes the data systems and shows a “comprehensive visualization” of where planes are going, how much traffic the system can handle and where congestion or weather could cause problems, according to the announcement. The idea is for specialists to then use the platform to fix problems before they arise.

The FAA says that during the initial test of the system in the Washington area this fall, the system helped reroute 180 flights and saved an average of 10 minutes and 80 gallons of fuel per flight by diverting around weather or other issues. The agency said that one flight cut 42 minutes off its trip from Tampa, Florida, to Dallas thanks to the system’s recommendations.

So SMART was expanded for use nationwide on Nov. 10 for the airspace of all 22 air traffic control centers. That was three days before the ground stop.

According to a government fact sheet on the tool, SMART can predict where air traffic is headed and identify other routes before an area becomes too congested. The system can also help reopen restricted airspace faster and better manage airport arrivals.

SMART is also supposed to help with weather-related issues faster.

Currently, the FAA blocks off portions of airspace affected by weather as “large, rectangular boxes,” according to the fact sheet, which doesn’t reflect how weather actually looks, with quick-moving thunderstorms and storm systems that are “often irregular in shape, with narrow gaps or corridors that aircraft could safely fly through.”

To help with this, SMART gives the FAA and airspace operators a “clearer picture of how weather will affect air traffic over time,” allowing air traffic controllers to plan around storms and use available airspace instead of unnecessarily shutting down large areas.

No other country has a working AI air traffic tool like the SMART system, according to Michael McCormick, a former FAA airspace manager and an associate professor at Embry-Riddle Aeronautical University. Eurocontrol, the European air traffic authority, has a system in testing and several manufacturers are prototyping similar tools.

Concerns around relying on AI

While government officials and ASI, the company behind SMART, have stressed that its AI system is designed to assist humans and not run flight control, its deployment has raised concerns about relying on artificial intelligence when human lives are at stake, given that even the most advanced models are prone to hallucinations — when models fabricate false information and present it as fact.

Mike O’Donnell, a former top FAA official who worked for the agency for 17 years and now does aviation safety consulting work, said the agency “has got to make sure that this is just a tool in the toolbox and not the actual machine that’s doing the work.”

The FAA, he said, “cannot afford to be wrong on this.”

“It’s good to use it with the proper guardrails and backstops to ensure that they’re not giving wrong information to controllers based on what (AI) has come up with,” O’Donnell said. “There’s got to be a way to verify and validate what it’s doing to make sure that it’s not the wrong information.”

He cautioned against putting too much trust in AI “until we know for sure that it’s proven itself time and time again. And that’s going to take a while.”

The AI system’s launch comes amid a longstanding shortage of air traffic controllers around the country. The National Air Traffic Controllers Association has been raising concerns about this for more than a decade, telling Congress in 2015 that the shortage was at a crisis level. It said then that the FAA had missed its hiring targets for five years in a row. The union represents about 14,000 controllers, roughly 3,000 of them trainees. That said, experts say no AI system can replace the human air traffic controllers who spend years training for this highly demanding and stressful job.

In an October interview with NPR, ASI co-founder and CEO Phillip Buckendorf said the SMART tool is designed for traffic managers who manage the entire U.S. airspace.

“The AI does the math, the AI does the computation,” Buckendorf told NPR, but it’s ultimately up to the human operator whether to take the system’s recommendation. “They make the decision, they are the decider, they are the judge.”

—Barbara Ortutay and Josh Funk, Associated Press

The solution to the data center energy crisis might be in your home

2 days 1 hour ago

What if the solution to the energy crunch exacerbated by data center expansion wasn’t building more power plants—and investing in gas turbines and dirty fossil fuel power—but instead, simply upgrading people’s homes? 

A number of advocates and clean energy wonks believe there’s another way forward. By creating choreographed networks of batteries and smart devices, called virtual power plants (VPPs), utilities can lower power usage and create a well of excess power they can unleash during periods of high demand. Batteries can store energy during off-peak hours to deploy later, and smart thermostats can cut air conditioning by a few degrees to moderate periods of high usage. All could be paid for by leveraging the billions of dollars being spent by the energy-hungry data center sector.

“This is a bridging technology, one that can help remake the American energy system,” says Chris Walker, vice president of power and platforms for Grid Alternatives. It’s one of the nation’s largest nonprofit clean energy providers, and it’s working on pilots seeking to prove this approach can work at scale. 

