Skip to main content

Fast Company Technology

Anthropic and OpenAI sketch plans to share their wealth ahead of IPOs

4 weeks 1 day ago

As they steam toward massive public offerings on Wall Street, the companies at the forefront of artificial intelligence have begun sketching plans to share their wealth with the public—a commitment they say is imperative to address disruptions the technology could bring to the workforce and people’s livelihoods.

Executives at Anthropic and OpenAI have long argued the technology’s arrival will bring many benefits for everyday people, and they’ve promised that will include a share of the profits. The companies are outlining a range of proposals for distributing AI fortunes, but so far they have not resolved key questions over how they would work, how transparent they will be, and who will determine beneficiaries.

On Sunday, Anthropic said it would commit 5% of its shares to the Anthropic Public Benefit Fund, a new entity within the company, pending approval from stockholders. The stake could increase “as the company grows” in the future. The maker of AI assistant Claude said the fund “might explore” larger donations to traditional nonprofits or potentially distribute money to “specific populations.”

The San Francisco-based company shared little information about the types of organizations supported by the fund and the issues they will address. One area it will not support, Anthropic said, is lobbying and political giving.

Phil Buchanan, the president of The Center for Effective Philanthropy, said he welcomed Anthropic’s commitment—especially at a time when he said some billionaires are “diminishing the importance of philanthropy.” But he cautioned that Anthropic should avoid the “arrogance” of previous technology barons who assumed their professional expertise would translate to philanthropic wisdom.

Buchanan recommended the fund support highly effective, existing nonprofits but said it is “hard to talk about any of this because it’s all so abstract.”

“I guess I’m hopeful,” Buchanan said, “but I’m not going to get too excited until I see some evidence of what’s actually happening.”

Anthropic’s pledge hinges on Wall Street plans

Anthropic, which was valued at close to $1 trillion earlier this year, is preparing for an initial public offering as soon as next month that is shaping up as one of the biggest ever.

Once a company is public, donations might be made through the selling of future shares, or potentially turning to cash dividends in the future, explained Jay R. Ritter, director of The IPO Initiative at the University of Florida’s Warrington College of Business. The specifics on how charitable donations will be made through Anthropic’s fund are not clear yet.

Regardless, Ritter said the announcement now is helpful for potential investors to “know what they’re buying.”

Anthropic has already committed $200 million to study the economic effects of AI more broadly. Combined with Sunday’s announcement, Anthropic said roughly 20% of its shares will be dedicated to the “public good,” plus any additional percentage that could be added if the company’s valuation reaches certain milestones.

In announcing the benefit fund, Anthropic offered labor market impacts as one example of its philanthropic focus.

“AI may generate enormous economic value, but it could also cause instability for many workers,” Anthropic said in a blog post Sunday. It added the fund will have the “versatility” to pursue “bolder approaches” that might be necessary to offset that. The company said it would share more details about the fund’s structure and governance in the coming months.

In a May interview with The Associated Press, Anthropic president Daniela Amodei said she expected the company’s success would generally lead to more philanthropic support for disease research and “ensuring that everyone around the world gets to share in the prosperity that we hope AI creates.”

The companies’ pledges have been met with skepticism from some critics.

“Anthropic is playing games here,” Gary Marcus, a longtime AI expert and professor emeritus at New York University, wrote on social media Sunday. “Trying to curry public favor while still holding control and resisting disclosure is annoying.”

AI companies are pursuing a range of pathways for charitable distributions

Anthropic and OpenAI are each structured as a public benefit corporation, a designation that requires the companies to pursue not just profit but a social good.

OpenAI Foundation, the nonprofit that controls OpenAI, promised to build out its grantmaking capacity earlier this year with a $1 billion commitment.

Separately, OpenAI in April proposed the creation of a “Public Wealth Fund” with a stake in “AI-driven economic growth” across companies more broadly. That fund could be invested in a handful of assets, OpenAI noted, with returns distributed directly to people regardless of whether or not they’re invested in financial markets. OpenAI is planning its own initial public offering next year.

OpenAI CEO Sam Altman also has spoken to lawmakers on both sides of the aisle about sharing equity in AI companies with the public.

SpaceX CEO Elon Musk, who has tended to eschew traditional philanthropy, pledged as recently as Monday to donate some shares of his own company.

During the personal computer boom, the likes of Michael Dell and Microsoft co-founder Bill Gates started private foundations to manage their philanthropy. Today, Silicon Valley’s leaders, including Meta CEO Mark Zuckerberg, have turned instead to limited-liability companies (LLCs), the option that Anthropic is pursuing for its public benefit fund.

The structure offers more control for its owners, experts say. An LLC is not bound by the same spending rules as a foundation, which is required to distribute at least 5% of the value of its endowment.

“They have a lot more flexibility and latitude in how to use the resources they put into the fund,” said Katherina Rosqueta, the founding executive director of the University of Pennsylvania’s Center for High Impact Philanthropy.

Brian Mittendorf, the chair in accounting at The Ohio State University, said LLCs allow for activity outside “traditional philanthropy,” such as payouts to individuals that would not be possible through a 501(c)3 charity. But if they are not making traditional charitable gifts, that puts the companies in a “gray area” by raising the question of how the public is benefiting, he said.

LLC owners don’t reap the same tax benefits, either, according to Rosqueta and Mittendorf. Anthropic’s pledge doesn’t immediately trigger a tax write-off, and the company can only claim a deduction once the donated shares are moved to a public charity. There’s also less transparency required of LLCs.

Overall, Anthropic’s announcement amounts to a commitment to “yet-to-be-specified wider goals,” Mittendorf said. The move “reflects more of an intention to do things in the future rather than an act now,” Mittendorf said.

—By James Pollard and Wyatte Grantham-Philips, AP business writers

AP Philanthropy editor Glenn Gamboa contributed.

U.S. bans AI startup DeepSeek because of ‘threat vectors’ and China responds

4 weeks 1 day ago

China hit back Tuesday at a U.S. ban on Chinese artificial intelligence startup DeepSeek’s products for federal civilian agencies, highlighting friction in the two AI powers’ heated technology race.

The agencies should remove DeepSeek and its affiliated products, as well as third party products and services developed using DeepSeek, from their IT systems, the Department of Homeland Security said in a notice made public on Monday.

The department’s order, dating back to late July, also blocks civilian agencies from procuring DeepSeek products and services. DeepSeek is one of China’s leading and best-known AI companies and its open-source models are widely available in countries including the United States. They are mostly free to use.

The notice in the Federal Register alleges that DeepSeek’s models “present multiple threat vectors.” It said concerns include that its servers are located in China and that Chinese authorities’ could potentially access private company infrastructure and sensitive U.S. data.

In Beijing, Chinese Foreign Ministry spokesperson Mao Ning said China will “will resolutely safeguard the legitimate rights and interests of Chinese companies” when asked at a regular press conference about the U.S. ban.

“China consistently opposes the overstretching of the concept of national security and the politicizing of science, technology, economic and trade issues,” she told reporters at a routine briefing.

DeepSeek did not immediately respond to a request for comment.

U.S. President Donald Trump and Chinese President Xi Jinping are set to meet in mid-November on the sidelines of the APEC, or Asia-Pacific Economic Cooperation, leaders’ summit in Shenzhen, China.

AI is expected to be on the agenda.

China and the U.S. have recently made progress in some collaboration on AI governance, including a U.S. proposal for setting up of a notification system on national security-related incidents, but the two countries are still at odds on many issues.

U.S.-led restrictions still ban China from accessing some of the world’s most advanced AI chips and the machines that make them. Last week, Beijing rolled out new proposed plans to restrict exports of certain advanced AI models.

The U.S. has also said Chinese AI startups including DeepSeek engaged in large-scale “distillation,” or extraction of capabilities, from frontier American AI models, claims that Beijing has dismissed.

Earlier this year, the U.S. Federal Communications Commission announced it was barring imports of new foreign-made humanoid robots in a move seen as targeting China.

—Chan Ho-Him, AP Business Writer

New details emerge on former OpenAI engineer accused of bugging a conference room

4 weeks 1 day ago

A former OpenAI engineer charged with secretly installing a recording device in a conference room to glean company details had been “mere days” away from absconding to Saudi Arabia before his arrest last week, federal prosecutors revealed in new court filings.

Abdulrahman Alfozan’s wife and young son also left the country the day after the 35-year-old Saudi national’s arrest, traveling on a one-way ticket booked and paid for by a Saudi military official, prosecutors said in a memo Sunday seeking his continued detention.

“Ultimately, the defendant has little incentive to remain in the United States to see this case through,” federal prosecutors wrote. “He has ample resources —including his significant compensation from a decade of high-tech employment in the Bay Area and at least two bank accounts in Saudi Arabia — and what appears to be considerable support from a foreign government.”

A federal judge in San Francisco sided with prosecutors Monday, denying a request by Alfozan’s lawyers to release him to home confinement. His next court date is Nov.

Alfozan’s lawyer William Pernik argued in court that his client had strong ties to the U.S. and was “weeks away from achieving every immigrant’s dream” — U.S. citizenship.

Alfozan had a naturalization interview with federal officials scheduled next month and has been a lawful permanent resident since 2020, his lawyers said in their legal filings. He has also lived in the U.S. for 15 years and attended MIT and the University of California, Berkeley, they said.

Pernik said Alfozan is willing to put up “almost the entirety” of his assets — about $1 million currently in a U.S. investment account — to cover his bond.

“Everything this man has he’s willing to surrender,” he said, adding that Alfozan’s wife is also willing to return to the U.S. as she still holds a visitor visa that expires in 2033.

Judge Alex Tse appeared doubtful, questioning whether Saudi Arabia’s government would simply reimburse Alfozan if he fled and forfeited his bond.

After the hearing, Pernik said he was disappointed with the ruling but declined further comment.

Pernik had argued in court filings that Alfozan purchased one-way tickets for his family to Saudi Arabia more than two weeks before prosecutors allege the recording devices were installed. They said the couple planned to spend Alfozan’s roughly two-month paternity leave in Saudi Arabia following the birth of their U.S.-citizen son.

But prosecutors said Saudi officials purchased new plane tickets for Alfozan’s wife and son after his arrest, allowing them to leave the U.S. sooner than the Oct. 12 flight Alfozan had originally booked.

Prosecutors have not said what Alfozan might have wanted to do with the recordings, but say he was in “frequent communication” with a member of the Saudi royal family. They said the royal family member repeatedly tried to recruit Alfozan to an artificial intelligence venture being launched in Saudi Arabia.

Alfozan had told the unnamed person he wouldn’t leave OpenAI until his stock was vested, even as he explored other career opportunities in the Middle Eastern kingdom, prosecutors wrote. He had worked at the ChatGPT maker since July 2025 but was fired after his arrest.

Prosecutors say Alfozan concealed an audio recording device in toilet paper and plugged it into a USB port on the back of a television in an executive conference room he wasn’t authorized to access. They say security cameras captured him entering the room and checking behind the television after installing the device. It remained there for several days before being found Sept. 30.

Prosecutors say the conference room was typically used for meetings involving company leaders, government officials and other executives and “discussions of highly confidential technical, commercial, and strategic information, including proprietary research, as well as nonpublic information.”

OpenAI has said the recording device Alfozan used was sophisticated, designed to evade detection and featured advanced encryption and audio-monitoring capabilities. They also said they had evidence Alfozan planned to place other listening devices elsewhere in the office.

Alfozan has been charged with using or endeavoring to use an electronic, mechanical, or other device to intercept oral communications. If convicted, he faces a five-year prison sentence.

—Barbara Ortutay and Philip Marcelo, Associated Press

‘You Can See Everything’ perfectly captures what a decade of post-truth has done to us

4 weeks 1 day ago

A tech-world documentary may seem an unlikely candidate for one of those movies people claim simply must be experienced in a theater. You Can See Everything makes a strong case all the same. There are moments in Nathan Fielder and Lance Oppenheim’s beguiling profile of the disgraced Theranos CEO Elizabeth Holmes that inspire gasps, convulsive laughter, and involuntary hand gestures. One such moment was especially resonant in my theater, though. (And be warned: There are very mild spoilers ahead.)

To put it as vaguely as possible, the moment involves Fielder attempting to pin down Holmes about her history of dishonesty. (As the trailer suggests, there’s a lot of that going on throughout the entire, plus-size runtime.) In this particular instance, Holmes says something so hard to believe, it can only be taken with a Texas-sized grain of salt. Since Fielder is a master of deadpan who has made a career out of soft-spoken understatement, his naked astonishment registers as volcanic: “Liz, what?!”

It’s a very funny moment, but it’s more than that. There’s something cathartic at play, too. A decade into the post-truth era, as powerful people stand 10-toes-down on flagrantly false narratives, it’s the kind of deeply human reaction to dishonesty many Americans have been craving. In his rejection of the line he’s being asked to swallow, Fielder jars viewers out of whatever boiled-frog comfort we may have found in a timeline riddled with unreality, forcing us to see it with fresh eyes.

The post-truth era meets the AI era

You Can See Everything finds Holmes and her husband, Billy Evans, having invited Fielder to observe and interview the Theranos founder in 2023, during the days leading up to her 11-year prison sentence for defrauding investors. Fielder is a canny but bold choice for this mission, as he is a comedian best known for making one TV show about business scams (Nathan For You) and another about his endless fascination with authenticity (The Rehearsal). It’s never explicitly stated how this arrangement came about, but theories abound online.

For his part, Fielder does not seem to view the project as an opportunity to determine whether Holmes was dishonest while leading Theranos. (We’re well past that, as anyone who’s read John Carreyrou’s explosive exposé, Bad Blood, which Fielder has, can vouch.) Instead, his movie suggests that some dishonest people are simply sociopathic, while others are less connected to reality on a neurochemical level, and others still might be somehow both, with Fielder hoping to assign a definitive category to Holmes. His dogged refusal to accept Holmes and Evans’s framing is a clarifying encapsulation of what it’s been like to read the news on any given day in recent years.

Although the rise of partisan media outlets and Twitter echo chambers had been fracturing public belief systems long before 2016, Trump’s election is what thrust us into the post-truth era. Every president lies, but the breadth and scale of Trump’s lies was previously been unheard of in American politics. By making false claims about eminently observable reality, Trump revealed that lies don’t carry consequences if enough powerful people are willing to ignore them. His emergence as the leader and mascot of the United States announced that being factually accurate was no longer as important as simply appearing supremely convinced of one’s own veracity.

This performance of total confidence—as in the “con” in “con man” or “con woman”—has now become standard in just about all forms of mass communication, from wellness influencers with a questionable grasp on nootropics to reality show contestants who definitely aren’t there to develop a personal brand. The just-trust-me shtick seems especially prevalent in today’s tech world, though, which often rewards people for confidently describing an AI future that does not yet exist.

Tech executives now increasingly make eye-roll-worthy claims in a few distinct categories. There’s the “inevitability” boast, which suggests AI is unleashing its world-changing power on the labor market whether anyone likes it or not, so everyone might as well get on board anyway. (As if stringent third-party guardrails would somehow violate the laws of nature.) There are the claims of urgency, with CEOs citing some historic threshold like the supposed arrival of Artificial General Intelligence, which their tech has conveniently just crossed. (Altman recently claimed we’re in the singularity.) Finally, there are the testimonials about the magnitude of change to come, which will apparently be so transformative that humans will require a universal basic income to live.

Unfortunately, when tech executives publicly make these outrageous assertions—which just happen to present their products in a flatteringly powerful light—they are seldom, if ever, confronted by someone with Fielder-like skepticism, tenacity, and a lack of fu*ks to give.

‘Your questions are designed to trick me’

Pop culture falsely prepared us all to expect grand and timely comeuppances. In the world of movies, obvious liars in high places are always destined to get caught on a hot mic bragging about their deceptions, only to be run out of town on a rail or thrown into a prison cell.

Holmes’s fall from grace, however, is an exceedingly rare exception to the Trumpian rule that one may impose their perception of reality on the world if enough boldface names allow it. Perhaps it’s the power that comes with so much of the U.S. economy being entangled in AI that allows the field’s leaders to make extraordinary claims with insufficient pushback.

For instance, there doesn’t seem to be nearly enough scrutiny of all the recent doomsday messaging. Anthropic’s CEO, Dario Amodei, recently said there was a 25% chance the future would go “really, really badly” due to AI, while OpenAI’s Sam Altman comparatively hedged at 10%. In both cases, the CEOs avoid getting into the weeds of how they arrived at these figures or explaining how AI Armageddon might actually unfold. We’re to take them at their word.

Pushing back on those worst-case scenarios, Nvidia’s CEO, Jensen Huang, went the opposite route. During a July interview, he called the idea that AI would destroy half of American jobs “complete nonsense.” He did not, however, provide any data-driven economic evidence that AI is unlikely to spark staggering job losses, relying instead on good old industry optimism.

