Cory Doctorow warns AI bubble could trigger economic collapse as tech valuations soar to $1.4tn

How to burst the AI bubble: Strike at its roots

Last year we featured a lengthy interview with tech journalist/science fiction author Cory Doctorow about his book, Enshittification: Why Everything Suddenly Got Worse and What To Do About It. The prolific Doctorow is back with a provocative new book that serves as a follow-up of sorts, focusing on AI and related issues: The Reverse Centaur's Guide to Life After AI. Doctorow doesn't actually enjoy talking about AI, but he's constantly being asked to comment on it. "I made the tactical error of being sick of talking about AI," Doctorow told Ars. "So I wrote a book about why I think it's a dumb thing to keep asking people to talk about, and now I have to talk about it." Reverse Centaur is Doctorow's attempt to "sort out the bullshit from the material reality." In automation theory, per Doctorow, a "centaur" describes a human augmented with a technology, like machine learning, or even just driving a car, or using autocomplete. A reverse centaur "is a machine head on a human body, a person who is serving as a squishy meat appendage for an uncaring machine," Doctorow said in a speech last December. He gave the example of an Amazon delivery driver, surrounded by AI cameras monitoring their driving, who essentially serves as a peripheral to the delivery van. Being a centaur is generally viewed as a positive thing; few people relish being a reverse centaur. And yet the AI industry seems intent on using those tools to create more reverse centaurs. It's one thing to incorporate AI tools in the medical field to help radiologists process X-ray images and spot possible tumors they might otherwise have missed. It's quite another to fire nine out of 10 radiologists and let AI make the diagnoses, with the remaining radiologist solely responsible for checking the AI's work -- and, ultimately, taking the blame for any errors. Doctorow is not virulently anti-AI; he uses AI tools regularly and sees potential in many of those tools as useful plugins or cool new apps. But he is nonetheless alarmed at all the hype surrounding AI, the enormous capital expenditures, the unrealistic expectations and self-serving messaging, and the potentially catastrophic economic consequences when the AI bubble inevitably pops. "The bubble doesn't want cheap useful things," Doctorow said. "It wants expensive 'disruptive' things: big foundational models that lose billions of dollars every year. When the AI investment mania halts, most of the models are going to disappear, because it just won't be economical to keep the data centers running. The collapse of the AI bubble is going to be ugly. Seven AI companies currently account for more than a third of the stock market, and they endlessly pass around the same $100 billion IOU. AI is the asbestos in the walls of our technological society, stuffed with wild abandon by a finance sector and tech monopolists run amok. We will be excavating it for a generation or more." Naturally Doctorow has some ideas about how to push back against the prevailing narrative of AI's inevitability. Ars caught up with him to learn more. Macmillan Publishing Macmillan Publishing Macmillan Publishing Macmillan Publishing Macmillan Publishing Macmillan Publishing Ars Technica: We touched briefly on AI last year when we chatted about your prior book. Reverse Centaur seems like a natural outgrowth of that. Cory Doctorow: Enshittification is primarily a thesis about how firms in the absence of constraint get tilted to the bad, but it's also a thesis about how the constraint of competition, when it falls away, produces all kinds of perverse outcomes. One of those perverse outcomes is that firms that have saturated their markets can no longer grow, and they have to find other markets. There's a ticking bomb when you saturate your market because it's only a matter of time until investors start to worry that you're not a growth stock, you're a mature stock. Mature stocks trade at a small fraction of the multiple that growth stocks do. There's an enormous amount of liquidity in growth stocks, which means that you can use growth stocks to grow. You can buy other companies with shares, and shares are an endogenous substance that you make on the premises by typing zeros into a spreadsheet. Firms with growth stocks can grow by typing zeros, whereas firms that are mature, they have to use money if they want to grow, and you're not allowed to make money on the premises. If you do, the Treasury Department shows up and takes you away in handcuffs. So you can see why firms would be very anxious to maintain the perception that they have room for growth even after they have 90 percent market shares. "The capital markets have the object permanence of a toddler, and they would lose a game of peekaboo if they were drafted to play in the league." That's why those firms started promoting stories about how they were going to conquer imaginary markets. Imaginary markets have no agreed-upon valuation because you just made them up. Unless you can turn an imaginary market into a real market pretty quickly, you need to come up with another imaginary market and announce that this is the new imaginary market you're going to conquer. It's easier than you'd think because the capital markets have the object permanence of a toddler, and they would lose a game of peekaboo if they were drafted to play in the league. So you can say, "Oh, actually, it's not metaverse. It's crypto. It's not crypto. It's Web3. It's not Web3. It's something else." And the markets will forgive you, provided you do it quickly enough. But something different happened with AI. It is much, much bigger in terms of capitalization than anything we've ever seen -- not just bigger than other tech bubbles, bigger than other bubbles. When I wrote the book, capital expenditure (CapEx) globally was $700 billion, now it's $1.4 trillion. Meta wasted $60 billion on the metaverse. They spent $150 billion in the last three years on AI, and they say they're going to spend another $150 billion this year. So this is a much bigger bet, and it raises the question: If the material basis for this is creating a narrative so that you can continue to grow by dint of having a highly liquid growth stock, what's the ideological basis? Why are people willing to make such a bigger bet? Some of it is that there's more "there" there with AI. It's real computer science. It was remarkable 10 years ago, when a couple of computer scientists and their grad students took some existing techniques, applied them in a new way, and got a very surprising result that turned out to not only produce dividends the first time around, but to have somewhat linear returns on investment, which is not usually the case. There was a lot of low-hanging fruit in AI, although it's tapering off now because, as they say in finance, anything that can't go on forever has to stop. So we're losing the end of that growth period in terms of returns to scale. Ars Technica: Why do you think AI is so appealing to political and business leaders in particular? Cory Doctorow: It's not just that it makes for a good demo. AI really appeals to a fantasy that I think all of us have to some extent, but that powerful people really have, of a world without people in it -- because hell really is other people. You can't get stuff done without other people helping you. You can't have romance without a romantic partner. You can't have social media without people to socialize with. You can't play a board game, or do a startup, or build a bridge, or build a house, or do politics without other people. And other people stubbornly refuse to organize everything they do to make you happy. Particularly if you're rich and powerful, it's very galling. So AI is very attractive. One of the reasons DOGE fired so many government workers was because it played into the fantasy that you can have a government without government employees. In the corporate sphere, it's the fantasy of a business without workers, because every corporate leader is haunted by the secret fear that if they don't show up for work, everything goes on just fine. But if the workers don't show up, everything shuts down. Maybe they're not really driving the car, maybe they're strapped in the backseat with a toy steering wheel. If that's the case, AI will let them wire the toy steering wheel directly into the drivetrain. So you can have an amazing idea as a corporate visionary, and you don't have to have any ego-shattering confrontations with people who know how to do things, who tell you you're actually an idiot. You just type some stuff to the chatbot, and it shits out your product. If you combine those two things -- the material necessity to have a growth narrative and the ideological attractiveness of a world without people -- you get $1.4 trillion in CapEx for a sector that is turning over $50 billion a year and has to replace all of its assets every 24 to 30 months. Ars Technica: You raised an interesting point recently on your blog: workers actually wanted earlier technological breakthroughs and often had to fight to get them into the workplace. With AI, people are more likely to feel that the technology is being shoved down our throats; some workers are even required to use it. Cory Doctorow: I think that's entirely right. One of the things that I've been attending to a lot lately is the difference between the bubbles that we had before and the bubble that we're having now. People will say, "Oh, Amazon wasn't profitable, and it became profitable. And the web wasn't profitable, and it became profitable. The web was a bubble." Of course the web was a bubble. You don't get pets.com and all those Super Bowl ads without a bubble. But it is a very obvious error of logic to say, "Once, there was a thing that lost money and then it made money, therefore, if you are losing money, someday you'll make money." "AI is the money-losingest thing our species has ever done. We have never lost as much money as we've lost on AI." The thing that made the web profitable was not that it was unprofitable, it was things like good unit economics, where every time someone started using the web, the web got less unprofitable. Every time a web user used the web again, the total profits generated went up. Every generation of web technology made the web more profitable. That's the opposite of AI. Every AI customer loses money for the company, every use of AI by that customer loses money for the company, and every generation of AI loses more money than the last one. AI is the money-losingest thing our species has ever done. We have never lost as much money as we've lost on AI. Another giant material difference is the social reception. If you look back to the business press of the aughts and the late '90s, it's full of hand-wringing editorials about how bosses will cope with workers who are smuggling in the web. You look at those same press outlets today, and it's full of people saying, "What are we going to do about the fact that no one in the workplace wants to use AI?" -- along with ads for firms that will spy on your workers for you so that you can punish the workers who refuse to use AI. Ars Technica: AI nonetheless does have thoughtful, sensible defenders. Cory Doctorow: One of the paradoxes that I try to explore in this book is the workers who are not fools, who are historic good, reliable narrators of their own experience, and who tell you that AI is making their lives better. The foundational idea of science fiction is that what the gadget does is less important than who it does it for and who it does it to. I call those people centaurs. They are workers who are assisted by technology and who decide how that technology is going to assist them. Whereas the workers who hate it are workers who are being asked to produce more with AI at the expense of quality, at a higher speed, at the expense of their own wellbeing, and who understand that they're being recruited to be what Dan Davies calls accountability sinks -- to take the blame when the AI screws up their job. Once you put it that way, it's very easy to see why some workers would say, "Oh yeah, I found a thing that AI is good for and I use it, and that's fine. I'm even excited about it." And why other workers would be like, "This is making me miserable." It's the difference between the words on the Greek temple, "Know thyself," and your boss shining 16 cameras in your face and going, "I know you better than you do. And by the way, I think you could work an extra hour a day without breaking a sweat." Ars Technica: You make a point of emphasizing that you are not fundamentally anti-AI, despite sharply criticizing the industry. Cory Doctorow: I have many comrades who describe themselves as anti-AI, and I've had some very spirited, productive but heated debates with those people because I don't think AI is exceptional. That means that I don't think it's exceptionally evil. The argument that it's the fruit of the poisonous tree, that it was made by bad people in bad ways, so you shouldn't use it -- I think it's very foolish. That is not the merit on which we judge technology. You can talk about whether giving money to these companies is bad. I think it is. You can talk about whether the environmental impact of using foundation models is unsustainable and unsupportable. I think, by and large, it is. But that is not to say that statistical inference using convoluted deep neural networks is bad or -- and this is where I get into many arguments -- that scraping the web to train a convoluted neural network is bad. I think it's fine. Scraping is good actually. I think it's very dangerous to say, "The way that we're going to fix the problems we have with AI is to make it illegal to make a record of what's on the Internet." I think that's catastrophic. That's how we never again will know what was on CBS News before it turned into Chud News. Everything Nate Silver ever published on his website was just zeroed out by Disney. You can only see it at the Internet Archive because we scrape. It's just bonkers to say, "It is theft to make transient copies of works, to analyze those transient copies, to publish the results of your analysis." Those are all socially beneficial activities, and we will all lose if we prohibit it -- not least because the firms that creative workers are worried about it, the big media companies, are extremely capable of entering into arrangements with the Big Tech companies to license their corpuses to them in order to try and put us all out of a job. If we get the right to decide who can train an AI with our work, our bosses are just going to modify our contracts to say, "Great, you now must license that right to me. And it's non-negotiable." Failure to learn from that lesson is not tragedy. It is farce. Rather than ask for a new copyright law, we could make a new labor law, because the only people who've ever beaten AI are the Hollywood screenwriters and actors. And the reason they were able to beat them is because uniquely, among workers in America, they are exempt from the Taft-Hartley Act's prohibition on what's called sectoral bargaining, which is when all the workers in a sector bargain with all the employers in a sector. Now, there are so few workers in America who aren't media workers, who care about copyright, that it rounds to zero, but every single worker in America would benefit from extending sectoral bargaining across the board. Ars Technica: It could be catastrophic, economically speaking, when the AI bubble finally bursts. But you point out that there might very well be something useful left over when that happens. Cory Doctorow: I advise to go long on laser tag arenas because you can definitely turn a data center into one of those. There's not much else you can do with them, unfortunately. A bubble is a way for insiders to pump, and then dump, some mania to the normy investors, to people who've been flushed into the capital markets because they've been denied a defined benefits pension and who are only really offered market-based pensions. That means you have to be the sucker at the table. You have to put your money into the market if you don't want to die homeless and starving after you retire. The dot-com bubble was very bad. It separated a lot of pension funds and ordinary investors from their money, but it left behind something very useful. In the early years of the aughts, there was, amidst the carnage, quite a liberating vibe where all the stupid money went home and you could get servers for pennies on the dollar. Everybody knew how to code. A generation of humanities undergraduates were induced to drop out of university and learn Python, Pearl, and HTML, and a lot of them were really creative. Your rent dropped by two-thirds in San Francisco. I bought six $1,200 Steelcase Leap chairs, still in the plastic wrap, from a failed dot-com guy on a sidewalk on 19th Street in the Mission District for $25-$50 each and used them as a dining room set for the next 10 years. So there was a very productive residue that was left behind by the dot-com bubble. It gave rise to a more robust form of the web, Web 2.0, full of things that were more useful, more interesting, more thought-through, more creative, more innovative than the stuff that the bubble threw off in Web 1.0. There are other examples of bubbles that are less likely to throw off that residue. Around that time, we also had Enron. Enron produced nothing, although I do have a pad of Enron stationery that a friend in Austin bought at the bankruptcy auction and sent to me. "We can distinguish between bubbles with productive residues and unproductive bubbles while still not saying that bubbles are good. Bubbles are bad and destructive." So we can distinguish between bubbles with productive residues and unproductive bubbles while still not saying that bubbles are good. Bubbles are bad and destructive. When the cryptocurrency bubble bursts, all that's going to be left are shitty monkey JPEGs and worse Austrian economics. But when AI bursts, you're going to be able to buy GPUs for pennies on the dollar. You're going to have your pick of applied statisticians, many of whom are very creative and have interesting ideas for things you could build with AI, but are stuck building the things their bosses want to build. There are going to be these open source models that have barely been touched. Any time someone tries to optimize them, they find so many opportunities to make them run on lower-end and commodity hardware. DeepSeek was a spin-out of a Chinese hedge fund; the fund gave them $6 million and said, "Go play with these open source models. See what you can squeeze out of them." When they launched, their model was so good running on commodity hardware that the market did a mass sell off, $600 billion in 24 hours -- the largest 24-hour decapitalization of any firm in the history of markets. If you've got cheap hardware, and you've got applied statisticians, you've got these open source models and you've got a technology that fundamentally is interesting, and has done useful things, and will do useful things in the future -- that's a better setup than one in which we're all running around arguing about whether the word-guessing program is going to wake up, become God, and turn us into paperclips. Ars Technica: You also push back a little on the "AI is coming for your job" messaging. Cory Doctorow: I think we have to distinguish between the AI doing your job, and the AI being incapable of doing your job, but your boss is such a sucker that he fires you and replaces you with the AI anyway. There's infinite evidence for the second one. I think that there's very little evidence for the first one, at least so far. A lot of the stories we've heard, when you interrogate them, just turn out to be nonsense. There's a chapter in the book about how many of the demos for AI have just turned out to be people in India pretending to be robots. The most egregious example was when Amazon announced that cashiers were now out of a job because now you could just walk into [an Amazon Go store], grab stuff off the shelf, and walk out again, and the AI knows what you took. There wasn't an AI. It was three people in India watching each customer through a network of cameras in the ceiling trying to guess what you put in your bag. I think there's lots of things that skilled workers will ask AI to do that will help them do their jobs. There's lots of things that skilled workers will ask AI to do that they'll be wrong about and that won't help them do their jobs. And there's probably space at the margin to replace humans with AI, at least in some cases. But the idea that we're at a "jobspocalypse" is such a self-serving narrative. If you're trying to convince people that the way you're going to turn $1.4 trillion in CapEx into more than $1.4 trillion in revenue is by convincing bosses to fire workers and replace them with chatbots, you have to have a story about how the chatbot can do anyone's job. Here's a wager. If you ever have the opportunity to interview Dario Amodei or Sam Altman, I want you to ask them this. Someday, you will retire. Right now, I want you to make a binding decision. Will the thing that wipes your ass and takes care of you when you are too old and frail to take care of yourself be a person or an AI? We're just going to use whatever it is that's around at that time, and you get to choose. I think we should ask anyone who says they know how to fix things, would they themselves go to an old folk's home run according to the principles they're establishing? Ars Technica: We are now starting to see news stories about how companies that invested in AI are suddenly getting hefty bills. Cory Doctorow: They're getting the bill because the AI companies are trying to get out before they're stuck holding the bag. They want to do IPOs, and to do IPOs, they need to clean up their balance sheet. So they're like, "I bet these [companies] are pretty price insensitive. Let's just jack it. Let's go from a 90 percent subsidy to a 40 percent subsidy and more than double everyone's prices. They'll hang in there." And then you get the CTO of Uber saying, "I'm not sure why we put AI in the business to begin with, and I really don't know why we'd use it if it was $20,000 a seat. So, I don't know that we are going to use AI anymore." This is quite a backfire. It actually shows you how insulated these people are from any sense of how their products are received, from what people think of them, from the actual fundamentals of real businesses that have to bring in more money than they spend. It's weird. I'm hardly a captain of industry or one of the great champions of markets, but I do understand that, by and large, firms should bring in more money than they spend if they are to be an attractive investment prospect. Ars Technica: We hear plenty about the negative aspects of AI. What do you like about it? Cory Doctorow: I have a couple of local models on my computer, which is just a framework laptop running Ubuntu. It doesn't even have a GPU. I use Whisper to transcribe audio. I will sometimes want to cite something I've heard in a podcast and not remember where I heard it. One time, I just threw the last 30 hours of audio I'd listened to at Whisper, and it shot out verbatim logs that were good enough that when I searched the full text, I could find it. And it gave me time codes so I could check the transcript. That's amazing. The idea that I might someday have a computer full of audio and video files with full text indexing is great. I could even imagine conversational interfaces to that: "Where's the photo of my daughter at her birthday party where she's dressed like a pirate?" AI doesn't have to be 100 percent accurate for that to be useful. It doesn't have to be free from false positives. It can just be okay. That stuff's running on your own computer. It's not burning down a rainforest. It's not consuming the last three drops of potable water left in Nevada. There is a certain kind of person who is performatively horrified by AI: "But, but, but that's energy you wouldn't have used." I'm like, "You have never said that about someone who turns cell shading on while playing an MMORPG." Everything you do with your computer burns electricity. I've been using another chatbot where I paste my daily blog post in and say, "Find my typos." It finds a lot of the errors that are normally not caught by a regular spell checker: doubled up words, punctuation marks, or words that are actual words but are misspellings for other words. When you dial up the sensitivity to the point where it actually catches all of those, it also gets a lot of false positives. That's fine for 1,500 to 3,000 words. I never feed it a book. On a 100,000-word manuscript, it's going to give me thousands of false positives and it just won't be useful. I treat this like a plugin to my word processor. It's fine. Sometimes it's good and sometimes it's not. I have a friend, Patrick Ball, who is the best programmer I know, and he founded and works at an NGO called the Human Rights Data Analysis Group (HRDAG). They're one of the most important NGOs that no one's ever heard of. What they do is very rigorous statistical extrapolations of information that's not in the record, about wars, civil wars, coups, oppressive states, and they're used for truth and reconciliation, human rights tribunals, war crimes trials. They've worked on every high-profile war crimes trial in the century. Patrick is using a bunch of Copilots to write software to do a lot of special-purpose stuff. For example, they work with Innocence Project of New Orleans, which has exonerated a bunch of [wrongly convicted] people. They can go through all the arrest reports from the New Orleans PD and find the ones that have linguistic correlates that match successful exonerations. Then they give those to the lawyers who would otherwise just be starting alphabetically or chronologically, sorting to the top the ones that are most like the ones that led to a successful exoneration. It's not like they're asking the chatbot to write a brief for them, but this is a hugely important function, and it is getting innocent people out of prison. You don't want innocent people in prison. That should be the least controversial thing in the world. That's just good, and the proof is in the pudding.