“We’re hearing more and more sort of advocacy for this approach, not least because you know the data center developers really need social license, but because communities are demanding actual benefits from this,” he adds.

Home renovations may seem like an inadequate solution to the massive, and growing, energy footprint of data centers. The demand is unprecedented; Goldman Sachs predicted that data center energy use would more than double from 31 gigawatts in 2025 to 66 gigawatts by 2027. (For comparison, 66 gigawatts is enough to power more than 50 million homes.) Increasing power generation by constructing new gas power plants is costly, both financially and environmentally, and takes time; gas turbine deliveries are backed up until 2028. It’s also broadly unpopular. Americans widely rejected data center development during the 2026 midterm election cycle, when energy costs were a constant refrain.

But by using the money set to be spent on the data center energy build-out, and instead electrifying homes, adding batteries and installing solar, advocates argue it’s possible to create a distributed power system. This web of decentralized generating and power storage sites can cut energy bills, quickly create excess grid capacity that can be utilized by data centers, and build a better electrical system. It’s also much quicker to get running. These virtual power plants can be stood up in months, and at less cost.

“You’re setting up a win-win solution where a hyperscaler is able to make investments in communities that make their lives better, and they’re able to get something valuable to them in return, which is speed to power,” says Annie Gilleo, vice president of the Ad Hoc Group, an energy industry consultancy. “If we’re aligning interests with the hyperscaler, the sky is the limit.” 

The virtual power plant experiment

A number of pilots and programs are testing how to dispatch and measure distributed power systems. When proven and scaled up, this is the kind of technology that may be able to support an additional data center without requiring new power generation. 

In April in Minnesota, the state’s public utilities commission approved a $430 million utility-owned, battery-based virtual power plant, called the  Capacity*Connect program. Xcel Energy, the local utility, will work with Sparkfund, a utility services company, to deploy 200 megawatts of small batteries across homes, local businesses, and nonprofits. These will help store energy to be used during periods of high demand. 

In Virginia, Google is partnering with Voltus, a distributed energy platform, to create a 1 gigawatt virtual power plant. Announced in June and recently expanded, this system will consist of a network of flexible power assets—batteries, smart thermostats, even adjustable industrial processes at factories—that can be tapped into during times of high demand. When the grid is under strain, Voltus sends signals to those assets to either release stored power or cut usage, freeing up capacity. The system is expected to save consumers energy, and it’s being tested in the PJM grid, which runs through the nation’s Middle Atlantic region and into the Midwest. PJM’s coverage area includes Virginia, a data center hotbed. 

“We really see virtual power plants as a critical short-term solution,” Ellen Zuckerman, head of energy market development for North and South America at Google, told The New York Times. “It’s a way to invest in our communities, really double down on our clean energy ambition, but also meet capacity at speed and scale.”

And in September, Pacific Gas & Electric announced it was working with Google and Rewiring America, a home electrification nonprofit, to enroll tens of thousands of customers across its Northern California user base—and their home batteries—in a program called SHARE (Smart Home Assets for Reliability and Efficiency). Proponents argue this can cut the energy bills of participating households and create a reliable grid resource. 

In addition, Grid Alternatives is collaborating with Google in Fort Wayne, Indiana, where Google operates the massive, $2 billion Project Zodiac data center. Consisting of 12 buildings spread across 700 acres, the center continues to draw community pushback and protests over its local and environmental impacts.

The partnership is funding energy retrofits, battery storage, and solar power for roughly a dozen homes, which should all be upgraded by the end of the year. A ribbon cutting will take place on December 1, says Walker, Grid Alternatives’ power and platforms VP. It’s a very small pilot aiming to gather data and prove a larger project can be possible.

“This is 12 homes, but we want to talk about 1,200 houses,” he says.

Unlocking power that’s already there

Research shows that efficiency and grid flexibility may be potent allies in meeting growing demand and even saving money. A recent study by the Brattle Group, an economic consultancy, found that improving system utilization by 10% could cut electricity rates by roughly 3.4%; that would add up to $100 billion in a decade. VPPs also cost much less to operate.

Many such systems are already in use. The U.S. Department of Energy estimated that about roughly 30 gigawatts of VPP capacity is deployed today, and the nation could hit 160 gigawatts by 2030. Instead of spending money to add generating capacity to the grid, and racing to do so with dirty sources of power, why not increase efficiency and put the power already being generated to better use?