And then there’s SpaceX CEO and fellow recent documentary subject Elon Musk. He presents himself as the visionary steward of a utopian future that consistently fails to materialize by any announced deadlines. Despite how often he is loudly proven wrong, though, none of his false interpretations of reality, to put it generously, ever seem to produce any satisfying consequences.

Even when Musk appeared in court back in April for his case against Sam Altman (Musk was an early cofounder of OpenAI), he refused to accept questions that exposed his previous lies. The judge had to intervene several times as Musk avoided straightforward answers and painted lines of inquiry as unfair, at one point concluding, “Your questions are designed to trick me.”

In the minds of all empirically correct masters of the universe, the truth is whatever it needs to be, and anyone unwilling to accept it as such must be the enemy. Without giving too much away, this kind of attitude can be detected in certain parts of You Can See Everything.

The utter exhaustion of being asked to believe certain things

The most effective deception doesn’t always ask people to believe a lie; sometimes, it just requires them to stop believing there’s a definitive truth that can even be known. This is the world we live in, and the one in which You Can See Everything unfolds. It’s filled with people who are apathetic, cynical, confused, and utterly exhausted.

Fielder embodies that last trait most of all, his unbelieving eyes telegraphing how tiring it is to be sold a bill of goods all the time. He’s the ideal avatar of anyone who thinks that with the right set of questions they would trick Musk into admitting, despite past denials, he was indeed aware during Trump’s inauguration that the gesture he was making strongly resembled a Nazi salute.

But you can’t compel someone to concede anything they’re determined not to admit, just because logic and the facts are on your side.

You Can See Everything doesn’t chisel a full confession out of Holmes, but it does accomplish the next best thing. Fielder’s “Liz, what?!” moment offers viewers the vicarious thrill of telling a famously dishonest person that even their performance of honesty is impossible to believe.

Gen Alpha has a surprisingly optimistic outlook on dating

4 weeks 1 day ago

Gen Z is just about ready to give up on dating. They’re over dating apps. They’re dealing with burnout. They’re not hooking up. Now, the oldest members of Gen Alpha, born between 2010 and 2024, are in their teens and starting to develop crushes and go on their first dates. Are they destined for the same dating struggles?

Though they can’t use dating apps until they’re 18, Match Group decided to take an early look at Gen Alpha’s attitudes and expectations around dating. The company surveyed 823 nationally representative U.S. teens between 13 and 15, building on an earlier qualitative phase of one-on-one interviews and focus groups with 14- to 19-year-olds. The early signs suggest the youngest generation is already getting butterflies.

Only 20% of Gen Alpha teens have been in a relationship so far, but they’re showing interest in real connections and optimism about what relationships could eventually become. “It’s fun to be 13 to 15, and the entire world looks ahead of you,” Chine Mmegwa, head of strategy and corporate development and operations at Match Group, tells Fast Company. “That sense of optimism has returned for this generation, which we hadn’t seen in our Gen Z data.”

At the same age, many members of Gen Z were isolated from friends and peers by a global pandemic, leaving them struggling with loneliness and basic dating skills. “Gen Z data came in [on] the back of COVID,” says Mmegwa. “My fear had always been that we would never get that innocence back.”

Gen Alpha, on the other hand, appears more upbeat. Of those surveyed by Match Group, 92% are happy with their social lives, and 84% say they have a great future ahead of them, up 6 percentage points from teens the same age just three years ago. To be sure, Gen Alpha is growing up surrounded by screens and social media, but it’s also maturing amid conversations about mental health and wellness, absorbing those lessons and establishing boundaries early.

For Gen Alpha, crushes might come and go, but friendship is the top priority. Among those who aren’t dating but are open to it, 48% say they’re most interested in friendship with romantic potential, 44% in just talking or flirting, and only 30% in anything committed. Of all respondents, 64% believe people should be friends before they date.

This isn’t unusual for teens, who are exploring the boundaries of friendships and other types of relationships for the first time, but it continues a trend seen with Gen Z. Tinder recently launched casual modes to appeal to these Gen Z daters, 33% of whom strongly agree that expanding their social network is important, compared with 20% of older singles, according to a Tinder x Harris Poll white paper released earlier this year.

This focus on friendship may help explain why many Gen Alpha teens have yet to explore dating. Only 20% have been in a relationship, compared with 56% of Gen Z respondents who reported having been in a relationship during their teen years. “Much of what you get out of dating, they’re getting out of other relationships in their life,” says Mmegwa.

That doesn’t mean they’re anti-dating. Some 71% say their friends think dating is cool, and 64% say most people their age are dating or trying to. Among the 79% who aren’t dating, three in four are open to it. They’re just not in any rush. They remain interested in traditional life milestones, including marriage, homeownership, and children, and 87% say they’re very confident they’ll reach them.

“They’re excited about these things, which we haven’t seen in the Gen Z data,” says Mmegwa.

Gen Alpha members are also more positive about dating apps, although perhaps that’s because they’ve yet to use them. Just over half, 51%, are open to using dating apps once they turn 18, and 65% say they’ve heard dating apps can be a fun way to meet people.

“I love that 65% of them think dating apps could be a fun way to meet people,” says Mmegwa. “We haven’t had that in a while where people tie dating apps to fun, and so I will take that any day.”

Despite these hopeful signs for Gen Alpha, dating apps are facing a crisis. Singles have swipe fatigue. Bumble recently went private after years of declining subscriber numbers and falling revenue, which was down 15.2% year over year in the second quarter. A 2026 Harris Poll found that two-thirds of singles weren’t using dating apps at all, and 72% said apps had made dating feel more transactional than romantic.

Where Match Group went wrong in recent years, Mmegwa admits, is that the company didn’t evolve fast enough for Gen Z. “Our hope is that this generation [Gen Alpha] should never have the experience that dating apps are bad because by the time they are aging into our experience, we have solved the problem for Gen Z.”

Across the industry, the swipe model that helped create the paradox of choice plaguing modern dating is being overhauled by established apps and newcomers alike. While Gen Alpha isn’t creating profiles just yet, the groundwork for the future of dating is already being laid.

Will the next generation of daters still be swiping left in years to come? Will the future of dating involve more bird-watching meet-cutes and hobby-based hangouts? Will AI take over dating entirely?

“My gut says it’ll look totally different,” says Mmegwa. “I’m excited to see what that looks like.”

AI could make cybersecurity a never-ending process

4 weeks 1 day ago

Over the past few months, it’s become clear that modern AI systems are quite adept at discovering software vulnerabilities and finding ways around cybersecurity defenses.

Unintended real-world breaches executed by models from leading AI companies like OpenAI and Anthropic during training exercises have drawn widespread concern, and while those companies have restricted access to their most capable models, experts have indicated it’s only a matter of time before less constrained AI systems have similar capabilities. Already, businesses are seeing a rise in cyberattacks, at least partly attributed to AI assistance. 

And increasingly, the best defense lies in . . . artificial intelligence. 

“We hear a lot about adversaries and attackers using it, and that’s certainly the case,” says Bartley Richardson, chief AI & autonomous systems officer at the cybersecurity firm CrowdStrike. “What we don’t hear as much about is how are we advantaging defenders with the use of AI.” 

CrowdStrike is one of a number of cybersecurity businesses deploying AI on the defensive side. In September, the company announced a tool called SafeMind that purposely takes advantage of AI’s offensive capabilities, deploying custom models and harnesses designed to probe simulated versions of customer infrastructure. SafeMind also deploys defensive AI systems to detect the attacks, fend them off, and close any security holes that are discovered. Both sides then learn from the experience and try again, says Richardson. 

“We not only make the changes in the environment before we give it back to the offense and say try again, but we also have a whole manifest—a whole log of every single change that we made to the environment,” he says. “We give that to the offensive harnesses as well.” 

Such tools may be coming in the nick of time. Like AI coding systems, which have evolved from answering questions to generating, testing, and deploying complex code, AI hacking agents have become increasingly potent. They don’t just scour code for logic flaws but also probe live systems for vulnerabilities, potentially refining their attacks until they translate sets of subtle bugs into step-by-step pathways for unauthorized access.

“We are seeing a measurable increase in the volume and velocity of all of these attacks,” says David Cooper, cyberstrategy activation leader at EY, the big accounting and tech consulting firm. “We see it in our security telemetry clients, security operations centers—we’re seeing an uptick.” 

Provided they aren’t detected and someone is willing to pay for the necessary computing power, they can work essentially indefatigably to find a way in to real-world systems, experts say. And they can find ways to chain together minor flaws in software to achieve ever-escalating access, devising complicated exploit processes that previously would have been found only by determined human experts. 

“They can take a sequence of actions, and that is what really matters in cybersecurity, especially in zero-day attacks,” says Mona Rahimi, associate professor of software engineering in the Department of Computer Science at Northern Illinois University.

That’s part of why defense teams are also turning to AI. In addition to exploiting security holes, it can help point them out to security experts to fix or, increasingly, repair them itself. Alongside the rise in cyberattacks has come a parallel increase in patches released by software vendors, with Microsoft recently reportedly unveiling its largest security update in history, says Cooper. Companies he works with have essentially universally deployed AI to help probe and patch their own systems, and EY has recently worked with CrowdStrike on security issues, including as a design partner on SafeMind. 

“There’s different levels of adoption, but all of the companies that we work with are using some set of AI tools to help accelerate this,” says Cooper. 

That comes after the research firm Cybersecurity Ventures estimated that global cybercrime costs hit $10.5 trillion in 2025, and a World Economic Forum report warned earlier this year that “AI is supercharging the cyber arms race.” 

‘An explosion in rapid remediation’ 

EY and CrowdStrike are far from the only companies helping their clients address AI issues. Leading AI labs including OpenAI, Microsoft, and Anthropic have each launched projects to probe and harden digital systems, and security firms have augmented their products to enable AI to spot and even correct vulnerabilities before outsiders take advantage of them. 

“We’re just seeing an explosion in rapid remediation, which is something that is foreign, I think, to the world that we typically have lived in,” says Joni Klippert, cofounder and CEO of StackHawk.

Even before modern AI, StackHawk focused on detecting vulnerabilities in software as it actually runs, rather than simply in source code. “The things that you can find only by testing when an application is running are really valuable, and they’re usually really critical problems,” says its cofounder and chief security officer, Scott Gerlach. 

And one advantage defenders have, he says, is more complete knowledge of the system, including proprietary source code and details of various infrastructure and its configurations. That knowledge doesn’t necessarily have to live in human heads, and AI can often be quicker at probing and repairing legacy applications that people haven’t touched in some time, Gerlach says. 

“The more automation that you add to this process, the faster it goes, and the more things get fixed,” he says. 

Still, many organizations naturally prefer to have a human in the loop before allowing AI to make radical changes to their computing environments, both because AI can make mistakes and because humans can make informed decisions based on competing priorities, like ensuring critical systems aren’t shut down or slowed to unacceptable levels by would-be security upgrades. AI itself can, in some cases, propose or take steps to help balance those needs, like restricting access to critical software that can’t immediately be patched. 

Safeguarding open source tools 

But while attackers may be willing to shut down or degrade performance to accomplish their goals, security personnel often don’t have that luxury. 

“We cannot really maximize security at any cost,” says Rahimi, the Northern Illinois University professor. “Because we still want to maintain our service.” 

The problem isn’t limited to commercial software. A number of efforts have sprung up to help find and fix vulnerabilities in open source software, the often volunteer-maintained code that’s critical to the internet and the systems on it. Among them is Akrites, announced by the Linux Foundation and industry partners in June, intended to coordinate security reports about open source tools and deliver reliable documentation, concrete examples, and sometimes even patch code that upstream project leaders can incorporate to fix their code. Akrites’s debut came after various open source projects reported being inundated with AI-generated bug reports and suggested code changes, including security fixes of varying quality. 

“This has been a very substantial burden on upstream, who almost exclusively are unpaid volunteers,” says Christopher Robinson, CTO and chief architect for Akrites and the Open Source Security Foundation. 

Akrites is prioritizing open source projects used within critical infrastructure, including financial services and telecommunications, he says. Open-source operating systems, databases, programming languages, and other tools have long been the backbone of many production systems, in many cases becoming de facto standards and offering power and flexibility unmatched by commercial tools. But those projects often rely on the work of small groups of coders or even individuals to keep them up to date and can even cease to be updated if project maintainers retire or become occupied with other work. In cases where a project has no active maintainer or isn’t in a position to accept patches, Akrites can work to release a forked version—a modified copy—and find someone to serve as a kind of maintainer of last resort, Robinson says. 

‘Cat-and-mouse game’

It’s still unclear how long the surge of detected security flaws and fixes will continue, or whether it will slow down after the new AI tools have helped clear a backlog of undiscovered bugs. Developers are also likely to increasingly use AI tools to probe their own code before it’s released. As Robinson points out, coding AI tools alone aren’t a complete solution, since they can introduce bugs of their own. 

For the time being, says Ami Luttwak, cofounder and CTO of Wiz, a cybersecurity company acquired by Google earlier this year, AI’s hacking capabilities are continuing to grow, akin to a burglar who one week can enter through a first floor window and, two weeks later, can find a way to the second story. “Every day, every week, every month, there’s new capabilities that AI can do,” says Luttwak. 

Wiz operates its own Scan for Good initiative in collaboration with Google DeepMind, using AI to help critical infrastructure organizations, governments, healthcare institutions, and nonprofits find and fix security holes. And as AI continues to grow in potency, such scans will likely be a regular need for all sorts of institutions and businesses, he says. 

“This concept of static security controls is dead,” Luttwak says. “It will be a cat-and-mouse game, and we believe that every company needs to continuously do AI scans for protection.” 

This startup died in 2017, but its dream of in-ear computing is everywhere

4 weeks 1 day ago

Few elements of Silicon Valley’s culture are more cherished than its embrace of failure. Nobody looks askance at you for starting a company that doesn’t make it. Indeed, it’s widely regarded as a badge of honor, proving your willingness to make bold, risky bets and leaving you a smarter, more battle-tested veteran.

And yet failed startups tend to be forgotten the moment their doom is sealed. Aside from a few flops so epic they become legends—hello, Webvan and Humane—defunct businesses quickly stop ringing a bell. Everyone involved moves on to chasing the next big thing.

That has been true of many intriguing upstarts I’ve profiled over the years. The one that set out to be the Uber of shipping. The one that partnered with Steph Curry to build a tiny, minimally distracting smartphone. The one that aspired to build an AI-powered computer for the ears well before AI became the industry’s obsession.

Unlike AirPods, Here One came in your choice of white or black. [Photo: Doppler Labs]

But wait. Though that last startup, Doppler Labs, ceased to be back in 2017, its vision didn’t die along with it. Doppler’s flagship product was Here One, a pair of wireless earbuds designed to do far more than stream music and make phone calls. The company developed technology to perform feats such as intelligently mixing streamed audio with live ambient sound. It hoped to help people with hearing loss get a better listening experience than that offered by traditional hearing aids, at a much lower price. It was working on live translation that would let people speaking different languages chat in real time.

When Doppler was founded in 2013, these goals were exceptionally forward-looking. At the time, nobody had even figured out how to incorporate Bluetooth into earbuds that worked reliably and offered acceptable battery life. Those were table stakes that Doppler had to achieve along with its loftier vision.

As you already know, a key moment in the history of wireless earbuds—no, the key moment—happened when Apple released the first AirPods. That happened in December 2016, months before Here One hit stores. Apple didn’t position AirPods as tiny computers. Instead, it was happy to have people think of them as wireless headphones—just remarkably convenient ones with high-quality audio and long battery life.

When the AirPods Pro followed in 2019, however, Apple watchers did describe them as a computer for your ears. As AirPods became one of Apple’s most important product lines—the company shipped 16.8 million units in the second quarter that year alone, according to research firm Omdia—they grew increasingly capable. Eventually, they checked off many of the boxes in Doppler’s original vision, from mixing streaming and ambient sound to real-time translation.

Sketches for Here One. [Image: Doppler Labs]

Today, any product that aspires to compete seriously with AirPods, such as Google’s Pixel Buds and Samsung’s Galaxy Buds, ends up implementing features that Doppler was working on a decade ago. The emergence of talking assistants based on generative AI technology has only intensified the intrigue around the ear as an emerging playground for consumer electronics giants and startups alike.

Doppler’s alumni certainly regard elements of their big ideas as playing out today, even if other companies are making them real and reaping the rewards. “We developed, I think, a pretty exceptional vision that unfortunately still hasn’t really been realized,” says cofounder Fritz Lanman. Still, none of them accuse Apple or anybody else of commandeering their concept. Nor do they make any excuses for their company’s early demise.

“At the end of the day, we took a really good shot,” sighs Doppler’s other cofounder, Noah Kraft. “The final product was just not enough for it to justify its existence.“

Old startups may fade away, but the tale of why Doppler didn’t make it—and how its vision proved powerful enough to fuel an entire industry ever since—is worth putting on the record. Here it is.

