This is a column about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

In the first two episodes of Platformer’s mini-series on AI and jobs, we heard from two tech leaders who pushed back on the idea that AI is about to leave most white-collar jobs worthless. Box CEO Aaron Levie argued that the “last mile” of human labor will resist efforts to automate it. And Google’s senior vice president of technology and society, James Manyika, explained how tech has improved at automating tasks but not jobs.

For our third episode, I wanted to highlight a contrasting view — someone who believes that AI really is on its way toward eliminating certain jobs. Boris Cherny is the creator and head of Claude Code — the agentic coding tool that Anthropic released last year and is, by most measures, the fastest-growing AI coding tool in the world. 

Cherny belongs firmly to the camp that believes the end of software engineering as we know it is already underway. He hasn't written a line of code himself in more than six months, and he says that for the kind of work he does, coding is effectively "solved."

Given that he is, by his own account, actively automating his own job, it's no surprise that Cherny sees the disruption arriving far faster than our first two guests. He told me that the title "software engineer" could start to disappear by the end of this year, dissolving into something closer to something like "builder" as the designers, product managers, and managers around him start shipping code of their own.

But his own jobs forecast is more optimistic — and in some ways similar to our first guests — than some of his more prominent comments about coding being “solved” might suggest. While companies may hire fewer engineers as we know them today, he argues, they’ll hire more of whatever “builder” role replaces them. 

"I don't think we're going to call them engineers," Cherny told me. "But if we talk about people writing code, or using agents to write code, I think there will be 100 times more engineers than there are today. That's my prediction."

Cherny was an unlikely candidate to lead a revolution in coding. He studied economics, dropped out of college to run a startup at 18, did a stint at a hedge fund, and spent five years as a principal engineer at Meta before arriving at Anthropic in September 2024. When he arrived, Cherny began to explore what the company's API could do. The product that became Claude Code began life as a tool to tell him what song he was currently listening to. After iterating it on a couple months, he released the first version of Claude Code internally — and 20 percent of Anthropic’s engineers began using it on the first day.

As always, note that my fiancé works at Anthropic. But I couldn’t imagine trying to understand the AI and jobs story in this moment without talking to the person behind Claude Code.

Highlights of our conversation are below, edited for clarity and length. Listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton.

And let us know what you think — we're new to podcast production, and welcome your feedback at casey@platformer.news.

Casey Newton: You joined Anthropic in September 2024, and my understanding is that no one told you to go build a coding product — you were just trying to learn the API. Can you tell us the origin story of Claude Code? I've read that it controlled your music.

Boris Cherny: All of these things are true. I joined this team called the Labs team, which built a bunch of cool stuff. We built Claude Code — I built that. A different person built MCP, someone built Skills, and two other people built the desktop app. That was essentially the size of the team. It was tiny. We built these features over the course of a few months, and because a lot of them were weird ideas, we had no idea whether they were going to work.

For a while, Anthropic's focus has been on the same kinds of things: enterprise, coding, and safety. We knew that somewhere in this journey we should probably build some kind of product. Early on, Anthropic didn't actually know if it wanted to build products at all, but if we were going to, it needed to be coding-related, because that helps us build better coding models and it helps us study safety. 

We didn't know what it should be, though. At the time, the coding products out there were all IDEs or IDE extensions, because the capability of the model — this was Sonnet 3.5 — was not very good yet. The best it could do was fancy autocomplete: you'd write a little code and it would complete the line. We had this feeling that there was a product overhang — this idea that you could build a product that does something the model is totally capable of doing, but no one has built the product that lets the model do it. And I'll tell you, it's still the same feeling today. The model can do all these things, and there's no product that lets it.

So I wanted to learn how to use the Anthropic API, and I built the cheapest possible thing — a little thing that ran in the terminal, so I didn't have to build a user interface or an app. I built it in a couple of days and started giving it to people to see whether they'd use it and how, just out of curiosity. Over the next few weeks, more and more people at Anthropic started using it. First it was the people who literally sat around me, then the next layer of the onion, and a few weeks in, a lot of Anthropic was using it every day. It was weird, because it was a little prototype in the terminal — the most "engineering" product possible. A lot of engineers don't want to touch a terminal, but they did, and they used it.

Newton: I've read that within five days of the initial release, half the engineering team was already using it. As that was happening, did you have a moment of thinking, "Okay, software engineering just changed forever"? Or were you still just iterating on the product?

Cherny: I was so focused on shipping. As soon as I got the idea, I spent every night and every weekend on it — it was the only thing I thought about, the only thing I worked on. I started having dreams about Claude Code, and that's still all I dream about: what should we do next, what do we build next. There's a chance now to zoom out, because a lot of people are using it and there's a lot to learn about how. But for a long time we were so focused on building that I didn't even have a chance to think about what it was.

Newton: Was there a moment when you did zoom out? It might be too minimizing to say you stumbled across it, but it does seem like there was a sense of accidental discovery. Was there a moment of "Oh gosh, this is different from the other things I've hacked on"?

Cherny: There was a lot of surprise. Broadly, we knew we wanted to build a coding product, but no one thought it would be in a terminal. The first moment was when Claude told me what music I was listening to. There were a couple of versions of this — we actually have a video demo I recorded, which we just donated to a computer museum— it's a weird historical artifact. I remember posting it in my Slack and getting two reactions, because no one understood that this would be it. I asked Claude what music I was listening to, and it wrote a little code to open my music player. It wrote the code in AppleScript, which I don't know, and I wouldn't have thought to write code to answer that question. It just did it, and I thought, this is surprising — it solved the problem in a way I wouldn't have as an engineer. 

Over the last year and a half there have been so many moments like that. I had one with Cowork recently. Every time we release a new model, I experiment with it to see the frontier of what it can do, because one of the hardest things about building on a model is that it's advancing so fast — you have to recalibrate every month, as I'm sure you know. I used Cowork to book a bunch of flights. Usually it works okay; this time it worked perfectly. Now, any time I travel, I use Cowork to book it. It booked eight flights and five hotels. The only mistake was that one of the hotels was way over budget — I think it was like $5,000 a night.

Newton: Cowork wants you to have a great time when you go on your stay.

Cherny: I said, "Please rebook this one." But otherwise it just worked for a couple of hours and did all of it. It was so cool. I feel that surprise every week, every month.

Newton: This feels like a moment to zoom out from the story of an initial discovery that spreads rapidly through Anthropic, and now Claude Code has become a default tool for a growing number of engineers. It's one of the products making the question of job automation feel really salient — at least for software engineers, but maybe for more people than that. During our first episode, Aaron Levie told me he didn't see jobs going anywhere — that there's always going to be a last mile of human work the software can't do. You have publicly predicted that the title "software engineer" could start to go away as soon as this year. So is Aaron wrong about this?

Cherny: There's a bunch of stuff that's true, and a bunch we don't know. The trends are exponential, and exponentials are very hard to think in. Honestly, anyone saying they know is guessing — some of these are educated guesses based on what we're seeing and on history. 

I think a few things are going to happen. One is that a lot of companies will need fewer engineers, because each engineer is more productive, so you don't need as many to do the same work. At the same time, a lot of companies will need many more engineers, because every engineer is more productive — the company can do more things, start more products, create more businesses. You see this with our team: we are constantly bottlenecked on good engineers. We're hiring as quickly as we can, and a lot of our customers are exactly the same. So I think both things will happen, and it depends on the company and the business.

There's another thing happening, where the roles are all blending together in an interesting way I don't think anyone would have predicted. Our manager, Fiona, hadn't coded in 15 years; she joined Claude Code and now she's coding. Kat, our product manager, codes. Megan, our designer, codes. Everyone on the team codes — you don't have to be an engineer anymore. If you project the trend a little, everyone who's not an engineer is going to code a little more, and engineers like me are going to code less. I haven't written a line of code in over six months; I'm building stuff all day. I see it all blending into one thing. Call it a builder, call it an engineer, call it a product manager — I don't know what the title is, but the role is changing.

Newton: So how we conceive of these roles is definitely going to change, but what that means for how many jobs are available at which companies is still unclear.

Cherny: Yeah, and history has a lot of examples. The tractor was invented in the 1890s — I was just reading about this. A guy named John Froelich invented it in Iowa. At the time, farm work was all horse-powered; you needed horses. Even though the tractor was invented in the 1890s, it wasn't until the '60s in the US that there were more tractors than horses. It took about 70 years. The number of tractors went up, the number of horses went down, and the lines crossed in the '60s. 

There were a bunch of reasons. The technology was magical — you could harvest a lot more crops, productivity was much higher — but if you were a farmer who wanted to learn to use a tractor, it took training. At the beginning, tractors were expensive, so in a lot of cases horses were still cheaper. And the tractors weren't very good at first: maybe you could use one for wheat but not for corn, so it took a long time for someone to make a tractor that worked for corn, and for okra, and for all the other things. That just took a while. What we're seeing right now is the same thing on a speed run — but we're seeing very similar issues.

Newton: This is the "AI as normal technology" argument: even as labs come up with incredibly capable models, people are slow to change, organizations are slow to change, and it takes time for these technologies to filter through companies. At the same time, people look at what's been reported about Anthropic's revenue and say — it doesn't seem like it's taking that long this time. So we're still trying to hone in on the actual rate of change.

Cherny: Here's a question for you: do computers make you more productive?

Newton: Yes. But "do they make me more productive" feels like a different question from "do I work less because of computers," if that makes sense.

Cherny: So because you can do more, you do more — you fit more into the same eight hours.

Newton: Absolutely. To be candid, I used to record one podcast episode a week in addition to writing a couple of newsletters. During this miniseries I'm experimenting with doing a couple of podcasts a week in addition to writing multiple newsletters, and AI is a reason I can do that. It's an incredible research assistant and podcast producer. I'm able to produce more. But I don't feel like I'm working less — and that's not a complaint, just how I'm navigating this moment.

Cherny: I feel the same way. I can do so much more — all this stuff I didn't get to before, because I didn't have enough hours in the day. There's another weird historical thing. In the '90s, when companies were adopting the personal computer — after the mainframe, after these big industrial computers that cost millions — they got miniaturized, so the average startup could just get computers. And there was a real question of whether computers made you more productive. People were complaining that they didn't. Now we look back and think, of course they do — I can't imagine going back to filing and paper. 

There's an awesome Harvard Business Review article from 1990, and the case it made was this: they studied companies adopting computers and found that some were getting more productive and some weren't. The difference was that the companies getting more productive were the ones that threw away all their paper — the filing cabinets, the pens, the desk drawers — and put a computer at the center of everything. The other companies still had teams writing everything by hand, with a computer in the corner used for one thing. The first category had big productivity gains; the second didn't.

It's similar right now. At Anthropic, we organize everything around Claude. When people join, if they have questions about how to write code or contribute to a codebase, the answer is: ask Claude. If the question is how to file an expense receipt, ask Claude. If it's when the next company holiday is, ask Claude. All this stuff you used to do manually, Claude is at the center of. The companies that are really getting it put Claude right at the center — not on the outskirts somewhere. You have to change all the business processes, and that takes time.

Newton: I've been reading about the Solow paradox, which is basically what you just described. It's an observation by an economist in the '80s that you could “see the computer age everywhere except in the productivity statistics.” Despite a very large build-out of computers, you weren't seeing people get much more productive. Eventually the gains did materialize, because companies reinvented their workflows around the new technology. So the question now is how quickly the economy can do that.

I want to ask a couple more fine-grained questions about software engineering, because I've heard you say coding is effectively solved and that you haven't written code in six months. Engineers push back on this, saying coding isn't only about typing — it's about judgment and taste and critical thinking, and agents can still be quite bad at those. What do you make of that critique? Are there parts of coding that remain unsolved?

Cherny: The critique is totally right. This is one of those things that gets taken out of context. The full quote is: coding is solved for the kinds of coding that I do. I work on pretty simple codebases — the Claude CLI is fairly new, the desktop and mobile apps are small, simple codebases. But we have so many enterprise customers now — some of the biggest enterprises. NASA is one of our customers. They have really big, really complicated codebases, and for them it's not solved yet; the model still makes mistakes, and its code isn't always perfect.

When you think about what engineers do, coding is a small percentage of it. It used to be that maybe 50% of my day was actually typing code, and the other 50% was talking to users, brainstorming and coming up with ideas, debugging, thinking through how something works, planning. There are all these things engineers do that aren't coding. So when I say coding is solved, I mean for the kind of coding I do — and coding is a small subset of what engineers do. You see this with the engineers at Anthropic, and more and more in the industry: when the model does the coding, they're freed up to do all the other stuff they actually enjoy, like talking to users and figuring out what's next.

Claude Code has been 100% written by Claude Code for over six months. That's true for Cowork and a lot of other products too, and we're hearing it more from customers. I was doing a talk for the latest Y Combinator batch earlier this week — a fireside. I used to start every talk by asking people to raise their hand if they used Claude Code. Now everyone does, so I stopped asking. Instead I ask people to raise their hand if 100% of their code is written by Claude Code. These are the most cutting-edge startups — usually a few people each — and half the hands went up. Then I asked people to raise their hand if none of their code is written by the model, and out of a couple hundred people, one hand went up. Everyone else was somewhere between 50% and 100%. So coding is getting solved for a bigger and bigger percentage of the code we write. Our team is an early indicator of what's happening in engineering, and engineering is an early indicator of what happens everywhere else. The shift started six months ago, and it's accelerating.

Newton: Let me ask about another fear: that in a world where engineers aren't writing as much code, people's understanding of their own profession will atrophy, which might be dangerous. You haven't written code in six months. Do you feel that atrophy starting? And how do you feel about it?

Cherny: There's one engineer on the team, Lena, who still writes C++ by hand on weekends, just for fun, because she still enjoys writing the code — and there's always room for that. But for me this is part of a much broader transition, and it's not about atrophy at all. Programming is always in flux. My grandpa programmed on punch cards in the Soviet Union 70 years ago, and for him that was programming — there was no JavaScript, no Python. The punch cards were pieces of paper a machine punched holes in; you'd feed it into a mainframe, it would process it, and a few lights would light up. Before that, programming was a room full of people — often women — doing math on paper, sometimes by hand. That was called programming. Then it became writing machine code, then assembly, then JavaScript and Python and Java. Now it's changing again: you talk to the agent. And I think it's about to change once more, where you talk to an agent that talks to agents that do the coding. It's always changed like this. It doesn't feel like atrophy to me; it feels like a sea change in the technology.

Newton: My feeling has been that using a graphing calculator probably caused some of my math skills to atrophy, but my solution is that I'll just keep using a calculator. I'm fine ceding some of that. Now, if over time the calculator becomes superintelligent and tries to undermine me in subtle ways, that would freak me out — but maybe we haven't crossed that bridge yet. 

Let me ask about your Claude swarm. You asked me earlier whether computers make me more productive. It seems clear Claude is making you more productive, but it doesn't seem to be reducing the amount of work you do. This feels important if we're curious what AI means for jobs: you believe companies will need fewer engineers, and yet you're never running out of things to do. Will there ever be a case where being more productive actually means you work less?

Cherny: There's a name for this — some paradox, I forget what it's called. I think it's really individual. Some of it is up to the company, because depending on the business there might be more or less need for people. But a lot of it is individual preference. When the washing machine came out — I'm going to give a historical analogy, because this is such a crazy technological change that I need history to anchor myself —

Newton: I love the stories.

Cherny: — to do a load of laundry took about five or six hours, and you had to walk something like 3,000 feet to get water. You'd collect it, boil it, put in the laundry, scrub it on a washboard, wring it out, and repeat for your entire family, maybe every day. It was a lot of work. The washing machine took about three hours off that, and it was one of the factors that let women enter the workforce en masse — because usually, not always, it was the women of the house doing this work, which kept them stuck at home. Suddenly there were three hours freed up every day, and different people could choose how to spend that time. For some it was hanging out with their kids or walking the dog or reading a book. But for a lot of people the answer was: I'm ready to enter the workforce, I want to work at a factory or an office. Because the time was freed up, you had a choice. It's similar now. Like any technology, it gives you more choice.

Newton: A couple of last questions on software engineering. I've been asking all our guests: if a 22-year-old just finished a CS degree this month and came to you and said, "Okay, now what?" — what do you say? Is there an entry-level job waiting for them, or do they need to think differently about the first part of their career?

Cherny: If you want to work at a company, you can totally still do that — there are entry-level jobs, there's a lot you can do. But if you're at all entrepreneurial, go start a startup. There has never been a better time in history to do it; it's the golden age. You and your agents can build a giant company. People are building billion-dollar companies with just a few people. Claude Code started as just a few of us. We have so many customers building really big businesses with one or two or three people. One person with the right idea has so much leverage. I couldn't imagine a better time to go into it.