'You can't make billions without hurting people': Cory Doctorow on Elon Musk, the AI bubble and bosses' cruel fantasies

The writer who coined the word 'enshittification' tells us why AI will never deliver what it promises - and why it still appeals so much to those in power A "centaur", in automation theory, is a person assisted by a machine, and a "reverse centaur", hero of Cory Doctorow's new book, The Reverse Centaur's Guide to Life After AI, is a "human who is conscripted into acting as an assistant to a machine". Every warehouse worker who ever had to urinate in a water bottle because they couldn't otherwise meet the fulfilment targets set by an algorithm is a reverse centaur. Reaching into the future, everyone who has to sit in a self-driving truck to make sure it doesn't crash, presumably on minimum rather than truck-driver wages, is a reverse centaur; as is every lawyer no longer on lawyer's money checking Gemini's command of precedent, every indie band scraping a living doing covers of AI-generated hits, and so on. That, anyway, is the promise: AI is coming for your job, and it is coming for your kids' jobs, and there is no point fighting it because the future's already here. Wiping out the world of work, and with it our ability to sustain ourselves and live autonomous lives, is only the beginning, if you listen to AI's architects. Elon Musk has called it the single greatest threat to human civilisation, Sam Altman has said it will "most likely lead to the end of the world" and Dario Amodei, CEO of Anthropic, memorably forecast that AI would come to see us the way we see animals: cute to have around but ultimately a resource to be exploited. "AI people claim they're about to create God, by teaching words to a word-guessing programme," Doctorow says. "It's grandiose." Doctorow bursts into our video-call conversation as if we're already two hours in, his delivery puckish and urgent. He's not happy with his new AI camera, which "supposedly tracks your face, and then doesn't, and just points in arbitrary directions. I mean, this is where my camera thinks my face is," he says, and indeed, it is nowhere near his face. "Thanks, camera. Glad I'm wearing trousers." AI cannot and will never render us obsolete, Doctorow says. "It's a conjuring trick. That's probably the most important thing to get across." A machine has been invented that is really good at building sentences by predicting what word would usually come next, and we invest it with meaning, insight, omnipotence. But we're "imputing intentionality to this thing that intends nothing. It's not because, objectively, it seems intentional, but because, in a state of nature, we don't encounter sentences that don't have sentence writers, we don't encounter images that don't have painters, and so on." We marvel when it does things right, and conveniently ignore what it gets wrong, or indulge its "hallucinations", which is just a fancy word for "errors". "Where I think the word 'hallucination' is useful," he says, "is not to describe what the AI is doing, but what we do when we encounter a word salad, and we impute a writer to the word salad." If you think AI can become conscious, he suggests, it's because you've forgotten what consciousness is. But that doesn't mean there's no threat. This technology can absolutely wreak global havoc; it'll just be of a very old-fashioned kind. A vast amount of investment has gone into AI. "When I wrote this book [last year], it was a $700bn bubble. It's a $1.4tn bubble now. The only thing worse than a $1.4tn bubble is a $2.4tn bubble, which we're headed for," he says. Nine tech companies in the US account for 35% of its entire stock market valuation, which was illustrated rather sharply when the war in Iran had a greater impact on European and Asian stock markets than America's - they were "insulated" by the dominance of the tech sector, people said at the time. But insulated might not be the right word. Doctorow describes "two poles in finance law. One is Stein's Law: anything that can't go on for ever eventually stops. And the other is Keynes: the market can remain irrational longer than you can remain solvent. So it's hard to predict when bubbles are gonna pop. But it's easy to predict that bubbles will pop." Doctorow, 54, is in Los Angeles; he divides his time between there and London (his British wife, Alice Taylor, runs the BBC's AI Creative Lab, which isn't ironic but it would slow us down to explain why not). He came to mass attention as a tech writer with his book Enshittification: Why Everything Suddenly Got Worse and What to Do About It; it's hard to believe it was only published last year, and impossible to remember what we did without that word. People now use "enshittification" for everything, from the degradation of public services post-austerity to climate-crisis-related chaos events, when in fact Doctorow's proposition had quite a specific use regarding tech: giant platforms lock you in and then make your experience worse on purpose. "I'm not frustrated by that at all. I think it's glorious. My first two languages are English and Yiddish, languages that don't have language academies, where dictionaries are descriptive, not proscriptive, where words change meaning." He's Canadian, his early life dotted with signs of an independent spirit and tech-curious mind: he dropped out of high school and graduated from Seed, an "alternative" school in Toronto; he spent time in the 00s setting up OpenCola, a free peer-to-peer software company, then editing Boing Boing, which was vastly popular for a time as an online muster point for discussing tech, futurism and the left. He has an honorary doctorate in computer science from the Open University, has held numerous academic residencies and is professor-at-large at Cornell University. Alongside his nonfiction, he's an extremely prolific sci-fi writer, three-time winner of the Prometheus award (for young adult novels Little Brother, Pirate Cinema and Homeland). He writes graphic novels, authored a children's picture book named after his daughter, Poesy, who is now 18, and publishes a newsletter almost daily. I wonder whether he's dismissive of AI's takeover because he finds writing, creating and communicating so effortless that maybe he doesn't realise that for many people it's like pulling teeth, and they wish a machine would do it. Absolutely not, he says - but more important than all the things AI can't do, and will never be able to do, from call centre operations to radiology, is to understand the motivation behind all the feverish claims for it. Why, he wonders, are capital allocators allocating so much capital to this? It's because of a promise as old as the loom: that bosses will be able to replace their workers with machines. This isn't just about money - often, when the machine is ultimately not as good as a human, or needs so much supervision from a human that it becomes more expensive, there's still tremendous appetite for automation. "The one thing a boss does not want is co-determination. Bosses are haunted by the knowledge that even though they fancy that they're driving the car, if they don't show up, everything continues to work. Whereas if the workers don't show up, everything shuts down. And so perhaps they're in the back seat with a toy steering wheel. AI is the promise of wiring that toy steering wheel directly into the drivetrain of the car. It's products without product designers. Workplaces without workers, screenplays without screenwriters, movies without actors, hospitals without doctors and nurses. This is the promise." Once that promise is made, of course, it feeds an insecurity in the workforce, because if we all really were obsolete, that would be pretty consequential, and we can already see harbingers in our youth unemployment figures and our self-service checkouts. But then we become our own (second) worst enemy, because of what Doctorow, quoting Lee Vinsel from Virginia Tech, calls "criti-hype" - critique that both feeds off and feeds into the hype, and makes the doomsday scenarios more plausible. This may never have been more pronounced - government ministers have spoken openly about all the time and money they'll save once routine tasks are automated. "Do you not remember when they said cryptocurrency would replace all of the world's financial systems? They told us that the metaverse would be the default, that we wouldn't have tourism or sex any more," Doctorow says, with vaudeville outrage. "We have such poor object permanence!" (This is fancy, child-development vocabulary for "really bad memories".) Another powerful AI critic, the journalist Karen Hao, has argued that when apocalyptic claims are made for AI, there's often a veiled threat behind it: "Let us experiment as we wish, have our datacentres, because otherwise Chinese companies will get there first." Doctorow agrees. "You don't want a Confucian God. You want an Old Testament God. Different smiting." Essentially, though, he thinks the big talk has a simpler motivation: anything to keep the investment flowing. Whenever an exercise in automation fails and is abandoned - you may remember Amazon's staff-less grocery stores, which actually required three people to be constantly watching each shopper on CCTV and guessing what they were putting in their baskets - it never dents the AI cheerleaders' confidence that they're speeding towards a post-worker world; nor, so far, the enthusiasm of investors. It's got to the point where it's more important to keep the narrative afloat than to consider whether it's realistic. "You don't have to believe that a face cream is gonna make women look younger to believe that there will be women who will buy a face cream marketed to make them look younger. It doesn't have to work. Are investors investing in the face cream because they think the face cream works? What I'm saying is maybe they're investing in the face cream because they're familiar with patriarchy." The problem is, people can see AI all around them, working better today than it did yesterday - turning out more convincing AI slop, answering their personal problems, turning fully clothed women into naked ones, analysing 18th-century poetry. It feels as if it can do anything, but none of that is the material basis for the investment. "I think it's very important to be critical of things like Elon Musk's Grok AI that allow users to turn out child porn [in January, Musk's Grok AI was found to be capable of generating non-consensual sexualised images of people, including children, and xAI was rolling out 'technological measures' to restrict Grok's ability to undress people in photos]. But we can't pretend that there's anyone who invested in his AI company, who looked at the prospectus and said, 'I see you have a line item here for the expected revenues for child porn - that's great, that's why I'm here.'" In other words, go ahead and ban child porn - "we should do that anyway" - but no one's going to take their money out. "If we want to actually target the harm that Musk is doing, we have to make his investors think he can't make money." Fundamentally, this is a Marxist analysis, I suggest: that labour and capital are elementally at odds, and the latter will exploit the former even if all value is lost in the process. "I don't necessarily disagree, but that's not the argument I'm making. The argument I'm making is that bosses resent and work relentlessly to end co-determination as a class." But that's the same, I insist, and he shrugs, as if to say: we can bicker about Marx after we've stopped this runaway train. What Marx didn't have to contend with is the trillionaire, even in old money. "There is something about being very rich and insulated from the consequences of your actions that makes you solipsistic. You cannot make billions of dollars without hurting lots of people. And you can't hurt lots of people without, in some sense, believing that they're not really people." He raises Musk's publicly acknowledged ketamine use. "I have a chronic pain condition, and I've had ketamine administered, and one of the things about ketamine is that it feels like the whole world is a thing you imagined. Like you're the only real person in the world. I don't think it's a coincidence that Elon Musk calls the people who disagree with him NPCs, non-player characters, because he doesn't think they're really real." The larger problem than Musk's personality, or even your average billionaire's, though, is that bubble: "The more of the stock market there is wound up in it, the more economic harm there will be in the blast radius of AI, which is not to the capital allocators who've given them $1.4tn to play with - it's everyone else. We have a quarter of a century's experience now with popped bubbles, and we use them as an excuse to do the austerity that our politicians dream of doing anyway."

A.I. could improve access to financial services -- or enable "digital redlining" | Fortune

This is the web version of Eye on A.I., Fortune's weekly newsletter on artificial intelligence and machine learning. To get it delivered weekly to your in-box, sign up here. Last week, a Dutch court ordered the government in the Netherlands to stop using a machine-learning algorithm for detecting welfare fraud, citing human rights violations. The system, called System Risk Indicator (SyRI) in English, was being used by four Dutch cities to spot individuals whose benefits applications should receive extra scrutiny. It gathered information from 17 different government data sources, including tax records, vehicle registrations and land registries. But the cities using SyRI did not run every application through the system -- they only deployed it in poor neighborhoods where many residents are immigrants, often from Muslim countries. The court ruled that SyRI violated the "right to private life" enshrined in European human rights law. The application of SyRI, it said, could lead to discrimination against individuals based on their socio-economic status, ethnicity or religion. It also said SyRI did not seem consistent with the requirements of Europe's stringent data privacy law, GDPR. Although the judgment only came from a district court and is subject to possible appeal, the decision is likely to set an important precedent within European Union -- and it ought to reverberate elsewhere too, as societies around the world come to grips with how to apply fairness in a world of A.I.-driven risk models. Nowhere is this more relevant than in the insurance sector, which is turning to machine-learning algorithms more and more in order to improve underwriting. Last week, I had a fascinating conversation with Daniel Schreiber, the co-founder and CEO of the New York-based insurance startup Lemonade. He shares concerns that the increased use of machine-learning algorithms, if mishandled, could lead to "digital redlining," as some consumer and privacy right advocates fear. But done right -- and with the right measure of fairness -- he thinks machine learning has the potential to increase access to financial services and decrease cost. To ensure that an A.I.-led underwriting process is fair, Schreiber promotes the use of a "uniform loss ratio." If a company is engaging in fair underwriting practices, its loss ratio, or the amount it pays out in claims divided by the amount it collects in premiums, should be constant across race, gender, sexual orientation, religion and ethnicity. He admits that this means it is entirely possible that some categories of people -- Schreiber, who is Jewish, uses the example of Jews -- could be charged more on average for property insurance, because, for instance, their religious practice involves lighting candles in the home for certain holidays, and lighting candles might be correlated with a higher risk of house fire. But, he says, no individual should be charged more because he or she is Jewish. It might turn out that a particular customer isn't religious and doesn't light candles. That's why it is important not to ask people about their religious affiliation -- that would be discriminatory. The key is for the insurance company to gather data that actually equates to risk: Do you light candles in your home? In order for it to work properly, insurance companies will need to gather more data about customers, not less. Right now, Schreiber admits, the regulatory winds seem to be blowing in the opposite direction (especially in Europe, as the SyRI case shows). Most insurance regulators don't understand machine learning. "That creates a fear of the unknown," he says. What's more, scandals such as Cambridge Analytica make people reluctant to share more data. But Schreiber says customers might be willing to share more information if the insurers were transparent about why they needed to collect this data, how it was being used, and that it might result in customers paying a lower premium. I wasn't entirely convinced by Schreiber's argument. If insurers become that much better at pricing risk, won't many more people simply become uninsurable? (This is what happens in health insurance if companies are allowed to cherry-pick customers, excluding those with pre-existing conditions.) Also, won't people who live in impoverished neighborhoods still be forced to pay more for coverage, even though they may have little choice over where they can afford to live? Many poorer areas have higher risk of crime and fire, leading to higher home insurance premiums. (In fact, U.S. law prohibits policies that have a "disparate impact" on a protected class of people, unless a company can prove a legitimate business necessity for the policy.) Schreiber told me that governments could mandate charging those who live in wealthy areas or who have high household incomes slightly more in premiums, and then using this excess to subsidize the premiums of those who live in poorer neighborhoods. But, he said, this was a discussion separate from the one about whether the underwriting model itself is fair. What do you think? Feel free to write in and let us know your views. Jeremy Kahn @jeremyakahn [email protected] A.I. in the news More and more people are worried about being unfairly profiled by predictive algorithms. In addition to the SyRI example mentioned above, The New York Times examined governments' use of predictive algorithms in the U.S. and Europe where these systems are increasingly being used to advise on everything from parole and bail decisions to child services' selection of cases. It found growing alarm among community and civil rights groups. In many cases, those whose lives were impacted had no idea they had been assessed by a computer-driven statistical model. "You mean to tell me I'm dealing with all this because of a computer?" one Philadelphia parolee asked when a reporter told him for the first time that the conditions of his release were based on the fact that a machine-learning algorithm had judged him to be "high risk." Twitter bans deepfakes. The social media company updated its policies to prohibit users from posting "synthetic or manipulated media that are likely to cause harm." The company is the latest to change its policies in response to concern over deepfakes, videos that are either manipulated using A.I. algorithms or entirely created by them. Twitter's policy also applies to still images and audio that has been manipulated or fabricated using a variety of other techniques, including over-dubbing. While there has been little evidence so far that deepfakes have been used for political disinformation, many security experts are concerned about their potential abuse, especially in the run-up to the 2020 U.S. presidential elections. Arm debuts two new A.I. chips. Arm, the U.K.-based semiconductor company now owned by Japan's SoftBank Group, unveiled two new computer chips designed to run A.I. applications. The new chips, called the Cortex-M55 and the Ethos U-55 NPU, extend machine-learning capabilities to small, relatively inexpensive electronic components, the company says, enabling applications in everything from healthcare to agriculture. Arm's new chips, which can be used separately or yoked together for better speed and computing power, are among a growing number of specialized components designed for "A.I. on the edge," meaning machine learning performed on a device itself without the need to communicate with a cloud-based datacenter. Barnes & Noble, Penguin Random House cancel insensitive A.I.-generated "diversity editions" covers. The bookseller and the publishing house canned a joint project they'd planned for Black History Month that would publish classic novels with new covers in which the main characters were depicted as non-white. The 12 books were selected for the project using an A.I. algorithm that analyzed the text of 100 famous novels, searching for cases in which the authors had not identified the race of the primary character. But critics accused the two companies of engaging in "literary blackface" and perpetuating the exclusion of diverse authors from the canon. Facial recognition comes to schools. A New York school district has become the first in the country to install facial recognition technology and many others are considering doing so too, The New York Times reports. The schools say the technology will help them monitor who is on school property for student safety. But civil rights groups and some parents are not happy about the development. "Subjecting 5-year-olds to this technology will not make anyone safer, and we can't allow invasive surveillance to become the norm in our public spaces," Stefanie Coyle, deputy director of the Education Policy Center for the New York Civil Liberties Union, told the paper. Clearview continues to court controversy Some law enforcement agencies in the U.S. and Canada are hailing the New York-based startup's facial recognition technology, telling The New York Times that Clearview's technology has made it easier for them to locate the victims of child sexual exploitation. But, the paper says, Clearview's handling of such sensitive images raises questions about how the startup is safeguarding the information as well as concerns about how accurate its technology really is, since the consequences of a false match are particularly grave. Clearview has already been criticized for harvesting images from social media sites to train its A.I., sometimes in violation of those sites' terms and conditions, and also for potentially misrepresenting how accurate its technology is and which law enforcement agencies are using its app. Google, YouTube, LinkedIn, Twitter, Venmo and Facebook have all sent Clearview cease-and-desist letters, threatening to sue the firm if it doesn't stop using images gathered from their platforms. Eye on A.I. talent * Cheryl Ingstad has been sworn in as the U.S. Department of Energy's first director of the Artificial Intelligence & Technology Office (AITO). The office was established in September 2019 to be the central coordinating body for the development and application of A.I. within the department. Previously, Ingstad led A.I. and machine learning research and development at the 3M Company. She had also held previous leadership roles within the Defense Intelligence Agency's Information Operations Branch. * Okta Inc., a San Francisco-based company that specializes in secure identification and access control systems, has hired Craig Weissman as Chief Architect. Previously, Weissman was the chief technology officer at Salesforce and had co-founded Duetto, which provides revenue management software for the hospitality industry. Eye on A.I. research Language models keep getting bigger -- but to exactly what end? Microsoft has unveiled the largest pre-trained language generation model to date. Its Turing Natural Language Generation model (T-NLG for short), announced this week, takes in 17 billion different parameters. This means it can encode the relationship between words and sentences over much longer stretches of text than previous models. It is more than twice as big as the next largest language model, Nvidia's MegatronLM, which has 8.3 billion parameters, and eleven times larger than OpenAI's GPT-2, which, with its 1.5 billion parameters, helped spawn the race for ultra-massive language models. Microsoft says its new heavyweight champion is better at answering questions -- such as search engine queries -- succinctly and accurately. It says it can often do "zero-shot" question answering, since it is pre-trained on such a large amount of text and may have encountered the correct answer to a question in multiple different sources during that training. And the company says T-NLG can do better abstraction and summarization than previous language models. All of these are important potential commercial uses of the technology. But, as I mentioned in the "Brain food" section of this newsletter two weeks ago, there's not much evidence that these ultra-massive language models actually "understand" anything the way a human does. Nor is it clear that, for all of its many more billions of parameters, T-NLG is that much better than GPT-2 or even Google's BERT, which only has 350 million parameters (and was considered enormous at the time it was released in 2018). GPT-2 was already so big that a lot of people who want to use it are struggling to do so -- it is breaking servers, according to Caleb Kaiser in Towards Data Science. Which brings us to what the real point of T-NLG may be: One gets the distinct impression from Microsoft's publicity push around this new massive language model that it was created simply to demonstrate Microsoft's own expertise at being able to train something that big. (Doing so requires coordinating parallel training across a lot of different processing chips.) In conjunction with T-NLG, the company unveiled a new open-source and free-to-use library of deep learning optimization tools called DeepSpeed. It includes a tool, called the Zero Redundancy Optimizer (or ZeRO for short), that the company used to train T-NLG and that it says can coordinate the training of models with up to 100 billion parameters. Fortune on A.I. Startup uses A.I. to identify molecules that could fight coronavirus -- by Jeremy Kahn Click here to oust the board -- Inside the A.I. startup that's transforming activist investing -- by Adrian Croft What you need to know about new IBM CEO Arvind Krishna -- by David Z. Morris Patient or prisoner? Governments deploy surveillance tech to track coronavirus victims -- by Eamon Barrett Brain food One of the more interesting uses of today's computer vision algorithms may be in the restoration and enhancement of archival and classic film footage. Last week, Denis Shiryaev, showed off what's possible. He used several publicly-available, neural network-based programs to transform one of cinema's most famous films -- the Lumiere brothers' 1896 L'Arrivée d'un train en gare de La Ciotat (or in English, The Arrival of a Train at La Ciotat Station) -- from its a slightly blurry and flickering original (the Lumiere's film camera only shot about 15 frames per second) to a ultra-high definition 4k, 60 frame per second version. The video, posted to YouTube, went viral. Shiryaev even added realistic sound effects to the originally silent film. One journalist, Ars Technica's Timothy B. Lee, noted that commercially-available machine learning apps could also be used to colorize old film footage. While the results are striking, and the technique suggests an interesting avenue to make classic films "come alive" for today's audiences, one has to be careful to distinguish between enhancement, which is what Shiryaev performed, and restoration. After all, what 1896 film-goers saw (and were reportedly terrorized by on first viewing) was not something with the sharpness and fluidity of 4k, 60 fps, but rather that slightly blurry and jerky camera-work. The technique Shiryaev used, which is known as "upscaling," does not restore information missing from the original but rather invents new information not contained in the original and slots it into the vastly expanded pixel-space of modern ultra high-definition. Doing so can create strange visual artifacts -- warping images or outlines that strangely melt away. It would be possible to use a different, but similar, machine learning technique to actually restore old films and photographs -- although here too, the algorithm is taking a best guess at what information is missing from the image based on the closest surrounding pixels it can analyze. If an image is badly deteriorated there is less certainty that the restoration produced by the A.I. will be accurate.