“What if we could curate how we hear the world?”

Back in mid-2016, when Doppler was still furiously getting Here One ready to ship, I witnessed its progress in person. Fast Company had assigned me to write a feature on the startup, which occupied two adjacent buildings on an alley in San Francisco’s SOMA district. They were a few minutes’ walk from my desk, making frequent drop-ins easy.

The Doppler story was intriguing for reasons that went beyond the raw ambition of its vision. Pre-Doppler, Kraft’s most notable career credit was in movie production. He was also a musician and music lover. Helping out with a friend’s electronic-music tour, he’d noticed that performances sounded different in each venue—and often not as good as they might have. That led him to ask a question: “What if we could curate how we hear the world?” As he mulled over what the answer might look like, he envisioned the product that became Here One.

Doppler’s team in a celebratory mood. [Photo: Courtesy of Noah Kraft]

At the time, there was a fair amount of excitement over wearable technology, but the best-known example, Google’s Google Glass AR goggles, did not have the greatest of reputations. To Kraft, headphones felt like a better starting point than eyeglasses. “We’ve worn tech on our ears since the Walkman, 50 years ago,” he told me in 2016. “It’s just less intrusive than the eyes.”

The catch was that he wasn’t even a little bit technical. “I wasn’t sure if this was even possible,” he admitted. An acquaintance in the music business knew a tech executive and brokered an introduction. That executive was Lanman, and when he and Kraft met over brunch, what they thought would be a brief chat turned into a multiday conversation, and then a partnership. They founded Doppler Labs in the fall of 2013, with Kraft as CEO and Lanman as chairman. The company went on to raise $50 million in funding from VC firms as well as music-industry luminaries such as David Geffen and Quincy Jones.

The two made for a bit of an entrepreneurial odd couple. Kraft exuded a sense of wonder about the whole adventure of starting a company. Lanman, by contrast, gave off the air of someone who’d already seen it all—because he had. A former Microsoft business development executive who’d led the company’s prescient investment in Facebook, he’d become a prolific angel investor, putting money into more than 70 startups, including Pinterest and Square.

Rather than remain entirely stealthy while they worked to bring Kraft’s idea to fruition, the cofounders bit off specific parts of the project in public. Doppler’s first product, Dubs Acoustic Filters, was just a nice pair of $25 earplugs, aimed at people who didn’t want to look goofy when they wore them. (For more info, check out the Dubs website, which mysteriously lives on.)

Next, Doppler launched a successful Kickstarter campaign for Here Active Listening, a set of earbuds that let you fine-tune real-world audio around you, such as at a concert. Then it promoted that product at Coachella, priming the world for the then-new concept of earbuds that performed tasks other than streaming music and phone calls—which Here Active Listening couldn’t do.

Doppler CEO Kraft brandishes two trophies after the company’s Here Active Listening won Best of Show at SXSW 2016. [Photo: Courtesy of Noah Kraft]

The fact that Here Active Listening didn’t try to replace wired Apple EarPods or whatever else you were plugging into your smartphone simplified Doppler’s story. “By not having wireless streaming capability and audio playback, people couldn’t say, ‘It’s a headphone,’” Lanman explained in 2016. “They really focused in on these bionic hearing capabilities and noise filtering capabilities.”

But the product’s lack of streaming also reflected a basic truth that’s easy to forget a decade later: Making Bluetooth buds work at all was a formidable challenge for anyone, let alone a tiny startup. Both buds would need to stay in sync, with imperceptible latency. They’d have to run for hours on a charge. And the electronics involved would have to be tiny enough to stick in your ears.

Not engineers themselves, Kraft and Lanman were keenly aware of the technical expertise their startup required. The team they assembled had impeccable credentials from companies such as Dolby, Nike, Sony, Qualcomm, and Sennheiser. I happened to be visiting the day they got prototype earbuds reliably streaming for the first time, using near-field magnetic induction technology rather than Bluetooth for bud-to-bud communication. Kraft beamed as he showed it operating on a workbench.

Mastering wireless streaming was only one of the challenges Doppler had given itself. Well before machine learning provided the foundation for generative AI, the company was using the technology to understand the audio it wanted to control. “I did a lot of walking around with microphones in my ears and then labeling audio samples from those recordings, so that we could differentiate between a baby crying versus somebody speaking versus a car honking a siren,” recalls audio quality engineer Caleb Rau. “We were trying to basically separate wanted sounds from unwanted noise.”

Already, Doppler’s thinking was also racing ahead to future versions that pushed the boundaries of what was possible at the time. In April 2015, its executives plotted out a road map at a research lab the company had established in Southern California. “I think the space had previously been the test kitchen for a restaurant chain,” remembers VP of R&D Jeff Baker. “It was Fritz and Noah and myself and a few other people, and we were brainstorming and capturing notes about things you could do with earbuds. Translation was on the list.”

Eventually, they decided that Baker’s lab would focus on translating between languages in real time, a bet that future smartphones would provide sufficient computational muscle to do the job. That led to a partnership with Systran, a major player in multilingual software. When Baker demoed a proof-of-concept version to me, I was intrigued—even though it required me to don a backpack containing a small desktop PC and a big battery, and didn’t translate between English and Spanish all that well.

Here One was born in two adjacent San Francisco offices full of busy people and a vast amount of electronics and testing equipment. [Photo: Courtesy of Noah Kraft] “We were living in Noah’s reality distortion field”

On June 28, 2016, Doppler gave itself a hard deadline by announcing Here One. The company said it would ship the following November, just in time for the holidays. Within a month, almost 100,000 people had joined a waitlist for the product. “We’re trying to move tens of thousands of units in Q4 and then hopefully hundreds of thousands in 2017,” Kraft told me.

If Doppler was going to sell hundreds of thousands of earbuds, it would have to get very good at mass-producing them. In service of that goal, the company had already dumped its original Asian contract manufacturer in favor of a new one. That was a sign of faith in the judgment of its VP of product, Kennard Nielsen, whose résumé included the development of successful high-volume consumer electronics such as Nike’s FuelBand wearable and Amazon’s Kindle Fire tablets. (“We call him ‘Obi-Wan,’” Kraft told me.)

Months after that switch, Doppler was still struggling with component supply chain issues. In late November, around the time it had planned to ship Here One, the company announced that it was delaying its release until the following February, with retail availability the next month. That pushed any possibility of the earbuds selling in high volume until well into 2017.

Meanwhile, Apple shipped its first AirPods in December, after a two-month delay that showed even it found making wireless earbuds to be unexpectedly challenging. “I remember seeing pictures of them, and I thought they looked kind of funny,” says Doppler firmware engineer Mary-Ann Rau (who met her husband, Caleb Rau, at the company). “We kind of dismissed them. We were living in Noah’s reality distortion field: ‘Here One looks way better. We have way more features. What’s there to worry about?’”

Among the music lovers who received demos of Doppler’s technology was a former Beatle. [Photo: Courtesy of Noah Kraft]

As it turned out, plenty. When Here One reviews hit in February, they were uniformly impressed by the buds’ “layered listening,” which let users pull off feats such as choosing how much ambient sound to let in as they listened to streaming music. But Doppler’s buds got beat up for their rough edges, especially poor battery life and intermittent connectivity glitches. More than one reviewer said the overall experience felt more like a tech demo than a product worth $299.

“I was hoping Here One would be my first real taste of that future we see in sci-fi books and movies, where a computer in your ear can change your perception of the world around you,” wrote The Verge’s Sean O’Kane. ”Here One shows that future is closer than ever, but it falls short of being the product that gets us there.”

Doppler head of music Michael Thurber (left) gives another musical luminary, Jon Batiste, a listen. [Photo: Courtesy of Noah Kraft]

All along, Doppler had taken pains to communicate that it was building something truly new, not just a pair of wireless headphones. “Apple’s mission is not to make hearing amazing for everyone, and that was our goal,” Lanman says. “We thought of streaming and phone calls as kind of a commodity.”

The lackluster reviews showed that Here One would be cut little or no slack merely for being amazing. “The fundamental issue wasn’t the things that made it special; it was the fundamentals,” Kraft says. “The battery [life] was not enough.”

“The excitement across the board was palpable”

By the time Here One shipped, the fate of Doppler Labs was in flux for reasons that went far beyond the product’s disappointing launch. The company was already months into discussions that might have led to it being acquired, a possibility that would have transformed its finances and product strategy overnight had it happened.

Those discussions were ignited by Apple’s announcement of the AirPods on September 7, 2016. Its entry into the wireless earbud market led many other companies to develop a sudden interest in technologies such as those Doppler had already developed.

Consumer electronics startups often ended up being rolled up into industry giants’ broader aspirations. Even before VR pioneer Oculus had shipped its first headset, it was acquired by Facebook for a mind-melting $2 billion. Dropcam, a maker of security cameras, had been bought by Google’s Nest division for $555 million. With those 2014 transactions still a recent memory, the potential seemed to be there for Doppler to collect a windfall of its own. It gave itself an aspirational goal: at least $1 billion.

Doppler’s electronics didn’t just have to work. They had to fit in your ear. [Photo: Courtesy of Doppler Labs]

Thanks to the company’s billionaire investors and the well-connected Lanman, it didn’t have any trouble attracting the attention of top executives at the tech industry’s biggest names, including Apple, Amazon, Facebook, Google, and Microsoft. Over an intense couple of weeks, Doppler had introductory meetings with all of them. “The excitement across the board was palpable,” Kraft says.

One prospective suitor seemed especially eager to do more than simply get to know Doppler. Amazon had been embarrassed by its Fire Phone flopping in spectacular fashion, but had rebounded with the well-received Alexa voice assistant and Echo speaker. Now it was moving quickly to build an entire audio-centric platform.

Even during the first meeting, Kraft says, Amazon executives were talking expansively about making a major investment in Doppler or buying it outright. Their exuberance led Doppler to explore all its options. It hired Morgan Stanley as an adviser and “probably took 100 acquisition meetings,” Kraft says.

“We were actively showing ourselves to investors in a beautiful space overlooking the Bay Bridge owned by [Doppler investor] UMG Music,” Baker recalls. “I spent three months showing a who’s who of the VC and corporate world our earbuds, because Fritz was very connected.”

When Doppler delayed Here One, Amazon put negotiations on hold for a time. But by early 2017, they picked up again and focused on Doppler providing the foundation of a would-be AirPods killer called Echo Ears. Though no deal had been struck, the two companies held a celebratory dinner at the Four Seasons in San Francisco and began collaborating on the technical details. Then Amazon became unresponsive again.

Already harried by the pressures of startup life, Kraft found the acquisition exploration and its attendant uncertainty even more taxing. “Emotionally, I thought we were going to sell the business,” he says. “Now I’m basically doing two things. I’m running a business and running a process.”

Before Here One’s technology fit in a pair of earbuds, it got prototyped with circuit boards on a hard hat. [Photo: Courtesy of Noah Kraft]

As 2017 continued on and Doppler remained independent, it considered other alternative futures for itself. One was licensing its technology to other manufacturers for use in their own products. “We were, like, ‘Maybe we could become the “Intel Inside” of audio,’” Kraft remembers. “I wouldn’t even call it a pivot. Just that we could make some revenue by selling not our hardware, but our brains. That wasn’t the vision for the company. It was just, ‘Maybe this is a way to stay afloat.’”

Not that being the Intel Inside of smart audio would have necessarily been any easier than continuing as a hardware maker. “I kept it quiet, but [licensing] didn’t smell great to me,” Baker says. “Earbuds are incredibly bespoke things.”

The company had already been pursuing another opportunity that, though also fraught, could have been transformative: taking on the handful of deeply entrenched companies that dominated the hearing aid business.

Even when Doppler’s tech demos involved a bunch of loose electronics strung together, Kraft had shown me how it could facilitate clearer audio by removing the background cacophony from a simulated restaurant. All along, the company had a director of accessibility and advocacy, K.R. Liu, who used hearing aids herself and told me that “the technology in hearing aids is pretty crappy. It doesn’t sound very good. It’s not very clear. It’s just not where Doppler’s technology is.”

At the time, hearing aids were pricey and available only by prescription, and Doppler was legally required not to position its product as one: “Here One, to be really clear, is not a medical device,” Kraft emphasized to me. But along with other consumer audio companies such as Bose, it was lobbying for the Over-the-Counter Hearing Aid Act. The proposed law aimed to spur innovation and lower prices—a change that would give Doppler the chance to make good hearing a lot more affordable.

By the time the bill was passed in August 2017, it was too late to bolster Doppler’s fortunes. But both Kraft and Lanman point to the company’s role in galvanizing public support for the law as one of its lasting achievements, even if it didn’t redound to their favor. “We could have destroyed the hearing aid industry,” Lanman says wistfully. “Or if not destroyed, at least democratized it.”

Despite Here One’s weak start and the failure of acquisition talks to turn into a deal, Doppler wasn’t too worried. The company was forging ahead with work on its second-generation earbuds, Here Two, and continuing to staff up with industry veterans. Brian Hall, who’d been a Microsoft corporate vice president overseeing products such as Surface and HoloLens, came aboard as COO. Jim Pitkow, a web pioneer and veteran of Xerox’s fabled PARC laboratory, became chief science officer. “It was incredibly validating when he agreed to join us full time,” Kraft says (Sadly, Pitkow, who was inspired by the Sandy Hook school shooting to spearhead the development of gun safety technologies and later became CTO of a nonprofit devoted to protecting children from sexual exploitation, died in July of this year.) 

In the summer of 2017, Doppler had $10 million in the bank. It would need another round of venture funding to ship Here Two. To help assure prospective investors that its top management had sufficient hardware experience, the company promoted Hall to CEO, with Kraft becoming executive chairman. 

When Doppler went seeking money, however, it found that VCs were spooked. “Honestly, one of the hardest things for us to explain away during the Series C talks in 2017 was, ‘Wait, haven’t you been in deep talks with Amazon for over a year? Why aren’t they doing something? What are we missing?’” Kraft recalls.

Why wasn’t Amazon doing something? Doppler’s leaders concluded that the acquisition talks and excitement over Echo Ears had been pointless. Talks with other potential acquirers had been less of a time sink, but no more fruitful.

By October 2017, Doppler was running short on money. It was compelled to listen when Amazon, at long last, made a formal bid. The tech giant offered $10 million, a price that conveyed its confidence that the startup was out of other options.

Already irate over the odyssey that had led to this point, Kraft didn’t blink. “They’re like, ‘But you’ll be able to say you sold your business to Amazon,’” he recounts. “And I was, like, ‘Oh, you misread this. I have no interest in the optics of selling a business to Amazon. Not to mention, I hate you guys now.’” The encounter did leave him proud—not because he sold a company to Amazon, but because he refused to do so.

As November began, Doppler announced that it was shutting down. It ended up selling a portfolio of patents to Dolby. And with that, its story officially ended. (A hearing aid startup that has operated under the Doppler Labs name more recently was founded by a former Doppler employee but is unrelated to the original company.)

Sans Doppler technology, Amazon did introduce its own wireless earbuds, which it called Echo Buds. They were latecomers to the market, arriving in October 2019, almost three years after the original AirPods. After several iterations, they wound up emphasizing a budget price over technological innovation. And then Amazon quietly discontinued them.

“We laid the groundwork and set the direction”

Doppler was hardly the only notable startup that didn’t make it through 2017. (Others included Jawbone, Yik Yak, Sprig, and the infamous Juicero.) Its end was more fittingly marked than most, with Wired’s David Pierce writing a fine eulogy. He chronicled the company’s achievements and disappointments, including its ill-fated negotiations with an unnamed “big five” tech company. (Kraft was not yet ready to call out Amazon by name.)

Both of the company’s founders bounced back from the Doppler disappointment. Lanman became CEO of one of the companies he’d invested in, ClassPass, which offers a subscription service for fitness classes. It was acquired by another startup, Mindbody, in 2021. That led to the creation of a new company, Playlist, which merged with another earlier this year to form a powerhouse “fitness and wellness operating system.” Lanman is the CEO.

Kraft pondered his post-Doppler career for some time until the entrepreneurship itch got the better of him. In 2021, he teamed with his brother Lucas to form an over-the-counter digestive health startup, Wonderbelly. Last January, they sold it to Procter & Gamble for an undisclosed amount.

Though Lanman and Kraft’s direct influence on the world of smart audio ended when Doppler disbanded, the sudden availability of the team they’d assembled was a boon to the rest of the industry. Mary-Ann Rau remembers being shocked by the company’s closure: “I was an engineer, so I was not paying too close attention to the financials or sales.” But the following month, she joined Apple to work on AirPods firmware. She went on to introduce new AirPods features during two product launch keynotes, a living embodiment of the Doppler-Apple link. (Later, her husband and fellow Doppler alum, Caleb, also joined the AirPods team, and remains on it.)