Newton: That's interesting, because the view we often get from the AI world is that model capabilities are advancing so quickly that maybe we won't even have companies in five years. But you think that, at least for now, there's still plenty of room to start a company.

Cherny: At least for the next few years. If you trace out the exponential, it gets really weird — there's a version where the idea of jobs doesn't make sense anymore, or companies don't, or software doesn't. But in the meantime there's so much to do. We're all here figuring out what the model means and what it can do, so you might as well be one of the people exploring the frontier.

Newton: Last one on engineering. Three years from now, do you think we'll see more engineers, fewer engineers, or will it be impossible to answer because we might not be calling them engineers anymore?

Cherny: I don't think we're going to call them engineers. But if we talk about people writing code or using agents to write code, I think there will be 100 times more of them than there are today. That's my prediction.

Newton: Let's return to Cowork, which you helped develop. It's the Anthropic product I actually use the most now. I use it as an editor on my columns, to help produce the podcast, even as a kind of financial advisor — you create projects, add some skills, and it can imitate lots of different roles in a workplace. As a non-technical person, I find the UI intuitive, because it mostly involves dragging and dropping documents into a box. Talk to me about the road ahead for Cowork, and whether you think it can solve for other jobs the way coding is now partially solved.

Cherny: Cowork is so exciting. We started building it when we saw people using Claude Code for things that aren't coding — someone installed the terminal just so they could do their tax returns in it. That's not what a terminal is for, but from a product standpoint it's amazing, because people clearly want this. The next few months are about figuring out how to make it work really well for people who aren't engineers, which is new for us, because most of our team works on coding and we just build for ourselves. 

With Cowork, it's useful for everything that isn't engineering — accounting, finance, legal. I used it to buy a clamming license so I could go clamming in Washington. I use it to book flights and concert tickets. The challenge is making it really good at all of that, and the way we do it is we use it all day, every day, and talk to customers all day, every day. I'd also expect it to keep getting better at running for long periods. A year and a half ago, Claude Code could run maybe 30 seconds before going off the rails. Now every night I have hundreds, sometimes thousands of agents running 5, 10, 20 hours. That's just how engineering is done now, and I think the same thing will happen with Cowork.

Newton: As we zoom out to the implications of a world where more jobs are disappearing or transforming — you've said you think this transition will be painful for a lot of people, and Anthropic is in a unique position: potentially a source of unemployment among software engineers and others. Does the company have an obligation to those people? Is this something government needs to pay attention to? What do we do?

Cherny: Like any technology, it's going to be mixed — good effects and bad effects, and we don't know the exact timing or mix; you can never predict it in the moment. I feel an immense obligation, as an engineer, that there's always more we should be doing to tell people what's coming, make sure they can use the tools, educate them, and bring them along. The team and I talk about this a lot. But broadly, this isn't a problem one company can solve — it's bigger than any company, and you really don't want one company to solve it, because it could be the wrong solution. This is a society-wide question we should be debating. What Anthropic is trying to add is economic reports and policy work — to make it obvious what we're seeing, so everyone else can decide what to do about it.

Newton: Our most recent guest, James Manyika from Google, studies technology at the level of economies and whole societies, and he's worried that what started as a digital divide — not everyone having equal access to the internet or a good laptop — is about to become an AI divide. The data so far shows the people getting the most out of AI are already near the top of the income ladder. Does Claude Code make that better or worse? Who do you see using it and not using it, and are there efforts to get it into the hands of people who haven't always had access to cutting-edge technology?

Cherny: There are a couple of programs to spread access, and Anthropic does this. The thing that continuously surprises me is that the people who get the most value out of Claude Code are not the people I'd expect. We just did a hackathon for the Opus 4.7 release, and the people who won were largely not professional engineers — there was an electrician, a doctor, a carpenter who used it to build an app. We saw the same thing with the 4.6 hackathon. A while ago it would have been mostly engineers, but the models are sophisticated enough now that it's often not engineers who learn to really harness them. We see this at big customers, too. As companies think about adopting AI tools, the biggest question is the business-process change — how you put Claude at the center. One approach that works best is to give everyone tokens and make everyone feel safe experimenting, because the ideas come from people you don't expect. It's not necessarily the senior engineer who was most productive before; it might be an accountant in the corner of the org, or a go-to-market person who built an internal dashboard that sped everything up or solved a problem no one realized the business had. It's important that people learn to use the tools, because the people who are best with today's tools aren't necessarily the ones who'll be best with tomorrow's.

Newton: Last question. In addition to everything you're doing with Claude Code and Cowork, you're amazingly omnipresent on X and Threads. I see you every day troubleshooting users' problems and offering tips. If you could automate that part of your job, would you? And how close are we to that world?

Cherny: I have automated it, but I prefer to do it myself.

The way I did it is I have a loop set up in Claude Code — I've actually moved it to a routine that runs every 30 minutes. Threads has an API, X has an API, so it aggregates the feedback and I look at it. But my favorite part of my job is interacting with people, even when they're telling me something is broken or could be 10 times better. That's still my favorite thing, because it lets us make the product better. People sometimes look back at a product and say it was a moment of genius — someone conceived it and built it and it was perfect. That's never how it works. Claude Code has so many flaws; it's so far from the product it could be. The only way to make it better is to listen to people, especially when they say something doesn't work, and to keep improving it. That's how good products get built, and it's the only reason Claude Code improves a little bit every day.

Newton: It strikes me that's also a really human part of your job. When you talk to someone using a product you made, it probably reminds you why you started in the first place. Those moments of connection feel important at a time when we're not always sure what value we'll bring to an AI-enabled workplace.

Cherny: That's right. We're all trying to figure this out together. We have hypotheses about where this is going, and we're building a bunch of stuff because we think we know — but I'm often wrong; not all my guesses are right. Good ideas come from all sorts of people, so you have to listen, try a bunch of stuff, and sometimes it works.

Newton: Boris, thanks so much for joining us.

Cherny: Thanks so much, Casey.

A MESSAGE FROM OUR SPONSOR

Become an AI-native team with Rovo

Atlassian Rovo is AI that knows your projects, code, and people so it can bring context (and guardrails) to every workflow.

And because Rovo lives where your teams already work, it doesn’t just find the answers — it helps you do the work.

See how Sprout Social is becoming an AI-native team with Rovo.

Learn more.

Following

The Pope drops 43,000 words about AI

What happened:  Pope Leo XIV delivered his first encyclical — a papal address to “all people of good will.” (Think of it as kind of holy Substack.) The encyclical, “Magnifica Humanitas,” concerns “safeguarding the human person in the time of artificial intelligence.”

The encyclical follows months of discussions between Silicon Valley and the Vatican; Meta, Google, and Amazon executives met with Vatican officials in April. In an unusual move, Pope Leo presented the encyclical alongside Anthropic co-founder Chris Olah.

The encyclical weighs in on AI and work, suggesting a “social criteria” for introducing automation and AI, alongside protections and re-training for workers.

While he acknowledged that “technology should not be considered, in itself, as a force antagonistic to humanity,” Pope Leo wrote that “the pursuit of greater profits cannot justify choices that systematically sacrifice jobs.”

Some of the encyclical mirrors terminology Anthropic has used to describe AI, such as one section that says AI systems can be unpredictable because engineers don’t directly code AI outputs — instead creating a framework within which AI “grows.”

Pope Leo’s encyclical addresses potential dangers of AI in warfare, writing that “it is not permissible to entrust lethal or otherwise irreversible decisions to artificial systems.”

The encyclical also calls for policies that protect children from violent, hypersexualized, or fake information generated by AI.

Chris Olah offered brief remarks on the encyclical. He emphasized that companies like Anthropic face “a set of incentives and constraints that can sometimes conflict with doing the right thing.”

“We need moral voices that the incentives cannot bend,” Mr. Olah said.

Why we’re following: It’s fascinating, and a bit surreal, to see the Pope engaging so directly with tech leaders on AI. (And to see His Holiness standing shoulder to shoulder with a tech executive.)

While it’s somewhat headache-inducing to imagine Meta lobbyists influencing the Vatican’s stance on AI, ultimately the Pope did not shy away from warning about the technology’s potential ill effects on work, war, and child safety.

Pope Leo’s teachings on AI will certainly have influence on the world's billion-plus Catholics. And his warnings about AI replacing jobs will at least get a listen from policymakers and tech executives. (Although some of those executives are probably fine with AI taking your job, regardless of what the Pope says.)

What people are saying: On LinkedIn, Turing Award winning computer scientist Yoshua Bengio responded to the Pope’s point that “Decisions about technology must never be separated from conscience and responsibility.” Bengio wrote, “I agree with this sentiment by Pope Leo XIV. The Vatican and other global institutions can and must play a role in the global dialogue on AI to raise public awareness and mobilize society for the challenges ahead.”

On X, investor and former Trump AI advisor David Sacks said that while “The Pope rightly warns that AI must serve human dignity,” he worries about what will happen if “we hand governments sweeping power over AI development in the name of safety,” including using AI to “surveil, and control citizens.” (You mean like the Trump administration is already doing, David?)

Others took note of the strange pairing of Pope Leo with a tech executive. X account @CoastalFuturist posted, “BREAKING NEWS: God joins Anthropic as member of technical staff.”

X user @Delicious_Tacos speculated that OpenAI might attempt to compete with Anthropic by backing an anti-pope, medieval monarch style. “It’s my honor to announce that the true pope is working with us from Avignon,” @Delicious_Tacos posted, alongside a photo of a smiling Sam Altman.

Billionaire and Democrat activist Tom Steyer wrote, “Pope Leo is right: The AI era demands moral clarity and urgent action from leaders everywhere. We can’t allow AI to be a technology that turns billionaires into trillionaires while putting millions out of work.”

Many X users, including an account called “Pope Crave,” were excited to see that Pope Leo included a Gandalf quote from the Lord of the Rings in his encyclical.

Blogger Matt Yglesias had some philosophical disagreements with the Pope: “It is not surprising to see that the Pope endorses a lot of superstitious mind/body dualism about artificial intelligence but it’s still wrong, even as he is also raising a lot of good points about some of the risks at play.”

Jesuit and X user Joseph Nolla wrote, “Pope Leo was a math major? I wonder what his encyclicals will be like,” attaching a passage where Pope Leo describes the Gospel as a “multifaceted polyhedron.”

@BicameralGrind replied with an excellent math pun: “Not only does he understand sin, he understands cos too.”

—Ella Markianos

Trump’s AI executive order fizzles

What happened: President Trump postponed a highly anticipated executive order on AI and cybersecurity Thursday after he spoke with a few oligarchs and decided that he “just hates regulation,” a source familiar with the matter told Axios.

“The whole thing was unnecessary” and “just something doomers wanted,” the source added. 

The executive order wasn’t drafted to be all that restrictive in the first place. According to a draft obtained by Politico, it was aimed at addressing concerns that advanced AI products could be weaponized to deploy destructive cyber attacks if criminals got a hold of them.

The order would have created a voluntary oversight system where AI developers could submit their products for a review by federal agencies before release. But it seems that even asking for voluntary cooperation with the industry was a bridge too far.

Why we’re following: There still isn’t any federal consensus on how powerful AI models should be regulated, weeks after Anthropic’s Mythos rattled Washington with its ability to identify previously unknown cybersecurity threats at scale. (Anthropic said last week that Mythos has since found "more than ten thousand high- or critical-severity vulnerabilities across the most systemically important software in the world.")

The draft order “was about as modest a frontier-AI intervention as the federal government could put on paper,” law professors Kevin Frazier and Alan Rozenshtein wrote. “Killing the order just leaves in place what [former White House AI adviser Dean] Ball has described as the ‘opaque and essentially lawless alternative: government access happening through back channels, on terms set case by case, with no stable rules at all. It may also slow AI adoption.” 

What people are saying: “I didn’t like certain aspects of it,” Trump told reporters in his usual style of speaking while saying nothing. “I think it gets in the way of — you know, we’re leading China, we’re leading everybody, and I didn’t want to do anything to get in the way of that lead,” he said.

Meta and Elon Musk sharply rebuked reports that Mark Zuckerberg and Musk had anything to do with the delay of the executive order, but acknowledged they spoke to the president at some point. 

“Mark didn't speak to the president until after the event had already been canceled,” Meta spokesman Andy Stone wrote on X. 

"I still don’t know what was in that EO and the President only spoke to me after declining to sign,” Musk wrote.

—Lindsey Choo

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and feedback on these changes: casey@platformer.news. Read our ethics policy here.

After a long day at Google I/O on Tuesday, I went out to dinner with a small group of reporters. Eventually the conversation turned to Daily Brief, an AI-powered feature that creates a personalized digest of your email, calendar, and Gemini activity for you every morning. The feature, which showed up in my Gemini app today, offers a quick tour through your inbox and calendar, attempting to highlight notable meetings and messages that need responses. And, like a handful of other AI features Google announced at its annual developer event, it sent a shiver down my spine.

After all, it’s not hard to imagine the brief soon incorporating summaries of emails like this one, elevating some newsletters to your attention while passing over others. You’ll always be able to click into the original message, of course — but will you? Some reporters at the table imagined that more sophisticated readers would be unlikely to use a feature like this, and for the moment that is almost certainly true. (For starters, this is a somewhat buried feature in the Gemini app, and there is likely a long way to go before Gmail opens to the Daily Brief.)

Over the medium term, though, I’d take the other side of that bet. A year ago Google launched AI Mode, a kind of side door in search leading to a more ChatGPT-like experience; as predicted, its rollout has been a boon to Google and a shattering blow to publishers. (Similarweb reports that nearly 69 percent of queries about the news no longer result in a click out to a website.) To users, the experience of Google feels largely the same. But the underlying bargain has changed: where once the company answered your question by directing you to a website, today it answers your question with its own large language model.

And increasingly, Google hopes you won’t need to ask your question at all. During the I/O keynote, the company unveiled a suite of agents that aim to anticipate your needs and meet them before you think to run a query. Whatever the situation may be, Google’s agents will monitor it for you, from local real estate prices to work deadlines. On some level, all of this will be at your direction — or, at least, your tolerance. But anyone who has ever been hypnotized for a few hours by auto-playing recommendations on YouTube knows that the distinction between the user’s choice and the algorithm’s often collapses over time.

Which is why I worry that the chief advantage of newsletter publishing over the past two decades — the direct access it gives you to an owned audience — will likely be eroded by AI. The more you come to trust the agent’s summary of your inbox, the less likely you are to dig for the original message. Over time, fewer and fewer newsletters will appear in your briefings. Eventually the agent will take the liberty of unsubscribing you from your least-read newsletters altogether — while alerting you to this in your daily briefing, of course, and offering to resubscribe for you if you would like. 

And at that point, would you bother?

In any case, this was more or less the world I was envisioning when I decided to adjust Platformer’s schedule and lean harder into original reporting, and the I/O announcements strike me as another step down that road. Two years ago at I/O, Google announced a path that would put the web into managed decline. This year, we got a preview of how that approach could extend to email publishing.

Notably, it may extend to YouTube as well. One of the less-discussed announcements at I/O this year was Ask YouTube: a new “conversational search” option for premium subscribers that lets you ask questions like “how do I teach my kid to ride a bike” and receive an answer that blends text and videos. In promotional materials, the feature highlights moments from several different videos. It looks useful, and I suspect it will be a hit. (I switched the feature on for my account here; so far, all my queries have returned an error reading “Sorry, I can't help with that.”)

While still an experiment, this feature is essentially AI Mode for video. And just as that mode reshaped the economics of web publishing, I suspect it will have real consequences for YouTube as well. Today there many large creators with channels dedicated to how-to videos, crafting, cooking, home repair, tutoring, and other educational subjects. Those channels make money by getting people to subscribe to them, watch mid-roll ads on their videos, and stay to the end.

Ask YouTube, on the other hand, might return a synthesized answer that stitches 30 seconds from three different videos — and resolves your question well enough that you never click into any of them. The old job of the YouTube search bar was to direct you to the best video for your query; its new job is to create a multimedia answer to your question, with video as a component.

And while the company could choose to offer new compensation to creators whose work is used in this way, it has said nothing about that so far.

Not all creators will be affected by these changes. Channels that focus more on entertainment are naturally insulated from whichever changes Google makes to YouTube’s search engine. But as with the changes to email, Ask YouTube highlights the way AI continues to centralize power away from smaller businesses and toward the platform.