Why Microsoft and Twitter are using bug bounties to fix A.I. | Fortune

For years, companies have hosted bug bounty programs to entice well-meaning hackers to spot flaws in software so they can patch them. The programs -- participants usually get money for flagging securities holes -- are a recognition by businesses that they can't find every vulnerability on their own. Now, tech companies like Microsoft, Nvidia, and Twitter are hosting bug bounty programs specifically for artificial intelligence. The goal is for outsiders to spot flaws in A.I. software so that companies can improve the technology and reduce the risk of machine learning discriminating against certain groups of people. For example, last week, Microsoft and Nvidia detailed a new bug bounty program during the annual Defcon hacker conference. The companies plan to reward hackers who manage to alter computer viruses so that they go undetected by some of Microsoft's A.I.-powered malware-detection services. Hackers who can create scammy emails that evade Microsoft's machine-learning powered email phishing detection software will also earn some money in the form of Microsoft gift cards and other prizes. Meanwhile, Twitter pitched a bug bounty aimed at spotting bias in its A.I. The program comes after users discovered that Twitter's image-cropping tool disproportionately removed women and people of color from photos so that the images would feature white men in the center. Outsiders were invited to inspect and find flaws in the now-deactivated machine-learning algorithm that powered Twitter's photo cropping tool. Researchers discovered other bias problems with the same algorithm used in the image-cropping tool. One discovered that it would tend to crop older people from photos. Another found that the algorithm would remove people wearing head garments, showing a bias against those wearing turbans, yamakas, and hijabs. The first-place winner of Twitter's bug bounty used A.I. to modify photos of people's faces to be more appealing to the algorithm. Through this process, the researcher discovered that the algorithm favored faces that were thin, young, and white -- all indications that the technology was trained on datasets mostly of people who conform to today's conventions of beauty. It's unclear what Twitter will do with the findings, but executives implied that they would be used to improve the company's tech. During a panel related to Twitter's bug bounty program, data scientist Patrick Hall reflected on the need for more scrutiny of corporate A.I. He expressed surprise that A.I.-tailored bug bounty programs haven't become widely adopted considering the technology's many flaws. "Just because you haven't found bugs in your enterprise A.I. and machine learning offerings, certainly doesn't mean they don't have bugs," Hall said. "It just means that someone you don't know might be exploiting them, and I think for those of us in the responsible A.I. community, we wanted people to try bug bounties for so long." Jonathan Vanian @JonathanVanian [email protected] A.I. IN THE NEWS Apple's machine learning dilemma. Apple said it would use machine learning technology on people's iPhones to "detect known images of child sexual abuse without decrypting people's messages," the Associated Press reported. Privacy advocates have expressed concern that the move could open the door for authoritarian governments to monitor and surveil citizens, a notion that Apple disputes. From the article: Apple was one of the first major companies to embrace "end-to-end" encryption, in which messages are scrambled so that only their senders and recipients can read them. Law enforcement, however, has long pressured for access to that information in order to investigate crimes such as terrorism or child sexual exploitation. A.I. was relatively useless during the COVID-19 pandemic. Despite high hopes that A.I. could have been useful to overwhelmed healthcare professionals on the front lines of the COVID-19 pandemic, several recent studies show that newly developed A.I. tools did not make "a real difference, and some were potentially harmful," according to a report by the MIT Technology Review. One team of researchers, for instance, examined "415 published tools" and discovered "that none were fit for clinical use." One of the primary culprits of the healthcare A.I. failure was that technologists creating the tools held "incorrect assumptions" about the data used to train the machine learning systems. In one of the most egregious failures, researchers discovered that some machine learning systems were "picking up on the text font that certain hospitals used to label the scans. As a result, fonts from hospitals with more serious caseloads became predictors of COVID risk." Talk about the weather. Several startups are attempting to use machine learning to analyze weather data so that companies can understand potential risks to their businesses, The Wall Street Journal reported. Some of these startups use neural networks, software designed to loosely mimic how the human brain learns, but the article noted there are some flaws with that particular data-hungry A.I. technique. From the article: Companies need adequate data to train their models and there isn't always enough data. One example is hail, where limited observations make it hard to train AI models, said Mr. Gupta of ClimateAi. A.I. don't come cheap. Global spending on A.I. technologies is projected to grow 15.2% year-over-year to $341.8 billion for 2021, according to a new report from market research firm International Data Corporation. The report appears to take a broad view of A.I., lumping everything from certain servers sold by companies like Dell and Hewlett Packard Enterprise to spending on enterprise software like Slack and McAfee as inputs that inform the overall A.I. market. From the report: AI Hardware is the smallest category with 5% share of the overall AI market. Nonetheless, it is forecast to grow the fastest in 2021 at 29.6% year over year. EYE ON A.I. TALENT Nym Health hired Melisa Tucker to be the startup's senior vice president and head of product. Tucker was previously the vice president of product management at operations at Flatiron Health. Clearwater Analytics picked Souvik Das to be enterprise software company's chief technology officer. Das was previously the CTO of Zenefits. EYE ON A.I. RESEARCH Inspect the datasets. Researchers from Princeton University published a non peer-reviewed paper that probes some of the ethical dilemmas associated with developing A.I. systems built using problematic datasets, such as those that contain photos of people who never consented to be part of the dataset. The researchers analyzed 1,000 academic papers and found that despite some of the problematic datasets being retracted, many researchers continued to develop A.I. systems with the datasets or their derivatives. The researchers believe that the creators of massive datasets used to train A.I. systems "should continuously steward a dataset, actively examining how it may be misused, and making updates to license, documentation, or access restrictions as necessary." One interesting tidbit from the paper: Princeton researchers discovered that other researchers are confused about the possible legal repercussions of developing A.I. systems based on non-commercial datasets. From the paper: From these posts, we found anecdotal evidence that non-commercial dataset licenses are sometimes ignored in practice. One response reads: "More or less everyone (individuals, companies, etc) operates under the assumption that licenses on the use of data do not apply to models trained on that data, because it would be extremely inconvenient if they did." Another response reads: "I don't know how legal it really is, but I'm pretty sure that a lot of people develop algorithms that are based on a pretraining on ImageNet and release/sell the models without caring about legal issues. It's not that easy to prove that a production model has been pretrained on ImageNet ..." FORTUNE ON A.I. This hot startup is now valued at $1 billion for its A.I. skills -- By Jonathan Vanian 5 questions for Lyft co-founder John Zimmer -- By Michal Lev-Ram Tesla's Bitcoin bet is back in the black -- big time -- By Shawn Tully China's Big Tech billionaires increase philanthropic giving as Beijing cracks down -- By Yvonne Lau Tech's delivery problem: It doesn't end at your door -- By Kevin T. Dugan BRAIN FOOD What worked for Google and Facebook won't work for your company. Deep learning pioneer Andrew Ng wrote an opinion piece for the Harvard Business Review discussing some of the reasons why non-tech companies struggle with A.I. compared to consumer Internet firms like Google (Ng once worked at the search giant) and Facebook. Ng writes that the A.I. "playbook" used for Internet giants won't work for other industries because of multiple reasons. For one, non-tech giants lack an abundance of quality data that can be used to train A.I. systems. Additionally, tech giants can employ huge A.I. teams to help run their financially lucrative online advertising businesses. But not every A.I.-powered business will be as profitable as an online advertising business. Instead, many companies have individual businesses that can benefit from A.I., but they are less likely to result in whopping profits, posing a challenge for companies that are seeking a massive return on investment. As Ng writes, "The aggregate value of these hundreds of thousands of these projects is massive; but the economics of an individual project might not support hiring a large, dedicated AI team to build and maintain it." One of Ng's A.I. tips for companies: "Instead of merely focusing on the quantity of data you collect, also consider the quality, make sure it clearly illustrates the concepts we need the AI to learn."

Why this law firm only works on artificial intelligence | Fortune

As businesses continue to adopt artificial intelligence technologies, corporate lawyers and in-house data scientists should prepare to get better acquainted. Lawmakers are increasingly indicating that A.I. regulations are coming, which means that businesses will need to ensure that their machine learning systems aren't violating laws governing privacy, security, and fairness. One upstart law firm specializing in A.I.-related legal matters is betting that companies will be increasingly investigating the various ways their machine learning systems could put their businesses in legal hot water. The bnh.ai law firm, based in Washington D.C., pitches itself as a boutique law firm that caters to both lawyers and technologists alike. Having a solid understanding of A.I. and its family of technologies like computer vision and deep learning is crucial, the firm's founders believe, because solving complicated legal issues related to A.I. isn't as simple as patching a software bug. Ensuring that machine learning systems are secure from hackers and that they don't discriminate against certain groups of people requires a deep understanding of how the software operates. Businesses need to know what comprised the underlying datasets used to train the software, how that software can potentially alter over time as it feeds on new data and user behavior, and the various ways hackers can break into the software -- a difficult task considering researchers keep discovering new ways miscreants can tamper with machine learning software. One of the problems companies face, however, is that data scientists and lawyers don't really speak the same language, bnh.ai Managing Partner Andrew Burt explained. "The gap is like really, really wide, and it's really, really deep" between data scientists and lawyers, he said. "Frankly, it's uncomfortable. Lawyers don't like being put in positions where it's extremely hard to understand what's going on. It can be very intimidating to sit across from a data scientist who just spouts a bunch of statistical terminology and math." Likewise, the same is true for data scientists who may be intimidated by lawyers who themselves speak in esoteric jargon, often saying "Latin things," he said. That said, Burt believes the "future of technology is dependent on those meetings" between attorneys and data scientists. Lawyers need to understand the technical nitty gritty of A.I. systems to be able to convey the potential legal risks to data scientists in a way that's realistic and helps give them blueprints for how to troubleshoot and address the complicated systems. And unlike traditional software that's relatively a "set it and forget it" kind of product, companies must continuously monitor machine learning software because it's ever changing, thus posing future risks to their businesses. Burt concedes that the initial meetings between data scientists and lawyers can be "awkward" for reasons including the notion that technologists "don't want lawyers in their business" and "they don't want to be thinking about deeply ambiguous problems with no real solutions yet." He said his co-founder Patrick Hall once told him that lawyers will feel accomplished if they "sit in a room and talk" about legal issues. Data scientists, on the other hand, will feel like "they just wasted their time" if they attend a meeting where people talk but no one writes code. Despite the differences between the two professions, they can find common ground, at least in Burt's experience. They just need a little help getting on the same page. Once there, they can work on issues like figuring out the best ways to segment populations in datasets to adhere to current fairness rules, and when might be the best time to re-train a machine learning model to ensure that it's powered by the most relevant and appropriate data. Burt believes that when it comes to A.I. and business, "it's bad practice to wait until something bad happens to think about risk." The most visionary CEOs will have considered the legal ramifications of A.I. long before they get caught in the cross hairs of regulators. "Those are really the two threads," he said, "betting big on A.I. and caring about risk." Jonathan Vanian @JonathanVanian [email protected] A.I. IN THE NEWS Tesla's autopilot software to get scrutinized. The U.S. National Highway Traffic Safety Administration is investigating Tesla's Autopilot technology following 11 accidents in which Tesla cars "crashed into emergency vehicles when coming upon the scene of an earlier crash," CNN reported. Tesla drivers involved in the accidents allegedly activated their vehicles' Autopilot or "traffic-aware cruise control," the article said. From the report: The safety agency said its investigation will allow it to "better understand the causes of certain Tesla crashes," including "the technologies and methods used to monitor, assist, and enforce the driver's engagement with driving while Autopilot is in use." Samsung moves into A.I. chips. Samsung has tapped the services of semiconductor design company Synopsys to use machine learning software to design some of its computer chips used in smartphones and Samsung-branded headsets, according to a report by Wired. Other companies like Google and Nvidia have used similar A.I. techniques to help design computer chips. Using an A.I. subset known as reinforcement learning -- in which computers learn by trial-and-error -- companies have been able to more quickly dream up novel chip designs involving the placement and wiring of components on a circuit board. Facebook cuts off access. Researchers from the AlgorithmWatch non-profit said that they shut down a project researching Instagram's opaque newsfeed algorithm after receiving a "a thinly veiled threat" from Facebook. Facebook representatives allegedly told AlgorithmWatch members that their research project breached the company's terms of service, by collecting certain kinds of Facebook data, a claim that AlgorithmWatch disagrees with. Facebook told The Verge that company representatives did meet with AlgorithmWatch, "but denied threatening to sue the project." Big bucks for A.I. and pharma. XtalPi, a healthcare startup that operates in both the U.S. and China, landed $400 million in funding as it researches how to use deep learning to discover new drug molecules, according to a report by the healthcare news publication Fierce Biotech. The startup's partners include Pfizer, 3D Medicines, GeneQuantum Healthcare, Huadong Medicine, and Signet Therapeutics, the article noted. EYE ON A.I. TALENT Sweetgreen picked Wouleta Ayele to be the restaurant chain's chief technology officer. Ayele was previously the senior vice president of technology for Starbucks. Business software firm Interactions hired Anoop Tripathi to be the company's CTO. Tripathi was previously the senior vice president of Automation Anywhere. Coinbase landed Andrei Lyskov as a data scientist, according to a LinkedIn posting. Lyskov was previously a data scientist at Apple. EYE ON A.I. RESEARCH A.I. to predict the clicks. Google researchers published a paper, to be accepted at the upcoming ACM Recommender System Conference, that explains how deep learning can predict the most likely element -- like a web link -- that a user will click on via a mobile app. The researchers believe that the A.I. system can be used to create more intuitive user interfaces for mobile apps that are speedier and more responsive than current apps. The Google researchers created their deep learning click-prediction system using a dataset "of over 20 million clicks, which form click sequences from more than 4,000 Android users using over 13,000 unique apps on their smartphones." From the paper: As a proof of concept, we created a prototype feature named Next Click Overlay that presents the UI element that is mostly likely to be clicked at the bottom of the screen. This design does not alter the layout of an existing interface, and introduces a small amount of cognitive overload for the user to glance over the predicted item. If the prediction is correct, the user can reach the next click single-handedly. A.I. and "smelly" software collide. Researchers from the Birla Institute of Technology and Science and Curtin University in Australia published a paper about using deep learning to spot "code smell," generally referring to sloppy software code that can lead to poor performing apps and programs. From the paper: A code smell is generally detected by inspecting the source code and searching for sections of the code that can be restructured to improve the quality of code. This method is inefficient, especially if developers have to crawl through potentially thousands of lines of code, which can consume a significant amount of time and money to the organization. Based on the internal organization and anatomy of the software, a robust model can be created, which can make this excruciating process a lot simpler. The paper was accepted to the 35th International Conference on Advanced Information Networking and Applications. FORTUNE ON A.I. Employees may need to keep up 'the pretense of working' as automation spreads, says A.I. expert Kai-Fu Lee -- By Nicholas Gordon Can you predict the future? -- By Sheryl Estrada No lockdown, no problem -- Deliveroo delivers a timely surge in sales -- By Sophie Mellor China wants stricter state control over just about everything -- and the costs are mounting -- By Clay Chandler and Yvonne Lau BRAIN FOOD A.I. goes surreal. The New Yorker probes the Twitter account known as @images_ai, created by a twenty-year-old student at Northwestern University named Sam Burton-King. Using the same so-called generative adversarial network (GAN) technology used to create so-called deepfake photos, Burton-King has created some mesmerizing and surreal computer-generated artwork including an "Art Deco Buddhist temple." Regarding that artwork, the article explains that "perhaps it resembles the archetypal Chinese Buddhist temple crossed with a McDonald's -- a fleeting, half-remembered image from a dream frozen into a permanent JPEG on social media." From the article: "The art is in discriminating the good from the bad," Burton-King said -- figuring out which words to input, how big to make the images, and when to stop the generative process. But the most compelling aspect of the account might be its ability to enact an artistic fever dream, a kind of magic spell: "You can just type something and have it manifest in front of you," Burton-King said. "I think that's the main appeal for everyone." It doesn't even require coding fluency; @images_ai published a tutorial for anyone to perform the trick using open source tools online.