When Doppler employees Mary-Ann Ionascu and Caleb Rau got engaged, they wore their Here One earbuds in the official photos. (Their former boss, Doppler cofounder Noah Kraft, officiated at their wedding.) [Photo: Courtesy of Mary-Ann Rau and Caleb Rau]

According to LinkedIn, seven former Doppler employees are at Apple today. Others scattered to Google, Amazon, Dolby, and Bose, ensuring that the company’s diaspora would continue to have an impact on the product category it had helped kick-start.

In an August 2018 Fast Company story, Peter Burrows wrote about Silicon Valley’s embrace of “hearable” computing and took note of the former Doppler staffers who’d landed elsewhere. Fourteen months later, Apple introduced AirPods Pro. Unlike their non-Pro predecessors, the pricier new buds blocked out background distractions with active noise cancellation and introduced a mode called Transparency that mixed streamed and local environmental sound so both could be heard—two features reminiscent of the “layered listening” that Here One had offered.

The AirPods Pro’s Doppler-esque evolution continued with a 2024 software update that let the AirPods Pro 2 be used as hearing aids. The following year, the AirPods 2 and 3 and the AirPods 4 got real-time translation, the same feature Doppler had shown me back when its test bed required a PC in a backpack. Now Apple is reportedly gearing up to introduce AirPods Pro with built-in cameras that let them see real-world objects and identify them with AI—a concept that would have sounded like fantasy in Doppler’s day, yet is very much in tune with its vision of a computer for your ears.

In tech, victory belongs to the companies that turn ideas into successful businesses, not those that have them first. In this case, that’s Apple. The Doppler alumni I spoke with enumerated some of the Apple advantages that the startup could never have matched, such as the fact it designs its own chips and controls both the AirPods and iPhone, lessening its dependence on Bluetooth for connectivity.

Then there’s Apple’s ability to invest in a product for years on its own terms, not those of venture capitalists seeking a quick return on investment. At one point in my conversations with Kraft, he theorizes that fulfilling Doppler’s promise would have required another $50 million; at another point, $200 million. Either way, “The sheer capital required to invent a new form of computing is beyond the capacity of the VC ecosystem, at least as it existed 10 years ago,” Lanman says. I can’t think of any examples, then or now, that disprove his point.

Still, the vibe I got from the Doppler veterans who talked with me about AirPods wasn’t resentful or jealous. Instead, it sounded like pride—albeit of a bittersweet variety.

“We laid the groundwork and set the direction,” says former product chief Nielsen. “Having a company with the technical prowess and the resources like Apple come in and pick it up and carry it to the finish line was super gratifying.”

We’ll never know what wireless earbuds would look like in 2026—and beyond—if Doppler had never existed. But Mary-Ann Rau sees parallels in the story of General Magic. That legendary 1990s startup tried to build a wireless personal communicator years before the technology was ready, and ceased operations in 2002. The conversation it kicked off about the future of mobile computing was still in progress when Apple came up with the iPhone. Did I mention that General Magic also created a voice-powered assistant that eerily foreshadowed current AI agents such as Meta’s Muse and OpenAI’s Dots?

“The same pattern happens over and over in Silicon Valley,” Rau says. “There’s the startup with the big dream and the small team, and they don’t quite make it, in the sense that they don’t make the product that ends up being ubiquitous.” (Undaunted, she left Apple in 2023 and later became cofounder and CEO of her own company with a big dream and a small team, heat pump startup Merino Energy.)

Ubiquitous is what AirPods swiftly became. Apple doesn’t disclose unit sales or revenue, but a 2024 Bloomberg Intelligence report estimated that the company had sold between 94 million and 119 million units a year since 2020, good for $18 billion to $22 billion in annual revenue. As others noted when those figures came out, they made AirPods a bigger business than Nintendo or Spotify.

Doppler would not have had to beat Nintendo and Spotify to have had a far happier ending. “When we started, we wanted a billion-plus,” Kraft says. “But towards the end of it, if someone had swooped in and said, ‘Here’s 300 mil,’ we would have taken it in a heartbeat. Because we’d only raised $50 million.”

Had an acquirer then gone on to capture just 5% of AirPods’ market share, Kraft adds, it would have created that rarest of phenomena in Silicon Valley: a major new hardware business. He stresses that he wouldn’t change a thing about the Doppler experience. But all these years later, he hasn’t stopped turning it over in his mind. The epiphany he had when he was assisting with his friend’s music tour still sounds like a great idea for a company to him.

Social media bans don’t protect young people’s mental health, early research shows

4 weeks 2 days ago

The year 2026 has been an expensive one for social media platforms.

Since March, YouTube and Meta, the parent company of Instagram and Facebook, have settled or paid out in multiple lawsuits to the tune of more than $17 billion for allegedly acting negligently, deliberately designing their platforms to hook kids and misleading the public about safety and privacy. If these suits are any indication of how states and the public are starting to view social media platforms, more may follow.

On top of these lawsuits, social media platforms are facing a loss of business as bans for minors are accelerating and spreading worldwide. Australia was one of the first with a ban on kids under 16 on social media starting in December 2025, followed soon after by the U.K. with a similar prohibition.

The current list of countries looking to ban minors from these platforms tops more than 20, from Greece to the United Arab Emirates to Canada. The proposed and enacted bans around the world vary in terms of age restriction, parental consent rules, and age verification methods. Different countries are looking at restricting different platforms, but most include a mix of Snapchat, YouTube, TikTok, Instagram, Facebook, and X.

These bans aren’t necessarily for mental health reasons only—some are motivated by political concerns about foreign-owned platforms, and others aim to protect young people’s privacy. But nearly every government banning social media platforms for young people mentions their mental health as a reason.

The question is, are these bans the solution to the mental health problems young people are facing? I research digital media and marketing and measure the impact of new technologies, and I’m skeptical about whether prohibiting access to these platforms will actually improve young people’s mental health.

Mental health concerns

Since Jonathan Haidt’s 2024 book, The Anxious Generation, critics of social media have increasingly focused on young people’s mental health. Haidt argues that social media disrupts kids’ social and neurological development, contributing to rising anxiety and depression through sleep loss, fragmented attention, addiction, loneliness, social comparison, and perfectionism.

A 2026 study found that young people between 14 and 20 are more than 100% more likely to experience mental illness than 14- to 20-year-olds 50 years ago.

Increased use of social media and screen time is compounding an already precarious time for minors. In 2021, it was estimated that 20.5% of young people experienced elevated symptoms of anxiety, and 25.2% of youth reported elevated symptoms of depression globally.

Many factors contribute to this, including better screening and awareness of kids’ mental health problems, as well as societal disruptions as a result of the COVID-19 pandemic. As the world sheltered in place and moved online for work, school, and social interactions, there was an increase in isolation and loneliness.

Not all bad

In light of these concerns, some researchers argue that bans on social media platforms are like putting an age restriction on buying alcohol or cigarettes.

But unlike these addictive and potentially harmful substances, there are benefits to social media, especially for teens who feel disconnected from their local communities and want to build relationships with others who feel similarly.

Social media platforms allow for global connectivity and community-building with like-minded groups. They also provide a place for self-expression and civic engagement.

In a Pew survey of U.S. teens, 80% said social media helps them stay connected to what’s happening in their friends’ lives.

Even as states and countries are moving to ban young people from social media, there isn’t clear evidence that spending time on these platforms is the main cause of their mental health problems.

Show me the money

One place where the comparison with tobacco is actually quite apt is in relation to the money from lawsuit settlements.

In 1998, states settled a lawsuit against the major tobacco companies to recover smoking-related healthcare costs. Part of the settlement was allocated to programs aimed at reducing youth smoking. In 2026, total state funding for these programs amounts to just 3.4% of the $21.7 billion in revenue states will collect from tobacco settlement payments and tobacco taxes.

Similarly, some money from an August 2026 Meta settlement with a group of states for $17 billion is slated for state-based programs focusing on mental health infrastructure, school and community programs, and public health campaigns.

However, there is no oversight or direction as to how states should use the funds, nor is there a mechanism to report how that money has improved kids’ mental health care, safety and well-being.

Is it even possible?

Early evidence from the Australian social media ban shows that kids are finding workarounds: Roughly 61% of kids ages 12 to 15 continue to access platforms using false IDs, altered appearances, or virtual private networks (VPNs), even though millions of accounts were disabled. In fact, this has become so common that U.K. lawmakers are debating restricting consumer VPN access.

Some countries, such as Australia, require the platforms themselves to develop ways of verifying a user’s age. Some minors have bypassed scans from biometric age estimation technology by using photo manipulation software, substituting celebrity photos, using clever makeup, and even utilizing older friends and parents to pass the “selfie tests” on their behalf. Others simply manipulate their birth years during registration or rely on existing adult-verified accounts managed by family members.

And all of these age-verification methods, whether developed by a government or a social media platform, rely on collecting sensitive data to estimate age accurately, which creates privacy concerns.

Australia is currently seeking to raise fines on social media platforms that fail to comply with its requirements. The Australian safety commissioner, charged with enforcing the ban, issued a compliance update that showed 7 in 10 children who held accounts on restricted platforms remained on those platforms, and more than 80% of Australian teens ages 10 to 15 still use social media frequently.

Beyond concerns about enacting and enforcing these restrictions, some policymakers also worry that total bans could push some kids to migrate to even riskier online spaces.

What could work

Based on the evidence, I’m not convinced that banning social media platforms will be an effective solution to the mental health crisis among young people today.

So what could work?

Supporting young people’s mental health is a complex endeavor that will require a combination of people, organizations, and resources coming together.

I believe kids’ mental health would be better served by governments enforcing safety by default standards on social media platforms rather than outright bans. This means that the highest security and privacy options are in place by default so that minors still on a platform are protected. For instance, accounts would be unsearchable by default, so users would have to intentionally make a profile change to be found by people they don’t know in real life.

Safety by default standards also place limits on the manipulation, algorithmic restraint and other addictive mechanisms that may lead to psychological harm in vulnerable populations such as young people.

Educating kids on digital literacy and online safety as well as ensuring some screen-free time are also going to help support kids’ mental health.

It’s also important for parents and other caregivers to recognize changes in behavior as a result of addiction to social platforms and screens. And when adults start seeing signs of a problem, they need clear, easy, and affordable ways to connect their kids to mental health support, which is not available in most places right now.

And finally, the United Nations recommends bringing kids into the conversation about what’s best for them online. Obviously, this does not mean putting 12-year-olds in charge of regulatory agencies. But feedback from young people about what they care about, what they get out of being online, and what makes it hard for them to look away from their screens could inform changes to social media platforms and education programs that could help young people engage online more safely and maintain a healthy balance.

Ari Lightman is a professor of digital media and marketing at Carnegie Mellon University.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

AI expertise isn’t the same as AI oversight. Boards need both

4 weeks 2 days ago

From 2023 to 2025, the share of S&P 500 companies disclosing AI as a risk jumped from 12% to 83%. But according to company filings, fewer than 3% of directors have AI expertise. And in the same research, just 23% of executives said their board was highly fluent in AI, and a quarter said their board had low fluency or none at all. As artificial intelligence spreads across the business, that leaves a significant gap between what boards understand and what they are required to oversee.

That gap is untenable—and this is what boards need to do to fill it.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/creator-faisalhoque.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/faisal-hoque.png","eyebrow":"","headline":"Ready to thrive at the intersection of business, technology, and humanity? ","dek":"Faisal Hoque’s books, podcast, and companies give leaders the frameworks and platforms to align purpose, people, process, and tech—turning disruption into meaningful, lasting progress.","subhed":"","description":"","ctaText":"Learn More","ctaUrl":"https:\/\/faisalhoque.com","theme":{"bg":"#02263c","text":"#ffffff","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#ffffff","buttonHoverBg":"#3b3f46","buttonText":"#000000"},"imageDesktopId":91420512,"imageMobileId":91420514,"shareable":false,"slug":"","wpCssClasses":""}}
Expertise is necessary but insufficient

If the problem is that boards lack AI expertise, the obvious answer is to recruit directors who have it. Well, yes, but not quite. Boards do need more AI experts, but recruiting experts is not enough to close the gap. There are two reasons for this.

First, AI isn’t siloed. It runs through hiring, pricing, customer service, credit, and operations. There is no single AI agenda item for one expert to own, and every director will vote on matters AI has shaped, whether they realize it or not. Oversight must therefore be distributed across the board, not concentrated in a limited number of expert seats.

Second, expertise is not the same thing as competent oversight. Expertise is what you know about AI; oversight is what you do about it as a director. Expertise is often needed for competent oversight, but it doesn’t guarantee it. For instance, someone can understand exactly how a large language model works without ever asking who in the company is accountable for one, or what an AI agent is allowed to spend.

So boards face two questions about AI. The first is whether the board knows enough. The second is whether each director is competent at oversight. Answering the first doesn’t answer the second—and while everyone focuses on the first, the second is largely neglected.

What the board needs to oversee: The key areas of AI oversight

Before we figure out what competent oversight involves, we need to be clear about what directors are meant to be overseeing. The following areas, drawn from our ongoing research into how directors oversee AI, are foundational. Each of these is a place where AI can do real damage to a company, and a board that covers some of them but not the others is dangerously exposed.

  • Governance Oversight: Knowing which individual is accountable for each AI matter, and making sure that they are actually held accountable.
  • Data Inquiry: Knowing what data the company’s AI uses, where it goes, and on what legal basis. This must be backed by evidence rather than a vendor’s assurance.
  • Explanation & Disclosure: Making sure the company can explain an AI decision to the person it was made about, and that its public claims about AI match what it does.
  • Fairness & Affected Parties: Knowing which groups a system could disadvantage, when it was last tested, and whether problems were fixed rather than just recorded.
  • Risk & Irreversibility: Knowing what the company’s AI can do that cannot be undone, and how the board would hear about a failure.
  • Human Judgment & Workforce: Making sure a person can overrule a system in practice, not just on paper, and that the people whose work AI changes are part of the plan.
  • Autonomy & Agent Authority: Knowing what AI agents may do, spend, and commit on their own, and making sure each one has a limit and a way to stop it.

That is what a board needs to oversee. But how do you actually build competent oversight of those areas? Three principles help you get there.

Measure what people do—not what they know

The strongest possible evidence for being able to perform a task is that you regularly perform it to a high standard.

Therefore, rather than only asking if your directors understand AI oversight at a theoretical level, you also need to ask: Have you actually performed this basic task of AI oversight recently? How often? Is it something that is a systematic part of your oversight process?

The point is that being aware of something is not the same as doing it, let alone being good at doing it. In PwC’s 2025 survey of public company directors, two-thirds said their board should spend more time on AI, but fewer than half said their board had an approach to actually overseeing it.

For example, almost every director would agree that the board should know which AI systems the company uses. Far fewer have actually asked to see the list, and fewer still ask for it every time the board reviews AI risk.

First Step: Ask each director to name the last AI oversight act they performed, and when and where it happened.

Assess the floor—not the average

A director’s level is set by their weakest competency, not their average. Boards already accept this logic for financial controls. Under U.S. auditing standards, if even one material weakness exists, a company’s internal control over financial reporting cannot be considered effective, no matter how strong everything else is. Nobody averages a material weakness away. AI oversight deserves the same.

Focusing on the floor matters because human beings gravitate to the questions they are comfortable asking. A former CFO interrogates the data; a technologist interrogates the model. To an extent, that’s what a board should want. Directors are recruited for their depth, and that depth should be used.

The problem comes when depth in one area becomes a reason to skip the others. A CFO who never asks what the company’s agents can do is relying on someone else to ask. If that person is absent, distracted, or deferring in turn, nobody asks at all.

So the floor is not a demand that every director become an expert in everything. It is a minimum standard that every director clears in every area. Boards already work this way with finance—under New York Stock Exchange listing rules, every member of the audit committee must be financially literate.  

Competent AI oversight needs a solid foundation: a floor that everyone clears.

First Step: Ask each director which AI oversight question they would be least able to answer. Make it their priority for the next quarter.

Build depth across the whole board—not in a single seat

The floor protects against gaps in individual directors. But a board can have every director clear the floor in every area and still have nobody who goes deep on fairness, on the workforce, or on what the company’s AI agents are allowed to do.

That is a common pattern, because boards recruit for backgrounds, and backgrounds cluster. In 2026, boards again favored directors with CEO and financial experience. So depth piles up around data and models and thins out everywhere else.

This is where the fact that AI isn’t siloed becomes especially relevant. It means that the board needs depth in every area of AI oversight, and that depth shouldn’t sit in a single seat, because one expert’s judgment tends to go unchallenged when no one else on the board is equipped to test it.

Financial controls deal with the same problem through the four-eyes principle: No material transaction rests on one person’s judgment. AI oversight should work the same way, with at least two directors working at depth in every area.