The implicit bargain of the open web was that Google would send you somewhere in exchange for your attention an an ad view. The place you ended up — a blog, a newsletter, a YouTube channel — got a chance to build a relationship with you, and a business around it. At I/O this week, the company sketched out a new arrangement: an agent that meets your needs before you ask, summarizing the work of others on its way to the answer. Undoubtedly these agents will prove to be quite useful. But they will also come at a cost.

On the podcast this week: Google CEO Sundar Pichai returns to the show to discuss the strengths and weaknesses of the company's models; his early experiments with agents; and how AI ought to be governed. It was my favorite interview I've had with Pichai to date, and I hope you will enjoy it.

Apple | Spotify | Stitcher | Amazon | Google | YouTube

Following

SpaceX files for an IPO

What happened: As SpaceX marches toward a highly anticipated IPO, the S-1 document it filed this week reveals a company with billions in losses and one man with near-total control.

Here’s a look at the company’s financials revealed in the filing: SpaceX had a net loss of $4.28 billion for the first quarter, up 711 percent from $528 million a year prior. The spiking losses are notable given that revenue grew a modest 17 percent over the past year, to $4.69 billion. A large part of the blame goes to xAI and Grok, whose costs far surpassed the building of literal rockets. (Last year the company's space division spent $3.83 billion, while the AI division spent $12.7 billion.)

On the plus side, SpaceX is getting a fresh boost from its frenemy. Anthropic is paying SpaceX $15 billion a year through May 2029 for access to its Colossus data centers, which it will use to serve its Claude models, according to the filing.

If there's one thing we learnt from the Elon Musk vs OpenAI trial, it’s that Musk loves control, and SpaceX isn’t any different. According to the filing, Musk has 85 percent of the voting power in the company, and will continue to control the company after the IPO. Musk also responded to earlier reports of that SpaceX has guaranteed Musk cannot be fired without his consent, writing on X: “yes, I need to make sure SpaceX stays focused on making life multiplanetary and extending consciousness to the stars, not pandering to someone’s bullshit quarterly earnings bonus!”

Why we’re following: Whatever you think of Musk, his company’s advances in rockets and satellites, his CSAM-filled conservative propaganda website, his gas-guzzling Colossus AI infrastructure project, and his plans for data centers in space — SpaceX will be an IPO unlike any other.

The company's IPO filing lists its mission as, “build[ing] the systems and technologies necessary to make life multiplanetary… and to extend the light of consciousness to the stars.” Its advertised total addressable market is the size of the entire US economy. Its business plans center around AI despite the fact so far xAI has produced little more than a middling chatbot that now faces regulatory investigations around the world.

The IPO still seems likely to be a huge success, given the company's strong record in building reusable rockets, the impressive growth of Starlink, and the enduring fandom around Musk and his companies. But we're increasingly curious what plans Musk has for xAI. By leasing out both Colossus 1 and 2 to rivals — and allowing his original founding team at xAI to fall apart — Musk seems to be ceding the race to build superintelligence to other companies.

Perhaps it's all a ploy to make the company's finances look healthier before an IPO. But the longer Musk is away from frontier model building, the harder it may be to get back into the game. (Just ask Mark Zuckerberg.)

What people are saying: On Musk’s social media site X — which is set to be part of the IPO — SpaceX’s S-1 filing was reviewed thoroughly, mostly for its entertainment value. 

Investor/podcaster Ben Gilbert wrote, “I love how heavily @SpaceX leaned into photography in their S-1,” the legal filing for their IPO. They went big: “The first 15 pages are all pictures, many of which have no explanation at all.” Gilbert said, “It just makes you... feel.”

Takes on the company's value proposition were more mixed. Tech blogger Gergeley Orosz wrote, “I don't offer investment advice, but when I see that SpaceX claims that their total addressable market is the size of the *complete* US economy…” Orosz said that given “they are saying they are now 1% SpaceX, 5% Starlink, 94% AI (???) — I sense AI washing.”

DeepMind engineer Susan Zhang wrote a poem for the occasion:

—Ella Markianos and Lindsey Choo

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and questions for YouTube: casey@platformer.news. Read our ethics policy here.

This is a column about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

Last week, I sat down with Aaron Levie, the CEO of Box, who made what I thought was a pretty strong case that jobs are just harder to automate away than AI companies keep telling us. Aaron's argument went something like: the last mile of human work — judgment, context, the messy parts — doesn't actually get automated anytime soon, and companies are about to have many more humans and agents working together, which means we can keep our jobs. I left that conversation pretty optimistic.

This week, I wanted to put that optimism in front of somebody who has spent his whole career trying to measure — sometimes at the level of whole economies — what new technology actually does to work. James Manyika is a senior vice president at Google and Alphabet, where he runs Google's research and labs operations, along with a team the company calls technology and society — an effort to consider the broader consequences of AI systems like Gemini and develop Google's strategy around them.

There's much to consider. As I discuss with Manyika in this conversation, seven in 10 Americans now oppose data center construction in their communities. And the message that many Americans have gotten about AI from the industry itself — first we'll take your job, and eventually we might kill you — clearly hasn't rallied many people to Big Tech's banner.

Given his role, it's no surprise that Manyika takes a more optimistic view: jobs are harder to automate than Silicon Valley often gives them credit for, he says, and the process will unfold more slowly than some of more aggressive predictions that radiate out of other frontier AI companies. (This is a view shared by some of Manyika's fellow Google executives: DeepMind CEO Demis Hassabis warned Wired today that it may be a mistake to replace software developers with AI tools. “I think it's a lack of imagination—and a lack of understanding of what's really going to happen,” he said.)

But unlike most Big Tech executives, Manyika developed his views during a long career outside Silicon Valley. A longtime McKinsey executive, Manyika co-authored a paper about the potential effects of automation on labor nearly a decade ago. He has since co-chaired the UN Secretary-General's high-level advisory body on AI and served as vice chair of the National AI Advisory Committee under President Biden.

And so when leaders at rival companies like Microsoft and Anthropic insist that a significant portion of white-collar work is about to disappear, Manyika is skeptical.

"Some of those predictions were made two years ago — that in two years, 50% of jobs would be wiped out," Manyika told me. "Well, two years is up. Let's take a look. And anybody who makes that prediction for two years from now, I'm willing to take the bet."

Highlights of our conversation follow. Highlights of our conversation are below, edited for clarity and length. We also hope you’ll listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton.

And let us know what you think — we’re new to podcast production, and welcome your feedback at casey@platformer.news. 

Casey Newton: You did your PhD in AI and robotics at Oxford decades before AI became the biggest story in the world. What did you believe or see back then that most people were missing?

James Manyika: I'll take you back even before that. I did my undergraduate degree at the University of Zimbabwe, and my undergraduate thesis was actually the first paper I ever published. Guess what it was on? AI — training and modeling an artificial neural network. There was a postdoc visiting from Canada who had worked as part of the Montreal–Toronto Geoff Hinton crew, and he suggested I should build a neural network for my undergraduate thesis. That was the very first thing I ever published, in 1993.

Newton: Well before folks like me were spending every waking hour reading and writing about this. What captured your interest?

Manyika: Two things. I grew up on Star Trek, so the idea of AI fascinated me. I watched 2001: A Space Odyssey. But I was also just intrigued by the idea that it would be possible to build systems that can do advanced cognitive tasks. So when I went to Oxford, I did my PhD in AI and robotics to continue pursuing this.

Newton: You've since spent a good portion of your career trying to measure how technologies change economies. You spent a long while at McKinsey, where you wrote a paper called "Jobs Lost, Jobs Gained" almost a decade ago. Now you're inside Google, where you can see what happens when these tools actually land in workplaces. When you look at the debate we're having right now about the future of AI and jobs, where do you land?

Manyika: It's such an exciting moment. The technology and its capabilities are expanding at an incredible pace. But when you try to translate that into what it might mean for work and jobs and occupations, I get a very mixed view. It's roughly what that paper said 10 years ago, which I think is still correct: there will be some jobs that decline, there will be jobs that grow, and most importantly — a third aspect — a lot more jobs will change.

Whether you're looking at the aggregate economy, the sectoral level, or by occupation, you get a different mix of those three things happening. But all three things will happen. The research hasn't changed very much. The debate that people have is, what's the mix of those three things? As opposed to, are these three things going to happen.

Newton: Let me name a dynamic that may be on some listeners' minds. You are now employed by one of the biggest beneficiaries of the current AI boom. How do you tell when you're hearing the labor economist in your head and when you're hearing the SVP at Google?

Manyika: I hear both things. Less the SVP at Google — more so the AI researcher and computer scientist in me is extraordinarily excited about the pace of the technology. That part of me thinks, "Oh my goodness, this is going to be extraordinary, and it's going to happen very, very quickly."

The labor economist part of me says, "Hang on a second — these things don't actually play out that quickly in the economy, and the dynamics are more mixed." So I almost hear two speeds proceeding here. I often think that as AI researchers, our community tends to overstate what happens in the labor markets based on what we're seeing on the technology frontiers. These are two very different conversations.

Newton: At the McKinsey Global Institute, you found that about 50% of tasks would be automatable through AI, but only about 10% of occupations would be fully automatable. A few generations of AI later, does that ratio still sound right?

Manyika: All of the pieces have been moving. 

At the task level, many more tasks are now possible to automate — that picture's moved pretty quickly. But if you look at the composition of occupations — the Bureau of Labor Statistics tracks somewhere between 850 and a thousand real occupations — and ask how many existing occupations have the majority, call it 90%, of their constituent tasks automatable, that number is still under 10%. Most researchers will still say that.

How many tasks look like they're going to be hard to automate? Partly because AI can't do them yet, or because of coupled tasks where the weak link slows down the combination. If you take two tasks and can automate one of them, but they need to be done in a coupled way, you'll only go at the speed of the weakest link. Most jobs have these couplings that make full automation very difficult.

One other thing that's moved is task duration. If you had asked in 2017, of the tasks possible to automate, some were very short — 30 seconds or a minute is about as long as you could predictably do a task in an automated way. Now we can do some of those tasks for up to four-plus hours. The task duration with reasonably predictable completion has made tremendous progress.

Newton: So what I'm hearing is that if you measured the tasks that are automatable now, that number is trending much higher than 50%. But at the same time, the number of jobs you could fully automate is stubbornly holding in roughly the same place it was 10 years ago. What is your best explanation for that divide?

Manyika: Part of the divide is that we now understand more fully that whole jobs have a much more complex mix of tasks, and this idea of weak links or coupled tasks matters a great deal in most occupations. If you look across the whole economy at most complex tasks, we can't automate most of them. So the question of which whole jobs you can automate more than 90% is still a relatively small number.

Most of the debate among labor economists is whether in the next decade that number is more like 2 or 3% or more like 9 or 10%. I don't think anybody who's looked at whole-job automation would say it's 50% or any of these extraordinarily large numbers.

That's why I come back to the view that three things will happen. Yes, there will be some job declines. But there'll also be jobs that grow — that's a function of existing jobs that grow in demand because the technology often changes the demand picture, and new jobs get created. We forget that David Autor and others have shown that if you went back to 1945 and compared to today, something like over 60% of the jobs we have in the economy today didn't exist back then, and many were introduced as a result of technological shifts that created the category.

But the biggest effect is the jobs-changed part. The nature of the job itself shifts. This is what happened with bank tellers. This is what happens with radiologists. We still have the category "bank teller," but I can guarantee what a bank teller does today is not what a bank teller in 1970 did.

Newton: I want to press on that optimistic picture. If that was all that happened in the next two or three years, I'd be breathing this deep sigh of relief. On the other hand, you have folks like Mustafa Suleyman at Microsoft saying he thinks that within 18 months, all white-collar work will be automatable. Dario Amodei at Anthropic is predicting very high unemployment as a result. When you look at statements like that, do you think those guys are wrong? Are they missing something?

Manyika: I'll just say: let's take the bet. Some of those predictions were made two years ago — that in two years, 50% of jobs would be wiped out. Well, two years is up. Let's take a look. And anybody who makes that prediction for two years from now, I'm willing to take the bet.

There's an extraordinary unevenness when it comes to things playing out in the labor markets that we forget. I live in San Francisco, and we've had driverless cars here for three years. But somebody in another city like Chicago has no idea what we're talking about. We often talk about the jaggedness of the technology — I think that's true. There's also jaggedness in the economy in terms of how this plays out. Don't get me wrong — I think it'll be faster than the Industrial Revolution, but it won't be as fast as the technology often suggests.

I'm happy to have this conversation a decade from now.

Newton: Let's bring it to Google. Since you've been there starting in 2022, what have you seen in terms of jobs lost, jobs created, jobs changed?

Manyika: A lot of jobs are changing. What software developers do is changing a lot. People now work with agents, they manage agents, they pose questions, they spend less time doing bug fixes.

Keep in mind that in this whole jobs conversation, we often forget what I'll call the demand elasticity of things. There are some activities where there's so much more we want to consume and do — we've just been limited by the ability to do that. Software development is one of those. The amount of software that could still be developed to build extraordinary things is very, very large. We haven't built all the software we're going to build. We haven't designed all the systems we're going to design.

That won't be true for every activity. There are some activities where the demand is quite frankly limited — there's only so much of something the economy needs. In those cases, you'll see trade-offs between jobs lost and jobs gained play out in a demand sense quite differently.

Newton: Even folks who agree with you on the big picture tend to worry a lot about the entry level. If you're talking to somebody graduating this June who says, "James, how do I approach the first part of my career?" — what are you telling them?

Manyika: The future is actually pretty exciting. The economies are going to grow, and there's going to be lots of opportunities. But what it will take to prepare for those opportunities is dramatically changing.

A decade ago, when people asked me what their kid should do, I'd ask how old the kid was. If they told me the kid was 18, I'd say they should learn to code. If they told me the kid was two, I'd say, "Hang on a second — you should think about what kind of skills are going to be important, because this AI thing is going to make a lot of progress."

What we said at the time was correct but may no longer be true — that coding was going to be important in the mechanical sense of churning out lines of code. Now the systems are able to do that. That doesn't mean computer science as a field has gone away. When I studied computer science as an undergraduate, the coding part was just one slice of what I had to learn. I had to learn algorithm design and so much more. We may need to go back to that, because we're finding that it's the more broadly educated, skilled computer scientists who are a lot more interesting than the ones whose only claim is the ability to generate lines of code.

One more thing. I was looking at the data the other day — the demand for software development jobs is actually going up. It's not that these jobs are going away, even in this moment. But the skills required are changing.

Newton: This is another tension in Silicon Valley I'm quite interested in. Just within the past week, I've talked with you, Aaron Levie from Box, and Nikesh Arora, the CEO of Palo Alto Networks. Both Aaron and Nikesh said, "Please send me more engineers. I don't have nearly enough engineers for what I want." At the same time, it's earnings call season, and other CEOs get on the call and say, "Well, we're getting rid of 5% of the workforce to prepare for the AI future."

Manyika: There's so much more going on than the AI effect. As somebody who's super excited about AI's impact on the economy, I'll say: not much has happened yet. I'm saying that both on the positive side and the negative side. On the positive side, at the economy level we've yet to see the productivity gains, which I'm looking forward to and excited about. But we also haven't seen much of the AI-driven labor impacts everybody's talking about.

There was a paper — I think it might have been the "Canaries in the Coal Mine" paper — and what I found interesting was that the sharp declines they showed happened around October 2022. ChatGPT didn't come out until November 2022, and adoption didn't really happen in the enterprise space until maybe 2023. So if the sharp declines in entry-level hiring happened in October 2022, you'd have to believe something else was going on. There's now been analysis showing a whole bunch of monetary effects in the labor markets, plus leftover hangover stuff from COVID. There may be a tiny sliver that's AI-driven, but a lot more of it is driven by other macroeconomic effects.

That's not to say we shouldn't worry about AI's labor market effects. We should. I just don't think they've happened yet at the scale anybody's concerned about.

Newton: Let's move further into speculation. When I talk with folks whose jobs are beginning to change due to AI, it seems like an increasingly large part of their job is reviewing AI output. Whereas once they spent a lot of time manually writing code, now they spend more time reviewing it. The thing about reviewing work is it doesn't always scratch that same creative itch. Could job change wind up being a different kind of job loss, because some jobs that once felt very creative now just feel like tedious drudgery?

Manyika: Yes and no. There's a fair amount of work that's now reviewing the outputs of these systems, making sure you're guiding them — "no, no, you're going down the wrong path, do this."