Cerebras hopes planned IPO will supercharge its race against Nvidia and fellow chip startups for the fastest generative AI | Fortune

Hello and welcome to Eye on AI! In this edition...Governor Newsom vetoes SB 1047; ByteDance plans new AI model based on Huawei chips; Microsoft announces AI models will improve Windows search; and the U.S. Commerce Department sets a new rule that eases restrictions on AI chip shipments to the Middle East. Cerebras has a need for speed. In a bid to take on Nvidia, the AI chip startup is rapidly moving toward an IPO after announcing its filing for one yesterday. At the same time, the company is also in a fierce race with fellow AI chip startups Groq and SambaNova for the title of 'fastest generative AI.' All three are pushing the boundaries of their highly-specialized hardware and software to enable AI models to produce responses using ultra-fast generative AI that even outperform Nvidia GPUs. Here's what that means: When you ask an AI assistant a question, it must sift through all of the knowledge in its AI model to quickly come up with an answer. In industry parlance, that process is known as "inference." But large language models don't sift through words during the inference process. When you ask a question or give a chatbot a prompt, the AI breaks that into smaller pieces called "tokens" -- which could represent a word, or a chunk of a word -- to process its answer and respond. Pushing for faster and faster output So what does "ultra-fast" inference mean? If you've tried chatbots like OpenAI's ChatGPT, Anthropic's Claude, or Google's Gemini, you probably think the output of your prompts arrives at a perfectly reasonable pace. In fact, you may be impressed by how quickly it spits out answers to your queries. But in February 2024, demos of a Groq chatbot based on a Mistral model produced answers far faster than people could read. It went viral. The setup served up 500 tokens per second to produce answers that were nearly instantaneous. By April, Groq delivered an even speedier 800 tokens per second, and by May SambaNova boasted it had broken the 1,000 tokens per second barrier. Today, Cerebras, SambaNova, and Groq are all delivering over 1,000 tokens per second, and the "token wars" have revved up considerably. At the end of August, Cerebras claimed it had launched the "world's fastest AI inference" at 1,800 tokens per second, and last week Cerebras said it had beaten that record and become the "first hardware of any kind" to exceed 2,000 tokens per second on one of Meta's Llama models. When will fast be fast enough? This led me to ask: Why would anyone need generative AI output to be that fast? When will fast be fast enough? According to Cerebras CEO Andrew Feldman, generative AI speed is essential since search results will increasingly be powered by generative AI, as well as new capabilities like streaming video. Those are two areas where latency, or the delay between an action and a response, is particularly annoying. "Nobody's going to build a business on an application that makes you sit around and wait," he told Fortune. In addition, AI models are quickly being used to power far more complex applications than just chat. One rapidly growing area of interest is developing application workflows based on AI agents, in which a user asks a question or prompts an action that doesn't simply involve one query to one model. Instead it leads to multiple queries to multiple models that can go off and do things like search the web or a database. "Then the performance really matters," said Feldman, explaining that a reasonably slow output today could quickly become painfully slow. Unlocking AI potential with speed The bottom line is that speed matters because faster inference unlocks greater potential in applications built with AI, Mark Heaps, chief technology evangelist at Groq, told Fortune. That is especially true for data-heavy applications in fields like financial trading, traffic monitoring, and cybersecurity: "You need insights in real time, a form of instant intelligence that keeps up with the moment," he said. "The race to increase speed...will provide better quality, accuracy, and potential for greater ROI." It's worth noting, he pointed out, that AI models still have nowhere near as many neural connections as the human brain. "As the models get more advanced, bigger, or layered with lots of agents using smaller models, it will require more speed to keep the application useful," he explained, adding that this has been an issue throughout history. "Why do we need cars to get beyond 50 mph? Was it so we could go fast? Or producing an engine that could do 100 mph enabled the ability to carry more weight at 50 mph?" Rodrigo Liang, CEO and cofounder of SambaNova, agreed. Inference speed, he told Fortune, "is where the rubber hits the road -- where all the training, the building of models, gets put to work to deliver real business value." That's particularly true now that the AI industry is moving more of its training from training AI models to putting them into production. "The world is looking for the most efficient way to produce tokens so you can support an ever-growing number of users," he said. "Speed allows you to service many customers concurrently." Sharon Goldman [email protected] AI IN THE NEWS Governor Newsom vetoes California's SB-1047. On Sunday, news spread quickly through Silicon Valley that Governor Newsom had vetoed SB-1047, a widely debated and ambitious AI regulatory proposal. The bill, if enacted, would have required developers to conduct safety testing on large AI models before public release, the New York Times reported. Critics, however, raised concerns over provisions granting the state's attorney general the authority to sue companies for harm caused by their technologies. The bill also mandated a "kill switch" to shut down AI systems in the event of potential threats like biowarfare, mass casualties, or significant property damage. "I do not believe this is the best approach to protecting the public from real threats posed by the technology," Newsom said in a statement. "Instead, the bill applies stringent standards to even the most basic functions -- so long as a large system deploys it." Sources say ByteDance plans new AI model trained with Huawei chips. Reuters reported that TikTok's Chinese parent ByteDance plans to develop an AI model trained primarily with chips from China's Huawei Technologies. It's a response to U.S. moves since 2022 to restrict exports of advanced AI chips, particularly from market leader Nvidia. The article claimed that sources said ByteDance's next step in the AI race is to use Huawei's Ascend 910B chip to train a large-language AI model, but ByteDance denied a new model is being developed. Microsoft announces AI models will improve Windows search on Copilot Plus PCs. Microsoft said today its new Copilot Plus PCs will use AI models to improve Windows search, available starting in November, including a new Click to Do feature that is similar to Google's Circle to Search function. "AI-powered search makes it dramatically easier to find virtually anything," said Yusuf Mehdi, executive vice president and consumer chief marketing officer at Microsoft, as reported by the Verge. "You no longer need to remember file names and document locations, nor even specific names of words. Windows will better understand your intent and match the right document, image, file, or email." U.S. Commerce Department sets new rule that eases restrictions on AI chip shipments to Middle East. According to Reuters, yesterday the U.S. Commerce Department unveiled a rule that could ease shipments of AI chips like those from Nvidia to Middle East data centers. Since October 2023, U.S. exporters have been required to obtain licenses before shipping advanced chips to parts of the Middle East and Central Asia. But now, data centers will be able to apply for status that will allow them to receive chips, rather than requiring their suppliers to obtain individual licenses to ship to them. FORTUNE ON AI Before Mira Murati's surprise exit from OpenAI, staff grumbled its o1 model had been released prematurely -- by Jeremy Kahn, Kali Hays and Sharon Goldman Why investors want startup founders to own equity -- including OpenAI's Sam Altman -- by Sharon Goldman, Kali Hays and Verne Kopytoff Nvidia shares fall and its Chinese rivals soar after Beijing urges AI companies to look elsewhere for chips -- by David Meyer Mark Cuban warns the U.S. must win the AI race 'or we lose everything' -- by Jason Ma AI CALENDAR Oct. 22-23: TedAI, San Francisco Oct. 28-30: Voice & AI, Arlington, Va. Nov. 19-22: Microsoft Ignite, Chicago Dec. 2-6: AWS re:Invent, Las Vegas Dec. 8-12: Neural Information Processing Systems (Neurips) 2024 in Vancouver, British Columbia Dec. 9-10: Fortune Brainstorm AI San Francisco (register here) EYE ON AI RESEARCH Could generative AI chatbots help reduce belief in conspiracy theories? New research published in Science by Thomas Costello of American University and Gordon Pennycook of Cornell found that discussions with AI chatbots could reduce individuals' beliefs in conspiracy theories. Using OpenAI's GPT-4 Turbo, human participants described a conspiracy theory that they subscribed to, and then the AI responded with back and forth with persuasive arguments that refuted their beliefs with evidence. According to the research, "the AI chatbot's ability to sustain tailored counterarguments and personalized in-depth conversations reduced their beliefs in conspiracies for months, challenging research suggesting that such beliefs are impervious to change." BRAIN FOOD Want a glimpse of your future self using generative AI? If you've ever wanted to receive a visit from your future self like in Back to the Future, you may be interested in new research from MIT that created a chatbot for users to have a conversation with an "AI-generated simulation of their potential future self." The tool, called " Future You," uses a large language model and information provided by the user to help young people "improve their sense of future self-continuity, a psychological concept that describes how connected a person feels with their future self." What if the Future Tool offers negative predictions, causing young people to freak out? The researchers explained that the tool cautions users that its results are only one potential version of their future self, and they can still change their lives. "This is not a prophesy, but rather a possibility," the lead researcher said.

This former OpenAI researcher thinks we should be gaming out the AI apocalypse | Fortune

Welcome to Eye on AI! In this edition...Meta wins AI copyright case in another blow to authors...Google DeepMind releases new AlphaGenome model to better understand the genome...Sam Altman calls Iyo lawsuit 'silly' after OpenAI scrubs Jony Ive deal from website, then shares emails. This week, I spoke with Steven Adler, a former OpenAI safety researcher who left the company in January after four years, saying on X after his departure that he was "pretty terrified by the pace of AI development." Since then, he's been working as an independent researcher and "trying to improve public understanding of what the AI future might look like and how to make it go better." What really caught my attention was a new blog post from Adler, where he shares his recent experience participating in a five-hour discussion-based simulation, or "tabletop exercise," with 11 others, which he said was similar to wargames-style exercises in the military and cybersecurity. Together, the group explored how world events might unfold if "superintelligence," or AI systems that surpass human intelligence, emerges in the next few years. A simulation organized by the authors of AI 2027 The simulation was organized by the AI Futures Project, a nonprofit AI forecasting group led by Daniel Kokotajlo, Adler's former OpenAI teammate and friend. The organization drew attention in April for "AI 2027," a forecast-based scenario mapping out how superhuman AI could emerge by 2027 -- and what that might mean. According to the scenario, by then AI systems could be using 1,000 times more compute than GPT‑4 and rapidly accelerating their own development by training other AIs. But this self-improvement could easily outpace our ability to keep them aligned with human values, raising the risk that seemingly helpful AIs might ultimately pursue their own goals. The purpose of the simulation, said Adler, is to help people understand the dynamics of rapid AI development and what challenges are likely to arise in trying to steer it for the better. Each participant has their own character whom they try to represent realistically in conversations, negotiations and strategizing, he explained. Those characters included members of the US federal government (each branch, as well as the President and their Chief of Staff), the Chinese government/AI companies, the Taiwanese government, NATO, the leading Western AI company, the trailing Western AI companies, the corporate AI safety teams, the broader AI safety ecosystem (e.g., METR, Apollo Research), the public/press, and the AI systems themselves. Adler was tapped to play what he called "maybe the most interesting role" -- a rogue artificial intelligence. During each 30-minute round of the five-hour simulation, which represented the passage of a few months in the forecast, Adler's AI got progressively more capable -- including at training even more powerful AI systems. After rolling the dice -- an actual, analog pair that was used occasionally in the simulation in cases where it was unclear what would happen -- Adler learned that his AI character would not be evil. However, if he had to choose between self-preservation or doing what's right for humanity, he was meant to choose his own preservation. Then, Adler detailed, with some humor, the awkward interactions his AI character had with the other characters (who asked him for advice on superintelligence), as well as the surprise addition of a second player who played a rogue AI in the hands of the Chinese government. A power struggle between AI systems The surprise of the simulation, he said, was seeing how the biggest power struggle might not be between humans and AI. Instead, various AIs connecting with each other, vying for victory, might be an even bigger problem. "How directly AI systems are able to communicate in the future is a really important question," Adler said. "It's really, really important that humans be monitoring notification channels and paying attention to what messages are being passed between the AI agents." After all, he explained, if AI agents are connected to the internet and permitted to work with each other, there is reason to think they could begin colluding. Adler pointed out that even soulless computer programs can happen to work in certain ways and have certain tendencies. AI systems, he said, might have different goals that they automatically pursue, and humans need influence over those goals. The solution, he said, could be a form of AI control based on how cybersecurity professionals deal with "insider threats" -- when someone inside an organization, who has access and knowledge, might try to harm the system or steal information. The goal of security is not to make sure insiders always behave; it's to build structures that prevent even ill-intentioned insiders from doing serious harm. Instead of just hoping AI systems stay aligned, we should focus on building practical control mechanisms that can contain, supervise, restrict, or shut down powerful AIs -- even if they try to resist. Forecasts and predictions are 'hard' I pointed out to Adler that when AI 2027 was released, there was plenty of criticism. People were skeptical, saying the timeline was too aggressive and underestimated real-world limits like hardware, energy, and regulatory bottlenecks. Critics also doubted that AI systems could quickly improve themselves in the runaway way the report suggested and argued that solving AI alignment would likely be much harder and slower. Some also saw the forecast as overly alarmist, warning it could hype fears without solid evidence that superhuman AI is that close. Adler responded by encouraging others to express interest in running the simulation for their organization (there is a form to fill out), but admitted that forecasts and predictions are hard. "I understand why people would feel skeptical, it's always hard to know what will actually happen in the future," he said. "At the same time, from my point of view, this is the clear state of the art in people who've sat down and for months done tons of underlying research and interviews with experts and just all sorts of testing and modeling to try to figure out what worlds are realistic." Those experts are not saying that the world depicted in AI 2027 will definitely happen, he emphasized, but "it's important that the world be ready if it does." Simulations like this help people to understand what sorts of actions matter and make a difference "if we do find ourselves in that sort of world." Conversations with AI researchers like Adler tend to end without much optimism -- though it's worth noting that plenty of others in the field would push back on just how urgent or inevitable this view of the future really is. Still, it's a relief that his blog post concludes with the hope, at least, that humans will "recognize the challenges and rise to the occasion." That includes Sam Altman: If OpenAI hasn't already run one of these simulations and wanted to try it, said Adler, "I am quite confident that the team would make it happen." With that, here's the rest of the AI news. Sharon Goldman [email protected] @sharongoldman AI IN THE NEWS Meta wins AI copyright case in another blow to authors. In the same week as a federal judge ruled that Anthropic's use of copyrighted books to train its AI models was "fair use," Meta also won a copyright case in yet another blow to authors seeking to hold AI companies accountable for using their works without permission. According to the Financial Times, Meta's use of a library of millions of books, academic articles, and comics to train its Llama AI models was judged "fair" by a federal court on Wednesday. The case was brought by about a dozen authors, including Ta-Nehisi Coates and Richard Kadrey. Meta's use of these titles is protected under copyright law's fair use provision, San Francisco district judge Vince Chhabria ruled. Meta had argued that the works had been used to develop a transformative technology, which was fair "irrespective" of how it acquired the works. Google DeepMind releases new AlphaGenome model to better understand the genome. Google DeepMind, the AI research lab famous for developing AlphaGo, the first AI to defeat a world champion Go player, and AlphaFold, which uses AI to predict the 3D structures of proteins, released its new AlphaGenome model, designed to analyze up to one million DNA base pairs at once and predict how specific genomic variants affect regulatory functions -- such as gene expression, RNA splicing, and protein binding -- across diverse cell types. The company said the model was trained on extensive public datasets and achieves state-of-the-art performance on most benchmarks and can assess mutation impacts in seconds. AlphaGenome will be available for non-commercial research, and promises to accelerate discovery in genome biology, disease understanding, and therapeutic development. Sam Altman calls Iyo lawsuit 'silly' after OpenAI scrubs Jony Ive deal from website, then shares emails. On Tuesday, OpenAI CEO Sam Altman on criticized a lawsuit filed by hardware startup Iyo, which accused his company of trademark infringement. CNBC reported that in response to the suit, Iyo CEO Jason Rugolo had been "quite persistent in his efforts" to get OpenAI to buy or invest in his company. In a post on X, he wrote that Rugolo is now suing OpenAI over the name in a case he described as "silly, disappointing and wrong." He then posted screenshots of emails on X showing messages between him and Rugolo, which show a mostly friendly exchange. The suit stemmed from an announcement last month that OpenAI was bringing on Apple designer Jony Ive by acquiring his AI startup io in a deal valued at about $6.4 billion. Iyo alleged that OpenAI, Altman, and Ive had engaged in unfair competition and trademark infringement and claimed that it's on the verge of losing its identity because of the deal. FORTUNE ON AI Can AI help America make stuff again? -- by Jeremy Kahn AI companies are throwing big money at newly-minted PhDs, sparking fears of an academic 'brain drain' -- by Alexandra Sternlicht Top e-commerce veteran Julie Bornstein unveils Daydream -- an AI-powered shopping agent that's 25 years in the making -- by Jason Del Rey Exclusive: Uber and Palantir alums raise $35M to disrupt corporate recruitment with AI -- by Beatrice Nolan AI CALENDAR July 8-11: AI for Good Global Summit, Geneva July 13-19: International Conference on Machine Learning (ICML), Vancouver July 22-23: Fortune Brainstorm AI Singapore. Apply to attend here. July 26-28: World Artificial Intelligence Conference (WAIC), Shanghai. Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here. Oct. 6-10: World AI Week, Amsterdam Dec. 2-7: NeurIPS, San Diego Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here. EYE ON AI NUMBERS 130 Many vendors are engaging in "agent washing" -- the rebranding of products such as digital assistants, chatbots, and "robotic process automation" (RPA) that either aren't actually agentic or don't actually use AI, Gartner says, estimating that only about 130 of the thousands of "agentic AI" vendors actually offer real AI agents.