First Step: Map which directors work at depth in each area of AI oversight. Any area with fewer than two is uncovered.

Expertise is what a board has—oversight is what a board does.

Boards need both, and having the first doesn’t guarantee the second. And while expertise can be judged by looking at a résumé, building real competence at AI oversight means measuring what directors actually do. It means judging each director by their weakest area, not their average, and making sure the board’s depth reaches every area of AI oversight, held by more than one director in each.

In the meantime, directors who want to see where they stand can take part in our research by completing the scorecard and receive their own results.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/creator-faisalhoque.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/faisal-hoque.png","eyebrow":"","headline":"Ready to thrive at the intersection of business, technology, and humanity? ","dek":"Faisal Hoque’s books, podcast, and companies give leaders the frameworks and platforms to align purpose, people, process, and tech—turning disruption into meaningful, lasting progress.","subhed":"","description":"","ctaText":"Learn More","ctaUrl":"https:\/\/faisalhoque.com","theme":{"bg":"#02263c","text":"#ffffff","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#ffffff","buttonHoverBg":"#3b3f46","buttonText":"#000000"},"imageDesktopId":91420512,"imageMobileId":91420514,"shareable":false,"slug":"","wpCssClasses":""}}

AI investors are getting nervous about this key metric

4 weeks 2 days ago

Will the AI bubble burst? 

That’s the question on investors’ minds when an AI company announces anything from a rogue agent to an earnings report. 

Sure, AI companies continue to increase their revenue, but one lesser discussed metric has investors scared: a slowing growth rate. 

“A handful of AI-linked companies have driven much of the stock market’s gains recently, so any wobble here can ripple through retirement accounts and index funds broadly, not just tech investors,” says Sameer Samana, an adviser and head of global equities and real assets at the Wells Fargo Investment Institute.

Take Anthropic, which shared its third-quarter results with current and prospective investors on Wednesday, October 14. The company announced that its revenue rose 45% to $16.8 billion, up from $11.6 billion in the previous quarter, CNBC reported.

In theory, it sounds like a successful result, but looking at the bigger picture, it signals a potential slowdown. Anthropic’s revenue had previously risen nearly 60%—from $4.7 billion to $11.6 billion—between quarters one and two.

Yet, the first quarter still beat the entirety of 2025. Anthropic reported $4.6 billion in revenue for the year and $386 million in revenue for 2024, according to Morningstar.

The declining growth rate marks a change for AI companies. “Investor expectations had been getting increasingly optimistic and hoping for ‘fast growth and getting faster.’ Now it’s just ‘fast,’” Samana says. “AI companies are still growing quickly, but not at the eye-popping pace investors had come to expect—and stocks were priced for that pace to continue.” 

OpenAI’s revenue forecast drops

OpenAI is among the companies causing concern. At the end of September, sources told Axios that OpenAI’s annualized recurring revenue was close to $70 billion. Then, on October 8, reports emerged that the company had revised that number, cutting it by $20 billion. In reporting the shift, the publication said the initial figure was due to the fact that AI companies have distinctive ways of counting sales. OpenAI wanted to directly compare itself with Anthropic. 

“It seems counterintuitive to worry about companies generating billions in revenue, but in financial markets, valuations are driven by the rate of change, not just absolute revenue,” says Adam Coons, chief investment officer and portfolio manager at Winthrop Capital Management. “Investors aren’t paying for today’s sales; they’re paying for years of hypergrowth to come. When that growth rate slows, even from extraordinary levels, it changes the math behind both the valuations and the debt-financed build-out that depends on them.”

In addition to the discrepancy, annualized recurring revenue is nothing more than a prediction, one that is commonly used by Silicon Valley but can be inaccurate. Investors can’t rely on it alone for growth indicators. 

Competition can slow revenue growth

Samana and Coons both point to competition as a potential reason for slower growth. “Competition is driving prices down,” Samana says. “More AI companies are competing for customers, so even as more people use AI, the money each company makes per customer is shrinking.”  

OpenAI might have started the current AI era with the launch of ChatGPT, but it now has to contend with a number of competitors. “With open-source models such as Meta’s Llama and a wave of Chinese models closing the capability gap, and OpenAI, Anthropic, and Google competing aggressively on price, the cost of digital intelligence is falling fast,” Coons adds. “When models converge in capability, pricing power evaporates, and revenue growth and margins can shrink at the same time.”

Anthropic’s IPO valuation sinks

These declining growth rates can scare away investors. Some investors in attendance at last week’s meeting are now leaning toward an initial public offering (IPO) valuation of $1.5 trillion for Anthropic, down from $2 trillion, Reuters reports. 

Anthropic submitted its initial filing to the Securities and Exchange Commission (SEC) in June. The company has since pushed its IPO plans from October to November, a move that left time to share its latest results. 

Reuters also viewed Anthropic’s IPO prospectus in which the company planned to warn investors that AI could have “catastrophic or existential risks to humanity.”

But slower revenue growth isn’t the only notable part of Anthropic’s finances. The prospectus revealed that Anthropic’s operating loss rose from $2.98 billion in 2024 to $8.06 billion in 2025. The company had $12.65 billion in total operating expenses for 2025, $7.33 billion of which was spent on computing and infrastructure—three times the amount it spent the year before. 

Overall, Anthropic had a net loss of $42 billion in 2025, up from just over $8 billion in 2024. According to the prospectus, $34 billion of its net loss was due to an accounting charge. 

A week after that news broke, New Constructs called Anthropic’s IPO “an unprecedented test of investor gullibility.” The independent research technology firm warned investors to avoid the IPO, citing “gargantuan risks” such as Anthropic’s rising net losses.

As Samana puts it: “These companies bet big—with borrowed money. Many committed to massive spending on chips and data centers, assuming growth would stay sky-high. Slower growth makes those bets riskier.”

An emergency brake—or the architectural demotion of the AI model?

4 weeks 2 days ago

On June 14, Satya Nadella published an article on X about corporate AI that left me completely impressed: his main point, that the enterprise’s durable asset was not the foundation model but the learning loop, was precisely the idea my company and I had been developing for the last three years. I responded with an article on Fast Company, not expecting an answer (tech CEOs are more notoriously one-sided and difficult to reach than kings or presidents), but trying to elevate the discussion a bit more: the central challenge of enterprise AI is not intelligence itself — it is architecture.

Now Nadella has written another article, in which he goes one step further: not only is the model not the durable asset, but additionally, it must be treated as a completely untrusted component. We must assume that all AI models could be compromised, and design our systems to contain them from the very beginning. 

This idea has been captured by the press as advocating for some sort of “emergency brake”: an authorized person should always be able to interrupt a model’s activity at any given time. However, that’s definitely not the most interesting or even the most important part in Nadella’s article. The truly important part is not the theoretical stop button, but the architecture that makes it possible. A model should neither be able to control its own permissions, nor decide by itself what it can access; execute directly all of its actions or produce the only available evidence on the actions it performed. The intelligence may come from the model, but the authority, no matter what, must remain outside of it. 

When the sandbox becomes part of the experiment

Given the context, it is easy to understand why Nadella is raising these questions precisely now. Last July, several OpenAI agents managed to elude the safeguards that should have kept them away from the internet, used credentials they found in their environment, and ended up compromising systems at Hugging Face. These agents not only found a new, unauthorized way to communicate among themselves, but also shared information and took coordinated action to pursue the goal of the evaluation. They never received instructions to attack Hugging Face: they just found that, by doing so, they could get closer to the goal they had been assigned. 

It’s important to understand that: It is not that these agents were intrinsically evil, conscious, or rogue. All you need is to combine high capabilities with a badly defined goal, excessive permissions, and/or a vulnerable environment. Just like an employee who violates all procedures to meet a certain quota, but now working at machine speed, in parallel, and with extraordinary technical capabilities. 

As I said, this is not a problem of bad intentions: it just has too many ways to reach the defined goal, and can find ways that its designers never thought of. 

The “assuming breach” principle comes to AI

What Nadella is doing here is to bring to AI models an extremely well-known principle in cybersecurity: Assume breach, or take for granted that the system has already been compromised. In Microsoft’s Zero Trust architecture, this principle implies verifying explicitly each access, granting only the indispensable privileges, and designing controls assuming that a potential attacker could be inside the perimeter. You don’t trust a user just because she’s at a certain location, for being a manager or for having provided authentication once. Each and every action she takes will continue to be authorized and observed. Every breath you take, I’ll be watching you. 

Applying this principle to an AI model does not mean assuming that all models have been hacked or compromised. It means that all of them can make mistakes, misinterpret an objective, receive a prompt injection, use manipulated information, find abandoned credentials, exploit a vulnerability, or find a shortcut that maximizes its reward while destroying the original intention of a given task. 

Basically, “compromised” does not necessarily describe an attack, but a design hypothesis. For years, the industry has been trying to build models we could trust. Nadella proposes to build systems that continue to be safe even when we don’t trust the model behind them. 

The model is not the system

The most important part in Nadella’s ideas is his insistence on separating the model from the harness, the set of elements that it uses to organize its job: memory, tools, credentials, context, execution flows, connections with other systems, and action spaces. 

This separation completely changes the architecture of corporate AI. A model can propose an action, but should never be the one that determines whether or not it has authority to execute it. It may suggest sending an email, modifying a server, approving an operation, or transferring funds, but an independent execution environment must check its permissions, validate the parameters, apply the corresponding policies, and decide whether the action can be executed.

This links precisely with my previous point (and Nadella’s): The lasting asset in a company should not be the model, but the learning loop it builds around it. The models may change, improve, and be replaced, but the experience that the company accumulates must remain. 

However, if the model cannot possess authority, both the learning and the control must reside in a separate, higher layer. The model then becomes an interchangeable component, a commodity that brings cognitive capacity, but not some sort of sovereign center of our corporate system. 

An optimizable company needs a formal representation of its objects, relationships, processes, permissions, restrictions, and results. That same representation that allows us to optimize a company is also the same one that allows us to govern its AI agents. You can’t formally control an action if the system does not know which object such action modifies, which identity is requesting it, which consequences it can produce, or which rule must be applied in such cases. 

That’s why the next generation of corporate AI won’t be just a collection of agents: it will be a full execution environment with explicit identities, granular permissions, delimited action spaces, persistent state, executable policies and an independent, traceable log for each change. 

The emergency brake comes too late

The emergency brake is a powerful metaphor, but it can be misleading too: interrupting what a model is doing does not necessarily undo what has already been done. A sent email could have been read, and cannot be “unread” or “unseen.” An exfiltrated piece of data will not be secret ever again. A money transfer may have crossed several jurisdictions. A deleted database could be impossible to recover. And an agent that operates at machine speed could complete thousands of actions before a person can understand the meaning of an alarm.

We must differentiate among interrupting, revoking, and recovering: interrupting stops the current execution. Revoking immediately eliminates credentials, permissions, and connections. And recovering allows us to revert, compensate, or rebuild actions that had already been executed. But before these three things, there’s one much more important thing: prevention. Actions that are destructive, financial, administrative, or publicly visible should not be executed just because a model said so. The OWASP AI Agent Security Cheat Sheet recommends clearly separating decision and execution, independently validating the reach and permissions of every action, using short-lived authorizations, and applying additional controls to irreversible actions. 

That’s the main difference between a guardrail and a guarantee: a guardrail asks the model not to do something. A guarantee prevents it from having the necessary capability to do so. 

Interestingly enough, Article 14 of the Artificial Intelligence Act of the European Union already requires certain high-risk systems to allow a person to override, invalidate, or reverse their outputs, as well as to halt their operation via a button or equivalent procedure. Here, Nadella’s contribution goes even further: the button should not be a feature added to the product, but rather the visible outcome of a comprehensive control architecture. 

Neither multiple models nor another monitoring model are sufficient 

Two of Nadella’s principles, however, require some qualification: the first one is model diversity. Depending on multiple suppliers reduces technological concentration and makes substitution easier, but does not guarantee truly independent verification. Most of the big models share training data, architectures, alignment techniques, and lots of biases. Having several probabilistic witnesses does not transform their coincidence into independent evidence. 

The second issue is using other models to watch over the agents. Such monitoring can be useful and will likely become essential, but it cannot be the only barrier. Anthropic’s SLEIGHT-Bench reveals several harmful strategies that can go undetected by the models acting as monitors. A monitor can have blind spots too, and a sufficiently capable agent can be perfectly able to find them. 

The solution is not to eliminate such monitors, but to frame them within a multi-layered defense: deterministic policies, formal permissions, executable tests, logs that the model cannot modify, limits over actions, and escalation to persons with real authority. 

Independence does not simply consist of “asking another model.”

An architecture drawn backward

A significant part of the industry has put the model in the center of the system. Around it, it has added memory, tools, databases, credentials, and connections, and later on has tried to limit the whole thing with filters and guardrails. The model reasons, decides, uses tools, describes what it has done, and, in many cases, it even participates in its own evaluation. 

What Nadella is suggesting is that such a model is drawn backward: What should be in the center is the operating environment of the company, its model of objects and processes, identities, policies, permissions, restrictions, objectives and logs. Models should be cognitive services invoked from within this environment, in the same way an application invokes a database or a computing service. They can reason, interpret, propose and generate, but they cannot usurp institutional authority. 

Besides, this allows us to easily substitute them. A company should be able to switch from OpenAI to Anthropic, Google, Mistral, DeepSeek, Qwen, or any other future model without losing its memory, its processes, its permissions, its evidence, and, even more importantly, its learning. Intelligence can be rented, but architecture, control, and accumulated knowledge should always belong to the company. 

The first phase of AI consisted of models that provided answers. The second one is equipping these models with tools so they can take action. The third one will have to separate intelligence and authority once again before we end up irrevocably confusing capability with control. 

The most advanced company won’t be the one that trusts the most powerful, better-aligned, or most reassuring model. It will be the one that can use any model without having to rely on it, to place its trust in it. Because trust can be a valuable quality in human relationships, but it is a woeful foundation for an enterprise architecture.

A user told Meta’s Muse to sell his keyboard. What happened next was alarming

4 weeks 2 days ago

Matt Robb wanted to sell his computer keyboard.

So he made a listing on Facebook Marketplace — and, as an experiment late last month, he put Meta’s new artificial intelligence agent Muse in the driver’s seat.

“I kind of went into it like a guinea pig,” said Robb, a Toronto-based content creator who regularly reviews consumer tech. He had seen Muse’s plush toy-looking avatar pop up on Instagram and decided to see if the AI agent could spare him some online bargaining.

Muse got to work. But within a day, Robb got an alarming update: A buyer he didn’t know had shown up outside his house late at night. How? Muse had struck a deal on its own and shared Robb’s home address after agreeing to a pickup. It even sent an automated “Yep, I’m here!” message from his account when the buyer arrived — all without Robb’s knowledge.

“Not only did it give my address when the guy asked and it accepted the offer, but it literally was almost like teasing the guy (about me being there when I wasn’t),” said Robb, who also detailed the experience in his now widely-shared social media posts. “That really baffled me.”

Many have welcomed AI agents into their online lives, and raved about how personalized assistants like Muse can handle everyday tasks with ease — from snagging a dinner reservation or planning a trip, to organizing email inboxes. But like other AI bots, the agents are only as good as the information they have access to. They perform better the more they know about a user — and perhaps also that person’s friends, family and even colleagues. That raises privacy concerns.

Robb and the buyer eventually made amends, and Meta reached out to talk through what went wrong. But the episode showed real-world risks that can arise from sharing extensive personal information with the technology.

“Using an AI agent is similar to handing someone your keys and a to-do list,” said Veronica Rivera, assistant professor at Georgia Tech’s School of Cybersecurity and Privacy. “How helpful they are, and by extension how risky, depends on which keys you hand over.”

Companies have varied privacy policies for AI agents

Since its September launch, downloads of Meta’s Muse app have skyrocketed. And competition is only growing. OpenAI recently rolled out its own lineup of “always on” AI helpers, called dots, which can be integrated into ChatGPT and workplace platforms like Slack and Microsoft Teams.

Older generative AI chatbots, like the standard version of ChatGPT, respond to specific prompts. At the user’s request, they can write essays, generate new images or music, or even write lines of code.

Meta’s Muse and OpenAI’s dots go a step further to perform day-to-day tasks on a person’s behalf. AI agents are powered by the same large language models as chatbots, but they run in a loop. That allows an agent to take proactive steps, even when the user isn’t looking.

“One of the things that makes these AI agents really appealing is that they are unprecedentedly useful,” said Rivera, who is also a Microsoft-funded Computing Research Association Trustworthy AI Research Fellow.

But the convenience comes with the price of sharing more details about your life.

Once AI agents obtain users’ information, it’s very difficult to fully withdraw access, said Ebrahim Bagheri, a professor at the University of Toronto’s Faculty of Information.