I actually like this moment. You have to be creative about what systems you should be building, what questions you should be asking, which experiments you want to run. If you've got 10 agents working for you, what different directions are you going to send them in? What orchestrations are you going to run? What kind of tournaments do you want to run for your different agents to experiment in different ways and compare the results? The nature of the creative problem-solving is different.

The creative part is exciting because it's going to be the scarce thing — the harder thing to do. I've seen this in our own teams. The teams in Google Labs creating these extraordinary new AI-first products spend some time reviewing what's come from these systems, but a lot of the time they're dreaming up new things to build.

Newton: Let's shift to what you called an "AI divide" at Davos in January. We recently talked here about survey data showing that the people getting the most out of AI are already the ones at the top of the income ladder. Do you think AI is going to accelerate income inequality, or is there a way it can close that gap?

Manyika: This is a fundamentally important question I worry a lot about. I'm excited about AI empowering people, economies, advancing science. But there's no law of economics or nature that says everybody will benefit, and everybody will have access. We've all seen enough instances in history where we do extraordinary things in the economy and in science and not everybody has access.

Questions around access to infrastructure, tools, and technology are going to be so important in terms of making sure we don't exacerbate already-existing divides. That's why it's important to build more accessible models. We try to cover the whole Pareto frontier — we have very capable Pro models, then Flash models, all the way to our Gemma models, which are open access, open weights. That's part of our attempt to expand access — for developers, for users, and for the power of these technologies. That's why we work hard on things like language access and access for people with disabilities.

Newton: Are there one or two policies you think would ensure these AI gains are broadly shared?

Manyika: Multiple. Access to training and improving literacy. Often, people who are concerned about AI change their minds when they've had the experience of actually using a tool, so literacy will go a long way. Policies that make it easy to make investments in communities — in schools, in science, in infrastructure — would also make a big difference.

I also think back to the question of work. I don't think we do enough training and skill building. I also don't think we do enough to help people in these transitions. When I ran the McKinsey Global Institute, one of the things that stuck with me was that one of the mistakes of the globalization era was that even though the so-called China shock in aggregate only impacted something like 4 million jobs — which in the scale of the US economy could be seen as not very large — if you're one of those people in those locations and occupations, it was everything. What we did in terms of safety nets, wage insurance, or transition support to help those workers, we didn't do anywhere near enough. That's a place where policy can make a big difference.

If you ask me what keeps me awake at night about AI and work, it's not job loss — quite honestly, for the next decade it really isn't. It's these questions about how we support the transitions that work and workers are going to need to go through, and how we make sure companies like us, as well as employers and policy makers, keep that in mind as we navigate through this moment. That's what I worry about.

Newton: There's a story as we record this today that 70% of Americans now oppose the construction of data centers in their communities. While they have some concerns about AI beyond the economic ones, I think a lot of the fear is rooted in the fact that if they were to lose their job today, there is not going to be some obvious government helper standing by to help them navigate into the next point in their life. It's interesting to think about what the conversation about AI would look like if the average American felt like there was a plan for them.

Manyika: That's one of the things we have to do. We also have to think about AI's growing energy footprint. We have to make sure that as we build this infrastructure, it doesn't increase the cost of energy for people. We've made a pledge about bringing our own energy and not raising the cost of energy. But having confidence that there is a plan — which involves everybody — to support the transitions and changes that are going to happen in the labor markets is very important. That's what we may be failing to do.

It doesn't help when we in the AI field talk about wiping out 50% of jobs. First of all, I don't think it's going to happen. The economic research and analysis mostly doesn't say that. We're probably impacting the possibilities of this technology having extraordinary impact by, quite frankly, scaring everybody — when in fact that fear is unfounded. Let's not confuse the pace of technological development with how quickly this plays out in the economy.

Newton: In 2023 at the UN, you co-led a project called "Governing AI for Humanity," which laid out an ambitious vision for global governance of AI. Three years later, have we made progress, or are we moving backwards?

Manyika: We haven't made much progress. Because of the nature of what that body was doing, we were engaged with all 190-some countries of the world, and attitudes towards AI were very different around the world. People in the so-called Global South — in Asia, India, Africa, Latin America — were generally more positive about AI and its possibilities. Of course, they had concerns around infrastructure and access. Western Europe and the US are probably the most negative, which is kind of interesting.

There are some basic things we should all agree on: this technology should be based on fundamental human rights and respect for international law; we should think about both possibilities and harms, not one or the other; and everyone should be able to participate in both the development and use of this technology. At the time, there was general agreement among the vast majority of countries on these basic principles. The question became: how do we put this into practice? That's the process that's taking much longer.

One thing I learned from that process that's fascinating: in discussing the risks from this technology, they're the ones you might expect — misapplication and misuse. But many members of the body, especially from the Global South, insisted on adding "missed use" as a risk.

Newton: As in not using it?

Manyika: Yes — not using it as being an actual risk. That actually got written into the recommendations. The idea was: there are some people who live in places, countries, and communities where the risk of not using AI is far greater than their current circumstances. They may live in a country, or even in the US in a county, where access to specialist doctors is very low. If you ever looked at the map of the US at the county-by-county level, I was amazed at how different access to expertise in healthcare is. There are some counties where certain types of expertise just aren't there, and some counties where you have 10x the number of oncologists.

For some communities and some countries, the risk of missed use is actually quite high. So often when I hear people say AI is not good enough, I ask: compared to what? If I have access to Stanford Medical Hospital — I live in San Francisco — yeah, it may not be as good as that. If I'm in a place where I don't have access to the world's best oncologist or the best doctor, I'd ask: compared to what?

Newton: You could apply the same questions to driverless cars.

Manyika: Yeah. If you say they're not safe, I'd say, compared to what? We already know in the case of Waymo, for example, that the incidence rate is 30 times better than your typical human driver. Now, this isn't to say AI is perfect. This technology still has lots of gaps, lots of issues, and we have to work on that. But when we think about the impact on society, we should think both about the risks and harms and the beneficial impacts. We should solve for both.

I often find that when I'm with a group of people who are doomers, I find myself trying to paint the other side — the exciting possibilities. And vice versa: when people are so optimistic that this is going to change everything, I say, hang on a second, what about these complexities? What about these risks? We should be comfortable enough to hold both things in mind.

Newton: Let's talk about something that's giving you optimism. You spend a lot of your time at Google working on science. You and Demis Hassabis wrote in Fortune in February that AI is already changing how more than 3 million researchers are working. What's happening in the near term that might actually change someone's mind about whether they want more AI in their life?

Manyika: Most people know of AlphaFold — Demis and John Jumper got the Nobel Prize for that. What's extraordinary is that as many as 3.5 million researchers in over 190 countries are now using it, working on a whole range of things. We have many AlphaFold-style breakthroughs in materials science, structural biology, and high-energy physics.

We're also seeing advances in already-useful technologies in health. We just published the results of a year-long study where we had over 200,000 patients in the UK with the National Health Service looking at breast cancer detection. The study, published in Nature, showed that we can do more accurate screenings assisted by AI, and we can detect earlier. We have examples like that in several cancer categories, tuberculosis, and diabetic retinopathy.

Other areas have to do with crisis response — flood forecasting, wildfire prediction. Our flood forecasting work, from a cold start two and a half years ago, now covers 150 countries where more than 2 billion people live, with six-plus days of advance notice.

Google has had a genomics program for the last 10 years. Most people don't know this. The tools built in that program enabled the completion of the human genome sequencing — the last 8% that hadn't been done since 2001 — and the creation of the first pan-genome reference. The science work has been proceeding quietly for a long time. We still need to be very thoughtful about the potential bio risks from this technology, but the potential for AI advancing science is already quite real, and there's even more of it ahead of us.

Newton: When you imagine the average knowledge worker in 2030 — or just one specific knowledge worker — what do you think their job is going to look like?

Manyika: Let me take science as an example. Many more scientists are going to be doing a lot of in silico research — exploring all the possibilities of structures, biological molecules, or target designs for drugs — and then doing validation in the lab. That changes the picture. Science has historically relied a lot on hypothesis. That's not going to go away, but we're going to have so much more of this kind of generative possibility, and then validating in the lab.

Scientists are also going to be able to connect ideas beyond their own disciplines. We published a paper on arXiv called "Co-Scientist" where you can compare ideas from multiple disciplines. As a scientist, you're no longer constrained to just what you know in your own field. You can explore ideas in many more papers, theories across the board, and then bring them all together with agentic systems and construct experiments.

Take AlphaFold. Scientists used to spend three or four years working on one protein, doing X-ray crystallography to figure out the structure. Now they can just look it up in AlphaFold. Sure, there's still validation to be done, but you're starting in a very different place. The range of diseases you can study is much wider. The reason we have researchers in 190 countries using AlphaFold is because before, if you were in a country where you didn't have a lab, you had to wait until somebody had figured out the structure of the protein you needed to study. Now you just look it up.

You're going to see that kind of change in the nature of work across many occupations. 

Newton: My last question was going to be: if I had you back in a year, what is the single number or fact we would track to see whether the worldview you presented today was valid? But I think we already have it, because you told me right at the top that we're not going to see 10% of jobs automated away in the next year.

Manyika: I don't think we will. I'm happy to take bets.

Newton: It is a bold bet, and it's what I want to hear. James, thank you so much for joining us today.

Manyika: Thanks for having me, Casey. If I could say one last thing — one of the challenges on the optimistic side is that a lot of what I mentioned in science and societal implications is going to be indirect for most people. You're going to get a flood alert on your phone telling you to get out of the way, and you'll say thank you to whoever sent it. You won't say, "This wasn't possible before AI, but now it is." You're going to get a great cancer screening and you'll say, "Great." Many of the beneficial impacts that are already becoming real, most people don't experience as directly as AI. We have to work on that part, too.

A MESSAGE FROM OUR SPONSOR

Become an AI-native team with Rovo

Atlassian Rovo is AI that knows your projects, code, and people so it can bring context (and guardrails) to every workflow.

And because Rovo lives where your teams already work, it doesn’t just find the answers — it helps you do the work.

See how Sprout Social is becoming an AI-native team with Rovo.

Learn more.

Following

It's AI everything at I/O

What happened: Google I/O began in Mountain View on Tuesday with as clear of a theme as the developer conference has ever had — in this case, all AI everything all of the time. Announcements at the event were expectedly Gemini-heavy, with new models and agents coming in a variety of modes, products, and prices.

As to the models: first, there's Gemini 3.5 Flash, the successor to Gemini 3 Flash, which launches today and is faster and more capable than its predecessor. (It's notable how the consumer hardware's longtime mantra of “thinner, lighter, faster” has ported so directly to AI.)

Then there’s Gemini Omni Flash, which Google describes as “like Nano Banana, but for video.” Omni Flash can generate AI videos using a wide variety of inputs — text, photo, or video. Google says it’s good for more precise video editing and creating consistent characters.

We were impressed by a video CEO Sundar Pichai shared depicting a marble run with accurate physics, a domain where models have historically struggled. 

Google envisions Gemini Omni Flash as an initial part of Gemini Omni, “a new agentic experience and custom tools” where models will eventually be able to “create anything from any input.” (Might we suggest Google begin work on Gemini Omni Alchemist, which would generate gold from a variety of inputs, including lead).

Gemini 3.5 Pro is set to come out next month.

Elsewhere, Daily Brief, an agent that gives you a “personalized morning digest,” is on its way — promising to soon fulfill our prophecy that AI is coming for Platformer’s job. While this feature has been within reach for those willing to set up Claude Code for many months now — we’ve dabbled with such agents ourselves — a more accessible, polished version could be more disruptive to inbox-based businesses than anything we've seen to date. (A more considered take on this one, which should also loop in the new "Ask YouTube" feature, to follow sometime soon.)

What else? Gemini Spark is Google’s OpenClaw competitor. It runs in the background 24/7 doing personal-assistant stuff like managing your inbox and planning your trips. (It's also hosted in the cloud, so you don’t even need to buy a Mac Mini.) 

Google says Gemini Spark “operates autonomously, but always under your direction.” (While managers of AIs and humans alike have long desired assistants that do everything independently but also exactly how they’d prefer, we’re not convinced Gemini Spark has squared that circle.) It arrives for people with the more expensive Google AI subscriptions next week.

Our take: Today marked Google's effort to take AI agents from their current niche market of “OpenClaw bros posting for X clout” into something resembling true product-market fit.

User interfaces matter a lot for AI tools. Before ChatGPT became very famous, OpenAI’s Playground provided access to similarly capable LLMs; but it was the intuitive ChatGPT interface, which puts you in conversation with a fine-tuned “AI assistant,” that launched the first wave of mainstream AI adoption. 

I think a lot of companies, including Google, are hoping to be the company that creates that moment for agents. Google is particularly well-positioned to do this. But we’ll have to see a lot of testing — and proof Gemini Spark won’t leak our social security numbers — before such dreams become reality.

What people are saying: On X, early testers of Gemini 3.5 Flash had good things to say. Wharton economics professor Ethan Mollick said 3.5 Flash was “Very fast for a flash model and very capable, though not as powerful as a full frontier model.”

Box CEO Aaron Levie found that 3.5 Flash has roughly 20% performance gains over Gemini 3 Flash for several of his company’s work task evaluations. “Incredible to see the continued performance gains,” he said. (On that evaluation, it’s about on par with Anthropic’s Claude Sonnet 4.6)

On Bluesky, Bloomberg opinion columnist Dave Lee was finding it “REALLY hard to follow the branding around Google's AI products.” He said, “I want to make an image. That's NanoBanana. Or is it Google Pics? No, it's Google Flow. And Gemini Omni is multi-modal, I think.” Which Omni? “That's Omni Flash. Or is it? Where am I? Is that Eric Schmidt? BOOOOOOOOO.”

AI policy researcher Dean Ball remarked on Google I/O’s eschewing of AI benchmark graphs in favor of demonstrating “real-world capabilities usable to actual consumers.”

Verge reporter Jay Peters was more skeptical about these capabilities. He wrote that it seems like Google “wants to do everything for you, all from a search box.” But “the fun of the internet is actually doing the work to find stuff, even if it’s sometimes frustrating, difficult, or time-consuming.” Given he’s “spent years honing my own email management system,” he’s not sure he wants Gemini to start doing it for him.

And if Google does al the web stuff for you… what will happen to those of us who create the rest of the Internet? “Google doing everything also means a lot of the web that Google relies on collapses under it,” Peters wrote. “If Google Search doesn’t send traffic to publishers or websites who need visitors to make money” — and instead people are, say, reading Daily Briefing — “what will Search learn from, and where will it point people to?”

In an interview with the NYT, Gary Rivlin author of history of GenAI AI Valley, said that Google “just have this reach that few, if any, companies on Earth have” — because so many people are familiar with their products, they is so much more surface area for consumers to start using their AI. “If I had to put a wager on the biggest winner of A.I., I would say it’s Google,” he said.

More I/O announcements: A major overhaul of the search box brings AI further into results. OpenAI announces it will support the SynthID watermarking standard. Antigravity (a Claude Code rival) hits desktops. Gemini has 900 million MAUs. Pics is an image generation and editing app for the enterprise. Voice-based conversational features for Docs, Gmail, and Keep (but not til summer.) A "universal shopping cart" that works across all Google products. Gemini for Science.

— Ella Markianos

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and your AI jobs bet: casey@platformer.news. Read our ethics policy here.

Musk’s claims were thrown out because they fell outside the statute of limitations, making for a somewhat anticlimactic finale to AI’s biggest courtroom drama to date.

Presiding judge Yvonne Gonzalez Rogers agreed with the jury. The court said “claims of breach of charitable trust and unjust enrichment are dismissed as untimely.”

Why we’re following: The ruling marks a big win for OpenAI. Had Musk won, the company faced the potential removal of CEO Sam Altman and President Greg Brockman, and up to $134 billion in damages.

But because the case was thrown out on timeliness grounds, jurors ultimately did not have to weigh in on whether OpenAI’s unusual transition from nonprofit to Big Tech behemoth was legal.

It’s also a significant loss for Musk — one of a string of recent high-profile losses in court. In February, a judge threw out his complaint that OpenAI was stealing trade secrets from xAI. In March, another judge tossed a suit from X accusing advertisers of mounting an illegal boycott against the company.

Later that month, a federal judge ruled that Musk would have to face a class-action suit from former Twitter shareholders over his failure to disclose the stake he was gathering in the company on the required timetable.

Musk was also found liable in a case alleging that he intentionally drove down Twitter’s stock price in 2022 as part of a scheme to get out of his $44 billion purchase.

This isn’t the end of the fight for Musk, whose lawyer said he would appeal. But Judge Gonzalez Rogers noted she was prepared to dismiss Musk’s appeal “on the spot.”