“We’re not going to be able to build power plants fast enough to serve demand,” says McGee Young, founder and CEO of WattCarbon, a marketplace that tracks demand-side energy resources. “How we fund and finance the deployment of distributed energy resources over the next 10 to 15 years is literally the trillion dollar question.”

Increasingly, there are policy and regulations that make this possible, says Gilleo of Ad Hoc Group. In addition, other groups have already started laying out a roadmap for where these programs could work: GRID Alternatives has built its Distributed Community Power Program on the back of work it did for the now-cancelled federal Solar for All program. Grid worked with Accenture, the technology consultancy, which created tools, including a searchable database of opportunities for distributed solar and other investments that can be deployed across Grid’s coalition of numerous partners and housing providers. The database can also help data center developers identify investable community-benefiting energy projects to support, including ones that could relieve grid congestion near their facilities.

The challenge of making it work at scale

The problem remains figuring out how to carry out the transfer of capacity. Does it work during energy peaks, and is it certain that there’s a balance of energy savings and utilization? Often, high power demand gets concentrated on a specific power line; are there enough other customers or energy usage on that line that can be freed up through these investments to feed data centers during peak usage? Gilleo was especially excited about the way the SHARE program created by PG&E, Google, and Rewiring America will study how these programs work, and within a year, provide data and insights into how to scale up the concept.

These may, at first, seem like simple shifts to the energy system. But effectively tracking the demand and usage across a real-time grid and accurately making sure everything is balanced requires extensive testing, regulations, and proof. Young at WattCarbon says that it takes a lot of convincing to get utilities, data center developers, and Big Tech to sign off on something like this.

“The lights are going to stay on because a whole bunch of random people are putting on solar panels and batteries and heat pumps,” he says. “That’s a terrifying idea.” 

Right now, as the tech industry formulates its response to the midterms and the drubbing felt by data center advocates, there is a renewed push for more community benefits. In October, Amazon announced its Built Together framework, which put $1 billion into benefits over the next five years, including energy affordability. Advocates want to see that money go toward more pilots and distributed energy projects. 

Gilleo hopes that soon, as distributed energy trials and pilots prove themselves, more of that kind of community funding gets diverted towards energy efficiency and a more responsive grid, not merely community investments. Research shows this kind of program and messaging work; a survey by the energy advocacy group Good Power found that two of the most effective promises data center developers can make are pledging to cut utility costs and use renewable power. 

 “I do think that we are at a pretty critical inflection point where we can do everything right, or we could not,” Gilleo says. “We have a really good chance to deliver a lot of benefits to communities and homeowners, but it does require us moving quickly to get the rules right.”

Voice AI is quietly transforming frontline work

2 days 1 hour ago

It’s no secret that AI chatbots and agents are answering employee questions, writing code and emails, and taking on other tasks in office environments.

But AI tools that communicate by voice, rather than through keyboards and screens, are also quietly taking off in frontline environments—from factory floors to retail stores. In some cases, AI that can understand spoken language is making it practical for workers to digitally access and share information in settings where protective gloves, outdoor weather conditions, or the need to operate other equipment have made traditional desktop or touchscreen devices difficult to use.

“With the advent of AI, voice became possible as an interface,” says Peter Moeckel, founder and CEO of Frontier Audio. “And frontline workers—they’re using their hands, they’re using their eyes for the job. That’s why the touchscreen, the keyboard, or the mouse obviously doesn’t work.”

[Photo: Frontier Audio]

Frontier Audio, which has raised $4.5 million in pre-seed funding, offers audio hardware and software to connect workers in industries ranging from manufacturing to aviation, including through contracts with the U.S. Air Force. The company’s devices come in a variety of form factors, including over-the-ear headphones with noise cancellation and shoulder-mounted microphones, similar to the walkie-talkie-style equipment many workers already use to stay connected. The equipment lets users talk to colleagues or AI, helping with tasks like reporting maintenance issues before they cause production problems, locating equipment such as carts in sprawling facilities, and receiving important alerts without needing to stop and look at a screen.

[Photo: Frontier Audio]

“We cannot add extra work,” Moeckel says. “They already have an immense amount of work to do, and we need to get out of their way. And our platform, a big part of making it seamless and enabling easy adoption is it working just like a normal radio.”