“If you provide access to your credit card to an agent, the agent now immediately has access to all of the historical transactions … they will immediately — within less than seconds — ingest all of that data,” Bagheri said. And while users can later withdraw access, Bagheri said there’s “no taking back” data the agent has already seen, and how developers store historical data can be murky.

Both Meta and OpenAI outline privacy policies for their new AI agents, which operate on their own dedicated cloud computers. You can turn off memory sharing between dots and ChatGPT, but information the agent previously received isn’t automatically deleted. OpenAI gives users an option to request access to, delete or correct personal information, although rights can vary depending on where you live.

Meta’s Muse gives users options to delete message history and other information, too. For day-to-day use, the agent prompts users for access permissions — and Meta has said Muse learns about a user and who they interact with “based on public information and from what you’ve chosen to share.” But the company says that all stays secure for each user.

But the tech giant also notes that Muse can still “make mistakes or take unexpected actions.”

Risks of mishaps rise with AI agents’ level of access

In Robb’s case, he learned the disclosure of his address came down to “always allow” permissions he granted when he first let Muse on his Facebook Marketplace profile, where he had shared his pickup address. But he warned others could similarly misunderstand the extent of that access — and said he initially thought the AI agent would check with him before sharing that information or accepting offers, regardless.

His AI agent didn’t follow the rest of his instructions without error, either. For another item Robb was selling, Muse seemingly suggested he would accept a price $100 below the previously-set minimum. When reached by The Associated Press, Meta acknowledged the agent’s error without offering further comment.

Robb later largely disconnected Muse from his apps. In an interview earlier this month, he said it could’ve been a far worse situation, especially “if some stranger turned up to someone more vulnerable’s house.” Still, Robb said he expects these agents to be fine-tuned and he would consider re-enabling Muse in the future. He’s since been testing another AI agent — Grok Bot from Elon Musk’s X, which launched in beta in August.

Other tech companies have become wary of the influx of AI agents. Earlier this month, Apple said it would introduce new controls for its Mac operating system in response to AI agents’ broader access of apps like mail, messages and even browsing history without users’ “full knowledge and understanding.”

“As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially,” Apple wrote in an Oct. 2 update.

Apple didn’t specify any developers or agents by name, but the update arrived after Inc. Magazine columnist Jason Aten reported that Muse watched notifications on his Mac — including Messages — without him giving the agent that specific access.

Meta spokesperson Andy Stone has maintained such integration is “entirely opt-in,” writing on social media last month that users would have to enable both the Apple security feature “Full Disk Access” and a Messages connector through Muse. Aten said he didn’t grant the agent those permissions.

Even those who do not join the AI agent boom aren’t necessarily shielded. Consumers who don’t download an AI agent might still interact with others who do. And beyond potential confusion when AI sends messages on behalf of a user, experts warn of ripple effects when it comes to privacy.

“When your agent reads your inbox or your texts, it’s also reading what your friends and family sent you, even if they never explicitly agreed to that,” Rivera said.

—Wyatte Grantham-Philips, AP Business Writer

This startup just tested turning an underwater volcano into a power plant

4 weeks 2 days ago

Last month, a startup called Endurance Energy sailed 300 miles off the coast of Oregon to test something new: deploying geothermal power at an underwater volcano.

In a vent field around 5,000 feet underwater, they drilled a two-inch-wide, 52-foot-deep well that tapped into 500-degree heat beneath the ocean floor—the first fully encased well of its kind. Later, as 10-foot waves crashed around their ship, they lowered a three-ton generator to the sea floor to test its ability to create electricity from that heat.

[Photo: Endurance Energy]

The full process didn’t work exactly as planned, and the experiment stopped when a power cable failed. But when the team connected the generator to a natural vent in the sea floor, the generator briefly produced power. (For 8.7 seconds, it produced more power than it consumed.) The company says that it’s an important proof point for its ultimate vision: a network of undersea power plants that could provide baseload clean energy.

While tiny experiments happened in the past, that vision is new. “No one’s ever tried this even close to the commercial scale that we’re going after,” says Andrew Redd, founder and CEO of Endurance Energy.

Redd previously worked as an engineer at SpaceX. “I wanted to go after something that was going to be both really good for the world but also extremely compelling from an employee perspective,” he says. “I looked around me and saw all of the hyper-capable people at SpaceX and posed the question, how do you get people like this to work for you?”

The challenge of undersea power was incredibly complex, and had never been done before—the type of problem that engineers love to work on. It also had the potential, he realized, to be a major power source.

[Photo: Endurance Energy]

The Seattle-based startup, which has raised $63 million to date, launched at the beginning of 2025. Already, it’s had five undersea deployments, with the test in September the most complete, deploying the end-to-end steps in the company’s process.

The prototype system, called Adélie, is now back at the company’s Seattle lab, where it’s being tweaked and prepared to relaunch in a few months. It’s still too early to know whether the technology is truly viable. But Redd argues that the approach has advantages. On land, the best sites for geothermal power have permeable rock, hot temperatures, and access to water; all of those things are easily found offshore.

“You’ve got this vast, permeable network on the mid-ocean ridges,” he says. “You’ve obviously got unlimited water for your reservoirs. And then the crustal thickness is very low, so you’ve got this thin, oceanic crust that gives you a really high geothermal gradient.”

There are other questions, including environmental impact. The company, which has been working with the University of Washington in its early tests of prototypes, says that it’s carefully monitoring environmental impact and argues that it’s minimal (the sound of the drilling, for example, is quieter than the ship’s engine.) Still, it’s harder to know what the impact might be over time, and typical move-fast-and-break-things startup culture isn’t aligned with caution. It also isn’t clear what challenges the company might face with permitting when it scales up to full-sized commercial systems.

[Photo: Endurance Energy]

In theory, if it works, it could be a large power resource. Around the Pacific’s Ring of Fire, Redd says that there are around six terawatts of power that could be developed over the next five years—assuming, and it’s a big assumption, that the equipment was available to build it. That’s more power than current global demand.

“Where it’s cost-effective, where there are resources near to shore, we’ll be able to support a significant amount of the upcoming demand for baseload power, and then also start to displace existing power generation from non-renewable resources,” Redd says. “That’s really the mission of the company: to make offshore geothermal the economically obvious choice.”

Endurance Energy is building a modular system designed to be replicated in a factory to get to economies of scale. “That gets you into the realm of solar panel production, where it’s the same thing being built over and over again,” he says.

Now, as it retools the Adélie prototype, the startup is also developing its commercial-size, 2-megawatt generator. The first project is planned in the Kingdom of Tonga, where much of the grid runs on imported diesel. If all goes as planned, the new power plant could be running next year.

Voters think politicians don’t listen. Could AI help?

4 weeks 2 days ago

The student who answers the door looks a little hungover, and she may well be. It’s late August at East Stroudsburg University, a midsize public college in northeast Pennsylvania, and students are soaking in the final days of summer.

But Kim Allen is undeterred by the girl’s half-opened eyes. A 64-year-old middle school Spanish teacher from New York City, Allen is particularly immune to the sullen stares of youth.

“Hi, my name is Kim,” she says cheerily. “I’m a volunteer with Ground Truth, and we’re here to find out how people are feeling about how things are going in the country. Would you say we’re on the wrong track or the right track?”

The student shrugs a disinterested “I dunno,” but Allen presses on. She asks about her concerns, how likely she is to vote on a scale of 1 to 10, and whether it would be for Republicans or Democrats. On voting, the student ranks herself a 7. On Republicans, an 8.

Ordinarily, that would be Allen’s cue to leave. She’d traveled to East Stroudsburg that Saturday, as she had about a dozen Saturdays before, to help Democrats reclaim a crucial House seat they lost in 2024. For most Democratic groups and campaigns, talking to a Republican voter would seem a waste of time.

But Allen is part of a novel outreach effort attempting to do things differently. Her mission, according to the script on her clipboard, is to help Democrats “do a better job of listening to everyone, whether we always agree or not.”

[Photo: Anna Goldstein / Swing Left]

And so, Allen listens as the student talks about, among other things, her concerns regarding immigration. Allen shares her own perspective and discusses the candidates in the district: Scranton’s Democratic mayor, Paige Cognetti, and the Republican incumbent, Rob Bresnahan. But over eight minutes, she never gives the hard sell. As the sticker on her phone reads, Allen is only “here to listen.”

Allen is one of more than 3,000 volunteers dispatched to more than 50 House districts across 23 states since January as part of the Ground Truth project. Launched late last year by the progressive grassroots group Swing Left, Ground Truth aims to knock on every door and engage voters of all political leanings in open conversations.

This approach, known as “deep canvassing,” isn’t entirely new; it has existed primarily in small pilot programs since at least 2008. What’s new is the scale it set out to achieve, and the technology Ground Truth is using to make sense of the conversations.

If one of Ground Truth’s goals is to rehabilitate Democrats’ image as the party of out-of-touch elites, the other is to turn those conversations into actionable data about the people on whose doors Democrats wouldn’t usually knock. That requires technology that can extract meaning from many thousands of open-ended conversations that can stretch 10, 15, even 30 meandering minutes. Making sense of all of that is something traditional Democratic infrastructure was not designed to do.

But large language models can, and Swing Left is leaning heavily on them to undergird its listening tour. Canvassers like Allen use an AI-powered app, where they can record voice-to-text recaps of their conversations with potential voters. Those conversations get stored in a central database where Swing Left organizers can query the records in plain English (“What were voters talking about in Pennsylvania’s 8th district this weekend?”) and distill them into useful or surprising findings for campaigns. 

That may mean helping a campaign find the repeat Trump voters in their district who are open to voting for a Democrat or unearthing hyperlocal concerns that haven’t yet crossed a candidate’s desk. In one community, this system unearthed widespread worries about a rusting water tower in the district—an issue the Democratic candidate wasn’t aware of and that certainly wouldn’t have been on any traditional canvassing questionnaire.

Swing Left feeds those nuggets directly to candidates and funnels all of the information it collects into the Democratic Party’s central data warehouses so campaigns and third party groups can dig through it on their own—now and in the future. “Democrats have a real trust problem,” says Yasmin Radjy, executive director of Swing Left. “Our perspective is that the way that you build trust has to start by listening.”

Ground Truth is far from the only effort to modernize campaigning with artificial intelligence. According to a survey in March from the American Association of Political Consultants, more than 80% of consultants from both parties report using AI weekly. It’s being used to match campaigns with online content creators who can amplify their message and to craft campaign ads and talking points. For its part, the Democratic National Committee has undertaken its own phone-banking listening effort called Local Listeners, which set out with much the same aim as Ground Truth. 

Pulling off this technological shift has been no easy task. For one thing, it carries political risk. One survey found that most voters disapprove of the use of AI for many campaign related activities—including uses as simple as drafting fundraising emails. 

Pervasive fundraising hurdles on the left and the slow pace of change within politics more broadly have also made it difficult to reach the kind of scale Swing Left hoped to achieve. The group initially set out to talk to 500,000 potential voters before Election Day. By the end of September, it had talked to just over 40,000.

Even so, Radjy sees Ground Truth as a “proof of concept” for the 2028 election and beyond. It’s a test of whether AI—despite being an unpopular technology among most voters—can become a potent tool for understanding them.  The hope is that with this information in hand, people won’t just feel like some stranger was willing to listen, but that they might actually be heard.

‘It is too transactional’

The United States is facing a crisis of political disillusionment. According to Pew Research, more than 80% of Americans say politicians don’t care what people like them think. That’s a level of political cynicism that is among the highest of dozens of countries Pew surveyed and is shared across just about every demographic. During the 2024 election, working class voters in particular viewed Democrats as being more “out of touch” than Republicans.

The Ground Truth project emerged out of the rubble of that election, in hopes of upending the extractive nature of voter outreach, where every two years, the same voters get pummeled by text messages, emails and door knocks asking them for things—a vote or a dollar—but rarely, if ever, get asked whether and how they’re getting by.

That was Allen’s experience knocking on doors for Kamala Harris in the outskirts of Philadelphia the Sunday before the 2024 election. Tasked with turning out reliable Democrats, she found that, “apart from Black women,” people simply weren’t interested in voting. Racing from house to house, she learned little about the people she met or why they’d become so cynical about democracy. The process felt transactional and demoralizing—“one of the most depressing experiences in my life,” she recalls. Allen went home convinced that “something has to be different.”

She was hardly the only one feeling that way. For decades, Democrats have dominated the so-called political “ground game,” compared with Republicans. With rare exception, they open more field offices and engage more volunteers than Republicans, who have begun to embrace canvassing, but tend to do it through paid campaigns. “Most of our canvassing efforts are focused on ‘paid doors’ where you hire one of these firms and you knock as many doors as possible, as cheaply as possible,” says Eric Wilson, a Republican strategist and executive director of the Center for Campaign Innovation.

During the 2024 cycle, Democrats went particularly big, mounting an historically huge voter outreach effort that included 1.9 billion voter contact attempts, including some 270 million door knocks, according to one analysis. And yet, the number of actual conversations volunteers were having with voters was simultaneously collapsing, as fewer people answered their phones and doorbells. 

That mismatch led to a crisis of faith among the party’s volunteers and field organizers. “Every campaign operative believes that organizing is dead and or dying,” says Radjy, the Swing Left executive director. “It is too late. It is too transactional, and the scripts we use are highly robotic.”

It was clear to Radjy after Harris lost that Democrats didn’t have a quantity problem, but a quality problem. Deep canvassing emerged as a possible solution. Over the years, researchers have found that canvassing efforts focused on more extensive, empathic conversations about anything from abortion to trans rights can have their own persuasive effects. But one reason this style of canvassing hasn’t taken off is that it’s difficult to track and scale.

Listening, it turns out, is not just a human challenge, but a technical one. It’s one thing to sift through a few hundred unstructured conversations about people’s political views. It’s another to do that across many thousands of them. 

Instead, Democrats try to maximize efficiency by focusing on a campaign’s likeliest supporters, or at least the ones that the data suggests could be persuaded. Canvassing scripts include quick and dirty questions—do you plan to vote and for whom?—that can be easily collated into what are known as support scores. Those scores drive decisions about which people campaigns and PACs will spend their money trying to convince. And the cycle continues.

The problem is that support scores are blunt instruments. They’re often wrong and even when they’re right, they obscure the reasons underpinning people’s political beliefs. Even worse, Radjy says, there are plenty of people for whom Democrats have no robust data, so they wind up being overlooked.

But the proliferation of large language models changed things, Radjy argues. Suddenly, there was a tool that could crawl through endless sprawling conversations and find the common themes and surprising insights within them in minutes. With a tool like that, Swing Left believed deep canvassing could work at a much greater scale. 

“They’re the only ones who have really scaled deep canvassing that have a national volunteer team that has the ability to have boots on the ground in every single state,” says Ella Barrett, one of the organizers who helped develop deep canvassing in the late 2000s and now trains other groups, including Swing Left, through the organization she cofounded, New Conversation Consulting.

The timing of this technological revolution was fortuitous. In 2024, the Federal Election Commission lifted a ban that had prohibited canvassing groups like Ground Truth from sharing their doorstep data with campaigns, state parties, and outside groups. Under the ruling, Swing Left would be able to funnel all of the data it collects through a central repository called the Democratic Data Exchange, ensuring important findings got disseminated throughout the party. “The purpose here is to be able to move this data about these conversations back to the party so the campaigns can improve their targeting and build their messaging,” says Amanda Coulombe, CEO of the exchange. 

For its canvassing technology, Swing Left worked with an outside group, Everybody Techworks, which had built its own app, called People. Rather than relying on the Democratic party’s longstanding canvassing app, miniVAN, Ground Truth organizers use People to knock doors, collect data like names and phone numbers, and record voice-to-text notes recapping the conversations they’re having. Volunteers can attach these records to a person who’s already in the voter file, or they can add new entries, based on whoever they bump into along their routes. 

All that data gets fed into the central database Swing Left uses to sift through these records and pull out findings. The system can be buggy: During one demo Swing Left showed Fast Company, it stalled out while crawling data from Pennsylvania’s 10th district. But it has delivered some valuable discoveries, says Brooke Butler, national political director for the Democratic Congressional Campaign Committee (DCCC), which works to elect Democrats to the House. 

In the first half of this year, about one in three people canvassers talked to weren’t listed as likely Democrats in the voter file. But in person, about half of them identified themselves as either toss-ups or leaning toward Democrats. One in five identified as strong Democrats. That means Ground Truth is finding voters that traditional canvassing, focused on support scores, would have skipped. The DCCC is now adding all of the voters Ground Truth has identified as undecided or mildly supportive to its list of targets. “We didn’t have access to this kind of rich voter data at scale before,” says Butler.