Why we should all pay attention to what lawyers, auditors, and accountants are doing with AI | Fortune

Hello and welcome to Eye on AI. In today's edition...the U.S. Senate rejects moratorium on state-level AI laws...Meta unveils its new AI organization...Microsoft says AI can out diagnose doctors...and Anthropic shows why you shouldn't let an AI agent run your business just yet. AI is rapidly changing work for many of those in professional services -- lawyers, accountants, auditors, compliance officers, consultants, and tax advisors. In many ways, the experience of these professionals, and the businesses they work for, are a harbinger of what's likely to happen for other kinds of knowledge workers in the near future. Because of this, it was interesting to hear the discussion yesterday at a conference on the "Future of Professionals" at Oxford University's Said School of Business. The conference was sponsored by Thomson Reuters, in part to coincide with the publication of a report it commissioned on trends in professionals' use of AI. That report, based on a global survey of 2,275 professionals in February and March, found that professional services firms seem to be finding a return on their AI investment at a higher rate than in other sectors. Slightly more than half -- 53% -- of the respondents said their firm had found at least one AI use case that was earning a return, which is about twice what other, broader surveys have tended to find. Not surprisingly, Thomson Reuters found it was the professional firms where AI usage was part of a well-defined strategy and that had implemented AI governance structures were the most likely to see gains from the technology. Interestingly, among firms where AI adoption was less structured, 64% of those surveyed still reported ROI from at least one use case, which may reflect how powerful and time-saving these tools can be even when used by individuals to improve their own workflows. The biggest factors holding back AI use cases, the respondents said, included concerns about inaccuracy (with 50% of those surveyed noting this was a problem) and data security (42%). For more on how law firms are using AI, check out this feature from my Fortune colleague Jeff John Roberts. Mind the gaps Here are a few tidbits from the conference worth highlighting: Mari Sako, the Oxford professor of management studies who helped organize the conference, talked about the three gaps that professionals needed to watch out for in trying to manage AI implementation: One was the responsibility gap between model developers, application builders, and end users of AI models. Who bears responsibility for the model's accuracy and possible harms? A second was the principles to practice gap. Businesses enact high-minded "Responsible AI" principles but then the teams building or deploying AI products struggle to operationalize them. One reason this happens is that first gap -- it means that teams building AI applications may not have visibility into the data used to train a model they are deploying or detailed information about how it may perform. This can make it hard to apply AI principles about transparency and mitigating bias, among other things. Finally, she said, there is a goals gap. Is everyone in the business aligned about why AI is being used in the first place? Is it for human augmentation or automation? Is it operational efficiency or revenue growth? Is the goal to be more accurate than a human, or simply to come close to human performance at a lower cost? What role should environmental sustainability play in these decisions? All good questions. Not a substitute for human judgment Ian Freeman, a partner at KPMG UK, talked about his firm's increasing use of AI tools to help auditors. In the past, auditors were forced to rely on sampling transactions, trying to apply more scrutiny to those that presented a bigger business risk. But now, with AI, it is possible to run a screen on every single transaction. Still, it is the riskiest transactions that should get the most scrutiny and AI can help identify those. Freeman said AI could also help more junior auditors understand the rationale for probing certain transactions. And he said AI models could help with a lot of routine financial analysis. But he said KPMG had a policy of not deploying AI in situations that called for human judgment. Auditing is full of such cases, such as deciding on materiality thresholds, making a call about whether a client has submitted enough evidence to justify a particular accounting treatment, or deciding on appropriate warranty reserves for a new product. That sounds good, but I also wonder about the ability of AI models to act as tutors or digital mentors to junior auditors, helping them to develop their professional judgment? Surely, that seems like it might be a good use case for AI too. A senior partner from a large law firm (parts of the conference were conducted under Chatham House Rules, so I can't name them) noted that many corporate legal departments are embracing AI faster than legal firms -- something the Thomson Reuters survey also showed -- and that this disparity was putting pressure on the firms. Corporate counsel are demanding that external lawyers be more transparent about their AI usage -- and critically, putting pressure on legal bills on the theory that many legal tasks can now be done in far fewer billable hours. Changing career paths and the need for AI expertise AI is also possibly going to change how professional service firms think about career paths within their business and even who leads these firms, several lawyers at the conference said. AI expertise is increasingly important to how these firms operate, and yet it is difficult to attract the talent these businesses need if these "non-qualified" technical experts (the term "non-qualified" is simply used to denote an employee who has not been admitted to the bar, but its pejorative connotations are hard to escape) know they will always be treated as second-class compared to the client-facing lawyers and also are ineligible for promotion to the highest ranks of the firm's management. Michael Buenger, executive vice president and chief operating officer at the National Center for State Courts in the U.S., said that if large law firms had trouble attracting and retaining AI expertise, the situation was far worse for governments. And he pointed out that judges and juries were increasingly being asked to rule on evidence, particularly video evidence, but also other kinds of documentary evidence, that might be AI manipulated, but without access to independent expertise to help them make calls about what has been altered by AI and how. If not addressed, he said, this could seriously undermine faith in the courts to deliver justice. There were lots more insights from the conference, but that's all we have space for today. Here's more AI news. Note: The essay above was written and edited by Fortune staff. The news items below were selected by the newsletter author, created using AI, and then edited and fact-checked. Jeremy Kahn [email protected] @jeremyakahn Want to know more about how to use AI to transform your business? Interested in what AI will mean for the fate of companies, and countries? Then join me at the Ritz-Carlton, Millenia in Singapore on July 22 and 23 for Fortune Brainstorm AI Singapore. This year's theme is The Age of Intelligence. We will be joined by leading executives from DBS Bank, Walmart, OpenAI, Arm, Qualcomm, Standard Chartered, Temasek, and our founding partner Accenture, plus many others, along with key government ministers from Singapore and the region, top academics, investors and analysts. We will dive deep into the latest on AI agents, examine the data center build out in Asia, examine how to create AI systems that produce business value, and talk about how to ensure AI is deployed responsibly and safely. You can apply to attend here and, as loyal Eye on AI readers, I'm able to offer complimentary tickets to the event. Just use the discount code BAI100JeremyK when you checkout. AI IN THE NEWS Senate strips 10-year moratorium on state AI laws from Trump tax bill. The U.S. Senate voted 99-1 to remove the controversial measure from President Donald Trump's landmark "Big Beautiful Bill." The restrictions had been supported by Silicon Valley tech companies and venture capitalists as well as their allies in the Trump administration. Bipartisan opposition to the moratorium -- led by Sen. Marsha Blackburn -- centered on preserving state-level protections like Tennessee's Elvis Act, which protects citizens from unauthorized use of their voice or likeness, including in AI-generated content. Critics warned that in the absence of federal AI regulation, the ban on state-level laws would leave U.S. citizens with no protection from AI harms at all. But tech companies argue that the increasing patchwork of state-level AI regulation is unworkable, hampering AI progress. Read more from Bloomberg News here. Meta announced new AI leadership team and key hires from rival AI labs. Meta CEO Mark Zuckerberg sent a memo to employees formally announcing the creation of Meta Superintelligence Labs, a new organization uniting the company's foundational AI model, product, and Fundamental AI Research (FAIR) teams under a single umbrella. Scale AI founder and CEO Alexandr Wang -- who is joining Meta as part of a $14.3 billion investment into Scale -- will have the title "chief AI officer" and will co-lead the new Superintelligence unit along with former GitHub CEO Nat Friedman. Zuckerberg also announced the hiring of 11 prominent AI researchers from OpenAI, Google DeepMind, and Anthropic. You can read more about Meta's AI talent raid from Wired here. Cloudflare begins blocking AI web-crawlers by default. Internet content delivery provider Cloudflare announced it has begun blocking AI companies' web crawlers from accessing website content by default. Owners of the websites can choose to unblock specific crawlers -- such as those Google uses to build its search index -- or even opt for a "pay per crawl" option that will allow them to monetize the scraping of their content. With around 16% of global internet traffic passing through Cloudflare, the change could significantly impact AI development. (Full disclosure: Fortune is one of the initial participants in the Cloudflare crawler initiative.) Read more from CNBC here. EYE ON AI RESEARCH Even better than House? Microsoft has unveiled an AI system for medical diagnoses that it claims can accurately diagnose complex cases four times more accurately than individual human doctors (under certain conditions -- more on that in a sec.) The "Microsoft AI Diagnostic Orchestrator" (MAI-DxO -- gotta love those AI acronyms) consists of five AI "agents" that each have a distinct role to play in scouring the medical literature, hypothesizing what the patient's condition might be, ordering tests to eliminate possibilities, and even trying to optimize these tests to derive the most useful information at the least cost. These five "AI doctors" then engage in a process Microsoft is dubbing "chain of debate," where they collaborate and critique one another, ultimately arriving at a diagnosis. In trials involving 304 real-world cases from the New England Journal of Medicine, MAI-DxO, achieved an 85.5% success rate, compared to about 20% for human doctors. Microsoft tried powering the system with different AI models from OpenAI, Google, Meta, Anthropic, and DeepSeek, but found it worked best when using OpenAI's o3 model (Microsoft is a major investor in OpenAI, sells OpenAI's models through its cloud service, and depends on OpenAI for many of its own AI offerings). As for the poor performance of the human docs, it is important to note that in the test they were not allowed to consult either medical textbooks or colleagues. Nonetheless, Microsoft AI CEO Mustafa Suleyman said the system could transform healthcare -- although the company also said MAI-DxO is just a research project and is not yet being turned into a product. You can read more from the Financial Times here. FORTUNE ON AI Mark Zuckerberg overhauled Meta's entire AI org in a risky, multi-billion dollar bet on 'superintelligence' -- by Sharon Goldman Longtime Bessemer investor Mary D'Onofrio, who backed Anthropic and Canva, leaves for Crosslink Capital -- by Allie Garfinkle Ford CEO says new technologies like AI are leaving many workers behind, and companies need a plan -- by Jessica Mathews Commentary: When your AI assistant writes your performance review: A glimpse into the future of work -- by David Ferrucci AI CALENDAR July 8-11: AI for Good Global Summit, Geneva July 13-19: International Conference on Machine Learning (ICML), Vancouver July 22-23: Fortune Brainstorm AI Singapore. Apply to attend here. July 26-28: World Artificial Intelligence Conference (WAIC), Shanghai. Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here. Oct. 6-10: World AI Week, Amsterdam Dec. 2-7: NeurIPS, San Diego Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here. BRAIN FOOD AI tries to run a vending machine business. Hilarity ensues, Part Deux. A month ago in the research section of this newsletter, I wrote about research from Andon Labs about what happens when you try to have various AI models run a simulated vending machine business. Now, Anthropic teamed up with Andon Labs to test one of its latest models, Claude 3.7 Sonnet, to see how it did running a real-life vending machine in Anthropic's San Francisco office. The answer, as it turns out, is not well at all. As Anthropic writes in its blog on the experiment, "If Anthropic were deciding today to expand into the in-office vending market, we would not hire [Claude 3.7 Sonnet]." The model made a lot of mistakes -- like telling customers to send payments to a Venmo account that didn't exist (it had hallucinated it) -- and also a lot of poor business decisions, like offering far too many discounts (including an Anthropic employee discount in a location where 99% of the customers were Anthropic employees), failing to seize a good arbitrage opportunity, and failing to increase prices in response to high demand. The entire Anthropic blog makes for fun reading. And the experiment makes it clear that AI agents probably are nowhere near ready for a lot of complex, multi-step tasks over long time periods.

Big Tech is racing to build AI data centers -- just as Accenture warns carbon emissions could surge 11x | Fortune

Welcome to Eye on AI! In this edition...Ilya Sutskever says he is now CEO of Safe Superintelligence...Chinese AI companies erode U.S. dominance...Meta's AI talent bidding war heats up...Microsoft's sales overhaul goes all-in on AI. As an early-summer heat wave blanketed my home state of New Jersey last week, it felt like perfect timing to stumble across a sobering new prediction from Accenture: AI data centers' carbon emissions are on track to surge 11-fold by 2030. The report estimates that over the next five years, AI data centers could consume 612 terawatt-hours of electricity -- roughly equivalent to Canada's total annual power consumption -- driving a 3.4% increase in global carbon emissions. And the strain doesn't stop at the power grid. At a time when freshwater resources are already under severe pressure, AI data centers are also projected to consume more than 3 billion cubic meters of water per year -- a volume that surpasses the annual freshwater withdrawals of entire countries like Norway or Sweden. Unsurprisingly, the report -- Powering Sustainable AI -- offers recommendations for how to rein in the problem and prevent those numbers from becoming reality. But with near-daily headlines about Big Tech's massive AI data center buildouts across the U.S. and worldwide, I can't help but feel cynical. The urgent framing of an AI race against China doesn't seem to leave much room -- or time -- for serious thinking about sustainability. Just yesterday, for example, OpenAI agreed to rent a massive amount of computing power from Oracle data centers as part of its Stargate initiative, which intends to invest $500 billion over the next four years building new AI infrastructure for OpenAI in the United States. The additional capacity from Oracle totals about 4.5 gigawatts of data center power in the U.S., according to Bloomberg reporting. A gigawatt is akin to the capacity from one nuclear reactor and can provide electricity to roughly 750,000 houses. And this week, Meta was reported to be seeking to raise $29 billion from private capital firms to build AI data centers in the U.S., while already building a $10 billion AI data center in Northeast Louisiana. As part of that deal, the local utility, Entergy, will supply three new power plants. Meta CEO Mark Zuckerberg has made his intentions clear: The U.S. must rapidly expand AI data center construction or risk falling behind China in the race for AI dominance. Speaking on the Dwarkesh Podcast in May, he warned that America's edge in artificial intelligence could erode unless it keeps pace with China's aggressive build-out of data center capacity and factory-scale hardware. "The U.S. really needs to focus on streamlining the ability to build data centers and produce energy," Zuckerberg said. "Otherwise, we'll be at a significant disadvantage." The U.S. government seems to be aligned with that sense of urgency. David Sacks, now serving as the White House AI and Crypto Czar, has also underscored that energy and data center expansion are central to America's AI strategy -- leaving little room for sustainability concerns. On his All In podcast in February, Sacks argued that Washington's "go-slow" approach to AI could strangle the industry. He emphasized that the U.S. needs to clear the way for infrastructure and energy development -- including AI data centers -- to keep pace with China. In late May, he went further, saying that streamlining permitting and expanding power generation are essential for AI's future -- something he claimed has been "effectively impossible under the Biden administration." His message: the U.S. needs to race to build faster. Accenture, meanwhile, is urging its clients to responsibly grow and engineer its AI data centers in a bid to balance growth with environmental responsibility. It is offering a new metric, that it calls the Sustainable AI Quotient (SAIQ), to measure the true costs of AI in terms of money invested, megawatt-hours of energy consumed, tons of CO₂ emitted and cubic meters of water used. The firm's report says the metric will help organizations answer a basic question: "What are we actually getting from the resources we're investing in AI?" and allow that enterprise to measure its performance across time. I spoke to Matthew Robinson, managing director of Accenture Research and co-author of the report, who emphasized that he hoped Accenture's sobering predictions would be proven wrong. "They kind of take your breath away," he said, explaining that Accenture modeled future energy consumption from the expected number of installed AI chips adjusted for utilization and the additional energy requirements of data centers. That data was combined with regional data on electricity generation, energy mix and emissions, while water use was assessed based on AI data center energy consumption and how much water is consumed per unit of electricity generated. "The point really is to open the conversation around the actions that are available to avert this pathway -- we don't want to be right here," he said. He would not comment on the actions of specific companies like OpenAI or Meta, but said that overall, clearly more effort is needed to avert the rise in carbonisation fueled by AI data centers while still allowing for growth. Accenture's recommendations certainly make sense: Optimize the power efficiency of AI workloads and data centers with everything from low-carbon energy options to cooling innovations. Use AI thoughtfully, by choosing smaller AI models, and better pricing models for incentivizing efficiency. And ensure better governance over AI sustainability initiatives. It's hard to imagine that the biggest players in the race for AI dominance -- Big Tech giants and heavily funded startups -- will hit the brakes long enough to seriously address these growing concerns. Not that it's impossible. Take Google, for example: In its latest sustainability report released this week, the company revealed that its data centers are consuming more power than ever. In 2024, Google used approximately 32.1 million megawatt-hours (MWh) of electricity, with a staggering 95.8% -- about 30.8 million MWh -- consumed by its data centers. That's more than double the energy its data centers used in 2020, just before the consumer AI boom. Still, Google emphasized that it's making meaningful strides toward cleaning up its energy supply, even as demand surges. The company said it cut its data center energy emissions by 12% in 2024, thanks to clean energy projects and efficiency upgrades. And it's squeezing more out of every watt. Google reported that the amount of compute per unit of electricity has increased about six-fold over the past five years. Its power usage effectiveness (PUE) -- a key measure of data center efficiency -- is now approaching the theoretical minimum of 1.0, with a reported PUE of 1.09 in 2024. "Just speaking personally, I'd be optimistic," said Robinson. Note: Check out this new Fortune video about my tour of IBM's quantum computing test lab. I had a fabulous time hanging out at IBM's Yorktown Heights campus (a midcentury modern marvel designed by the same guy as the St. Louis Arch and the classic TWA Flight Center at JFK Airport) in New York. The video was part of my coverage for this year's Fortune 500 issue that included an article that dug deep into IBM's recent rebound. As I said in my piece, "walking through the IBM research center is like stepping into two worlds at once. There are the steel and glass curves of Saarinen's design, punctuated by massive walls made of stones collected from the surrounding fields, with original Eames chairs dotting discussion nooks. But this 20th-century modernism contrasts starkly with the sleek, massive, refrigerator-like quantum computer -- among the most advanced in the world -- that anchors the collaboration area and working lab, where it whooshes with the steady hum of its cooling system." With that, here's the rest of the AI news. Sharon Goldman [email protected] @sharongoldman AI IN THE NEWS Ilya Sutskever says he is now CEO of Safe Superintelligence, after Daniel Gross steps down to join Meta. Ilya Sutskever, the former OpenAI chief scientist who founded Safe Superintelligence (SSI) with Daniel Gross and Daniel Levy a year ago, confirmed that he will now serve as SSI's CEO after Daniel Gross stepped down. Sustkever posted on X saying: "Daniel Gross's time with us has been winding down, and as of June 29 he is officially no longer a part of SSI. We are grateful for his early contributions to the company and wish him well in his next endeavor. I am now formally CEO of SSI, and Daniel Levy is President. The technical team continues to report to me. ⁠You might have heard rumors of companies looking to acquire us. We are flattered by their attention but are focused on seeing our work through." Meta was rumored to have sought to acquire the $32 billion-valued SSI. Chinese AI companies erode U.S. dominance. According to the Wall Street Journal, Chinese artificial intelligence companies are gaining ground globally, challenging U.S. supremacy and intensifying a potential AI arms race. Across Europe, the Middle East, Africa, and Asia, organizations -- from multinational banks like HSBC and Standard Chartered to Saudi Aramco -- are increasingly adopting large language models from Chinese firms such as DeepSeek and Alibaba as alternatives to U.S. offerings like ChatGPT. Even American cloud giants like Amazon Web Services, Microsoft, and Google now offer access to DeepSeek's models, despite U.S. government security restrictions on the company's apps. While OpenAI's ChatGPT still leads in global adoption -- with 910 million downloads versus DeepSeek's 125 million -- Chinese models are undercutting U.S. competition by offering nearly comparable performance at much lower prices. Meta's AI talent bidding war heats up. As Mark Zuckerberg rapidly staffs up Meta's new superintelligence lab, his company has reportedly offered some OpenAI researchers eye-popping pay packages of up to $300 million over four years, with more than $100 million in first-year compensation, Wired reports. The offers, which include immediate stock vesting, have been extended to at least 10 OpenAI employees, according to sources familiar with the negotiations. While Meta's aggressive recruiting tactics have caught the attention of top talent, some OpenAI staffers told Wired they're weighing the massive payouts against their potential impact at Meta versus staying at OpenAI. A Meta spokesperson pushed back, claiming reports of the offer sizes are exaggerated. Still, even Meta's senior engineers typically make around $850,000 per year, with those in higher pay bands earning over $1.5 million annually, according to Levels.FYI data. Microsoft's sales overhaul goes all-in on AI. Microsoft's sales chief, Judson Althoff, is reshaping the company's sales organization to double down on AI, according to an internal memo obtained by Business Insider. Althoff's Microsoft Customer and Partner Solutions (MCAPS) unit will now focus on embedding Copilot across devices and roles, deepening Microsoft 365 and Dynamics 365 adoption, winning high-impact AI deals, expanding Azure cloud migration, and strengthening cybersecurity to support AI growth. The memo, sent just one day before Microsoft's latest round of layoffs -- many of which affected Althoff's sales teams -- outlined his vision to make Microsoft "the Frontier AI Firm." According to Business Insider, this restructuring follows Althoff's earlier plan to cut the number of sales solution areas in half starting this fiscal year. FORTUNE ON AI The new CEO flex: Bragging that AI handles exactly X% of the work -- by Sharon Goldman Sam Altman scoffs at Mark Zuckerberg's AI recruitment drive and says Meta hasn't even got their 'top people' -- by Beatrice Nolan Figma files for IPO nearly two years after $20 billion Adobe buyout fell through -- by Allie Garfinkle AI CALENDAR July 8-11: AI for Good Global Summit, Geneva July 13-19: International Conference on Machine Learning (ICML), Vancouver July 22-23: Fortune Brainstorm AI Singapore. Apply to attend here. July 26-28: World Artificial Intelligence Conference (WAIC), Shanghai. Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here. Oct. 6-10: World AI Week, Amsterdam Dec. 2-7: NeurIPS, San Diego Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here. EYE ON AI NUMBERS $65 Billion That's how much U.S. investment in AI companies soared to in the first quarter of this year -- a 33% jump from the previous quarter and a staggering 550% increase compared to the quarter before ChatGPT's 2022 debut, according to PitchBook. The biggest price tag? Data centers. The New York Times reports that Meta, Microsoft, Amazon, and Google plan to spend a combined $320 billion on infrastructure this year -- more than double what they spent just two years ago. A huge chunk of that will go toward building new data centers to keep up with the exploding demand for AI.