Frontier Audio is far from the only company building voice-powered AI for the workplace, often combined with radio services. Motorola Solutions offers AI features integrated into its police radios. Samsara, which builds connectivity and analytics tools for a variety of frontline operations, announced last month that voice commands have become the primary interface for people interacting with AI agents on its platform. Its voice bots can communicate with utility crews working on power lines, truck drivers arriving at unfamiliar worksites, and other workers in situations where they can’t look at a screen.

MangoApps, which offers AI and intranet-style tools geared toward frontline workers, enables users to access information about scheduling, company policies, and other matters through either text-based chat or voice in their preferred language, typically from their existing phones. Andy Tolton, vice president of marketing at MangoApps, says many businesses realized during the COVID-19 pandemic that they had limited ways to stay in contact with employees outside of traditional office roles.

“For the vast majority of businesses, their intranet is really a desktop paradigm,” he says.

While workplace AI tools have often been built chiefly with office workers in mind, businesses are now realizing they can offer multilingual access to company data and processes to field employees who, in some cases, were still using clipboards and paper forms until recently. That’s thanks to improvements in audio processing that make speech recognition and freeform conversations less frustrating, even in noisy environments, alongside advances in the ability of AI models to deliver accurate information and take action on behalf of humans.

“The speech AI technology has improved vastly, [as has] the knowledge that AI can access and be trained on,” Tolton says. “That, plus its ability to act, has increased exponentially in the last two to three years.”

Still, as in office environments, rolling out workplace AI for frontline workers isn’t as simple as flipping a switch. Businesses need to ensure that the devices workers use to communicate with AI, whether smartphones or radio-style equipment, are practical for their work environments. And whether accessed by voice or text, AI systems typically need to be integrated with existing data sources and processes for collecting employee input, while also being reliable and secure.

“In order for AI to work, you need to share data not only internally, but also with vendors. And you might also have to utilize sensitive data from your customers,” says Dayan Rodriguez, corporate vice president of global manufacturing and mobility at Microsoft.

In March, Microsoft announced new tools to build and deploy voice-powered AI agents through its Azure cloud platform. The company also offers security safeguards, including for manufacturing-specific technology that can be integrated with artificial intelligence. Microsoft points to examples of voice interfaces used by customers like Gulf Air, the national carrier of Bahrain, and electronics retailer MediaMarktSaturn, which has deployed Azure-powered voice assistants that sales employees can use to look up information about products and inventory while helping customers.

At least as critical as the technology itself, though, is ensuring that employees actually embrace it and can use it to do their jobs better, Rodriguez says. That’s part of the reason to give workers access to voice interfaces or whatever other systems make it easiest to interact with workplace tools. But it’s also important to monitor how those tools are used to make sure the investment pays off.

“Change management is more critical, in my opinion, than anything else,” he says.

How to disable 5 hidden location trackers killing your iPhone’s battery

2 days 6 hours ago

Do you always feel like somebody’s watching you?

You probably know your mapping app needs your location to give you directions. But your iPhone may be letting dozens of other apps ping your exact GPS coordinates in the background, draining your battery and quietly building a profile of everywhere you go.

Retail apps want to know when you walk into their competitors’ stores. Weather widgets constantly refresh your coordinates on your home screen even when you haven’t tapped them in a week. All of this silent radio activity burns through your battery before dinner time.

You don’t have to put your iPhone in airplane mode to stop it. You just need to shut down the unnecessary trackers hiding deep inside your system settings.

Kill precise location for apps that don’t need it

Your weather app needs to know what city you’re in, but it doesn’t need to know what side of the street you’re standing on.

Go to Settings > Privacy & Security > Location Services. Scroll through your apps and look for anything that isn’t a map or ride-sharing service. Tap each app and toggle off Precise Location at the bottom of the screen.

Restrict access to ‘While Using’

While you’re in that same menu, make sure no app is set to Always. Change every app’s permission to While Using the App.

This ensures a fast-food app or social media network can’t track your morning commute while your iPhone is locked in your pocket.

Shut down hidden Apple system services

This is where the real battery drains hide. Scroll to the very bottom of the Location Services menu and tap System Services.

You’ll see a massive list of internal trackers. You can safely turn off Location-Based Apple Ads, Routing & Traffic, and iPhone Analytics.

Your iPhone will still work perfectly, but it won’t constantly log your daily routine to send back to Cupertino.

Delete your ‘Significant Locations’ history

Inside that same System Services menu, look for a tab called Significant Locations. Apple uses this to track the places you visit most often to provide predictive traffic routing.