What people are saying: OpenAI’s lead attorney, William Savitt, told reporters, “It’s not a technical decision, it’s a substantive one.” 

Savitt added, “It says: You brought your claims too late, and you did it because you were sitting on them to use them as a weapon of a competitor who can’t compete in the marketplace.” (The jury decision didn’t really say that second part.)

Unsurprisingly, Musk has been posting about the ruling on X, the social media platform that’s set to be part of an IPO this year now that it has been absorbed into Musk’s Frankenstein-esque SpaceX-xAI-X corporation. 

Demonstrating characteristic levels of decorum, Musk called Hon. Yvonne Gonzalez Rogers “the terrible activist Oakland judge” and said the ruling was “a free license to loot charities.” (Notably, the trial showcased evidence that Musk himself repeatedly tried to turn OpenAI into a for-profit, including by folding it into Tesla).

Musk wrote that the ruling was on “a calendar technicality,” and, “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity.”

Altman hasn't posted about the trial at all Monday, instead talking up recent performance improvements to ChatGPT.

On Bluesky, Wired reporter Paresh Dave posted an amazing photocopy of “Juror #3 in Musk v Altman suggesting the difficult questions,” noting “They went unaddressed in this specific trial.”

Perhaps at the next one?

— Casey Newton contributed to this item.

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and juror questions: casey@platformer.news. Read our ethics policy here.

In March, when Jay Graber stepped down as Bluesky’s CEO, I noted that it had not been growing for some time. But an important aspect of that story, and one I did not include in that column, is that text-based social networks as a category don’t appear to be growing, either.

On Thursday the market-research firm Apptopia shared new data on usage of the former Twitter and its clones. And while third-party estimates often vary from the companies’ own internal usage metrics, over time I have found that they are useful in understanding platforms’ basic trajectories.

In these cases — at least according to Apptopia — the trajectory basically looks like a down arrow.

Start with Meta’s Threads, which will turn three in July. Daily active users on the platform have declined in seven of the past eight months, Adam Blacker wrote in a blog post on Apptopia. After peaking in October 2024 — just before the US presidential election — daily users are now down 61 percent, Blacker wrote. Global monthly users have held up better, at 388 million — but that’s still down from an estimated 400 million in January of this year.

In the first quarter of 2024, Threads had actually surpassed X in US daily users, according to Apptopia. But while X usage has been generally flat to down, Threads lost users more quickly, the company said.

Meta pushed back strongly on Apptopia’s report, calling it “objectively false.” The company pointed me to a report from another market research firm, eMarketer, which in January predicted that Threads users would grow 19.6 percent this year and would surpass X in users by the end of 2027.

What about Bluesky? Apptopia’s report says the left-leaning network “has effectively collapsed as a competitive threat.” It reports that daily users are down 96 percent from January 2025.

Like Meta, Bluesky pushed back on the report when I asked. “Apptopia's numbers don’t match what we see internally,” a company spokesman told me. “While our data do show a decline in DAU over the same time period, that drop is far less pronounced — down roughly 25 percent since January 2025, not 96 percent.” 

The Apptopia report comes at a time when Bluesky users are conducting one of their periodic re-evaluations of what anyone is doing over there. The latest round kicked off when independent journalist David Roberts, who writes the climate newsletter Volts, said that Bluesky is mostly irrelevant to US politics.

“It's hard to avoid the conclusion that Bluesky has been a net negative for US politics,” Roberts wrote. “They corralled everyone on the left into a little glass fishbowl where they shout at one another & everyone else ignores them. Meanwhile, all the pols & institutions stayed on X & are being dragged farther right.”  

Writing about the subsequent pile-on Roberts received in Garbage Day, Ryan Broderick suggested another cause for Bluesky’s woes: even though the network doesn’t penalize links in posts the same way that its competitors do, for the most part users do not seem interested in reading whatever is attached to them. 

“Its users just don’t click on anything,” Broderick wrote. “It’s a cultural problem with no real fix. Regardless of what kind of publisher or creator we’re talking about, no one I’ve met is getting any meaningful traffic from Bluesky.”

Whether this is a trait of the Bluesky user base or simply reflects a declining interest in text generally is hard to tell. But Broderick’s experience largely tracks with mine: even a post with hundreds or thousands of likes on Bluesky typically does not drive significant traffic to Platformer.

Bluesky told me that change is coming.

“Rapid, spiky growth followed by churn from a peak is a normal growth pattern for newer social networks, not a collapse,” the spokesman said. “We grew incredibly quickly through the fall of 2024 and into early 2025, the starting point for Apptopia’s dataset. Since then, we’ve been expanding our team and building many of the basic systems needed to support that sudden network growth. With those pieces falling into place, we’re shifting more focus to building new experiences for our users this year. We're confident in where the app is headed.”

The company also noted that Bluesky has yet to run any paid marketing, unlike its rivals. “Our numbers are entirely organic,” the spokesman said. (For better and for worse!)

As for X, Apptopia found that daily users grew 28 percent in the United States in the first quarter of this year. Average time spent in the app per daily user is around 30 minutes, compared to nine on Threads, the report said.

One jarring stat from the report: Apptopia estimates that fully 75 percent of X users are men — a gender ratio typically seen only in C-suites, presidential cabinets, and tech podcast audience numbers. The overwhelming maleness of X is apparent even on a casual browse of the app, of course, but it seems like a grim sign for the overall health of the network. (Apptopia estimates the percentage of men on Threads at 54 percent, for whatever that’s worth.)

Apptopia attempted to spin these numbers as a sign that X is resurgent. “What we might be watching is the text-social category consolidating back to a single app after two years of fragmentation,” Tom Grant, the company’s vice president of research, said in the blog post. 

I’m less certain. Traffic to publishers sent from social networks is collapsing across the board. And the bigger picture is that text-based social networking is a rounding error compared to video. Telecom provider Ericsson estimated that by the end of 2025, 76 percent of all mobile data traffic was video. 

All of which challenges the idea that a text-based network will ever again plausibly serve as a global town square. The dream of early Bluesky and Threads was that you might be able to fully substitute for Twitter by re-creating it without Elon Musk in charge. And there were reasons to think that was plausible: Meta had Instagram’s enormous social graph to build on, and Bluesky had an open-source ethos that proved attractive to people who wanted an alternative to mainstream platforms.

A few years later, though, Twitter looks less like a product you can clone and more like a moment in time. The mass audience has now moved fully to video; the personal audience now lives in the group chat. Professionals are getting their information from newsletters and podcasts. 

Social products never stop evolving, and perhaps one day we’ll see a resurgence of text. In the meantime, though, social products appear to be fragmenting into smaller and stranger audiences. And the Twitter clones aren’t really competing with each other anymore. They’re competing with TikTok, Instagram, and YouTube — and losing.

On the podcast this week: Kevin and I discuss the Trump administration's apparent about-face on AI safety. Then, Palo Alto Networks CEO Nikesh Arora drops by to talk about cybersecurity in the Mythos era. And finally, the Hot Mess Express returns.

Apple | Spotify | Stitcher | Amazon | Google | YouTube

Following

OpenAI and Apple’s friends-to-enemies arc

What happened: OpenAI and Apple’s two-year relationship has become strained and might even escalate into a legal fight, Bloomberg reported.

In 2024, the two tech juggernauts inked what seemed like a mutually beneficial deal where ChatGPT would be integrated across iOS, iPadOS, and macOS. For OpenAI, the agreement promised to grant access to a giant customer base. For Apple, it offered a way to bring high-quality AI to its customers without having to get Siri to work properly. The arrangement was supposed to be so good for both parties that it was reportedly structured without cash payments in either direction.

But OpenAI executives now allege that Apple failed to hold up its end of the bargain, according to Bloomberg. “They basically said, ‘OpenAI needs to take a leap of faith and trust us,’” an executive with knowledge of the matter told Mark Gurman, but the potential for revenue never materialized — and “hasn’t come close to happening.”

Why we’re following: It's unusual to see a big company criticize Apple so directly — and even more unusual for it to sue over a partnership deal.

OpenAI’s complaints come amid Apple's plans to open up the iPhone to other AI models — a version of Siri powered by Google’s Gemini is expected to debut this year. OpenAI insists that isn't what's driving the legal threat. OpenAI said it didn't want to work with Apple on the new models because of how the last deal went down.

Meanwhile, Apple has expressed concerns over OpenAI’s privacy practices, not to mention its aggressive recruiting of its hardware engineers for its new division led by former Apple design chief Jony Ive.

What people are saying:  “It is pretty funny to be like "oh damn the deal we cut sucks, time to sue you",” New York Times tech reporter Mike Isaac posted on X.

“One fireworks show ends, another begins…,” wrote Spyglass’s MG Siegler on Bluesky.

— Lindsey Choo

Americans hate data centers

What happened: New Gallup polling found that 71 percent of Americans would oppose a data center getting built in their area. Americans are now more reluctant to put a data center in their backyard than they are about even a nuclear power plant, which only 53 percent say they would oppose. 

The survey, conducted in March this year, showed that 70 percent of Americans worry about data centers’ environmental impacts, while about 1 in 5 worried about their impact on local quality of life.

The issue remains fairly bipartisan: 75 percent of Democrats say they’d oppose a local data center project; 63 percent of Republicans say the same.

Why we’re following: Data centers have moved from "who cares" to "absolutely not" in the public imagination with record speed.

There’s clear distaste for the projects on both sides of the aisle. And it's become a more salient issue for voters.

Growing opposition to data centers could challenge the AI boom, as big cloud providers struggle to get data centers approved in the municipalities where they want to build. Tech executives usually respond by threatening to build them elsewhere. But the more that opposition grows, the fewer options they are going to have.

What people are saying: On X, Senior Fellow at the Abundance Institute Kevin Frazier wrote, “Until AI stakeholders take this backlash seriously, all positive visions of AI for science, healthcare, affordability, you name it...will be delayed.”

Sen. Bernie Sanders, who has proposed a legislated moratorium on data centers, posted, “71% of Americans oppose new AI data centers near them. They're right.” He said, “So, why has Congress done nothing to address their concerns?” His opinion: “Big Tech has spent $300 million on the midterm elections to ensure Congress does nothing to regulate them.”

— Ella Markianos

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and Threads posts: casey@platformer.news. Read our ethics policy here.

This edition of Platformer is about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

One obvious reason for the public’s rapid turn against AI is the fear that it will someday take their jobs. It’s a fear the AI industry has encouraged them to have: tech CEOs issue regular warnings about AI-related job loss — and it’s already starting to show up in Silicon Valley.

In March alone, tech companies announced nearly 46,000 layoffs — the worst single-month total in more than a year — with a growing number of executives citing AI as a factor in their thinning headcounts. Anthropic's Economic Index shows the share of work-related AI conversations climbing into nearly every white-collar profession. And a steady drumbeat of research suggests that entry-level work — the rung of the ladder most exposed to LLMs — is showing the earliest signs of disruption.

In one sign of how seriously the tech industry is taking this, Google DeepMind workers in the United Kingdom just voted to unionize.

At the same time, AI has been notoriously difficult to find in the productivity statistics. Amazon says it will hire about the same number of software engineering interns in 2026 as it has in recent years. Openings for software engineering roles are currently the highest they have been in the last three years. 

So what gives? Are jobs disappearing, or just transforming? Are workers becoming less essential to their bosses, or more? Are we witnessing the beginning of a massive disruption, or just another hype cycle?

These are the questions we’re setting out to answer in a new mini-series on Platformer. Over the next seven weeks, I'll be talking with CEOs, operators, and academics watching this transition up close. In each episode, we’ll consider the AI and jobs story in the kind of depth that often isn’t possible in a single news story. And we’ll also bring data: my colleague Ella Markianos will join me at the top of each podcast to review the latest surveys, research and news stories that speak to the intersection of tech and labor.

For our first episode, I wanted to talk to someone I've known about as long as I've known anyone in Silicon Valley: Aaron Levie, the CEO of Box. Aaron was the first person who explained software-as-a-service to me when I moved here in 2010, drawing diagrams on a whiteboard in the Box office with the kind of patience usually seen in a teacher showing kindergartners how to spell. 

Sixteen years later, he remains an enthusiastic explainer of the SaaS world. It helps that Box has a good story to tell — the company’s stock has held up materially better than most of its SaaS peers over the past year, even as a chorus of investors, founders, and posters have warned that traditional enterprise software is about to be eaten by AI agents.

As you’ll hear, Levie is not in that camp. In our conversation, he makes a careful — and at times provocative — case for why he thinks both the "SaaSpocalypse" and the broader narrative of mass AI-driven job loss are wrong. He argues that agents will multiply the number of workers using business software rather than eliminate them; that the "last mile" of human work is far more durable than people assume; and that the engineer of the future is more likely to work at a pharma company than at Meta.

“If you or I go and vibe-code something, we think we've replaced the engineer, replaced the accountant, replaced the lawyer,” Levie told me. “But then you actually look — that was the first 80% of the job. The extra 20%, it turns out, is all the value creation of that profession. All the expertise and domain knowledge is in that last 20%, not the text that got generated.”

Highlights of our conversation are below, edited for clarity and length. We also hope you’ll listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton.

And let us know what you think — we’re new to podcast production, and welcome your feedback at casey@platformer.news. 

Casey Newton: Aaron Levie, welcome to Platformer.

Levie: Hey, good to be here.

Newton: Aaron, you and I first sat down in 2010 —

Levie: We were so young.

Platformer: We were so young. Back then you were sort of early gray, but now you're just like — normal gray. I think running a public company will probably do that to you.

Levie: The problem is, I've been like this for 13 of those years. It would be one thing if this only happened in the past six months, but it's actually been like this since I was 24.

Newton: Maybe there are more reasons to be gray today, or maybe not — we'll get into it. 

That first time I met you, I have this core memory — because I had truly been in Silicon Valley for what felt like weeks when I came down to the Box office.

Levie: Didn’t you come from, like, the Arizona Star Tribune or something?

Newton: The Arizona Republic. I'd been covering local government. And then one day I said, "What's going on with computers? That seems interesting." And now here I am. But I needed people to explain it to me, and that's where you came in. As I recall, you explained the software-as-a-service business model to me on a whiteboard. So my first question: if you were explaining your business to a reporter like that today, how much of that whiteboard would look the same and how much would be totally different?

Levie: If my recollection serves, a lot of it was trying to compare the on-prem days to cloud, and why cloud was such a big deal. My predictive capabilities were pretty locked in, maybe short of AGI. The whole idea was that software was going to move from your data centers to the internet, and in the process, the real power is that it becomes available to way more companies — businesses of all sizes, lines of business that never could have used software before, end users. This was the phase of consumerization of IT. So that played out.

Now we're in the next frontier of what software is going to look like. A lot of the core architectural components hold. If you're running a global supply chain at a Fortune 500 company, you want deterministic systems and software that power your ERP. If you're at a large B2B like Salesforce, you want a clear set of business logic around how your CRM works, and how your internal workflows around sales automation work. If you're managing documents for a government agency or a pharma company or a law firm or a large bank, you want to make sure you can secure that data, protect it, govern it, ensure it's in a safe place and available to the right people. All of that is staying the same.

What has completely changed is the interaction patterns on those systems — where the interaction is coming from. And what you can now do with all that data. The big idea is that in the future, if today maybe 90% of activity on this software is humans interacting with the interfaces, probably three years from now it'll be 90/10 the other direction. Agents will be interacting with these systems, talking to the data, pulling up data from these tools. And maybe 10% will be you going and browsing and looking through the software yourself.

The interesting thing — and this is going to be the open debate for the industry — is in that 90/10, did the human side go down by 90%, or did we just have a 10x increase, where agents are now leveraging these tools? My argument is more the latter: agents are this explosion of new workers all using these systems, which makes the technologies even more valuable. You have all these new workers on these digital platforms that need data, that need to be secure, that aren’t leaking information in the wrong way. So you still need those guardrails — but now you've got a massive multiplier of what people can do with their data, because you have agents that can run in parallel.

Newton: Right, that makes sense to me. There's this really interesting challenge —

Aaron: By the way — podcast over?

Newton: Yeah. That's all the time we have. I really want to thank you for joining us. I think we all learned a lot. 

No — let's throw in a few more questions for the super fans, because you just introduced what seems like a possibly profound change in the business model for what you all do. SaaS companies have gotten used to selling by the seat. You have 10 employees, you want 10 of them to be able to use Box, you pay a monthly fee. And it seems like that business model is under a lot of pressure in a world where maybe I don't have 10 people in those jobs anymore. What I need is a business outcome. So how are you navigating that? Do you think this seat-based business model survives in SaaS?