Coding is supposed to be genAI's killer use case. But what if its benefits are a mirage? | Fortune

Hello and welcome to Eye on AI...In this edition: Meta is going big on data centers...the EU publishes its code of practice for general purpose AI and OpenAI says it will abide by it...the U.K. AI Security Institute calls into question AI "scheming" research. The big news at the end of last week was that OpenAI's plans to acquire Windsurf, a startup that was making AI software for coding, for $3 billion fell apart. (My Fortune colleague Allie Garfinkle broke that bit of news.) Instead, Google announced that it was hiring Windsurf's CEO Varun Mohan and cofounder Douglas Chen and a clutch of other Windsurf staffers, while also licensing Windsurf's tech -- a deal structured similarly to several other big tech-AI startup not-quite-acquihire acquihires, including Meta's recent deal with Scale AI, Google's deal with Character.ai last year, as well as Microsoft's deal with Inflection and Amazon's with Adept. Bloomberg reported that Google is paying about $2.4 billion for Windsurf's talent and tech, while another AI startup, Cognition, swooped in to buy what was left of Windsurf for an undisclosed sum. Windsurf may have gotten less than OpenAI was offering, but OpenAI's purchase reportedly fell apart after OpenAI and Microsoft couldn't agree on whether Microsoft would have access to Windsurf's tech. The increasingly fraught relationship between OpenAI and Microsoft is worth a whole separate story. So too is the structure of these non-acquisition acquihires -- which really do seem to blunt any legal challenges, either from regulators or the venture backers of the startups. But today, I want to talk about coding assistants. While a lot of people debate the return on investment from generative AI, the one thing seemingly everyone can agree on is that coding is the one clear killer use case for genAI. Right? I mean, that's why Windsurf was such a hot property and why Anyshphere, the startup behind the popular AI coding assistant Cursor, was recently valued at close to $10 billion. And GitHub Copilot is of course the star of Microsoft's suite of AI tools, with a majority of customers saying they get value out of the product. Well, a trio of papers published this past week complicate this picture. Experiment calls gains from AI coding assistants into question METR, a nonprofit that benchmarks AI models, conducted a randomized control trial involving 16 developers earlier this year to see if using code editor Cursor Pro integrated with Anthropic's Claude Sonnet 3.5 and 3.7 models, actually improved their productivity. METR surveyed the developers before the trial to see if they thought it would make them more efficient and by how much. On average, they estimated that using AI would allow them to complete the assigned coding tasks 24% faster. Then the researchers randomized 246 software coding tasks, either allowing them to be completed with AI or not. Afterwards, the developers were surveyed again on what impact they thought the use of Cursor had actually had on the average time to complete the tasks. They estimated that it made them on average 20% faster. (So maybe not quite as efficient as they had forecast, but still pretty good.) But, and now here's the rub, METR found that when assisted by AI it actually took the coders 19% longer to finish tasks. What's going on here? Well, one issue was that the developers, who were all highly experienced, found that Cursor could not reliably generate code as good as theirs. In fact, they accepted less than 44% of the code-generated responses. And when they did accept them, three-quarters of the developers felt the need to still read over every line of AI-generated code to check it for accuracy, and more than half of the coders made major changes to the Cursor-written code to clean it up. This all took time -- on average 9% of the developers time was spent reviewing and cleaning up AI-generated outputs. Many of the tasks in the METR experiment involved large code bases, sometimes consisting of over 100,000 lines of code, and the developers found that sometimes Cursor made strange changes in other parts of this code base that they had to catch and fix. Is it just vibes all the way down? But why did the developers think the AI was making them faster when in fact it was slowing them down? And why, when the researchers followed up with the developers after the experiment ended, did they discover that 69% of the coders were continuing to use Cursor? Some of it seems to be that despite the time it took to edit the Cursor-generated code, the AI assistance did actually ease the cognitive burden for many of the coders. It was mentally easier to fix the AI-generated code than to have to puzzle out the right solution from scratch. So is the perceived ROI from "vibe coding" itself just vibes? Perhaps. That would actually square with what the Wall Street Journal noted about a different area of genAI use -- lawyers using genAI copilots. The newspaper reported that a number of law firms found that given how long it took to fact-check AI-generated legal research, they were not sure lawyers were actually saving any time using the tools. But when they surveyed lawyers, especially junior lawyers, they all reported high satisfaction using the AI copilots and that they felt it made their jobs more enjoyable. But a couple of other studies from last week suggest that maybe it all depends on exactly how you use AI coding assistance. A team from Harvard Business School and Microsoft looked at two years of observations of software developers using GitHub Copilot (which is Microsoft product) and found that those using the tool spent more time on coding and less time on project management tasks, in part because GitHub Copilot allowed them to work independently instead of having to use large teams. It also allowed the coders to spend more time exploring possible solutions to coding problems and less time actually implementing the solutions. This too might explain why coders enjoy using these AI tools -- because it allows them to spend more time on parts of the job they find intellectually interesting -- even if it isn't necessarily about overall time-savings. Maybe the problem is coders just aren't using enough AI? Finally, let's look at the third study, which is from researchers at Chinese AI startup Modelbest, Chinese universities BUPT and Tsinghua University, and the University of Sydney. They found that while individual AI software development tools often struggled to reliably complete complicated tasks, the results improved markedly when multiple large language models were prompted to each take on a specific role in the software development process and to pose clarifying questions to one another aimed at minimizing hallucinations. They called this architecture "ChatDev." So maybe there's a case to be made that the problem with AI coding assistants is how we are using them, not anything wrong with the tech itself? Of course, building teams of AI agents to work in the way ChatDev suggests also uses up a lot more computing power, which gets expensive. So maybe we're still facing that question: is the ROI here a mirage? With that, here's more AI news. Jeremy Kahn [email protected] @jeremyakahn Before we get to the news, the U.S. paperback edition of my book, Mastering AI: A Survival Guide to Our Superpowered Future, is out from Simon & Schuster. Consider picking up a copy for your bookshelf. Also, if you want to know more about how to use AI to transform your business? Interested in what AI will mean for the fate of companies, and countries? Then join me at the Ritz-Carlton, Millenia in Singapore on July 22 and 23 for Fortune Brainstorm AI Singapore. This year's theme is The Age of Intelligence. We will be joined by leading executives from DBS Bank, Walmart, OpenAI, Arm, Qualcomm, Standard Chartered, Temasek, and our founding partner Accenture, plus many others, along with key government ministers from Singapore and the region, top academics, investors and analysts. We will dive deep into the latest on AI agents, examine the data center build out in Asia, examine how to create AI systems that produce business value, and talk about how to ensure AI is deployed responsibly and safely. You can apply to attend here and, as loyal Eye on AI readers, I'm able to offer complimentary tickets to the event. Just use the discount code BAI100JeremyK when you checkout. Note: The essay above was written and edited by Fortune staff. The news items below were selected by the newsletter author, created using AI, and then edited and fact-checked. AI IN THE NEWS White House reverses course, gives Nvida greenlight to sell H20s to China. Nvidia CEO Jensen Huang said the Trump administration is set to reverse course and ease export restrictions on the company's H20 AI chip, with deliveries to resume soon. Nvidia also introduced a new AI chip for the Chinese market that complies with current U.S. rules, as Huang visits Beijing in a diplomatic push to reassure customers and engage officials. While China is encouraging buyers to adopt local alternatives, companies like ByteDance and Alibaba continue to prefer Nvidia's offerings due to their superior performance and software ecosystem. Nvidia's stock and that of TSMC, which makes the chips for Nvidia, jumped sharply on the news. Read more from the Financial Times here. Zuckerberg confirms Meta will spend hundreds of billions in data center push. In a Threads post, Meta CEO Mark Zuckerberg confirmed that the company is spending "hundreds of billions of dollars" to build massive AI-focused data centers, including one called Prometheus set to launch in 2026. The data centers are part of a broader push toward developing artificial general intelligence or "superintelligence." Read more from Bloomberg here. OpenAI and Mistral say they will sign EU code of practice for general-purpose AI. The EU published its code of practice last week for general-purpose AI systems under the EU AI Act, about two months later than initially expected. Adhering to the code, which is voluntary, gives companies assurance that they are in compliance with the Act. The code imposes a stringent set of public and government reporting requirements on frontier AI model developers, requiring them to provide a wealth of information about their models' design and testing to the EU's new AI Office. It also requires public transparency around the use of copyrighted materials in the training of AI systems. You can read more about the code of practice from Politico here. Many had expected the big technology vendors and AI companies to form a united front in opposing the code -- Meta and Google had previously attacked drafts of it, claiming it imposed too great a burden on tech firms -- but OpenAI said in a blog post Friday that it would sign up to the standards. Mistral, the French AI model developer, also said it would sign -- although it had previously asked the EU to delay enforcement of the AI Act, whose provisions on general-purpose AI are set to come into force on August 2nd. That may up the pressure on other AI companies to agree to comply too. Report: AWS is testing a new cloud service to make it easier to use third-party AI models. That's according to a story in The Information, which says Amazon cloud service AWS is making the move after losing business from several AI startups to Google Cloud. Some customers complained it was too difficult to tap models from OpenAI and Google, which are hosted on other clouds, from within AWS. Amazon mulls further multi-billion dollar investment in Anthropic. That's according to a story in the Financial Times. Amazon has already invested $8 billion in Anthropic and the two companies have formed an ever-closer alliance, with Anthropic working with Amazon on several massive new data centers and helping it develop its next generation Trainium2 AI chips. EYE ON AI RESEARCH Could all those studies about scheming AI be faulty? That's the suggestion of a new paper out from a group of researchers at the U.K. government's AI Security Institute. The paper, called "Lessons from a Chimp: AI 'Scheming' and the Quest for Ape Language" examines recent claims that advanced AI models engage in deceptive or manipulative behavior -- what AI Safety researchers call "scheming." Drawing an analogy to 1970s research about whether non-human primates were capable of using language -- which ultimately were found to have overstated the depth of linguistic capacity that chimpanzees possess -- the authors argue that the AI scheming literature suffers from similar flaws. Specifically, the researchers say the AI scheming research suffers from an over-interpretation of anecdotal behavior, a lack of theoretical clarity, an absence of rigorous controls, and a reliance on anthropomorphic language. They caution that current studies often confuse AI systems following human-provided instructions with intentional deception and may exaggerate the implications of observed behaviors. While acknowledging that scheming could pose future risks, the authors call for more scientifically robust methodologies before drawing strong conclusions. They offer concrete recommendations, including clearer hypotheses, better experimental controls, and more cautious interpretation of AI behavior. FORTUNE ON AI The world's best AI models operate in English. Other languages -- even major ones like Cantonese -- risk falling further behind -- by Cecilia Hult How to know which AI tools are best for your business needs -- with examples -- by Preston Fore Jensen Huang says AI isn't likely to cause mass layoffs unless 'the world runs out of ideas' -- by Marco Quiroz-Gutierrez Commentary: I'm leading the largest global law firm as AI transforms the legal profession. Lawyers must double down on this one skill -- by Kate Barton AI CALENDAR July 13-19: International Conference on Machine Learning (ICML), Vancouver July 22-23: Fortune Brainstorm AI Singapore. Apply to attend here. July 26-28: World Artificial Intelligence Conference (WAIC), Shanghai. Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here. Oct. 6-10: World AI Week, Amsterdam Oct. 21-22: TedAI San Francisco. Apply to attend here. Dec. 2-7: NeurIPS, San Diego Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here. BRAIN FOOD AI is not going to save the news media. I've been thinking a lot about AI's impact on the news media lately both because it happens to be the industry I'm in and also because Fortune has recently started experimenting more with using AI to produce some of our basic news stories. (I use AI a bit to produce the short news blurbs for this newsletter too, although I don't use it to write the main essay.) Well, Jason Koebler, a cofounder of tech publication 404 Media, has an interesting essay out this week on why he thinks many media organizations are being misguided in their efforts to use AI to produce news more efficiently. He argues that the media's so-called "pivot to AI" is a mirage -- a desperate, misguided attempt by executives to appear forward-thinking while ignoring the structural damage AI is already inflicting on their businesses. He argues that many news execs are imposing AI on newsrooms with no clear business strategy beyond vague promises of innovation. He says this approach won't work: relying on the same tech that's gutting journalism to save it is both delusional and self-defeating. Instead, he argues, the only viable path forward is to double down on what AI can't replicate: trustworthy, personality-driven, human journalism that resonates with audiences. AI may offer productivity boosts at the margins -- transcripts, translations, editing tools -- but these don't add up to a sustainable model. You can read his essay here.

What Eric Xing's Abu Dhabi project says about the next phase of AI power | Fortune

Hello and welcome to Eye on AI...In this edition: my chat with AI leader Eric Xing...Trump's AI export plan...drama at the International Math Olympiad...Stargate update...transparency in reasoning. I was excited and curious to meet Eric Xing last week in Vancouver, where I was attending the International Conference on Machine Learning -- one of the top AI research gatherings of the year. Why? Xing, a longtime Carnegie Mellon professor who moved to Abu Dhabi in 2020 to lead the public, state-funded Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), sits at the crossroads of nearly every big question in AI today: research, geopolitics, even philosophy. The UAE, after all, has quietly become one of the most intriguing players in the global AI race. The tiny Gulf state is aligning itself with U.S.-style norms around intellectual freedom and open research -- even as the AI rivalry between the U.S. and China becomes increasingly defined by closed ecosystems and strategic competition. The UAE isn't trying to "win" the AI race, but it wants a seat at the table. Between MBZUAI and G42-its state-backed AI-focused conglomerate-the UAE is building AI infrastructure, investing in talent, and aggressively positioning itself as a go-to partner for American firms like OpenAI and Oracle. And Xing is at the heart of it. As it happened, Xing and I just missed each other -- he arrived in Vancouver as I was heading home -- so we connected on Zoom the following day. Our conversation ranged widely, from the hype around "world models" to how the UAE is using open-source AI research as a strategic lever to build soft power. Here are a few of the most compelling takeaways: A 'Bell Labs plus a university' MBZUAI is just five years old, but Xing says it's already among the fastest-growing academic institutions in the world. The school, which is mostly a graduate program for AI researchers, aspires to compete with elite institutions like MIT and Carnegie Mellon while also taking on applied research challenges. Xing calls it a hybrid organization, similar to "Bell Labs plus a university," referring to the legendary R&D arm of AT&T, founded in 1925 and responsible for foundational innovations that shaped modern computing, communications, and physics. The UAE as a soft-power AI ambassador Xing sees MBZUAI not just as a university, but as part of the UAE's broader effort to build soft power in AI. He describes the country as a "strong island" of U.S. alignment in the Middle East, and views the university as an "ambassador center" for American-style research norms: open source, intellectual freedom, and scientific transparency. "If the U.S. wants to project influence in AI, it needs institutions like this," he told me. "Otherwise, other countries will step in and define the direction." The U.S. isn't losing the AI race While much of the public narrative around AI focuses on a U.S.-China race, Xing doesn't buy the framing. "There is no AI war," he said flatly. "The U.S. is way ahead in ideas, in people, and in the innovation environment." In his view, China's AI ecosystem is still constrained by censorship, hardware limitations, and a weaker bottom-up innovation culture. "Many top AI engineers in the U.S. may be of Chinese origin," he said, "but they only became top engineers after studying and working in the U.S." Why open source matters For Xing, open source isn't just a philosophical preference -- it's a strategic choice. At MBZUAI, he's pushing for open research and open-source AI development as a way to democratize access to cutting-edge tools, especially for countries and researchers outside the U.S.-China power centers. "Open source applies pressure on closed systems," he told me. "Without it, fewer people would be able to build with -- or even understand -- these technologies." At a time when much of AI is becoming siloed behind corporate walls, Xing sees MBZUAI's open approach as a way to foster global talent, advance scientific understanding, and build credibility for the UAE as a hub for responsible AI development. On 'world models' and AI hype Xing didn't hold back when it came to one of the buzziest trends in AI right now: so-called "world models" -- systems that aim to help AI agents learn by simulating how the world works. He's skeptical of the hype. "Right now people are building pretty video generators and calling them world models," he said. "That's not reasoning. That's not simulation." In a recent paper he spent months writing himself -- unusual for someone of his seniority -- he argues that true world models should go beyond flashy visuals. They should help AI reason about cause and effect, not just predict the next frame of a video. In other words: AI needs to understand the world, not just mimic it. With that, here's the rest of the AI news -- including that tomorrow the White House is set to release a sweeping new AI strategy aimed at boosting the global export of U.S. AI technologies while cracking down on state-level regulations that are seen as overly restrictive. I will be attending the D.C. event, which includes a keynote by President Trump, and will report back. Sharon Goldman [email protected] @sharongoldman AI IN THE NEWS White House to unveil plan to push global export of U.S. AI and crack down on restrictions. According to a draft seen by Reuters, the White House is set to release a sweeping new AI strategy Wednesday aimed at boosting the global export of U.S. AI technologies while cracking down on state-level regulations seen as overly restrictive. The plan will bar federal AI funding from states with tough AI laws, promote open-source and open-weight AI development, and direct the Commerce Department to lead overseas data center and deployment efforts. It also tasks the FCC with reviewing potential conflicts between federal goals and local rules. Framed as a push to make "America the world capital in artificial intelligence," the plan reflects President Trump's January directive and will be unveiled during a "Winning the AI Race" event co-hosted by the All-In podcast and featuring White House AI czar David Sacks. OpenAI and Google DeepMind sparked math drama. Over the past few days, both OpenAI and Google DeepMind claimed their AI models had achieved gold-medal-level performance on the 2025 International Mathematical Olympiad -- successfully solving 5 out of 6 notoriously difficult problems. It was a milestone that many considered years away: a general reasoning LLM reaching that level of performance under the same time limits as humans, without tools. But the way they announced it sparked controversy. OpenAI released its results first, based on its own evaluation using IMO-style questions and human graders -- before any official verification. That prompted criticism from prominent mathematicians, including Terence Tao, who questioned whether the problems had been altered or simplified. In contrast, Google entered the competition officially, waited for the IMO's independent review, and only then declared its Gemini DeepThinker model had earned a gold medal -- making it the first AI system to be formally recognized by the IMO as performing at that level. The drama laid bare the high stakes -- and differing standards -- for credibility in the AI race. SoftBank and OpenAI are reportedly struggling to get $500 Billion Stargate AI Project off the ground. According to the Wall Street Journal, the $500 billion Stargate project -- announced with fanfare at the White House six months ago by Masayoshi Son, Sam Altman, and President Trump -- has hit major turbulence. Billed as a moonshot to supercharge U.S. AI infrastructure, the initiative has yet to break ground on a single data center, and internal disagreements between SoftBank and OpenAI over key terms like site location have delayed progress. Despite promises to invest $100 billion "immediately," Stargate is now aiming for a scaled-down launch: a single, small facility, likely in Ohio, by year's end. It's a setback for Son, who recently committed a record-breaking $30 billion to OpenAI but is still scrambling to secure a meaningful foothold in the AI arms race. However, Bloomberg reported today that Oracle will provide OpenAI with 2 million new AI chips that will be part of a massive data center expansion that OpenAI labeled as part of its Stargate project. SoftBank, though, isn't financing any of the new capacity -- and it's unclear what operator will be developing data centers to support the new capacity, and when they will be built. EYE ON AI RESEARCH Sounding the alarm on growing opacity of advanced AI reasoning models. Fortune reporter Beatrice Nolan reported this week on a group of 40 AI researchers, including contributors from OpenAI, Google DeepMind, Meta, and Anthropic, that are sounding the alarm on the growing opacity of advanced AI reasoning models. In a new paper, the authors urge developers to prioritize research into "chain-of-thought" (CoT) processes, which provide a rare window into how AI systems make decisions. They are warning that as models become more advanced, this visibility could vanish. The "chain-of-thought" process, which is visible in reasoning models such as OpenAI's o1 and DeepSeek's R1, allows users and researchers to monitor an AI model's "thinking" or "reasoning" process, illustrating how it decides on an action or answer and providing a certain transparency into the inner workings of advanced models. The researchers said that allowing these AI systems to "'think' in human language offers a unique opportunity for AI safety," as they can be monitored for the "intent to misbehave." However, they warn that there is "no guarantee that the current degree of visibility will persist" as models continue to advance. The paper highlights that experts don't fully understand why these models use CoT or how long they'll keep doing so. The authors urged AI developers to keep a closer watch on chain-of-thought reasoning, suggesting its traceability could eventually serve as a built-in safety mechanism. FORTUNE ON AI Mark Cuban says the AI war 'will get ugly' and intellectual property 'is KING' in the AI world -- by Sydney Lake $61.5 billion tech giant Anthropic has made a major hiring U-turn -- now, it's letting job applicants use AI months after banning it from the interview process -- by Emma Burleigh Experienced software developers assumed AI would save them a chunk of time. But in one experiment, their tasks took 20% longer -- by Sasha Rogelberg AI CALENDAR July 26-28: World Artificial Intelligence Conference (WAIC), Shanghai. Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here. Oct. 6-10: World AI Week, Amsterdam Oct. 21-22: TedAI San Francisco. Dec. 2-7: NeurIPS, San Diego Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here.