If you authenticate and open it, you will likely see a deeply detailed, borderline creepy map of your daily life. Tap Clear History at the bottom, then toggle the feature completely off at the top.

Turn off background app refresh

Apps love to use your location to download fresh data while you aren’t looking. Go to Settings > General > Background App Refresh.

You don’t need to turn the master switch off completely, but you should absolutely go down the list and toggle it off for social media, retail, and gaming apps that have no business running in the background.

AI is making software easier to build. Keeping it running is another story

3 days 2 hours ago

These days, it almost feels like old news: Developers can now use AI tools to generate code, build features, and make changes in a fraction of the time they once needed. The problem now surfacing across businesses is that once software goes into production, teams still need to monitor, maintain, and fix it. A feature that takes minutes to build can introduce problems that take engineers hours to investigate.

Google’s 2025 DORA report, based on responses from nearly 5,000 technology professionals, found that greater AI adoption was associated with faster software delivery but reduced stability. While the researchers didn’t establish that AI-generated code directly caused more outages, their findings suggest that companies can produce software faster without necessarily becoming better at maintaining it.

That raises a critical question for companies investing heavily in AI coding tools: If development keeps getting faster, how much of the work involved in running software can also be automated? A growing group of companies is working to find out, developing AI systems that can investigate failures, identify possible causes, and, in some cases, take corrective action without waiting for an engineer.

MORE SOFTWARE, MORE TO MANAGE

Ishay Yaari, cofounder and CEO of DataAgent, a company developing autonomous remediation software, says the operational side of software development hasn’t advanced at the same pace as coding.

“More code, more frequent deployments, and more infrastructure changes naturally create more potential failure points,” he tells Fast Company.

For site reliability engineering (SRE) teams, that means managing applications and infrastructure that may be changing more frequently. When something fails, engineers must figure out what changed, identify the cause, and restore service.

Consider an application that begins crashing after an update. The cause could be faulty code, an incorrect configuration, insufficient computing resources, or a problem with another service. While monitoring tools may detect the failure immediately, engineers still need to examine logs, review recent changes, and investigate several possibilities before deciding what to fix.

Yaari says companies often respond to growing complexity by adding monitoring tools and collecting more data. That can help engineers see what’s happening, but it can also mean more alerts to investigate without necessarily improving how quickly problems are resolved.

The pressure isn’t necessarily coming from poor-quality AI-generated code. Even reliable code can be difficult to manage when applications, infrastructure, and dependencies are changing frequently.

In a recent analysis of AI-assisted development, Google’s DORA researchers found that developers saved time generating code but spent more time reviewing and verifying it. They also emphasized the importance of testing, feedback, and established engineering practices in turning faster development into better results.

For companies, that means the benefits of AI coding tools may depend partly on what happens after developers finish writing the code, particularly the ability of SRE teams to keep applications running reliably.

GIVING AI THE KEYS TO PRODUCTION

Software monitoring and observability tools have traditionally helped engineers detect problems and understand their causes. Now companies are trying to automate more of the response.

DataAgent is pursuing that approach. Its platform is designed to monitor applications and cloud infrastructure, identify potential issues, and help investigate them. Yaari wants AI to take on more of the operational work engineers currently perform manually.

Some failures have familiar remedies. An application might repeatedly crash because it doesn’t have enough memory. A deployment might fail because of an incorrect configuration. Or a new release might introduce errors that disappear when the system returns to an earlier version.

Companies already use automation to address certain problems through predefined rules. AI could extend that capability by examining the circumstances, identifying likely causes, and selecting an appropriate response.

Yaari says DataAgent’s platform is designed to understand both the infrastructure supporting an application and the application itself, including how services interact, what the application is expected to do, and how its behavior changes after an update.

The aim is to recognize problems that don’t fit a predefined rule. But identifying a likely solution is one thing. Allowing software to make changes to a live application is another.

Consider a deployment that introduces errors. Rolling back to the previous version may restore service, but if the update also changed a database or introduced new dependencies, reversing it could create additional problems. That’s why the ability to act matters as much as the ability to diagnose.

HOW MUCH SHOULD AI BE ALLOWED TO FIX?

In its 2026 State of AI-First Operations report, PagerDuty found that 44% of surveyed organizations required human involvement when AI was used to fix problems affecting customer-facing systems, while 43% required it when coordinating incident responses across teams.

Even companies adopting AI for operational work appear reluctant to give it unrestricted authority over critical systems.