Levie: You posited the scenario that's most open for debate, which is: did the people go away? In the math I laid out, the people stayed the same number, but the agents multiplied on top of the platforms. There will be some software categories where the literal seats are not as relevant because you don't have as many people doing the work. I would actually argue that for a large portion of software categories, that won't be the case. You'll have the same number or more people, but you'll also have 10 times the number of agents as people. So it's a multiplicative effect of more people — or the same number of people, or maybe a minor reduction — and then vastly more agents.

The part that's not being priced in by the market is, is that scenario playing out? If I look at our software consumption internally at Box, there aren't a large number of cases I can make for many of our software products to reduce the number of people that exist as seats. But there are a lot of cases for a lot more agentic use cases on that software.

To take an objective example that's not Box: if I look at Salesforce, we're actually going to have more sales reps at the end of this year than we had at the start. That's more seats within the Salesforce universe. At the same time, I can imagine 10 to 100 more agent use cases on the Salesforce platform than I could have two years ago. Those agents might not be roaming around the interface of Salesforce — they'll show up inside Claude Cowork or Codex or ChatGPT. The agent will be interacting via a different interface, but the underlying seat that says "Aaron is a user in this platform, with this level of access to this type of data" doesn't necessarily go away.

We're already seeing this with our customers: you want a seat for the person because you want some kind of stateful representation of what data that person has, what their entitlements are. But then an agent might do an unbounded level of consumption on the software — where I, as a person, can only click so many things per day, but an agent can do that at 100x the scale. So the seat gives me the ability to use my information across these other agents. But then at some scale, there's so much data being used that there's a consumption model on top. This is why I think you're going to have a stacking business model in software: humans still have seats, but agents will be a consumption pattern on top of that.

Newton: As a CEO, I'm imagining you're looking at all the SaaS you guys buy to run your business. I imagine you might be happy if you didn't have to pay for all those seats and could just have agents do it. So when you look at your own spending on SaaS, your feeling is truly, "I'm happy to keep spending for all the seats"?

Levie: There's a difference between happy and practical. I'd always like our IT spend to be less, but I'm extremely practical about how technology works. The bear case of software is a confusing amalgamation of multiple issues people have — it's a Rorschach test of "what do you hate about software?" Some people say, "What we're going to do is vibe-code CRM systems." Others say, "We're just not going to have employees, it'll just be agents." Others say, "We just don't need all the features of these SaaS systems, agents will do those." Some I'm sympathetic to, some I'm not.

The one I'm extremely not sympathetic to: we have no projects internally that I've approved to vibe-code a replacement to an existing SaaS service. If I look at the stack of our ERP system, HR system, CRM system, document management system, it would do us no good to spend our time and IT resources trying to replicate functionality that's already doing its purpose — especially at a moment when I'm about to get 10 times the value from those systems with agents using that data. If I have to both transition a system that's homegrown and figure out the next set of use cases, you'll just halt your ability to innovate.

And a minor aside: if you did a word cloud of the past two to three weeks in AI, one of the biggest words would be cybersecurity. Not the Mythos part — the "we leaked customer data, the credentials, the secrets of our system got leaked, we downloaded a package that was exploited." Think about if the entire economy was trying to rebuild their own version of Salesforce or Workday or an ERP system, and any one of those events happened. Now the entire economy has to halt and do upgrades, or handle the maintenance and ongoing improvement of these technologies. That's just not very logical economically.

The part I am sympathetic to: there's some software where, as you use agents more, some of the value proposition goes more into the agentic layer than the software layer. In those cases, you'd compress the value proposition of that software, and at the next renewal you might not spend as much. But conversely — for every scenario where that happens, there's another scenario where agents add more value on the system you're using. So the net vendor actually has more leverage in the future. You might save on one part of the stack but end up re-spending it on a different part because of all the upside.

Newton: What we're really getting at is the skepticism the enterprise software market is facing right now. The reason I wanted to talk to you first is that Box has been facing this kind of skepticism in various ways its whole existence. You had to survive a very early pivot from being a consumer company to an enterprise-focused one. You had to convince people the cloud would be a safe and profitable place to be. And you faced a lot of skepticism about whether Box might just be a feature rather than a company. Now you have AI come along introducing this fresh wave. Maybe the most accelerationist version of that argument is that every company is now a feature, and the only thing that matters is going to be the frontier labs. So to what extent is this SaaSpocalypse story just the latest incarnation of a story that's never been true, and to what extent is this AI moment truly something different?

Levie: The market is somewhat parsing the different outcomes — not perfectly, but there's some discerning behavior. If you took Wall Street as one metric and looked at our stock, it's held up better than most. One of the reasons is that the thing not really under debate is that your most agentic, vibe-coded enterprise future still has to store the data somewhere. You still have to secure and govern the important information of whatever the workflow is. You can vibe-code the creation of the contract, but the contract still has to get stored somewhere, governed somewhere, still has to have a retention policy, access controls.

The part I'm excited by is that becomes meaningfully more important in a world of agents. When I think about the use of data in the enterprise, what all these agents really want to do is access data. They want to read data, write data, know context about your organization — your best practices, your policies, your customer relationships, your research. All of that sits inside your enterprise data, and most of it sits inside unstructured data in the form of business content. So we're firmly on the side of: bring on all the AI humanly possible, because those agents are all going to be working with enterprise content that still needs to get stored somewhere.

A customer comes to us and says, "We want to automate our entire insurance claim process." A tremendous amount of enterprise content goes into an insurance claim. When they do that automation — maybe they build the agent on Anthropic, maybe on OpenAI — that agent still needs to talk to all the data in their enterprise. So they have to upgrade their infrastructure.

There'll be winners and losers in software and SaaS, as has been true of every era of disruption.

Newton: Tell us about some of the losers. You don't have to name —

Levie: Rather not.

Newton: I know you'd rather not. But basically what you're saying — and I believe this — is that your business has access to this very rich, valuable data, and that data is not, for the foreseeable future, going to be stored at one of the frontier model companies. So you have a lot of value to create based on that unique advantage. I'm guessing there are other SaaS companies that don't have that same advantage. As you're scanning across the market, is there a business where you're like, "I don't want to be in that business in a world of agentic AI"?

Levie: I'll give you a framework, but I'm not going to name names. The factors you want are: Do you have some degree of business logic or workflow in the system, because the agent still needs to do that, even in an agentic world? Do you store data? Are you the natural place for the information to get stored? Do you have a set of domain experience and context that the next training run of the agent doesn't just replace? Is there an element of security, governance, trust that matters a lot? Are there network effects? This is one reason Slack has been incredibly durable: we're already communicating in Slack, so agents naturally show up in Slack, as opposed to "I'm going to agentically do Slack."

There's probably one more element: how much does the system benefit from a world of multiple agents needing the data, as opposed to one agent needing the data? Because that points to whether the enterprise wants to put all your value into one of the labs, or whether it needs to be a different layer that everything talks to. This is why you still see companies like Databricks and Snowflake growing quite well. You don't really want to move your data around constantly. You want it abstracted from where the agent is, so you can structure it, secure it, govern it, and then let all the agents talk to it.

So you probably could have a quotient for “how durable is the platform, based on those factors?”

Newton: Right. The Levie formula. 

Here's what I'm taking away: if you have a to-do list app for teams, get out of that business.

Levie: I would say that business is actively pivoting.

Newton: Yeah, I actually think I know one that might be. Box has talked about AI for a long time, but I'm curious for you personally — we do have these momentary enthusiasms in Silicon Valley. I think it's fair to say both of us had a crypto phase. I'm imagining your AI phase started earlier but took longer to reach fruition. Did you have a moment of conversion where you saw a paper, a product, something where you're like, "Okay, I need to start taking this really seriously"?

Levie: There have been three moments. About eight years ago, vision models were getting good enough that you could give the vision model a document and it could OCR it, or give it an image of a retail product and it could classify it properly. That was a big deal. But the problem was you had to train individual models for every domain — this was just before the transformer. If you wanted to do document classification in legal, you needed a different model than for financial equity research analysis. So you never had the takeoff moment.

Big deal number two: ChatGPT. We were following GPT-2 and GPT-3, and we had a hackathon where somebody did GPT-2 inside of a document, but it was producing garbled text. It could maybe type ahead five extra words — not going to game-change your productivity. ChatGPT was the first time in this era where it was like, "Okay, this is a very big deal. We're going to be able to wire up these LLMs and connect to your data."

The most recent moment, the past year-plus, was marked by Claude Code — but really these more agentic patterns. The LLM runs in a loop, the agent has access to a set of tools — on your computer or in the cloud — and you can hand off long-running tasks. The efficacy has been improving exponentially. This might just be the final form factor of AI: an agent that could run for a minute or a month, has access to any data you need, has access to all the tools you work with, can act as you or as its own entity. It's just an LLM constantly running in a loop, making decisions, and you intervene to steer or review. This appears to be the architecture of the future of AI.

Newton: I've come to believe that basically everyone should have the experience of building a website using an AI agent of their choice. A lot of things clicked for me when I started watching a computer use itself. You recently posted about how, somewhat strangely, AI doesn't seem to be helping any of us work less. You mentioned that you'll start working on something with an agent that you think is simple and you lose three hours to it. Was that a real project? Can you share what it was?

Levie: It's so pedestrian I'm embarrassed to share, but I was going into a city a week later and needed to map out a bunch of customers I should be visiting. I was using Perplexity Computer, which does some pretty good workhorse stuff, and I gave it the task: "Rank-order all the top 50 companies in this region. Get me the LinkedIn of every single CIO of those companies, so I can make sure — okay, who have I connected with?" I didn't even know what I'd do next, but I wanted to get a good map. That task took maybe 15 minutes — you prompt it once, you get back some data you don't really like, you re-prompt, it does better, you do it a third time, and boom, you're off to the races.

So maybe 15 to 30 minutes of AI work. But then the very next thing — I had to do something with this data. I spent the next two hours emailing all the people and filling up my calendar more. It was the kind of task where I thought, "Oh, use AI to accelerate this thing" — and it kind of worked too well, to the point where I had created more work for myself. It would almost have been better if it came back with full hallucinations, because then I could have just gone to bed and been like, "Well, that failed." But it worked. And by 10 PM I was like, "Well, now I'm going to feel bad if I don't actually do anything with the tokens I just exhausted."

It's a small anecdote, but I think this is happening everywhere. You're like, "Oh, I'm going to tell AI to write this little web app." Then you're like, "Well, I built it. Now I've got to add this other feature. I should probably get it hosted somewhere. I might need to change this one thing." It's basically If You Give a Mouse a Cookie applied to the economy. You just start building up more and more work.

The part I find really interesting — and this tangentially relates to why I think the job-loss argument is wrong — is that people will find there are way more tasks they could be doing that they just never could do before, because the fixed cost of starting the task was too high. AI made it easy to get going. They lit up the project, did the research, the analysis, reached out to the customer. That kicks off a cycle of downstream work, or a new set of constraints that start to emerge.

To bring this home: the task worked so well that I idly wonder, should I have a full-time person just doing this with an agent? They'd use an agent, but I don't want to do this every night for the rest of my life between 9 PM and midnight. It might be valuable enough that it's worth a person to do this for me — in which case, it actually created a job because of my experimentation with this AI I was doing for fun.

Newton: I love that story. When I built my own website, it was absolutely If You Give a Mouse a Cookie, because of course after the website was done, I said, "Well, I need to host it." And then, "Well, it should probably have a blog." So I added a blog. Then, "Why isn't it telling me the current weather in San Francisco? I have to solve this problem for the three annual visitors to this website." It was super fun. I don't regret the time. I also didn't, I think, create a lot of value for myself.

On the flip side, I run what I sometimes think of as a somewhat fake business, in the sense that people pay me to email them. So it's a very strange, uncomplicated business. I have a bookkeeper and an accountant, and they send me a monthly email letting me know how things are going, but I've never really done real financial analysis. And then Claude Cowork shows up, and I can just start chucking spreadsheets into it. For the first time since I started my company five years ago, I'm like, "Tell me about my business."

Levie: "What's our revenue?"

Newton: Yeah, completely. It pushed me to make some changes, including, by the way, starting this series of conversations that you were the first person on. So: what advice do you give for people when they're like, "Okay, I'm bought in, I'm going to try this stuff. But how do I know when I'm just spinning my wheels versus creating something of durable value"?

Levie: This one is hard to answer generically. But this will be a defining question of the next decade. If you have access to abundant intelligence, but it's not free — it's abundant but not free — how do you allocate the spend across the organization?

On one hand, it's easier the higher up in the organization you go, because you have all the data about what the company does well and what doesn't. On the other hand, it's somewhat easier lower in the organization where the work is actually happening, because you can self-identify the work that matters. The problem is, both of those have issues: the direct user might think their work is the most important for the tokens, and the person high up might not know about the new innovation somebody has.

The world was a little easier with scarce intelligence, because everything was kind of slow, and it had to be slow. Now with abundant intelligence, you could have everybody running around spinning up agents doing lots of work — maybe 70% of which is not valuable. But you don't know which is the 30/70 until you've done the whole set of things.

Newton: Is one potential solution to just put up a leaderboard of who has burned the most tokens in a given month and reward them somehow?

Levie: (jokingly) That is emerging as the best practice. The alternative is just token-maxing, and you're good. But token-maxing aside — there's a tool set I'm sure 10 startups are working on right now. They'll see this and they'll all pitch us, but it'll be a good business for a few of them. There's a new kind of ERP, HR, finance system that lets you have a heat map of where the tokens are going, and the rough value allocation of what that produced. Right now it's a comical idea — "Oh, you're going to have to treat tokens like headcount." But we're only going to be able to apply compute rationally to the most important areas of the business. Right now it might be 1% of your company spend, but if in three years it's 10% of the spend, we don't joke around with where 10% of our labor force is going. The same will be true of your tokens.

Newton: Let's talk about jobs. You recently posted about a kind of Gell-Mann amnesia for AI. Gell-Mann amnesia, of course, is where you read about something you know a lot about and notice the obvious errors, but assume that same source is credible on other topics. You wrote: "People use AI for their job and see all the various things they have to do in the last mile, but then look at someone else's job and think that AI will eliminate it immediately." Why are some people so quick to think AI can automate away a whole job?

Levie: This is amazing technology — the coolest technology I've ever played with in my life. So it's deceptively cool. You're like, "Oh my God, I think I just did my taxes," or "Oh my God, I just built this amazing marketing website in five minutes." Then we look at the output and we're like, "Gosh, that must totally replace the job of XYZ profession."

There are a few core flaws with that. What is that profession doing for all the hours in their day? How much of it is just doing the final calculation of your taxes, versus getting all your data in order, reviewing all the work, asking you questions back and forth, knowing the right questions to ask, dealing with the three missing things you didn't even remember to add — but if an AI system had done it, it would have totally glossed over? That's what the profession does. The automation of one or two or five of the steps are just individual tasks.

In the case of development, you or I can tell Claude Code, "Generate me the XYZ product." And we could be like, "Wow, that must automate the engineer out of existence." Well, the code quality is probably horrendous. The ability to ask it to do 40 other things over a 12-month period is going to stack in complexity. The moment you actually want to get that software hosted, make sure there's no downtime, ensure you have a good distributed system — that's already 100 times more complex than just prompting the code to get written. The moment there's a security event, somebody has to wake up and respond. I can name 30 other things a developer has to do.

A lot of people say the job of the engineer was never to write code, it was to do X. But no — they were writing code most of the time in the prior world of work. The problem is they were highly constrained by how much code they could write in a day, and they were automatically bottlenecked from doing the other things their job could be. So what is the future engineer? It's to understand what you're trying to build, to make sure it gets built properly, to ensure there are no security issues, to ensure it gets released, to ensure it's high quality.

If you or I go and vibe-code something, we think we've replaced the engineer, replaced the accountant, replaced the lawyer. But then you actually look — that was the first 80% of the job. The extra 20%, it turns out, is all the value creation of that profession. All the expertise and domain knowledge is in that last 20%, not the text that got generated.

And the converse: I'll use AI for analysis of a market I'm thinking about. If I just took the output and ran with it, I know it would not work, because I know it's missing context — either I didn't give it, or I know something else about a different trend. But somebody else might see that and say, "Wow, Aaron's job is incredibly easy, AI just gave me the answer of what he's going to go do." And I'm like, "No, my job is way harder, I promise."