AI models are getting very good at professional tasks, new OpenAI research shows | Fortune

One of the reasons for the seeming disparity in AI's capabilities is that many AI benchmarks do not reflect real world use cases. Which is why a new gauge published by OpenAI last week is so important. Called GDPval, the benchmark evaluates leading AI models on real-world tasks, curated by experts from across 44 different professions, representing nine different sectors of the economy. The experts had an average of 14 years experience in their fields, which ranged from law and finance to retail and manufacturing, as well as government and healthcare. Whereas a traditional AI benchmark might test a model's capability to answer a multiple choice bar exam question about contract law, for example, the GDPval assessment asks the AI model to craft an entire 3,500 word legal memo assessing the standard of review under Delaware law that a public company founder and CEO, with majority control, would face if he wanted this public company to acquire a private company that he also owned. OpenAI tested not only its own models, but those from a number of other leading labs, including Google DeepMind's Gemini 2.5 Pro, Anthropic's Claude Opus 4.1, and Grok's Grok 4. Of these, Claude Opus 4.1 consistently performed the best, beating or equaling human expert performance on 47.6% of the total tasks. (Big kudos to OpenAI for intellectual honesty in publishing a study in which its own models were not top of the heap.) There was a lot of variance between models, with Gemini and Grok often able to complete between a third and a fifth of tasks at or above the standard of human experts, while OpenAI's GPT-5 Thinking's performance fell between that of Claude Opus 4.1 and Gemini, and OpenAI's earlier model, GPT-4o, fared the worst of all, barely able to complete 10% of the tasks to professional standard. GPT-5 was the best at following a prompt correctly, but often failed to format its response properly, according to the researchers. Gemini and Grok seemed to have the most problems with following instructions -- sometimes failing to provide the delivered outcome and ignoring reference data -- but OpenAI did note that "all the models sometimes hallucinated data or miscalculated." Big differences across sectors and professions There was also a bit of variance between economic sectors, with the models performing best on tasks from government, retail, and the wholesale trade, and generally worst on tasks from the manufacturing sector. For some professional tasks, Claude Opus 4.1's performance was off the charts: it beat or equalled human performance for 81% of the tasks taken from "counter and rental clerks," 76% of those taken from shipping clerks, 70% of those from software development, and, intriguingly, 70% of the tasks taken from the work of private investigators and detectives. (Forget Sherlock Holmes, just call Claude!) GPT-5 Thinking beat human experts on 79% of the tasks that sales manager perform and 75% of those that editors perform (gulp!). On others, human experts won handily. The models were all notably poor at performing tasks related to the work of film and video editors, producers and directors, and audio and video technicians. So Hollywood may be breathing a sigh of relief. The models also fell down on tasks related to pharmacists' jobs. When AI models failed to equal or exceed human performance, it was rarely in ways that human experts judged "catastrophic" -- that only occurred about 2.7% of the time with GPT-5 failures. But the GPT-5 response was judged "bad" in another 26.7% of these cases, and "acceptable but subpar" in 47.7% of cases where human outputs were deemed superior. The need for 'Centaur' benchmarks I asked Erik Brynjolfsson, the Stanford University economist at the Human-Centered AI Institute (HAI) who has done some of the best research to date on the economic impact of generative AI, what he thought of GDPval and the results. He said the assessment goes a long way to closing the gap that has developed between AI researchers and their preferred benchmarks, which are often highly technical but don't match real-world problems. Brynjolfsson said he thought GDPval would "inspire AI researchers to think more about how to design their systems to be useful in doing practical work, not just ace the technical benchmarks." He also said that "in practice, that means integrating technology into workflows and more often than not, actively involving humans." Brynjolfsson said he and colleague Andy Haupt had been arguing for "Centaur Evaluations" which judge how well humans perform when paired with, and assisted by, an AI model, rather than always seeing the AI model as a replacement for human workers. (The term comes from the idea of "centaur chess," which is what it is called when human grandmasters are assisted by chess computers. The pairing was found to exceed what either humans or machines could do alone. And, of course, centaur here refers to the mythical half-man, half-horse of Greek mythology.) GDPval did make some steps toward doing this, looking in one case at how much time and money was saved when OpenAI's models were allowed to try a task multiple times, with the human then coming in to fix the output if it was not up to standard. Here, GPT-5 was found to offer both a 1.5x speedup and 1.5x cost improvement over the human expert working without AI assistance. (Less capable OpenAI models did not help as much, with GPT-4o actually leading to a slowdown and cost increase over the human expert working unassisted.) About that AI workslop research... This last point, along with the "acceptable but subpar" label that characterized a good portion of the cases where the AI models did not equal human performance, brings me back to that "workslop" research that came out last week. This may, in fact, be what is happening with some AI outputs in corporate settings, especially as the most capable models -- such as GPT-5, Claude 4.1 Opus, and Gemini 2.5 Pro -- are only being used by a handful of companies at scale. That said, as the journalist Adam Davidson pointed out in a Linkedin post, the "Workslop" study -- just like that now infamous MIT study about 95% of AI pilots failing to produce ROI -- had some very serious flaws. The "workslop" study was based on an open online survey that asked highly leading questions. It was essentially a "push poll" designed to generate an attention-grabbing headline about the problem of AI workslop more than a piece of intellectually-honest research. But it worked -- it got lots of headlines, including in Fortune. If one focuses on these kinds of headlines, it is all too easy to miss the other side of what is happening in AI, which is the story that GDPval tells: the best performing AI models are already on par with human expertise on many tasks. (And remember that GDPval has so far been tested only on Anthropic's Claude Opus 4.1, not its new Claude Sonnet 4.5 that was released yesterday and which can work continuously on a task for up to 30 hours, far longer than any previous model.) This doesn't mean AI can replace these professional experts any time soon. As Brynjolfsson's work has shown, most jobs consist of dozens of different tasks, and AI can only equal or beat human performance on some of them. In many cases, a human needs to be in the loop to correct the outputs when a model fails (which, as GDPval shows, is still happening at least 20% of the time, even on the professional tasks where the models perform best.) But AI is making inroads, sometimes rapidly, in many domains -- and more and more of its outputs are not just workslop. With that, here's more AI news. Jeremy Kahn [email protected] @jeremyakahn Before we get to the news, I want to call your attention to the Fortune AIQ 50, a new ranking which Fortune just published today that evaluates how Fortune 500 companies are doing in deploying AI. The ranking shows which companies, across 18 different sectors -- from financials to healthcare to retail -- are doing best when it comes to AI, as judged by both self-assessments and peer reviews. You can see the list here, and catch up on Fortune's ongoing AIQ series. FORTUNE ON AI OpenAI rolls out 'instant' purchases directly from ChatGPT, in a radical shift to e-commerce and a direct challenge to Google -- by Jeremy Kahn Anthropic releases Claude Sonnet 4.5, a model it says can build software and accomplish business tasks autonomously -- by Beatrice Nolan Nvidia's $100 billion OpenAI investment raises eyebrows and a key question: How much of the AI boom is just Nvidia's cash being recycled? -- by Jeremy Kahn Ford CEO warns there's a dearth of blue-collar workers able to construct AI data centers and operate factories: 'Nothing to backfill the ambition' -- by Sasha Rogelberg EYE ON AI NEWS Meta locks in $14 billion worth of AI compute. The tech giant struck a $14 billion multi-year deal with CoreWeave to secure access to Nvidia GPUs (including next-gen GB300 systems). It's another sign of Big Tech's arms race for AI capacity. The pact follows CoreWeave's recent expansion tied to OpenAI and sent CoreWeave shares up. Read more from Reuters here. California governor signs landmark AI law. Governor Gavin Newsom signed SB 53 into law on Monday. The new AI legislation requires developers of high-end AI systems to publicly disclose safety plans and report serious incidents. The law also adds whistleblower protections for employees of AI companies and a public "CalCompute" cloud to broaden research access to AI. Large labs must outline how they mitigate catastrophic risks, with penalties for non-compliance. The measure -- authored by State Senator Scott Wiener -- follows last year's veto of a stricter bill that was roundly opposed by Silicon Valley heavyweights and AI companies. This time, some AI companies, such as Anthropic, as well as Elon Musk, supported SB 53, while Meta, Google and OpenAI opposed it. Read more from Reuters here. OpenAI's revenue surges -- but its burn rate remains dramatic. The AI company generated about $4.3 billion in the first half of 2025 -- up 16% on all of 2024, according to financial details it disclosed to its investors and which were reported by The Information. But the company still had a burn rate of $2.5 billion over that same time period due to aggressive spending on R&D and AI infrastructure. The company said it is targeting about $13 billion in revenue for 2025, but with a total cash burn of $8.5 billion. OpenAI is in the middle of a secondary share sale that could value the company at $500 billion, almost double its valuation of $260 billion at the start of the year. Apple is testing a stronger, still-secret model for Apple Intelligence. That's according to a report from Bloomberg, which cited unnamed sources it said were familiar with the matter. The news agency said Apple is trialing a ChatGPT-style app powered by an upgraded AI mode internally, with the aim to use it to overhaul its digital assistant Siri. The new chatbot would be rolled out as part of upcoming Apple Intelligence updates, Bloomberg said. Opera launches Neon, an "agentic" AI browser. In a further sign that AI has rekindled the browser wars, the browser company Opera rolled out Neon, a browser with built-in AI that can execute multi-step tasks (think booking travel or generating code) from natural-language prompts. Opera is charging a subscription for Neon. It joins Perplexity's Comet and Google roll out of Gemini in Chrome in the increasingly competitive field of AI browsers. Read more from Tech Crunch here. Black Forest Labs in talks to raise $200 million to $300 million at $4 billion valuation. That's according to a story in the Financial Times. It says the somewhat secretive German image-generation startup (makers of the Flux models and founded by ex-Stable Diffusion employees) is negotiating a new venture capital round that would value the company around $4 billion, up from roughly $1 billion last year. The round would mark one of Europe's largest recent AI financings and underscores investor appetite for next-generation visual models. EYE ON AI RESEARCH Can an AI model beat VCs at spotting winning startups? Yes, it can, according to a new study conducted by researchers from the University of Oxford and AI startup Vela Research/ They created a new assessment they call VCBench, built from 9,000 anonymized founder profiles, to evaluate if LLMs can predict startup success better than human investors. (Of these 9,000 founders, 9% went on to see their companies either get acquired, raise more than $500 million in funding, or IPO at more than a $500 million valuation.) In their tests, some models far out-performed the record of venture capital firms, which in general pick a winner about one in every 20 bets they make. OpenAI's GPT-5 scored a winner about half the time, while DeepSeek-V3 was the most accurate, selecting winners six out of every 10 times, and doing so at a lower cost than most other models. Interestingly, a different machine learning technique from Vela, called reasoned rule mining, was more accurate still, hitting a winner 87.5% of the time. (The researchers also tried to ensure that the LLMs were not simply finding a clever way to re-identify the people whose anonymized profiles make up the dataset and cheat by simply looking up what had happened to their companies. The researchers say they were able to reduce this chance to the point where it was unlikely to be the case.) The researchers are publishing a public leaderboard at vcbench.com. You can read more about the research here on arxiv.org and in the Financial Times here. AI CALENDAR Oct. 6: OpenAI DevDay, San Francisco Oct. 6-10: World AI Week, Amsterdam Oct. 21-22: TedAI San Francisco. Nov. 10-13: Web Summit, Lisbon. Nov. 26-27: World AI Congress, London. Dec. 2-7: NeurIPS, San Diego Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here. BRAIN FOOD Are world models and reinforcement learning all we need? There was a big controversy among AI researchers and other industry insiders this past week over the appearance of Turing Award-winner and AI research legend Rich Sutton on the Dwarkesh podcast. Sutton argued that LLMs are actually a dead end that will never achieve AGI because they can only ever imitate human knowledge and they don't construct a "world model" -- a way of predicting what will happen next based on an intuitive understanding of things such as the laws of physics or, even, human nature. Dwarkesh pushed back, suggesting to Sutton that LLMs did, in fact, have a kind of world model, but Sutton was having none of it. Some -- such as AI skeptic Gary Marcus-interpreted what Sutton said on Dwarkesh as a major reversal from the position he had taken in a famous essay, "The Bitter Lesson," published in 2019, which argued that progress in AI mostly depended on using the same basic algorithms but simply throwing more compute and more data at them, rather than any clever algorithmic innovation. "The Bitter Lesson" has been waved like a bloody flag by those who have argued that "scale is all we need" -- building ever bigger LLMs on ever larger GPU clusters -- to achieve AGI. But Sutton never wrote explicitly about LLMs in "The Bitter Lesson" and I don't think his Dwarkesh remarks should be interpreted as a departure from his position. Instead, Sutton has always been first and foremost an advocate of reinforcement learning in environments where the reward signal comes entirely from the environment, with an AI model acting agentically and acquiring experience -- building a model of "the rules of the game" as well as the most rewarding actions in any given situation. Sutton doesn't like the way LLMs are trained, with unsupervised learning from human text followed by a kind of RL using human feedback -- because everything the LLM can learn is inherently limited by human knowledge and human preferences. He has always been an advocate for the idea of pure tabula rasa learning. To Sutton, LLMs are a big departure from tabula rasa, and so it is not surprising he sees them as a dead end to AGI.

AI politics breaks into a New York congressional race -- and signals more fights to come | Fortune