Yaari agrees that some decisions should remain with people. He points to actions such as deleting large amounts of data, disabling security controls, or exposing sensitive information as examples that should require approval.

“The goal is not to remove humans from SRE; it is to remove humans from the repetitive operational work that machines can handle safely,” he says.

Companies must determine which systems an AI agent can access, what changes it can make, and when it must request approval. They also need records of what the agent did, particularly when an attempted fix fails.

Another consideration is the cost of monitoring and observability software. Applications generate vast amounts of logs, performance measurements, and other operational data. Companies pay to collect, store, and analyze that information, often across several platforms. Adding AI can create further expenses if it requires transferring large amounts of data to external services.

Yaari argues that more analysis should happen closer to where applications run. DataAgent is pursuing a distributed approach, with agents deployed within customers’ infrastructure to reduce the amount of data sent elsewhere.

Its published documentation describes local monitoring alongside cloud-based AI analysis and guided incident investigation. Yaari’s responses describe broader autonomous remediation capabilities. The company has not provided independent performance data establishing how much operational work or cost those capabilities can eliminate.

WHAT HAPPENS TO THE ENGINEER?

Yaari expects engineers to spend less time investigating repetitive incidents and more time designing systems, setting operational policies, and deciding what automated tools should be allowed to do.

“The fundamental shift is from people operating infrastructure to people managing autonomous infrastructure,” he says.

That could allow smaller engineering teams to support more applications, but it would also require them to understand how automated systems make decisions—and what happens when those decisions are wrong.

Engineering skills may evolve as a result. System architecture, security, reliability, and the ability to investigate unusual failures could become more important as routine operational work becomes easier to automate.

Google’s DORA research found that organizations with strong engineering practices were better positioned to benefit from AI-assisted development. The same lesson may apply to AI-powered operations. Companies will still need reliable processes for testing changes, responding to incidents, and evaluating whether automated fixes work.

The software industry has made considerable progress in helping developers produce code faster. Now it’s trying to bring some of that speed to keeping applications running.

If AI eventually handles both, engineers may spend less time performing individual tasks and more time deciding what the systems they build should do, how they should behave, and when a person needs to intervene.

The chip war never ended

3 days 3 hours ago

It’s no secret that the world runs on computer chips. 

This became clear during the pandemic, when a race for even basic chips needed to operate cars and other devices set back manufacturing timelines across the world. Beltway policymakers are reminded of just how important semiconductors are to the world every time China seems to move toward invading Taiwan, which raises the specter that the country could subsume control of TSCM, the most prominent computer chip contract manufacturer. 

Now Jensen Huang’s exit from Nvidia, which comes after reports of at least three known extramarital affairs—one of them with an alleged spy for the Chinese government—has thrown the world’s supply of computer chips back in the spotlight. With stocks falling and no obvious succession plan, it’s not clear what the future of Nvidia is right now, or how it might jostle the broader semiconductor ecosystem. 

“Many companies that have in the past relied on Nvidia will be watching very carefully to see whether they can rely on Nvidia going forward, and if not, they might turn to competitors,” says Chris Miller, a Tufts University professor and author of the 2022 book Chip War, which documents how the world came to depend on these tiny components.

Fast Company chatted with Miller on November 12 about Washington, D.C.’s response to the Huang scandal, Nvidia’s future, and the prospect of keeping advanced semiconductors out of countries like China. This interview has been edited for length and clarity. 

You’ve followed the chip industry for quite a long time. Are you surprised by what’s happened over the last few weeks? 

Jensen was an extraordinary CEO for 33 years, since Nvidia’s founding. [He] deserves a lot of the credit for the company’s extraordinary rise, but also deserves a lot of the blame for a series of poor decisions that have led not just to his ouster, but to a period of crisis for the company. Jensen really was central to Nvidia’s success, and he ran the company in a pretty unique way. He had 60 different people reporting to him directly, so there wasn’t a sort of standard corporate hierarchy, which worked really well for a long time. 

But it means that now that he’s gone, it’s not obvious who can fill his shoes. There’s not really a deputy who was waiting in the wings to take over. 

I’m curious about the history of espionage attempts like what’s being alleged here. 

There is some history during the Cold War, for example. There were efforts by spies from the Soviet Union or Eastern European countries to get access to sensitive technology by getting close to semiconductor executives. But I think the scale of this is kind of unique. First, that it happened to such a central company in the U.S. tech ecosystem, the most valuable company in the world. That’s somewhat unique. 