There's a different axis people need to think about. If you took today's static work, maybe it would be: you'll get the first 90%, then we're going to automate the next 9%, then the next 0.9%. But there's a dynamic part of the equation. The market is starting to ask more from the provider, because they now know what's possible. So just as you automated the first 90%, the market shifts on you, and that 90% is now the new 50%. The demands of what you ask an engineer to do go up tenfold, because you're like, "I think you can do that thing way faster now, so I'm going to give you a much bigger project." You have this dynamic system: our needs and demands are growing as a result of what we can automate.

Newton: I love the idea that AI will let us reenvision what our jobs could be in a more expansive, creative vision. My fear is that the last mile of human supervision will turn out to be kind of boring. We're already starting to see this in some jobs — there's a piece that's automated, and my job is now just to review AI output, and that is just pure drudgery. Does that complicate the picture you just painted?

Levie: It adds a wrinkle. The big question is: are there jobs in the future? The answer is yes. Now the question is, do we want those jobs of reviewing the output of AI agents? Maybe we all just opt out of the economy because we don't want that job. So interesting philosophical question of what's the new way to get fulfillment and creativity out of these jobs. You see burnout of engineers on X, basically saying, "All my job is is to review slop from the AI." There's a limit to how fun that is.

But engineering is a unique job compared to the rest of the economy. Most of an engineer's day is to think about a problem, think about a system, write code — and that code is text. You're just writing a lot of text, and somebody else reviews the text and you ship it. So if the agent did the writing and reviewing, then all you're really doing is reviewing the text and shipping it.

But go talk to an investment banker, or a lawyer above paralegal, or a doctor or a nurse, or a pharma researcher — they would love to get out of the toil. They'd say, "I don't want to spend 15 hours generating a corporate pitch deck for this client — that is just me moving images around on a PowerPoint, doing some Google searches to find market trends, and pasting that in. I'd love to automate that. The job I should be doing is getting in front of my clients and making sure we're delivering unique value and insights to them."

That's why I'm not overly worried about this. We're seeing some hyper-accelerated dynamics in engineering that don't always relate to other forms of work. If you talk to a doctor and ask, "How much do you enjoy typing up the patient notes after the patient meeting?" — they want to automate the heck out of that. They want to be done with that part of the job. Mostly it'll be a net positive change as people are able to get rid of the stuff they hate doing, and the demands of the job evolve in a way that makes it much more exciting.

Newton: It's the sort of thing I love to hear. I would love to not live through a massive disruption where we see super high unemployment. I also can't help but note we saw almost 46,000 tech layoffs announced in March alone, with AI sometimes cited as a potential cause. So would you put a number to it? On the software engineering front, do you think in three years we have about as many software engineers as we have today, or more? Or do you think there's a bigger shift?

Levie: I think we're going to have more. I don't want to be unsympathetic to people who really will face these changes. But big picture: if you were a CS grad of the past two decades from a top 25-to-50 CS school, by and large you were trying to go to a tech company — in Silicon Valley or a couple other places. So most of the software talent in the world of this cohort ended up building software for consumers. We were building ad apps, ride-sharing, enterprise software (thank God). We've accumulated a lot of engineers on that kind of work, and some of those companies have over-hired.

Who's the loser of that equation? Every other company on the planet, because they couldn't compete with Google and Facebook and Microsoft for that top engineer. They couldn't automate things in the life sciences process, or the supply chain, or automotive AI systems. I don't know how much software you've used from companies that aren't in the Valley, but if you log into your bank and you're happy, you're a totally rare person. If you look at most car console designs of any car that's not from two companies, you can imagine how unusable these systems are. That correlates to the fact that those companies couldn't overstaff with all the top engineers and designers.

Now what happens? All of a sudden, what was maybe a 30- or 50-engineer problem previously, Claude Code and Codex come in, and now it's a 5- or 10-engineer problem. For the first time ever, those companies are able to take on work that wasn't possible before. They can bring automation to all the systems and workflows they couldn't have afforded or justified.

So in some cases of tech, you'll see a temporary dislocation. At the exact same time, the thing you should be tracking is the number of engineering jobs opening up at traditionally non-Silicon Valley tech companies — small businesses, consulting firms, life sciences, manufacturing.

And as a force multiplier, you're going to have a number of new types of engineering jobs where the job is entirely about how to deploy agents inside the firm to automate work. I did this for fun just to make sure I wasn't full of shit. If you go to the Eli Lilly careers page, as one does, they have this job title called "lab automation software adviser." That person is an engineer whose job it is to bring automation through AI to the lab process.

Think about how many hundreds of thousands or millions of jobs will look like that in the future. My job is to take the innovation coming from AI-land and apply it to this particular business process in my organization. You're kind of like an FDE — a forward-deployed engineer — but for that company. Those will be the people who would have gone to Meta or Google five years ago. They're going to now work in pharma, banking, manufacturing. And those are actually incredibly stimulating jobs. You're not just building an app, you're automating drug discovery.

Newton: So you're saying even in the future, even just with my small newsletter business, I might one day be able to fulfill my lifelong dream — which is to compete directly with Palantir and just build a vast surveillance and analysis program?

Levie: If you so choose, you can. What's the fastest-growing role at something like Anthropic or OpenAI? It's these FDE roles.

Newton: Forward-deployed engineers.

Levie: Yeah, you need humans to go implement this stuff inside the organization. Those are the engineers of the future, if you're not just building software that is an application.

There's a funny article out of the FT — I don't know how much I trust their views on technology — that lawyers are being inundated by clients asking them questions because they went to AI, and the lawyers have to verify the advice is good, or review the contract that got written. What we're doing is lowering the barrier for everybody to participate in these things in a touristy way. I can be a tourist in software development, in legal, in healthcare. But that eventually still needs to get verified, or the work actually has to get done that last mile, and that eventually moves into needing some kind of semi-expert.

This is why I don't think we yet know what degree you should go into in college — I don't think any of the degrees are off the table. You should totally go into CS if you're really excited about software development. You just shouldn't expect to go build a little app that you press a button on. You should expect that you're going to use CS skills to go do clinical trial automation at a pharma company.

Newton: I think that's a great place to land, because it leaves me with a feeling I have so rarely when thinking about the tech-enabled future lately, which is optimism. So thank you, Aaron, for giving us a jolt of that. It'd be great to check in with you again in a year and see if the picture still is as rosy. Aaron, thanks so much for joining us.

Levie: Thanks, man.

A MESSAGE FROM OUR SPONSOR

Become an AI-native team with Rovo

Atlassian Rovo is AI that knows your projects, code, and people so it can bring context (and guardrails) to every workflow.

And because Rovo lives where your teams already work, it doesn’t just find the answers — it helps you do the work.

See how Datasite is becoming an AI-native team with Rovo.

Learn more.

Following

Google fights an AI-generated zero-day

What happened: Google says it found an AI-generated zero-day exploit that could have triggered a “mass exploitation event.”

Google found “prominent cyber crime threat actors” were planning to conduct a “mass vulnerability exploitation operation.” The exploit would have enabled users to bypass two-factor authentication on a popular open-source tool.

Google’s Threat Intelligence Group says it tipped off the affected software maker and worked with them to prevent the attack.

The company said it is confident that a Python script the attackers created was AI-generated — because of its characteristic “textbook” AI coding style, and also because, despite the code working, there were some hallucinations (lol).

Google said AI is increasingly helpful for discovering vulnerabilities of the type used in this exploit — which was previously done via manual review from human experts. Unlike traditional software, AI can sniff out code that has high-level logical flaws. AI can find types of errors that “appear functionally correct to traditional scanners but are strategically broken from a security perspective.”

Why we’re following: The exploit Google found is confirmation of a growing threat: AI can find and exploit vulnerabilities that previously only humans could. 

What’s more, hackers are starting to experiment with going beyond just discovering vulnerabilities, to actually using AI to orchestrate cyberattacks in real time.

The company’s research has found actors associated with China and North Korea doing “sophisticated” experimentation with AI, it said. They’ve also seen actors experimenting with orchestrating attacks autonomously and using OpenClaw to refine attacks.

What people are saying: On X, John Hultquist, Google Threat Intelligence’s chief analyst, wrote, “I think most of us are surprised we have not found more evidence” of bad actors using AI to discover exploits. “We believe this is the tip of the iceberg. Other AI-developed 0days are probably out there.”

He added, “If criminals are doing it, then state actors with significant resources probably are too.”

OpenAI needed “big computer”

What happened: Sam Altman appeared in court today for the trial where Elon Musk is seeking to remove Altman and co-founder Greg Brockman from their jobs at OpenAI.

The NYT’s Mike Isaac writes about the contrast between Altman and Musk’s testimony: while “Musk openly sparred with opposing counsel,” Altman “has taken a different tack. His answers have been terse, quiet and noncombative.”

Musk’s lawyer Steven Molo questioned Altman about several executives and board members who said under oath that Altman had lied to them. 

At one point, Molo asked “are you completely trustworthy?” Altman replied, “I believe so.”

In possibly our favorite detail so far, Altman described an evening meeting at Tesla, during a time in which Elon Musk was trying to fold OpenAI into his car company. The meeting apparently contained “a long, long period of time with Elon showing us memes on his phone.”

Elsewhere at trial, Microsoft CEO Satya Nadella testified that Elon Musk never told him about concerns that Microsoft’s investment in OpenAI violated OpenAI’s charitable commitments. (Musk is suing Microsoft in addition to OpenAI).

OpenAI co-founder Ilya Sutskever said in court that his stake in OpenAI is now worth about $7 billion. He gave judge Yvonne Gonzalez Rogers a memorable explanation of the difference between the AI of OpenAI’s early days and now: “I would describe it as the difference between an ant and a cat.”

Sutskever told Judge Gonzalez Rogers why OpenAI needed outside investment after Musk’s departure. “If there is no funding, there is no big computer,” Sutskever said. “You don't need the biggest computer, but you need a big enough computer.” And, well, “If you don't have a big enough computer, it was not going to work.” 

Why we’re following: Platformer would like to start a petition to replace all mention of “compute” with “big computer” from now on.

The evidence that Musk was trying to fold OpenAI into his company isn’t great for his case that Altman and Brockman “stole a charity.” Although he apparently could have managed his time better while he was doing it! But in fairness to Musk, we cannot ourselves plead innocent to repeatedly showing coworkers memes on our phones.

What people are saying: On X, programmer and investor Paul Graham, who chose Altman as his successor at startup incubator Y Combinator, wrote: “One of the things Musk vs Altman shows is how much more promising AI is than anyone expected.” Graham added, “Sam could have started it as a for-profit company. His life would be much simpler now if he had. But he didn't realize in 2015 that AI would warrant more than you can raise in donations.”

—Ella Markianos

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and job automation arguments: casey@platformer.news. Read our ethics policy here.

This is a column about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

In February 2025, Vice President JD Vance took the stage at the Paris AI Action Summit to share the administration’s views on AI regulation. “The AI future is not going to be won by hand-wringing about safety,” he warned. Excessive regulations might “kill a transformative industry just as it’s taking off,” Vance said, and suggested that AI companies asking to be regulated might simply be trying to crush their future competitors.

Vance’s remarks reflected the idea, then common among Trump officials, that fears about AI capabilities are dramatically overstated. David Sacks, the White House’s AI and crypto czar, has referred to a “doomer industrial complex” enacting a “sophisticated regulatory capture strategy based on fearmongering.” Michael Kratsios, who leads the Office of Science and Technology Policy, has complained that international efforts to govern AI “maintain a general atmosphere of fear.”

The administration has backed up its rhetoric with a lobbying push intended to block most state-level AI regulation. Axios reported last month that Trump officials are pressuring Republican lawmakers in Nebraska and Tennessee to weaken or abandon bills in their respective states that would introduce safety and transparency requirements for AI companies.

Which is what makes the administration's latest move so striking. Trump is quietly reviving a Biden-era idea his own officials once mocked — pre-release government review of powerful new AI models.

Here are Tripp Mickle, Julian E. Barnes, Sheera Frenkel and Dustin Volz in the New York Times:

The administration is discussing an executive order to create an A.I. working group that would bring together tech executives and government officials to examine potential oversight procedures, according to U.S. officials, who declined to be identified in order to discuss deliberations over sensitive policies. Among the potential plans is a formal government review process for new A.I. models. [...]

The working group is likely to consider a number of oversight approaches, officials said. But a review process could be similar to one being developed in Britain, which has assigned several government bodies to ensure that A.I. models meet certain safety standards, people in the tech industry and the administration said.

The Biden administration had issued its own executive order that instructed AI companies to perform safety testing and share the results with the government before releasing new models. Trump revoked the order on his first day of his second term. Three days later, he issued a new order titled “Removing Barriers to American Leadership in Artificial Intelligence” that effectively ended safety testing requirements.

What changed? Mythos. Anthropic’s latest large language model, now available in preview to a small number of companies, has proven capable enough at developing cybersecurity exploits that the government believes it poses national security risks. The White House now opposes the company’s plan to expand access from roughly 50 companies to 120 for security reasons. (It also says it worries Anthropic doesn’t have enough compute available to serve the model to government customers; Anthropic denies this.)

All of this is complicated, of course, by the fact that the Trump administration has also sought to designate Anthropic as a “supply chain risk” because it refused to amend its contract with the Pentagon to enable “all lawful use” of its technologies. While continuing to defend that designation in court, the administration has simultaneously been working to expand access to Mythos throughout the government. 

Trump officials are now in the nonsensical position of trying to help agencies get around the legal roadblock they themselves set up to stop them from using Anthropic’s models. One set of officials is working to phase out the use of Anthropic models over the next six months; another is working to expand agencies’ access to its technology throughout the government.

In the meantime, the rest of the industry now faces a regulatory environment that looks awfully similar to the one Democrats had implemented under Biden: a world where they submit their models to the government for review before releasing them widely. On Tuesday, Google, Microsoft and xAI all said that they would give the government early access to their models. The reviews will be handled by the US Commerce Department’s Center for AI Standards and Innovation.

Before Trump 2.0, by the way, that body was known as the US AI Safety Institute. Its name changed last June. “For far too long, censorship and regulations have been used under the guise of national security,” Commerce Sec. Howard Lutnick said at the time. “Innovators will no longer be limited by these standards.”

Less than a year later, the administration’s sneering dismissal of safety concerns has transformed into something that resembles a mild panic. The National Security Agency is now using the model to look for vulnerabilities in Microsoft products — and, one assumes, contemplating the fact that foreign nations will soon be using similarly capable technology against US critical infrastructure, if they aren’t already.

​​Meanwhile, the public backlash against data centers and other symbols of AI power is putting the Trump administration increasingly at odds with its own base. And the government’s half-baked AI sales pitch to the general public, which has amounted to little more than “get rich and beat China,” has failed to resonate much beyond the venture-capitalist offices where it was originally conceived.

One result of this is that Trump’s effort to place a moratorium on most state-level regulations of AI now seems even less likely to pass than it was before. Another likely effect of accelerationists’ declining influence is that we’ll see a push for expanded export controls on powerful chips to China. (Sacks, who recently left his job as AI czar for a role on the Council of Advisors on ​Science and Technology, had been a vocal proponent of loosening those controls.) A less likely but welcome development would be that the US re-engages with the United Kingdom, Japan, Korea and other allies to develop a shared strategy toward governing more powerful models.

Still, less democratic possibilities exist as well. Critics of the White House’s plans to subject frontier models to safety evaluations worry that the Trump administration will use any licensing regime for censorship — denying releases to models whose output is deemed “woke,” for example, or simply to pressure companies into doing other favors for the administration. Imagine Brendan Carr’s Federal Communications Commission, but for AI. Some level of worrying there is warranted.

But after Vance’s speech in Paris, I noted here the dangerous negligence of an AI policy that amounted to little more than “let’s see what happens.” A year later, the administration has come to realize that all those AI safety concerns were no mere hand-wringing. The models are getting more capable — and more dangerous. What Sacks once dismissed as the doomer industrial complex now includes a growing number of federal agencies and Trump administration officials.

And while they should have taken these fears seriously all along, I will settle for the administration taking them seriously now.

A MESSAGE FROM OUR SPONSOR

Become an AI-native team with Rovo

Atlassian Rovo is AI that knows your projects, code, and people so it can bring context (and guardrails) to every workflow.

And because Rovo lives where your teams already work, it doesn’t just find the answers — it helps you do the work.

See how Sprout Social is becoming an AI-native team with Rovo.

Learn more.

Following

The OpenAI-Elon Musk trial enters week two

This week in court: 

OpenAI co-founder Greg Brockman said he didn’t want Elon Musk to be OpenAI’s CEO because “he did not – and I believe does not – know AI,” in federal court today. Brockman added that he and co-founder Ilya Sutskever “did not think that he was going to spend the time required to actually get good at it.”