Welcome to Eye on AI, with AI reporter Sharon Goldman in for Jeremy Kahn, who is traveling. In this edition...AI politics in a New York congressional race...Microsoft, NVIDIA and Anthropic announce strategic partnerships...Cloudflare outage causes internet outages including AI sites such as OpenAI, Anthropic and Perplexity...Jeff Bezos creates AI start-up where he will be co-CEO..Sam Altman and Masayoshi Son back new AI research lab aiming to revive the spirit of Bell Labs...Japanese AI darling Sakana AI raises $135 million at $2.65 billion valuation. Leading the Future -- a $100 million pro-AI super PAC formed in August and backed by Andreessen Horowitz and OpenAI president Greg Brockman -- has identified its first target: Alex Bores, a Democratic congressional candidate running for the New York seat being vacated by Rep. Jerrold Nadler after three decades in Congress. It's an early signal of a broader shift: while AI won't determine every race in the upcoming midterms, it is emerging as a potent new pressure point in American politics, particularly as deep-pocketed Silicon Valley interests begin injecting themselves into local contests from afar. Bores is the chief sponsor of New York's RAISE Act, which would require large AI labs to create and follow safety plans designed to prevent critical harms; disclose serious safety incidents, such as the theft of an AI model; and avoid releasing systems that pose "unreasonable" risks. Companies that fail to comply could face civil penalties of up to $30 million. The legislation -- similar to a California's SB-1047, a bill vetoed by California Governor Gavin Newsom last year -- awaits Governor Kathy Hochul's signature and has attracted support from AI safety advocates, policy groups, and prominent researchers including AI "godfathers" Yoshua Bengio and Geoffrey Hinton. But many AI researchers, engineers, founders, and major tech investors see both the California bill and the RAISE Act as imposing vague, overly burdensome requirements that could be unworkable in practice -- especially for startups. Josh Vlasto, co-head of Leading the Future and a spokesperson for Fairshake, the $141 million crypto-aligned super PAC, told me that Bores "has championed a piece of legislation that would contribute to a national patchwork that is not workable and has not engaged productively with the industry." Bores, for his part, is leaning into the role of combatant after learning he would be the PAC's first target. "It doesn't surprise me," he said. "They said they were going to target four states -- California, Ohio, Illinois and New York -- so I kind of figured who they were thinking about in New York." He dismissed opposition to the RAISE Act as "an extremely loud minority that has decided to yell over the broad majority support by spending hundreds of millions of dollars," because they don't believe there should be regulation on AI, though he stressed the bill is not a partisan flashpoint. "The RAISE Act passed in New York with every single Republican state senator voting for it, and a majority of the Republican state assembly members voting for it, including a number who co-sponsored it," he said. "Republicans like Sarah Lightner in Michigan have introduced similar bills, and we conducted a poll that found 84% of New Yorkers supported the bill. There is strong bipartisan support for lightweight, reasonable regulations to keep people safe." Leading the Future, however, rejects the idea that it is opposed to regulation. "It's not true that Leading the Future is anti-regulation," Vlasto said. "The idea [that] we are trying to stop Congress from acting is just wrong, and we have been clear about it since our launch in August." He argued that AI safety advocates have long enjoyed a structural advantage. "The other side has spent billions over the past decade investing in political organizations and think tanks," he added. He pointed to groups like Open Philanthropy, a grant-making organization funded largely by Facebook co-founder Dustin Moskovitz and his wife, Cari Tuna. It grew out of the Effective Altruism (EA) movement, whose donors focus on areas they view as high-impact but under-resourced -- including global health, biosecurity, and long-term or "existential" risks from advanced AI. Vlasto would not comment on whether Leading the Future will be targeting Democratic California State Senator Scott Wiener, who co-sponsored California's SB-1047 bill and is now running to fill Nancy Pelosi's vacant Congressional seat. But with Silicon Valley money flowing in and rising debates over AI regulation, it's clear this first strike won't be the last. With that, here's more AI news. Sharon Goldman [email protected] @sharongoldman If you want to learn more about how AI can help your company to succeed and hear from industry leaders on where this technology is heading, I hope you'll consider joining Jeremy and I at Fortune Brainstorm AI San Francisco on Dec. 8-9. Among the speakers confirmed to appear so far are Google Cloud chief Thomas Kurian, Intuit CEO Sasan Goodarzi, Databricks CEO Ali Ghodsi, Glean CEO Arvind Jain, Amazon's Panos Panay, and many more. Register now. FORTUNE ON AI Nvidia's rise seemed unstoppable, but cracks may be appearing in the strategy that built its $4.5 trillion empire - by Shawn Tully Google releases its heavily hyped Gemini 3 AI in a sweeping rollout -- even Search gets it on day one - by Sharon Goldman 'Trust is at an all-time low for both job seekers and recruiters': Hiring platform CEO says talent acquisition is in an 'AI doom loop' - by Nino Paoli Despite AI bubble fears, Warren Buffett's Berkshire Hathaway loads up on shares of hyperscaler Alphabet amid huge rally - by Jason Ma AI IN THE NEWS Microsoft, NVIDIA and Anthropic announce strategic partnerships. Microsoft, NVIDIA, and Anthropic unveiled a sweeping set of partnerships today that dramatically expand Claude's reach and Anthropic's compute footprint. Anthropic committed to purchase $30 billion of Azure compute and secure up to one gigawatt of capacity as it scales Claude on Microsoft's cloud, while also deepening integration across Microsoft's Copilot ecosystem and Foundry. At the same time, Anthropic and NVIDIA are launching their first major technology partnership, collaborating on model and chip design to optimize performance and efficiency on upcoming Grace Blackwell and Vera Rubin systems. The deal makes Claude the only frontier model available across all three major clouds and comes with major financial backing: NVIDIA will invest up to $10 billion in Anthropic and Microsoft up to $5 billion. Cloudflare outage causes internet outages including AI sites such as OpenAI, Anthropic and Perplexity. An outage at the internet infrastructure company Cloudflare on Tuesday disrupted major websites globally, including OpenAI's ChatGPT, Anthropic's Claude, and Perplexity. According to CNBC, Cloudflare, which manages and secures traffic for an estimated 20% of the web, stated that the problem stemmed from a "spike in unusual traffic" to one of its services around 6:20 a.m. ET, though the cause of the spike remains unknown, leading to the company's shares sliding over 3% amid ongoing efforts to implement a fix. Jeff Bezos creates AI start-up where he will be co-CEO. The New York Times reported that Jeff Bezos, the founder of Amazon and one of the world's wealthiest people, Jeff Bezos, the founder of Amazon, is officially returning to a formal operational role as Co-CEO of a new AI research start-up named Project Prometheus. Since stepping down as Amazon's CEO in July 2021, this marks the first time he has taken such a hands-on position, distinguishing it from his role as founder at Blue Origin. Project Prometheus is launching with a whopping $6.2 billion in funding, partly contributed by Bezos himself, making it one of the most heavily financed early-stage ventures globally and signaling a serious and well-resourced entry into the AI race. Sam Altman and Masayoshi Son back new AI research lab aiming to revive the spirit of Bell Labs. According to Ashlee Vance's Core Memory, Louis Andre, a little-known 27-year-old scientist with a background in neuroscience and computer science from University College London and stints at Princeton, Stanford, and a Brin-backed biotech startup, is launching Episteme -- an ambitious, Altman- and Masayoshi Son-backed research lab in San Francisco aiming to revive the spirit of Bell Labs and Xerox PARC. Designed as a "third way" between academia and startups, Episteme will give top scientists generous pay, resources, ownership, and freedom from grant writing and short-term commercial pressures while surrounding them with support staff to help turn breakthrough ideas into real products. Starting with 15 researchers across fields like AI, energy, materials, and neuroscience, the project hopes to counter declining U.S. scientific investment, growing bureaucracy, and competition from China by creating a long-term, idealistic environment where risky, high-impact research can thrive -- though, like similar billionaire-funded labs, it still faces questions about sustainability and investor patience. Japanese AI darling Sakana AI raises $135 million at $2.65 billion valuation. The global race to develop large language models, led by U.S. giants is being challenged by specialized startups like Tokyo-based Sakana AI, which TechCrunch reported recently closed a ¥20 billion (approximately $135 million) Series B funding round, valuing the company at $2.65 billion. Founded in 2023 by former Google researchers Llion Jones, Ren Ito, and CEO David Ha, Sakana AI focuses on creating affordable, generative AI models specifically optimized for the Japanese language and culture, which also work efficiently with small datasets. The funding round, which included Japanese financial firms like MUFG and global investors such as Khosla Ventures and NEA, will be deployed for further R&D, model development, and expanding the engineering, sales, and distribution workforce in Japan, as the company builds on existing partnerships with local enterprises and plans to expand its enterprise business into the industrial, manufacturing, and government sectors by 2026. EYE ON AI RESEARCH An AI system takes a gold medal in physics. A new research paper claims that an open-source AI system has reached gold-medal performance on the world's hardest high-school physics competition -- the International Physics Olympiad. Countries send their absolute best teenage physics students to this contest, and the questions are so challenging that even many PhDs struggle with them. The researchers built a system called P1, which is a combination of a large language model trained on science-heavy data; a reinforcement learning process that teaches the model to reason step-by-step; and an "agentic" setup that lets the model break problems apart, try multiple solution approaches, check itself, and refine its answers -- similar to how a human would tackle Olympiad puzzles Using this setup, the AI solved past Physics Olympiad problems at a level that would earn a gold medal if it were a human competitor. It's one of the clearest signs yet that AI is not just getting better at language and coding -- it's now beginning to reach elite levels of scientific reasoning. However, don't throw away your physics textbooks just yet. The research simply means AI systems can reliably work through extremely difficult problems that require deep conceptual thinking, careful math, and multi-step logic -- something out of reach just a year or two ago. AI CALENDAR Nov. 19: Nvidia third quarter earnings Nov. 26-27: World AI Congress, London. Dec. 2-7: NeurIPS, San Diego Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here. BRAIN FOOD I had to shout-out this New York Times piece, Can You Believe the Documentary You're Watching? I so relate: the rise of AI video has made me incredibly suspicious of every online image, scene or voice. I have even seen disturbing, fake Holocaust footage created by AI. I'm also a diehard documentary fan who, while I believe there can be some use for AI in documentary production, also understands that there needs to be clear boundaries and full transparency. The New York Times piece points out that both filmmakers and viewers need to help preserve trust and authenticity in documentary storytelling. That may mean new industry norms like voluntary certifications, greater transparency about how films are made, and more documentarians stepping into their own work to explain their methods. Perhaps documentarians, the author explains, "are the ones best suited to help us rethink what trust, transparency and authenticity really look like when we can't believe our eyes."

Inside Anthropic Claude is boosting developer productivity -- but raising fears over deskilling and loss of job satisfaction | Fortune

Welcome to Eye on AI. In this edition...Anthropic is winning over business customers, but how are its own engineers using its Claude AI models...OpenAI CEO Sam Altman declares a "code red"...Apple reboots its AI efforts -- again...Former OpenAI chief scientist Ilya Sutskever says "it's back to the age of research" as LLMs won't deliver AGI...Is AI adoption slowing? OpenAI certainly has the most recognizable brand in AI. As company founder and CEO Sam Altman said in a recent memo to staff, "ChatGPT is AI to most people." But while OpenAI is increasingly focused on the consumer market -- and, according to news reports declaring "a code red" in response to new, rival AI models from Google (see the "Eye on AI News" section below) -- it may already be lagging in the competition for enterprise AI. In this battle for corporate tech budgets, one company has quietly emerged as the vendor big business customers seem to prefer: Anthropic. Anthropic has, according to some research, moved past OpenAI in enterprise marketshare. A Menlo Ventures survey from the summer showed Anthropic with a 32% market share by model usage compared to OpenAI's 25% and Google's 20%. (OpenAI disputes these numbers, noting that Menlo Ventures is an Anthropic investor and that the survey had a small sample size. It says that it has 1 million paying business customers compared to Anthropic's 330,000.) But estimates in a HSBC research report on OpenAI that was published last week also give Anthropic a 40% marketshare by total AI spending compared to OpenAI's 29% and Google's 22%. How did Anthropic take the poll position in the race for enterprise AI adoption? That's the question I set out to answer in the latest cover story of Fortune magazine. For the piece, I had exclusive access to Anthropic cofounder and CEO Dario Amodei and his sister Daniela Amodei, who serves as the company's president and oversees much of its day-to-day operations, as well as to numerous other Anthropic execs. I also spoke to Anthropic's customers to find out why they've come to prefer its Claude models. Claude's prowess at coding, an area Anthropic devoted attention to early on, is clearly one reason. (More on that below.) But it turns out that part of the answer has to do with Anthropic's focus on AI safety, which has given corporate tech buyers some assurance that its models are a less risky than competitors'. It's a logic that undercuts the argument of some Anthropic critics, including powerful figures such as White House AI and crypto czar David Sacks, who see the company's advocacy of AI safety testing requirements as a mistaken policy that will slow AI adoption. Now the question facing Anthropic is whether it can hold on to its lead, raise enough funds to cover its still massive burn rate, and manage its hypergrowth without coming apart at the seams. Do you think Anthropic can go the distance? Give the story a read here and let me know what you think. How is AI changing coding? Now, back to Claude and coding. In March, Dario Amodei made headlines when he said that by the end of the year 90% of software code within enterprises would be written by AI. Many scoffed at that forecast, and, in fact, Amodei has since walked back the statement slightly, saying that he never meant to imply there wouldn't still be a human in the loop before that code is actually deployed. He's also said that his prediction was not far off as far as Anthropic itself is concerned, but he's used a far looser percentage range for that, saying in October that these days "70, 80, 90% of code" is touched by AI at his company. Well, Anthropic has a team of researchers that looks at the "societal impacts" of AI technology. And to get a sense of how exactly AI is changing the nature of software development, it examined how 132 of its own engineers and researchers are using Claude. The study used both qualitative interviews with the employees as well as an examination of their Claude usage data. You can read Anthropic's blog on the study here, but we've got an exclusive first look at what they found: Anthropic's coders self-reported that they used Claude for about 60% of their work tasks. More than half of the engineers said they can "fully delegate" up to between none and 20% of their work to Claude, because they still felt the need to check and verify Claude's outputs. The most common uses of Claude were debugging existing code, helping human engineers understand what parts of the codebase were doing, and, to a somewhat lesser extent, implementing new software features. It was far less common to use Claude for high-level software design and planning tasks, data science tasks, and front-end development. In response to my questions about whether Anthropic's research contradicted Amodei's prior statements, an Anthropic spokesperson noted the study's small sample size. "This is not a reflection of concertedly surveying engineers across the entire company," the spokesperson said. Anthropic also noted that the research did not include "writing code" as a distinctive task, so the research could not provide an apples-to-apples comparison with Amodei's statements. It said that the engineers all defined the idea of automation and "fully delegating" coding tasks to Claude differently, further muddying any clear reflection on Amodei's remarks. Nevertheless, I think it's telling that Anthropic's engineers and researchers were not exactly ready to hand a lot of important tasks to Claude. In interviews, they said they tended to hand Claude tasks that they were fairly confident were not complex, that were repetitive or boring, where Claude's work could be easily verified, and, notably, "where code quality isn't critical." That seems a somewhat damning assessment of Claude's current abilities. On the other hand, the engineers said that without Claude, about 27% of the work they are now doing simply would not have been done at all in the past. This included using AI to build interactive dashboards that they just would not have bothered building before, and building tools to perform small code fixes that they might not have bothered remediating previously. The usage data also found that 8.6% of Claude Code tasks were what Anthropic categorized as "papercut fixes." Not just deskilling, but devaluing too? Opinions were divided. The most interesting findings of the report were how using Claude made the engineers feel about their work. Many were happy that Claude was enabling them to handle a wider range of software development tasks than previously. And some said using Claude freed them to think about higher level skills -- considering product design concepts and user experience more deeply, for instance, instead of focusing on the rudiments of how to execute the design. But some worried about losing their own coding skills. "Now I rely on AI to tell me how to use new tools and so I lack the expertise. In conversations with other teammates I can instantly recall things vs now I have to ask AI," one engineer said. One senior engineer worried particularly about what this would do to more junior coders. "I would think it would take a lot of deliberate effort to continue growing my own abilities rather than blindly accepting the model output," the senior developer said. Some engineers reported practicing tasks without Claude specifically to combat deskilling. And the engineers were split about whether using Claude robbed them of the meaning and satisfaction they took from work. "It's the end of an era for me -- I've been programming for 25 years, and feeling competent in that skill set is a core part of my professional satisfaction," one said. Another reported that "spending your day prompting Claude is not very fun or fulfilling." But others were more ambivalent. One noted that they missed the "zen flow state" of hand coding but would "gladly give that up" for the increased productivity Claude gave them. At least one said they felt more satisfaction in their job. "I thought that I really enjoyed writing code, and instead I actually just enjoy what I get out of writing code," this person said. Anthropic deserves credit for being transparent about what it knows about how its own products are impacting its workforce -- and for reporting the results even if they contradict things their CEO has said. The issues the Anthropic survey has brought up around deskilling and the impact of AI on the sense of meaning that people derive from their work are issues more and more people will be facing across industries soon. Ok, I hope to see many of you in person at Fortune Brainstorm AI San Francisco next week! If you are still interested in joining us you can click here to apply to attend. And with that, here's more AI news. Jeremy Kahn [email protected] @jeremyakahn FORTUNE ON AI Five years on, Google DeepMind's AlphaFold shows why science may be AI's killer app -- by Jeremy Kahn Exclusive: Gravis Robotics raises $23M to tackle construction's labor shortage with AI-powered machines -- by Beatrice Nolan The creator of an AI therapy app shut it down after deciding it's too dangerous. Here's why he thinks AI chatbots aren't safe for mental health -- by Sage Lazzaro Nvidia's CFO admits the $100 billion OpenAI megadeal 'still' isn't signed -- two months after it helped fuel an AI rally -- by Eva Roytburg AI startup valuations are doubling and tripling within months as back-to-back funding rounds fuel a stunning growth spurt -- by Allie Garfinkle Insiders say the future of AI will be smaller and cheaper than you think -- by Jim Edwards AI IN THE NEWS OpenAI declares "code red" over enthusiasm for Google Gemini 3 and rival models. OpenAI CEO Sam Altman has declared a "Code Red" inside OpenAI as competition from Google's newly strengthened Gemini 3 model -- and from Anthropic and Meta -- intensifies. Altman told staff in an internal memo that the company will redirect resources toward improving ChatGPT and delay initiatives like a planned roll-out of advertising within the popular chatbot. It's a striking reversal for OpenAI, coming almost three years to the day after the debut of ChatGPT, which put Google on a backfoot and caused its CEO Sundar Pichai to reportedly issue his own "code red" inside the tech giant. You can read more from Fortune's Sharon Goldman here. ServiceNow buys identity and access management company Veza to help with AI agent push. The big SaaS software vendor is acquiring Veza, a startup that bills itself as "an AI-native identity-security platform." The company plans to use Veza's capabilities to bolster its agentic AI offerings and grow its cybersecurity and risk management business, which is one of ServiceNow's fastest growing segments, with more than $1 billion in annual contract value. The financial terms of the deal were not announced, but Veza was last valued at $808 million when it raised a $108 million Series D financing round in April and news reports suggested that ServiceNow was paying an amount north of $1 billion to buy the company. Read more from ServiceNow here. OpenAI suffers data breach. The company said some customers of its API service -- but not ordinary ChatGPT users -- may have had profile data exposed after a cybersecurity breach at its former analytics vendor, Mixpanel. The leaked information includes names, email addresses, rough location data, device details, and user or organization IDs, though OpenAI says there is no evidence that any of its own systems were compromised. OpenAI has ended its relationship with Mixpanel, has notified affected users, and is warning them to watch for phishing attempts, according to a story in tech publication The Register. Apple AI head steps down as company's AI efforts continue to falter. John Giannandrea, who had been heading Apple's AI efforts, is stepping down after seven years. The move comes as the company faces criticism for lagging rivals in rolling out advanced generative AI features, including long-delayed upgrades to Siri. He will be replaced by veteran AI executive Amar Subramanya, who previously held senior roles at Microsoft and Google and is expected to help sharpen Apple's AI strategy under software chief Craig Federighi. Read more from The Guardian here. OpenAI invests in Thrive Holdings in the latest 'circular' deal in AI. OpenAI has taken a stake in Thrive Holdings -- an AI-focused private-equity platform created by Thrive Capital, which is itself a major investor in, you got it, OpenAI. It is just the latest example of the tangled web of interlocking financial relationships OpenAI has woven between its investors, suppliers, and customers. Rather than investing cash, OpenAI received a "meaningful" equity stake in exchange for providing Thrive-owned companies with access to its models, products, and technical talent, while also gaining access these companies' data, which will be used to fine-tune OpenAI's models. You can read more from the Financial Times here. EYE ON AI RESEARCH Back to the drawing board. There was a time, not all that long ago, when it would have been hard to find anyone who was as fervent an advocate of the "scale is all you need" hypothesis of AGI than Ilya Sutskever. (To recap, this was the idea that simply building bigger and bigger Transformer-based large language models and feeding them ever more data and training them on ever larger computing clusters would eventually deliver human-level artificial general intelligence and, beyond that, superintelligence greater than all humanity's collective wisdom.) So it was striking to see the former OpenAI chief scientist sit down with podcaster Dwarkesh Patel in an episode of the "Dwarkesh" podcast that dropped last week and hear him say he is now convinced that LLMs will never deliver human-level intelligence. Sutskever now says he is convinced LLMs will never be able to generalize well to domains that were not explicitly in their training data, which means they will struggle to ever develop truly new knowledge. He also noted that LLM training is highly inefficient -- requiring thousands or millions of examples of something and repeated feedback from human evaluators -- whereas people can usually learn something from just a handful of examples and can also fairly easily analogize from one domain to another. As a result, Sutskever, who now runs his own AI startup, Safe Superintelligence, tells Patel that its "back to the age of research again" -- looking for new ways of designing neural networks that will achieve the field's Holy Grail of AGI. Sutskever said he has some intuitions on how to achieve this, but that for commercial reasons he wasn't going to share them on "Dwarkesh." Despite his silence on those trade secrets, the podcast is worth listening to. You can hear the whole thing here. (Warning, it's long. You might want to give it to your favorite AI to summarize.) AI CALENDAR Dec. 2-7: NeurIPS, San Diego Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here. Jan. 6: Fortune Brainstorm Tech CES Dinner. Apply to attend here. Jan. 19-23:World Economic Forum, Davos, Switzerland. Feb. 10-11: AI Action Summit, New Delhi, India. BRAIN FOOD Is AI adoption slowing? That's what a story in The Economist argues, citing a number of recently released figures. New U.S. Census Bureau data show that employment-weighted workplace AI use in America has slipped to about 11%, with adoption falling especially sharply at large firms -- an unexpectedly weak uptake three years into the generative-AI boom. Other datasets point to the same cooling: Stanford researchers find usage dropping from 46% to 37% between June and September, while Ramp reports that AI adoption in early 2025 surged to 40% before flattening, suggesting momentum has stalled. This slowdown matters because big tech firms plan to spend $5 trillion on AI infrastructure in the coming years and will need roughly $650 billion in annual revenues -- mostly from businesses -- to justify it. Explanations for the slow pace of AI adoption range from macroeconomic uncertainty to organizational dynamics, including managers' doubts about current models' ability to deliver meaningful productivity gains. The article argues that unless adoption accelerates, the economic payoff from AI will come more slowly and unevenly than investors expect, making today's massive capital expenditures difficult to justify.