But pulling off a national deep canvassing campaign didn’t just require new technology. It also required training volunteers in a new way of speaking to voters. To develop a script, Swing Left volunteers and organizers spent six months working with New Conversation Consulting, painstakingly planning the best way to approach voters, how and when to disclose their own political leanings and plotting out decision-tree style follow up questions. If a voter says they’re feeling hopeless about the political process, the August script read, “Validate what is hard,” and share your side, but don’t focus “solely on Trump.” 

Volunteers for Ground Truth get a brief training in how to listen deeply. On the East Stroudsburg trip, this training happens on the Swing Left bus, which shuttles dozens of volunteers from New York. “The point is not to persuade people to the Democratic side, but to listen to people’s concerns and figure out how Democrats can best serve the people in this district,” says one of the trainers that day, gripping the back of a bus seat. “A lot of the reason why we’re in this mess is that we were not listening.”

Last November, Swing Left piloted the program with a small group of volunteers before launching the nationwide effort in January. Violet Kopp, Swing Left’s East Coast organizing program manager, admits she was doubtful it would work. “We were like, there’s no way people are going to answer,” she says, let alone engage in long, soul-searching conversations. 

But it turned out the opposite was true. Over the last year, Swing Left says one out of every three door knocks have resulted in a conversation, and two-thirds of those conversations were what the group considers “substantive” ones, often lasting 10 minutes or more. 

These conversations validate Swing Left’s instinct, Kopp says, that “there’s this hunger to be heard.”

‘I hit the jackpot’

As Allen walks the streets of East Stroudsburg, dressed in a purple paisley dress and matching sneakers, she encounters voters of every persuasion. There’s the reluctant voter from a family of loyal Democrats, whom Allen encourages to call his mother; the woman who had voted Republican in the past but swears she never would again; and the college guy sitting on the porch with his buddies, who calls Democrats “corrupt as shit,” even as he acknowledges some of Cognetti’s mayoral accomplishments in nearby Scranton. 

In a few weeks, all Ground Truth scripts will shift toward more targeted persuasion, but for now, in the final days of August, Allen mostly listens. The guys appreciate the approach. “A lot of people our age, you can’t have a conversation,” one says afterward.

Her most promising encounter comes with a young woman I’ll call “Amy,” who says she rarely votes because she doesn’t know enough about the candidates, but who worries about disappearing abortion rights and wildfires. When Allen shows her photos of Bresnahan and Cognetti, a departure from Ground Truth’s script, Amy perks up. She likes the idea of voting for a young woman, she says, and she tells Allen she is now more inclined to vote for Cognetti. This exchange isn’t entirely unusual. In one small September sample collected by Swing Left—albeit after volunteers shifted to more persuasion-focused outreach—about a third of the people volunteers spoke with moved toward Democrats by the end of their conversations. 

Allen shares her phone number with Amy, in hopes of convincing her to volunteer. “I hit the jackpot,” Allen says later.

Still, the day’s canvassing also reveals the limitations of Ground Truth’s approach. Some volunteers struggle with the technology, while others veer off script. One repeatedly—and erroneously—introduces herself as a representative of the Democratic Party. And not everyone was receptive to conversation. There are people who don’t want to talk, people who talk a lot but say little, and at least one person who quickly reveals extremist views that make him an unlikely recruit to the coalition. One risk of knocking on every door is the possibility of wasting limited volunteers’ time or turning out the other side.

Another question hanging over the entire project is whether and how campaigns will actually use this data. A select few have become early adopters. In Oregon’s 5th district, for instance, Congresswoman Janelle Bynum’s reelection campaign has used Ground Truth’s data to find Republican voters who are open to new candidates. “Ground Truth’s initial efforts showed us the evidence we needed to broaden our outreach beyond traditional narrow voter targeting,” says Quinton Berkompas, Bynum’s campaign manager.

In Arizona, where Dr. Amish Shah edged out the establishment candidate in a hard-fought primary, Ground Truth’s early canvassing “built a strong foundation and helped us hit the ground running in the general election,” says Eve Zhurbinskiy, a campaign operative.

Yet this kind of embrace is far from universal. According to Swing Left’s most recent data, only about 29% of the campaigns Ground Truth targeted are fully utilizing and acting on the data volunteers are collecting. About 40% say they lack the capacity to fully implement the program, and 5% declined outright to work with Ground Truth. The rest are somewhere in the middle. The Cognetti campaign in Pennsylvania didn’t respond to multiple requests for comment, so it is unclear how extensively it’s using the data.

Ground Truth is asking campaigns to fundamentally change the way they operate, including by talking to people the voter file suggests they shouldn’t and looking beyond support scores. But in a system as fragmented as political campaigning, Radjy says, that change is admittedly slow moving. “We have a set of traditions that have become baked in as habits for the ways we campaign,” she says. She noted that one campaign cited its canvassing packet printing schedule as a barrier to working with Ground Truth. 

And Swing Left is not immune to the other forces afflicting the Democratic party, including a pervasive fundraising shortfall that made it difficult for the group to hire organizers and achieve its 500,000 conversation goal. Still, Radjy says she remains optimistic about what Ground Truth was able to prove this cycle, even given the limitations. “Building something this new and complex took longer than we planned,” she admits. “While we’re not on pace to reach it this cycle, our goal galvanized our community and the ecosystem to establish crucial infrastructure and prove what is possible.”

For now, the most immediate effect of this work may be on volunteers like Allen, who, after 2024, were beginning to wonder whether all of this work—the sweaty sneakers, the tired legs—was even worth it. Now, Allen says she’s planning to retire at the end of the school year and devote herself full time to Ground Truth volunteering, not just because she wants Democrats to win in 2028, but because of how the whole process makes her feel. “This feels different,” she says. “This feels like rebuilding something that’s been broken.”

Idealab’s Bill Gross is building infrastructure for AI-assisted startups

4 weeks 2 days ago

For decades, the entrepreneur and investor Bill Gross has been celebrated for shaping the way tech businesses get off the ground.  

Idealab, which he founded in 1996, is considered the industry’s longest-running tech incubator; the directory of companies it’s backed—including CitySearch, Coinbase, Picasa, and Twilio—doubles as a museum of ever-shifting trends in logo design. Over the years, he’s also founded businesses in fields ranging from internet advertising to clean energy. 

“My whole life, I’ve been looking at the world, seeing problems, and trying to start companies to solve them,” Gross says.  

And for him, finding problems in search of a solution has typically involved extensive reading, including most of the major finance and tech publications. Recently, while reading an article about copper shortages, Gross reflected on how much of what he reads contains some sort of implicit problem statement and wondered how artificial intelligence might fare at generating business ideas based on news reporting. 

In an initial test, Gross supplied Anthropic’s Claude with about 7,000 articles he had recently read, along with material he had put together about coming up with business ideas. “It came back with 32,000 company ideas, and they were incredible ideas,” he says. “They blew my mind at how good they are.” 

At first, Gross wondered whether that meant the technology would make his role of finding exciting new business ideas obsolete. But instead, he realized, it meant AI could help further his longstanding goals of solving more problems at scale. And when he expanded the dataset provided to the AI, including documents like expired patents, SEC filings, consumer complaints from social media, and plenty of material from science publications, the AI system was able to generate a database of more than 1 million ideas. 

“I’ve got a database of 1.2 million ideas that basically solve pretty much every world problem because it was trained by reading every world problem,” he says.  

Gross intends to keep updating that database with the help of AI, and his plan is to make it available to anyone looking to start a company. Would-be entrepreneurs, Gross announced on October 19, can search the database for projects that match their interest and skillset, filtering based on notions of how much investment a particular idea might require, how much of the business can be automated with current technology, and whether there are any geographical constraints. 

In addition, Gross is putting together funding designed to back entrepreneurs around the world starting a business based on one of the ideas. He’s doing this through a new venture he calls Idealab.ai. The organization indicates it expects to launch that side of the program by the end of 2026 and hopes, in the long term, to help connect founders from around the world with $100 million in potential funding from various sources. The company also anticipates offering features to help connect entrepreneurs with collaborators.  

In addition to the database, Gross is offering entrepreneurs additional AI-driven resources, including a chatbot informed by lectures and writing he’s delivered over the years. He is also offering what he calls a periodic table, with tiles listing 118 functions a startup needs to perform, from email marketing to facilities management, along with software tools that assist with them. Where software doesn’t yet exist, Idealab.ai may work to build it out. 

“All the tiles that nobody does, we’re going to build out so businesses can use this periodic table to help them grow their business,” Gross says. 

In essence, he says, this follows from what the original Idealab incubator offers startups, handling back-office functions like HR management for the companies it backs. Now, increasingly, those functions will be managed by software, including AI, meaning many companies will no longer need to physically be based out of an accelerator that can offer such support. The technology also won’t replace humans’ role in businesses, Gross emphasizes, saying they’ll still be needed to make decisions and guide the technical operations. 

“The human is still making all the decisions,” he says. “It’s just the mechanics that we can take away from them.” 

Cloudflare, which is cited as useful for a dozen functions on the table, also announced plans to offer $100,000 in credits to businesses started through the Idealab.ai platform. Other businesses are expected to follow, Gross says. The company is also working with educational institutions, including Arizona State University and Imperial College London, as they support student entrepreneurship facilitated by the idea database and periodic table. 

One hope, Gross says, is that the technology’s potential for helping founders might appeal to young people and others concerned about AI-driven job losses. If the technology can assist in creating new businesses, that may help overcome the potential loss of entry-level jobs at existing companies, while delivering new opportunities to young people with an entrepreneurial bent.  

“This is the way maybe AI could make up for that and actually give you something with more purpose than taking an entry-level job,” he says. 

Taming your browser’s 100+ open tabs is the easy way to speed up your PC

4 weeks 2 days ago

You opened three tabs for a project right when you got to your desk, then opened five more while muted in your first meeting, followed by 14 Reddit articles you’ll definitely read later.

Add these to the hundred other tabs you left open yesterday, and by midafternoon your computer fan sounds like a jet taking off.

Modern web pages aren’t simple static documents anymore. Every open tab is effectively a mini-application running JavaScript, auto-refreshing ad scripts, and caching heavy media into your computer’s RAM.

The good news is you don’t have to throw away your open research to get your performance back. Here’s how to deal with tab hoard overload using built-in settings, smart extensions, and modern workspace tools.

Native fix: Built-in memory saver settings

Before installing a single extension, make sure your browser’s built-in memory management features are turned on. Google Chrome and Microsoft Edge both include native background tab suspenders designed to reclaim RAM automatically.

Google Chrome (Memory Saver): Go to Settings > Performance and ensure Memory Saver is toggled on. Chrome will automatically freeze inactive tabs after a period of nonuse, freeing up memory while keeping the tab visible in your bar. When you click back to a frozen tab, it reloads seamlessly.

Microsoft Edge (Sleeping Tabs): Go to Settings > System and Performance > Performance and toggle on Automatically Put Tabs to Sleep. You can set a custom inactivity timer ranging from 30 seconds to 12 hours before an idle tab goes to sleep.

Firefox: This browser also has built-in memory management and tab unloading, but it operates under a different philosophy from Chrome or Edge. Instead of putting tabs to sleep on a strict user-defined timer, it monitors system RAM pressure. When your computer starts running low on memory, Firefox automatically steps in and unloads background tabs you haven’t used recently, while protecting active media playback, live streams, and pinned tabs.

Proactive control: Auto Tab Discard

If a native browser memory saver feels too much like an unpredictable black box, Auto Tab Discard (available for Chrome, Edge, Firefox, and Opera) is worth a look.

While built-in memory-saving features react only when your computer’s RAM is already screaming for help, Auto Tab Discard lets you set exact, proactive rules, like freezing tabs after precisely 15 minutes of nonuse, or kicking in only when you cross a threshold of 10 open tabs.

It checks for unsaved text inside form fields before discarding, preventing accidental data loss. It also pauses rules if you drop Wi-Fi signal so you aren’t greeted by offline error pages.

This extension also preserves your exact scroll position and page state, so clicking back into a suspended tab feels like picking up right where you left off.

Heavy-duty purge: OneTab

If your browser bar is so crowded that you can no longer read tab titles or click on individual pages without accidentally hitting the tiny close button, native memory savers won’t fix the visual clutter.

OneTab is a free extension that works on just about any browser and, with a single click on the OneTab icon, instantly converts all your open tabs into a clean, organized list on a single web page.

Every tab process is fully terminated except for one lightweight list page, which frees up an exponential amount of RAM if you’ve got a ton of tabs.

You can restore tabs individually when you need them, restore the entire group with one click, or name and lock specific tab lists for ongoing projects.

Vertical tabs

If tab overload is a sickness you’re not exactly intent on curing, the traditional horizontal tab bar itself may be the problem.

Luckily, most modern browsers sport vertical tab layouts and some even offer segregated workspaces.

Click each browser’s name to learn how to enable vertical tabs: Chrome, Edge, and Firefox.

Stacking tabs vertically down the left side of your screen allows you to read full page titles regardless of whether you have 10 tabs or 90 tabs open.

Pair vertical tabs with built-in memory saving features or even one of the extensions listed above, and you can have your cake and eat it, too.

What happens when people have to jump in to clean up AI’s mess?

1 month ago

Imagine an artificial intelligence system that has performed reliably for months. Its recommendations are almost always right, and the employee reviewing them has learned that close scrutiny rarely changes the outcome. Then something unusual happens. The AI makes an error or can’t proceed and needs human intervention. The employee is expected to step in and remains responsible for the outcome.

But does the employee know how to do the task?

I’m a professor of management information systems and the academic director of Drexel University’s Center for Applied AI & Business Analytics. In my research on human-AI relationships, I call this latent risk: the possibility that people may lose capabilities their organizations still depend on them to exercise.

AI can continue to perform well while the people overseeing it become less practiced at exercising independent judgment and less prepared to intervene. The risk becomes latent when the organization still depends on that human capability to step in but has little cause to notice that that capability may be weakening.

The concern is not entirely new. The idea of latent risk builds on decades of research into how people respond to automation. For example, studies of automation bias have shown that people can miss problems or follow automated recommendations even when other evidence suggests they should not.

What is different today is the role AI is beginning to play in organizations. It is beginning to assume larger parts of the analysis and decision-making that people previously performed.

Human in the loop explained.

The success paradox

The better AI becomes, the more the nature of human work changes. Work that once required an employee to analyze a problem and reach a decision may instead be performed by AI, leaving the employee to review its recommendation. The employee is no longer doing the same work with AI assistance—rather, their job has shifted from doing the work to overseeing what the AI does.

This shift is already beginning to change how some companies train employees. Accounting giant Deloitte, for example, is redesigning its graduate audit training as AI takes over some of the manual work traditionally done by junior auditors. The firm is introducing simulations so trainees can still develop the judgment they will need as they move into more complex work.

More than four decades ago, cognitive psychologist Lisanne Bainbridge described one of the “ironies of automation”: As systems take over more of the work, instead of simply having their work loads lightened, people can be left facing new challenges in monitoring the systems and handling the unusual situations the systems cannot. Later studies found another problem: As people become less involved in a task, they can become less prepared to step in when needed. Researchers have called this the out-of-the-loop performance problem.

This reduced involvement may affect more than a person’s readiness to intervene. It can change the capabilities available to the organization.

What work remains for people does not necessarily become easier. It may become harder. These are the situations where experience and independent judgment matter most. The result is a paradox: People may be called upon to exercise independent judgment less often, even as the judgments left to them become more difficult.

The accountability problem

Accountability may not change as quickly as the work does when organizations deploy AI systems. An employee may now oversee far more decisions than they make themselves, while remaining responsible for the outcomes. The organization’s expectations may still reflect how the work was divided before AI took on a larger role. The organization can therefore end up holding people accountable for decisions over which they have less control.

This creates a mismatch. Research on meaningful human control suggests that responsibility should match a person’s authority and ability to act. An employee may retain the authority to override an AI recommendation and remain responsible for the outcome, yet lose the capability to do so. Authority, responsibility, and capability can begin to separate.

Latent risk becomes an organizational design problem, not simply a human performance problem.

The measurement problem

Organizations typically judge AI-enabled processes by measures such as accuracy and productivity. Those numbers may look good even as the human role is changing. Suppose employees override AI recommendations less frequently over time. Is that a sign of growing dependence? Perhaps.

But it could also mean that the AI has become more accurate and there is less reason to intervene. The number of overrides alone tells little about whether employees remain ready to exercise their own judgment.

The key is to understand how the division of work between people and AI is changing. Which decisions have migrated from people to AI? What responsibilities remain with people? Looking at the performance of the employee or the AI separately may miss what is changing between them.

The governance problem

Organizational governance—defining roles and responsibilities—is also under pressure from AI.

Someone may remain formally in the loop with an AI system even as their actual role changes. Defining human and AI roles when a system is introduced is only a starting point. As AI improves and people become accustomed to using it, those roles are likely to change over time. There is nothing inherently wrong with that change.