The second thing that’s unique is that the affair in question was both with a major customer, which shouldn’t have happened, as well as an [alleged] Chinese intelligence asset. There were a series of failures of judgment for a person who was one of the most important people in the U.S. and really in the entire world’s technology sector, and that combination does make this case unique.

I’m wondering about unexpected winners or companies that stand to benefit from this happening, especially other companies in the GPU market.

There are going to be competitors of Nvidia that will see this as an opportunity to win market share. AMD, for example, produces GPUs like Nvidia. In addition, many of Nvidia’s key customers, firms like Amazon—which has a vast data center business—buy Nvidia GPUs, but also design their own chips that have comparable performance characteristics. 

Many companies that have in the past relied on Nvidia will be watching very carefully to see whether they can rely on Nvidia going forward, and if not, they might turn to competitors.

Some leaders on the left are arguing this scandal is made worse because there’s so much market concentration in the semiconductor sector, that we’re too dependent on a small set of companies. Does that ring true to you?  

The market’s concentrated right now, but it’s a market that is also highly competitive. It’s worth remembering the market for AI chips scarcely existed five years ago when ChatGPT was launched. . . . I think what we’re going to see over the next couple of months is that if Nvidia does face a prolonged period of distraction and chaos, competitors to Nvidia are going to win market share.

I would look at this as actually a quite competitive market with a bunch of different players that all have pretty advanced capabilities.

Is it feasible to wall off technology from China? I know Nvidia was even considering putting location tracking in some of its chips to keep them away from certain places, even if they’re sold on the gray or black market.

One of the ironies of this is that Nvidia, in general, and Jensen in particular, thought they could play a big role in dictating Washington’s technology policy and were naive to the fact that China was playing them by [allegedly] turning Jensen’s mistress into an intelligence agent. 

This is going to lead to a major reduction in Nvidia’s influence in Washington. Its influence in the White House is declining as well, and I think Washington will respond with tighter controls over where chips can be sent and which types of companies are allowed to buy large numbers of these chips when it comes to installing them in foreign countries.

The FAA’s IT troubles are far bigger then AI

3 days 20 hours ago

U.S. airlines are still reeling from a nationwide ground stop issued on Friday that interrupted weekend travel and wreaked havoc on schedules. The culprit, the Federal Aviation Administration says, was simply outdated tech.

The ground stop was followed by eager speculation on social media about what might have caused the issue. Some pointed a finger at a new AI system that the FAA integrated earlier in the week; others blamed foreign adversaries. At the same time, a range of tech companies that contract with the FAA, like Microsoft,were quick to distance themselves from the agency’s problems. 

Later Friday evening, Transportation Secretary Sean Duffy told reporters the outage occurred after an old analog system called time-division multiplexing became overloaded. This is an arcane system of processing data signals over a network. The overload, in turn, impacted the telecommunications systems used by terminal radar approach controls, or TRACONs, the air traffic control facilities that help direct flights at airports. Duffy noted that systems that had already been upgraded based on internet protocols did not experience the outage. “This is all the old legacy equipment,” he said. “All of the new equipment that’s been deployed did not go down.”

Officials, at least for now, insist there’s no evidence anything nefarious happened. Instead, the Department of Transportation is, once again, begging Congress for more money to overhaul FAA tech. 

Duffy urged Congress to provide additional funding to modernize the nation’s air traffic control system, arguing that the issue should transcend partisan politics. He noted that the FAA had received $15 billion for modernization efforts over the past 25 years, with little to show for it. “It’s not red, it’s not blue, it’s not Democrat, it’s not Republican, it’s American,” he told reporters. “Give us the money to fix it.”

Indeed, the FAA’s tech problems are way bigger than AI. Instead, they have a lot more to do with the challenges of updating technology while it’s being used. 

The pilots guidance system is ancient and isn’t designed to handle upgrades

This isn’t the first time the FAA has experienced a major outage that’s brought flights across the country to a halt. Back in 2023, the Notice to Airmen system (NOTAM), which alerts pilots to potential hazards in American airspace, went offline after contractors accidentally deleted files, disrupting its core technology.

There’s been an effort to improve the NOTAM system, but it’s moving at a snail’s pace. The FAA has brought in private industry for help and engaged in an effort to modernize the system, but that process has only brought to light myriad other problems. For one, at least according to contract paperwork obtained last year, NOTAMs weren’t exclusively being stored in a virtual environment or in an entirely digital format. Another aspiration the FAA was working on, at least at the time: making sure that upgrades to some systems didn’t disrupt, well, all of air traffic control.