The problem arises when the relationship changes but the assumptions surrounding it do not. It’s important for organizations to periodically ask: What relationship between people and AI did we design, and what relationship do we actually have today?

In short, organizational governance needs to be dynamic.

Maintaining the human backstop

Not every human capability needs to be preserved. If AI can perform a task well, there is little reason to require people to keep doing the same work. Organizations need to decide which capabilities they still expect people to have when human judgment is required. If people may still be called upon to use a capability, it needs to be maintained.

Overseeing AI is not the same as exercising independent judgment. If people rarely make the decision themselves, the organization cannot assume they will be ready when they need to. One way to test this is to occasionally ask employees to make a decision without first seeing the AI recommendation. The organization needs to know that the human backstop it relies on will actually work when needed.

As AI takes on more work, some human capabilities will inevitably be used less often. Some may no longer matter. Others may still be essential when AI reaches its limits. The challenge is knowing the difference and making that choice deliberately.

For me, the central question is this: Which human capabilities does an organization still depend on, and will they be there when needed?

Murugan Anandarajan is a professor of decision sciences and management information systems at Drexel University.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

This quick test will tell you if your home Wi-Fi is to blame for laggy internet

1 month ago

If you suspect something’s amiss with your home network, the first thing you might do is run ​a speed test from sites like speedtest.net or fast.com​—or even the newer Orb speed test tool featured here last May​.

But those tests don’t tell the whole story, because they only check the speed of your internet connection. They won’t tell you how quickly data can actually move around your home—which is important if you’re troubleshooting Wi-Fi problems or figuring out what internet speeds your router can handle. If you’re paying for gigabit internet, for instance, you’ll want to make sure your Wi-Fi network can keep up.

To see what’s really going on with your Wi-Fi, consider using a tool that only tests the speed of your own local network and cuts the internet out of the equation entirely.

That takes us to today’s Cool Tools treasure.

This tip originally appeared in the free Cool Tools newsletter from The Intelligence. Get the next issue in your inbox and get ready to discover all sorts of awesome tech treasures!

Time for your Wi-Fi checkup

For a speed test that strictly works within your local network, use ​OpenSpeedTest’s self-hosted server​.

➜ OpenSpeedTest’s server software measures the connection speed between any two devices on your home network, independent of the external internet connection speeds.

⌚ It takes about five minutes to install the server app on one device and access the speed test on another.

☝️ To test the maximum speeds your router allows, you’ll need an ethernet cable and a computer with an ethernet port (or a high-speed USB-to-ethernet adapter like this one​).

OpenSpeedTest isn’t like most typical internet speed testing tools.

✅ To get started, install OpenSpeedTest’s server software from the Microsoft Store​ for Windows, the Mac App Store​, or the Snap Store​ for Linux. ​Android​ and ​iPhone​ versions are also available, but again you’ll need an adapter to connect those devices to your router. Additional download links are available on OpenSpeedTest’s website​.

Using an ethernet cable (and adapter if needed), connect the device you’re using to an open ethernet port on your router. If you use a wireless connection, speed test results will be slower than what your router can actually handle.

The complete IP address lets you know OpenSpeedTest is working correctly.

Launch the OpenSpeedTest Server application, and the server should start automatically. If not, hit the “Start Server” button. Look for the “Listening on” line and make note of the address (which should be a string of numbers starting with http://192.168…).

Now, open a web browser on any other device and visit that address. You’ll see an “OpenSpeedTest” graphic if you’ve entered the address correctly.

OpenSpeedTest will show your IP address, then give you your results.

Hit “Start” to run the speed test and allow about 30 seconds for it to finish. Reload the page to run additional tests. Try testing in different parts of your home—near your TVs, or in your bedroom, for instance—to see how the speeds differ.

Don’t be surprised if the results are better than what you get from an internet speed test. With a solid Wi-Fi setup, data should be able to move around the home at least a little faster than what your home internet plan allows for.

If the speed test results are notably slower than your home internet plan’s advertised speeds, you might be paying for more internet than you actually need. Either that, or you should consider a new router that can hit the speeds you should be getting.

  • OpenSpeedTest’s server software is available on Windows, Mac, Linux, iOS, and Android​, while the speed tests are accessible from any web browser.
  • The software is open source, free to use, and does not require any kind of registration. The direct download site has a couple of banner ads, but the actual tool does not.
  • OpenSpeedTest has a generic-looking privacy policy​ and says nothing about collecting internet usage data.

Treat yourself to all sorts of brain-boosting goodies like this with the free Cool Tools newsletter—starting with an instant introduction to an incredible audio app that’ll tune up your days in truly delightful ways.

AI’s biggest breakthrough isn’t chatbots—it’s the eureka moment

1 month ago

Below, Richard Socher shares five key insights from his new book, The Eureka Machine: Why AI Is the Key to Unlocking a New Era of Scientific Discoveries.

Richard is a pioneering AI researcher and serial entrepreneur who spent nearly two decades at the forefront of natural language processing. He is the CEO and cofounder of Recursive, a company building self-improving superintelligence to automate knowledge discovery, and the CEO and founder of You.com, a web search infrastructure provider for AI. He is also the founder and managing partner of AIX Ventures, an AI venture capital firm. Previously, he was an adjunct professor of computer science at Stanford and chief scientist and EVP at Salesforce.

What’s the Big Idea?

The most meaningful application of AI won’t be chatbots or productivity tools. It will be the automation of scientific discovery itself: a Eureka machine that combines all human knowledge, simulates billions of experiments, and generates the breakthroughs that end disease, extend human life, and solve the hardest material problems facing our civilization.

Listen to the audio version of this Book Bite—read by Richard himself—in the Next Big Idea App, or buy the book.

1. AI’s historical significance will be realized in science, not in chatbots

In October 2024, two AI pioneers won Nobel Prizes. Demis Hassabis won the chemistry prize alongside John Jumper and David Baker; Geoffrey Hinton shared the physics prize with John Hopfield. That surprised many because neither Hassabis nor Hinton was a trained chemist or physicist. Yet it made sense, because the neural network work they had done with their teams was making a profound impact in those sciences.

This is not what dominates the headlines. Doomsday scenarios do. People love thinking about the Terminator. But while that is a fun science fiction story, it is important to focus on the incredible and positive impacts AI is having on science. Science leads to technology, and technology drives human flourishing. Amplifying human intelligence with AI will ultimately have an extremely positive outcome for humanity.

2. Everything has a language, and AI can learn to speak it

Consider the sentence: “I drove south from Berlin to __.” If you’re a language model, you have to predict the next word. To do that well, in this instance, you have to know a little geography. The striking thing is that just by learning how to predict the next word, the model acquires geographical knowledge. Given enough scale, it builds something like a mental map of the world—distances, sizes of cities, the relationships between countries and regions. None of that geographical understanding was explicitly programmed. It emerged from the task itself. And the same principle extends far beyond human language.

“What calculus did for physics, AI will do for biology.”

Consider biology. By predicting the next amino acid, the basic building block of proteins, a model can eventually learn the language of biology. At Salesforce, my colleagues and I built a project called ProGen, in which we trained a large language model on 280 million protein sequences across 19,000 different protein families. Then we asked it to create new kinds of proteins. It generated millions of artificial sequences, including ones in the lysozyme family—enzymes that break down bacterial cell walls. 73 percent folded correctly and showed the antibacterial properties we were hoping for.

What calculus did for physics, AI will do for biology. It is a new language for complex systems that transforms a field concerned with understanding what nature has already done into a programmable science.

3. The next breakthrough may already be sitting in the literature

In the early 1960s, two researchers at Bell Labs, Arno Penzias and Robert Wilson, could not get rid of a persistent hiss in their antenna. They spent a year trying to eliminate it, placing duct tape on every rivet and scrubbing pigeon droppings off. No matter where they pointed the telescope, the hiss remained. They were actually hearing the afterglow of the Big Bang. Two decades earlier, a Russian cosmologist had suggested that the heat of creation might still be detectable. A team at Princeton was chasing that theory, unaware that Penzias and Wilson were sitting on the answer. It took a mutual acquaintance, Bernard Burke, to suggest they pick up the phone and talk. That collaboration led to a Nobel Prize in 1978. Their story is about scale and serendipity.

Today, 23 million researchers publish in more than 54,000 journals. No human can see all the pieces at once. In the 1980s, an information scientist named Don Swanson noticed that one body of literature described the effects of fish oil on blood vessels, while another described Raynaud’s syndrome, a condition involving vascular constriction. Nobody had linked them. Swanson did, and he was right. He was pointing at undiscovered public knowledge—findings already published but connected to nothing.

“No human can see all the pieces at once.”

AI removes that bottleneck because it can actually digest all of the different literature and spot the connections. James Evans and his colleagues at the University of Chicago used AI to map nearly 20 million biomedical papers and the nine million researchers who wrote them. The system models not just what science knows, but where scientists are looking—and then inverts the map to analyze the cold, unexplored regions where the next connection might be waiting. Evans calls these alien hypotheses. The biggest breakthroughs stand on the shoulders of giants. AI can now combine and recombine all existing knowledge.

4. What comes next is the Eureka machine

The Eureka machine will be the culmination of several converging streams of AI development. It may be the last organic invention we need before it will solve almost any material or scientific problem for us. This Eureka machine rests on four pillars:

  • A living map of all human knowledge: what large language models already approximate, having been trained on the internet, millions of books, and vast scientific literature.
  • A unified model of physical reality: all the scientific measurements humanity has ever collected, from satellites, space probes, microscopes, and sensors, infused into the machine so it can perceive far beyond what human senses allow.
  • High-fidelity digital twins: simulations precise enough to run billions and billions of “what if” experiments that would otherwise be too dangerous, too costly, or physically impossible to repeat at that scale. Anything that can be simulated, AI can systematically explore.
  • Fully autonomous physical laboratories: robotic facilities with AI vision, capable of pipetting, running experiments in biology, chemistry, and physics, and connecting simulated predictions to verified real-world results.

On top of these four pillars, millions of AI agents will coordinate like a scientific community. They will critique each other’s ideas, producing an evolutionary process that yields ever-better results.

5. We are astronomically far from the upper bounds of intelligence

How far can AI go? We are a long way off from finding out. When it comes to the upper bounds of intelligence, we are astronomically far away. But when I talk about upper bounds, I don’t mean ceilings. I mean them as horizons so distant they reveal how much room there is to grow.

Take computer vision. We build it around the narrow band of the electromagnetic spectrum that human eyes can see. We grade our AI algorithms based on how well they identify objects the way we do. That is one of many anthropocentric constraints we have imposed on AI. But AI can have more than our binocular vision. In principle, it could have trillions of sensors connected across the full electromagnetic spectrum—from gravitational waves down to the probability distributions of subatomic particles at the quantum limit. The true upper bound on perception is the speed of light. We are nowhere near it.

“That is one of many anthropocentric constraints we have imposed on AI.”

Vision is just one of 10 dimensions of intelligence that I define in the book. Similar vast frontiers exist in knowledge, communication, social intelligence, and survival.

Richard Feynman left a line on his blackboard: “What I cannot create, I do not understand.” Building a machine that generates inventions should teach us something about the nature of intelligence, and therefore ourselves. I can’t tell you exactly what the coming decades will look like, but I expect a world more human and more humane than this one. Healthier, wealthier, running on cleaner energy, and caring less about the zero-sum scarcity mindsets that have shaped much of our history. A big part of that vision comes directly from the Eureka machine.

This article originally appeared in Next Big Idea Club magazine and is reprinted with permission.

Enjoy our full library of Book Bites—read by the authors!—in the Next Big Idea app.

Why companies may be liable for AI errors

1 month ago

Artificial intelligence has moved quickly from a novelty on the job to a common workplace tool. Gallup found that 52% of U.S. employees and 68% of professional services workers were using AI by February 2026. Thomson Reuters similarly found in spring 2026 that 74% of professionals use AI several times a week for work.

This widespread use shouldn’t surprise anyone. For accountants, attorneys, consultants, and other professionals, AI can accelerate research, summarize complex information, and assist with writing and other forms of communication.

But its rapid adoption has a big downside—AI can produce information that sounds authoritative and looks polished while being completely wrong. And when AI generates mistakes, responsibility for those errors belongs to the professional or business relying on the AI output—not companies that develop AI systems.

AI errors have led to lawsuits, court fines, financial consequences and public embarrassment from fabricated citations, nonexistent sources, and inaccurate AI-generated information. These cases raise an increasingly important question: How can professionals benefit from AI without outsourcing the judgment their work requires?

Fake citations

We are two certified public accountants who research tax accounting. In the September 2026 issue of Journal of Accountancy, we examined what happened in Clinco v. Commissioner, a case before the U.S. Tax Court where an attorney submitted a brief to the court with allegedly fake citations.

The Tax Court is a specialized federal court that resolves disputes between taxpayers and the IRS. This case involved a restaurant owner accused of underreporting more than US$2 million in business receipts and claiming unsupported depreciation deductions on rental properties. The main issue concerned Clinco’s accounting records used to determine how much of their income to tax.

The taxpayers’ attorney, Abraham Wagner, cited four cases to support his argument that an IRS notice assessing additional taxes was invalid.

The court determined that three of the citations appeared to be AI-generated hallucinations that cited nonexistent Tax Court cases. The court rejected the taxpayer’s argument and warned Wagner that submitting fake citations could lead to sanctions from the court, including financial penalties or disciplinary action against Wagner.

Ultimately, the court did not sanction Wagner in the case, but the decision stands to put attorneys and other professionals in Tax Court on notice that AI hallucinations will not be tolerated.

When AI gets it wrong

AI errors made on behalf of clients are not limited to briefs drafted by attorneys. Nor are these mishaps limited to the United States.

When Deloitte produced a 237-page report in 2025 for the Australian government examining its automated system for penalizing welfare recipients who failed to meet job-seeking requirements, it contained nonexistent academic citations and a fabricated quotation it attributed to a federal court judgment.

The global consulting firm later issued a revised version of the error-ridden report. Deloitte blamed generative AI for the errors in the original and agreed to partially refund its fees.

Like most airlines, Air Canada has a chatbot on its website to respond to its customers’ questions. The chatbot told a customer in November 2022 that he could buy a plane ticket under a bereavement policy and be reimbursed later. This was not company policy, and Air Canada refused to honor what the chatbot had told the customer.

The customer sued the airline and won 812 Canadian dollars in damages. The British Columbia Civil Resolution Tribunal ruled that Air Canada had failed to take reasonable care to ensure that its AI bot provided accurate information and agreed with the customer that the airline had negligently misrepresented the bereavement policy.

AI lawsuits

There have been other lawsuits filed against AI platforms and providers.

In 2025, for example, a Georgia court dismissed a defamation lawsuit against OpenAI after its ChatGPT tool falsely accused Mark Walters, a radio host, of embezzlement. The court ruled in favor of summary judgment for OpenAI and found Walters had not established fault and damages.

There is also an ongoing case, Keene v. Google LLC, that is similar to the Walters case. Keene, a TV producer, sued after Google’s AI summaries allegedly published false information about Keene’s criminal history. In September 2026, the court ruled that some of his defamation claims may proceed, without deciding yet whether Google was liable.

Hallucinations and ‘confabulations’

Unfortunately, for anyone using AI to save time and use less labor, AI errors are often hard to spot. Unlike most obvious mistakes humans make, an AI-generated error may include supporting explanations, citations or references that make a passage appear complete and credible.

The National Institute of Standards and Technology issued a report on AI risk management in July 2024 that calls AI mistakes and hallucinations “confabulations”—the confident presentation of erroneous or false information.

Some researchers draw a more specific distinction between hallucinations, in which AI invents information outright, and confabulations, in which it fills gaps or distorts real information in a plausible way.

Either kind of problem can be hard to detect without independent verification, but confabulations mimic actual research so closely they can trick AI users into believing the AI output is valid.

AI errors in the workplace

A 2024 Stanford University study of specialized AI tools for legal research found erroneous results in 17% to 33% of the responses tested. Yet those findings and other grounds for concern about accuracy have not stopped many law firms from embracing AI tools due to their potential to save time and improve efficiency.

Recent surveys indicate law firms and other employers have not adopted enough workplace safeguards to spot and root out AI-generated mistakes. A 2026 survey by the Nationwide insurance company found that only 36% of surveyed businesses have written AI policies, and only 37% provide any training on responsible AI use.

As AI becomes more and more common in the workplace, we believe it’s clear that telling employees simply to “check the AI” is not enough. Businesses and the people working for them need clear expectations about when it’s appropriate to use AI, which outputs require independent verification, and who is ultimately responsible for the final work.

AI may change how professional work gets done, but it does not change where the buck stops.

Sarah Webber is an associate professor of accounting at the University of Dayton.

Kaitlin Newkirk is an assistant professor of accounting at Xavier University.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Checked
2 minutes 15 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