“Everyone will say, ‘Well, there’s backups in place,’” Shawn Pruchnicki, a former pilot who is a professor at Ohio State University’s Department of Aviation, told FedScoop at the time. “But we’ve also seen time and time again where those things have been said and things have gone horribly wrong.”

Poor AI rollout

Like a lot of other federal agencies, the Federal Aviation Administration has been inching toward deploying artificial intelligence for a host of applications. Back in 2023, the Department of Transportation, of which the FAA is part, said ChatGPT was being used for certain purposes, which the FAA said at the time was incorrect. Since then, though, the FAA has begun to embrace AI: An inventory of use cases released by the government references several uses, including AI-powered pilot voice transcription, systems for detecting drones, and meteorology.

But the FAA hasn’t been the best at communicating to the public how it plans to use AI safely, without becoming susceptible to the risks of the technology. 

Many people, early on, suspected that Friday’s outage might have been related to the SMART system, an AI-based airspace planning tool the agency introduced earlier in the week. The FAA explains that the tool is designed to “analyze airline schedules, weather, airport capacity, airspace conditions, and operational constraints to predict traffic flows and identify potential conflicts before they occur.” SMART doesn’t appear to have had anything to do with the ground stop, though. 

Bad wires and old hardware

The deficiency of the FAA’s timeworn infrastructure was on display in April 2025, when a burned copper wire triggered a communications blackout at Newark Liberty International Airport. For 90 seconds, air traffic controllers lost radar and radio contact with planes approaching one of the nation’s busiest airports. The incident triggered weeks of delays and cancellations. Another communications failure followed in May. 

Newark’s hiccup was the sign of an endemic problem: Some of the FAA’s copper wiring dates back to the 1960s, and according to a September 2026 report from the Government Accountability Office (GAO), the FAA had replaced fewer than half of its aging copper wires, 2,560 out of 5,170.

And it’s not just wires. Back in 2007, the FAA began a technological overhaul known as NextGen, which aimed to reduce its reliance on ground-based radar in favor of satellite navigation and digital communications. Nearly two decades later, that effort remains unfinished, due in part to years of inadequate funding. Just earlier this month, the FAA proudly announced it had finally begun retiring floppy disks.

“What we’re doing is hard and it hasn’t happened in the past because it is hard. But I’ll tell you by doing nothing, you inject way more risk into the system,” Duffy said on Friday. “To use 1980s technology, 1990s software . . . 1970s radar that we don’t even make parts for anymore, to have an old network configuration from the 1980s versus re-networking the air traffic control system through the FAA, I believe that should be a priority,” he said.

The FAA’s tech is the problem no one wants to fix

Duffy came into office as transportation secretary promising to overhaul the FAA’s aging technology. In May 2025, he unveiled an ambitious plan to rebuild the nation’s air traffic control system by 2028 and replace critical infrastructure like radars and radios. Congress approved an initial $12.5 billion for the effort, but the price tag quickly jumped. Just two months later, Duffy told lawmakers that the full overhaul would cost an estimated $31.5 billion. Even his 2028 deadline comes with an asterisk. It covers only the first phase of modernization, with no timetable yet established for the second, which includes new software to track aircraft and manage air traffic. 

The GAO put it most damningly in its September 2026 report, writing that the FAA “doesn’t have a lifecycle cost estimate or master schedule for related projects,” which “could end up causing air traffic disruptions or prolonging reliance on severely outdated systems.” These tech problems only exacerbate the FAA’s other endemic challenges, including overworked air traffic control offices and a controller recruitment crisis.

Meanwhile, the $30 billion Modern Skies Act, sponsored by Senator Eric Schmitt of Missouri and aiming to fund upgrades on everything from air traffic control software to airport terminals, has yet to advance out of committee. And on November 13, the White House asked Congress for another $30.1 billion for aviation infrastructure, including $10.1 billion for air traffic control technology.

When asked whether Americans could be sure that the issues seen Friday wouldn’t occur again, Duffy was blunt: “I can’t build things if I don’t have money.”

Checked
3 minutes 21 seconds ago
Fast Company Technology
Fast Company inspires a new breed of innovative and creative thought leaders who are actively inventing the future of business.
Subscribe to Fast Company Technology feed