Brockman told jurors that Musk called a predecessor to ChatGPT “stupid,” and said that “kids on the internet could do a better job of it,” which raised concerns within OpenAI about his ability to run the company.

During discussions about a potential for-profit conversion, Brockman says Musk demanded a majority stake, saying he needed $80 billion to start a city on Mars. When Brockman pushed back, Musk allegedly said he could start another AI company tomorrow with “one Tweet.”

Musk, who also owns AI company xAI, is suing OpenAI for unlawful enrichment. He claims his original charitable donation to OpenAI should not have contributed to the for-profit venture OpenAI eventually created. OpenAI claims the suit is a “jealous” bid to attack a competitor to Musk’s xAI.

During his time on the stand, Brockman got grilled about his personal journal, which included such musings on OpenAI’s for-profit conversion as: “Financially what will take me to $1B?”

Musk lawyer Steven Molo asked why — if Brockman’s goal was a mere billion dollars — he hasn’t donated the rest of his $30 billion stake to OpenAI’s nonprofit. “It takes 30 billion dollars to get you out of bed in the morning?” Molo asked. Brockman said Molo was twisting his words.

In one of the trial’s more operatic twists, Brockman testified that when then-OpenAI board member Shivon Zilis had twins, she didn’t initially tell him Musk was the father. The trial proceedings previously revealed that Zilis secretly funneled information about OpenAI to Musk. Brockman said he found out Musk was the father of her children through public reporting — and that Zilis told him at the time that her relationship with Musk was “platonic” and that the children were born via IVF.

Elsewhere at trial, Brockman testified that OpenAI will spend $50 billion on compute this year.

Why we’re following: Musk is asking the court to remove Brockman and OpenAI CEO Sam Altman from their leadership positions, and is seeking as much as $134 billion in damages, which he says he will donate to the non-profit foundation that controls OpenAI.

While the stakes for OpenAI’s future are high, we are admittedly more attuned to the various petty dramas that are unfolding in court. (Just two days before the trial began, after Brockman rebuffed Musk’s text suggesting that the parties settle, Musk responded: “By the end of this week, you and Sam will be the most hated men in America. If you insist, so it will be.”)

In any case, let Brockman’s experience be a reminder to all of us to never, ever, write your diary in a Google Doc.

What people are saying: The “takeaway from Greg Brockman[‘s] testimony at Elon vs. OpenAI trial today is that no grown man should have a diary,” wrote Sources’ Alex Heath.

Meanwhile, Musk agreed to pay $1.5 million to settle SEC allegations that he deceived Twitter shareholders when he failed to disclose his growing stake in the company, which the SEC alleged led to an artificially low stock price.

Fascinatingly, notorious Silicon Valley fraudster Elizabeth Holmes congratulated Elon on his Twitter settlement, writing, “I had an SEC settlement too” (you don't say). She added, “Elon's $1.5M settlement is basically a parking ticket. No admission. No criminal conviction,” concluding, “Big win for @elonmusk.”

—Ella Markianos

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and feedback on these changes: casey@platformer.news. Read our ethics policy here.

Hi from New York, where I spent this week talking with chief financial officers, chief operating officers, and IT leaders about how the AI moment is affecting their businesses. In general I find that people are open to the idea that it can help them; skeptical it has done much for their businesses so far; but also committed to continuing to experiment lest their rivals figure it out before they do.

Those conversations were on my mind as I read this week’s earnings reports from the tech giants, which offer us a fresh chance to check in on the AI bubble. And while this has been true for a while now, these numbers should put to rest one of the primary battle cries of the AI backlash: that “nobody wants this.”

Of course, many people (and more every day!) hate AI. But their bosses are buying it — and in record numbers. Google Cloud revenue, which is one imperfect but useful proxy for AI spending, was up 63 percent over the last quarter. Meta, which has invested heavily in building AI tools to make ads more effective, reported a 33 percent increase in revenue.

Amazon revenue was up 17 percent, and said its chips business is growing in the triple digits; Microsoft was up 18 percent, and said its AI revenue is up 123 percent year over year.

Last year those four companies unnerved some investors by spending a record $410 billion on capital expenditures. And with three of those companies updating their guidance this week, the big four now plan to spend 77 percent more than they did last year — $725 billion.

That’s because the biggest players are already selling about as much AI as they can make, and are scrambling to build or rent the infrastructure that will allow them to sell more. OpenAI said today that it has signed contracts for 10 gigawatts of AI computing capacity three years ahead of schedule. (And also that its models can’t stop talking about goblins, but let’s set that one aside for now.)

Notably, the market has more trust in some companies than others. Google stock was up nearly 10 percent today to an all-time high. Meta stock, meanwhile, was down more than 8 percent. 

Of the big four, Meta currently has the weakest case that its own services will generate the level of AI demand that would necessitate spending of up to $145 billion this year on infrastructure. CFO Susan Li told investors that access to compute “will be critical to determining the quality of the models we develop, the types of products we can introduce, [and] how productive we can be as an organization.” But it’s clear investors have not been particularly impressed with Muse Spark, and after losing nearly $84 billion on its Reality Labs side quest to date, the company has a trust deficit. A rare decline in users isn’t helping, either. 

At the same time, investors instantly snapped up $25 billion in Meta bonds — albeit at a higher risk premium than an earlier sale six months ago.

On the whole, though, the market is telling an uncommonly simple story: demand for AI is increasing exponentially, and the hyperscalers are investing heavily to make more of it. 

All of this remains very difficult for Ed Zitron to understand. But after years of speculating about what kind of bubble this is — including by me — a relatively clear story is coming into focus. It is not a momentary-enthusiasm-deflating-once-reality-sets-in kind of bubble. Instead, it’s looking more and more like a railroad bubble: one where massive overbuilding and speculation may lead to periodic crashes and flameouts, but also one where the resulting product transforms commerce and productivity.

Which is why, in the end, I think the leaders I spoke with this week have it right: saying yes to gradual iteration, and no to betting the farm. There’s value to be found in there somewhere, but for the moment the playbooks are still being written. 

On the podcast this week: Kevin and I discuss OpenAI's big reset (and trial with Elon Musk). Then, Dr. Adam Rodman returns to the show to discuss the latest with AI and medicine. And finally, Talkie co-creator David Duvenaud stops by to discuss his AI model trained exclusively on pre-1930 data — and whether he can use it to predict the recent past.

Apple | Spotify | Stitcher | Amazon | Google | YouTube

Following

The government can't decide what to do about Mythos

What happened: The federal government’s relationship with Anthropic grows more arcane by the day. The White House is working on guidance for agencies to work around Anthropic’s supply-chain risk designation — which the administration itself is responsible for — to access Claude Mythos, the company’s highly cyber-capable model. An anonymous source described the plan to Axios as a way to “save face and bring ‘em back in.” 

The Trump administration is also writing a memo that will guide national security agencies on AI use, Bloomberg reports. The memo will encourage agencies to use multiple AI providers to avoid the risk of relying on a single (and potentially woke) vendor. It will also tell companies not to interfere with the military’s chain of command. The memo is seemingly based on the issues the DoD had with Anthropic, although it’s currently ambiguous whether these policies meant to guide the further use of Anthropic’s systems, or to encourage agencies to phase Anthropic out.

Secretary of Defense Pete Hegseth definitely still hasn’t warmed up to Anthropic. In a Senate hearing yesterday, he called CEO Dario Amodei “an ideological lunatic who shouldn’t have a sole decision-making over what we do.”

For now, though, the National Security Agency has gotten access to Mythos, and is using it to test Microsoft software for vulnerabilities.

Meanwhile, the White House opposes Anthropic’s plan to expand Mythos access to 70 additional companies and organizations — which would bring the total number of organizations with Mythos access to about 120. Administration officials told Anthropic they had security concerns with the initiative. (So that wasn’t just “fear-based marketing”?)

And there are plenty of other cyber models for those in the know to test, enjoy, and worry about. The UK AI Security Institute found that OpenAI’s GPT-5.5 demonstrates cyber capabilities comparable to Mythos’. And Anthropic released Claude Security, a cyber product built on its Opus 4.7 model, to enterprise customers.

Why we’re following: The White House appears to be impressed with Mythos’ cyber capabilities — impressed enough to try to get around its own rules to keep using it. And impressed enough to do something that it has historically opposed: restrict the use of AI!

It seems AI cybersecurity is a policy priority for the administration — although it appears to come second to its top policy priority, which is never admitting to being wrong. Which means Anthropic’s models might play an increasingly important role in America’s national security, at the same time the company is officially labeled as a national security liability.

GPT-5.5’s cyber capabilities make the story more interesting. Typically, when one lab makes a big leap in AI capabilities, that leap is then followed by other top developers — and eventually by open-weight AI models. Well, now here’s OpenAI coming a few weeks after Anthropic with another highly capable cyber model — and this one is generally available.

The White House has no guarantee that every developer of cyber-capable models will make like Anthropic and ask the US government for permission before deploying its models more widely. If the administration believes that models of Mythos’ cyber capability level should be restricted, they’ll probably want a regulatory lever that goes beyond asking companies informally to restrict their models.

What people are saying: Former White House AI advisor Dean Ball wrote on X, “Perhaps the White House should consider ‘issuing guidance’ to the DoW to put an end to this stupid, manufactured crisis by… withdrawing the supply-chain risk [designation].” 

Blogger Matt Yglesias joked about the White House’s predicament: “When you fucked up but don’t want to admit it.”

Ball did agree with the White House asking Anthropic to not expand Mythos access to new organizations, saying “I suspect the White House is making the right call.” But, he added, “this is the opposite of a tenable strategy, like trying to erect a dam against a tsunami.” Given the increase in AI capabilities from multiple model-makers, “There is no way to stop the diffusion of capabilities like Mythos within the next 6-18 months.”

On the other hand, OpenAI safety researcher Boaz Barak thought Anthropic’s original plan was the right one: “Assuming they have the compute, expanding access is good.” But he thinks it would be even better to make a restricted version of Mythos “generally available.” 

Barak thinks that more access to cyber models will be good for both politics and cybersecurity: “Wide iterative deployment helps avoiding concentration of power,” among specific companies or political actors, and would help cyber defenders in “discovering issues early.”

—Ella Markianos

Musk takes the stand

Week one of the high-profile OpenAI trial is off to a fairly anticlimactic start. Elon Musk, an early investor in OpenAI, alleges that he was “manipulated” by OpenAI into donating to what he believed would always remain a nonprofit. When OpenAI developed commercial ambitions, he says, the company effectively “stole a charity.” 

On the witness stand, Musk reiterated how upset he was that OpenAI allegedly deceived him — and said he was a “fool who provided them free funding to create a startup.” (When asked earlier about the potential tax breaks he received as a result of providing the “free funding,” he accused OpenAI’s attorney of asking questions “designed to trick me.”)

OpenAI maintains that Musk knew about the company’s for-profit plans and supported its decision, but created a competitor after the founders refused to give him control of the venture, and is only now suing them in an effort to slow them down.

OpenAI’s lead counsel showed emails from September 2017 between Musk and OpenAI executives in which Musk demands to have more voting power than his cofounders by having the right to pick four members of the company’s board. “I would unequivocally have initial control of the company, but this will change quickly,” Musk wrote in a message.

Notably, Musk admitted on the stand that xAI has used tech from OpenAI to build its models, which is not allowed by OpenAI’s terms of service. He said it was “partly” true that xAI has distilled OpenAI tech, and said “generally A.I. companies distill other A.I. companies.”

Meanwhile, one topic you will not hear about throughout this trial: whether AI is an existential threat to humanity. The judge put a stop to the bickering after both Musk and OpenAI’s lawyers started arguing over the issue.

“I suspect that there are a number of people who do not want to put the future of humanity in Mr. Musk’s hands,” the judge said. “But we’re not going to get into that.” 

—Lindsey Choo

Also reading:

Meta threatens to pull its apps out of New Mexico.

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and bubble discourse: casey@platformer.news. Read our ethics policy here.

This column touches on AI. My fiancé works at Anthropic. See my full ethics disclosure here.

I. 

As I do most mornings, I began work yesterday by checking my Signal messages. Along with the usual unwanted PR pitches and messages from people in the middle stages of AI psychosis, I had received a genuinely great tip. It was a story squarely in our coverage area that, if properly covered, could draw attention to a pressing issue on tech platforms and put pressure on it to change.

The tip should have filled me with excitement. Instead, though, I felt something closer to dread. When was I going to begin making the many phone calls needed to verify this information? How could I find time to meet a source or two in person? Did I have all the sources I would need, or would I need to somehow develop some more?

Since I began writing a daily newsletter in 2017, I have always faced some version of these pressures. My historical approach has been to report scoops whenever I can, and fill out the rest of the time by writing news analysis — bolstering it whenever I can with extra details of original reporting. As longtime Platformer readers will know, in practice this meant that the balance of what we publish here has leaned toward news analysis. A daily publishing cadence leaves enough time for synthesis and sense-making, but not for deep digging and phone tag.

For most of the past decade, I’ve liked this arrangement. I began writing in the aftermath of the 2016 US presidential election and the growing backlash against tech companies, and the glut of coverage benefited from a publication dedicated to a daily close reading of the news. When I started publishing a roundup of links related to the intersection of tech and democracy, I felt like I was doing something genuinely novel on my beat.

Fast forward to today, and the world of link roundups feels much more crowded. A generation of tech writers filed out of the newsrooms where they grew up and began to write for audiences of their own. Newsletters, which were once an afterthought in media, are now a central pillar of many publishers’ strategies. But the ongoing collapse of the web and related struggles at big media companies means that there is now less tech journalism overall. The need for sense-making is greater than ever, but due to a half-decade of layoffs and shuttered publications, there is less and less journalism to make sense of.

Meanwhile, improvements in artificial intelligence over this year have resulted in systems that further encroach on the work we do here. In January, I wrote about the experience of building an automated daily briefing of link summaries for myself; I have been using it all year to look for story ideas. It does about as good a job as I do in finding stories of interest, and it does so automatically while I sleep.

Link aggregation was never the highest-value work we did here. But I do think that its value has decreased significantly over the past year, and will decline further as more people begin using personal agents to write news digests for them. (Already, it seems that a staggering percentage of Substack posts are AI-generated in whole or part; they are arbitraging the fact that you are not yet doing this yourself.)

And to crawl a bit further out onto a limb, I expect some smaller set of people — but in particular the executives, policy professionals, and communications team that Platformer has long written for — to begin relying on AI for news analysis as well. Over the past year, chatbots have gotten sharper at responding to questions about the implications of this or that news story — how it changes the competitive landscape, for example, or how regulators might respond.

For the moment, chatbots carry far less authority on these subjects than the domain experts who often write paid newsletters about them. But having been a reporter since 2002, my experience has been that the internet is working continuously to deskill and replace you. It doesn’t require much of a leap in imagination on my part to imagine a day where your current lineup of morning and afternoon newsletters is largely replaced by an agent-written briefing that has been exquisitely tuned to your professional concerns — and, unlike this newsletter, instantly respond to your questions about its findings.

If Platformer’s three pillars are original reporting, news analysis, and link aggregation, then, it would seem that one of those has already been commoditized and the second may be on its way.

II.

We've been trying to evolve to keep pace with these changes. But I've come to believe that we need to move faster.

In September, in my annual anniversary post, I mentioned that I wanted to take more time off from writing to report. You all were universally supportive of the move, reminding me that you are paying for quality rather than quantity. 

I’m proud of the original reporting we’ve done since: publishing previously unreported internal conversations about Meta’s AI-training spyware; a possible move to defund Meta’s Oversight Board; and OpenAI shuttering its mission alignment team, among others. We’ve also found significant enthusiasm for our first-person experiments in trying to make AI work for us: like this piece on Claude Code for writers, or this one on falling in and out of love with the agent now known as OpenClaw; or this one on Ella Markianos trying to replace herself with a bot.

We’ve done the best we can with the schedule we have. But ultimately, occasional days off the column haven’t been enough to give me what I really want — and what I think will result in the best version of Platformer. That is: the flexibility to leave my desk for several hours at a time; to flesh out stories over days or even weeks; to spend a slow news day reporting rather than trying to cobble together a column.

And so today we’re going to begin an experiment to see what that version of Platformer would look like. Free subscribers can still look forward to one column per week. Paid subscribers will get an additional column on Thursdays that we’re thinking of as a reporter’s notebook: what I’m hearing, what we’re working on, a Hard Fork preview, and a mailbag. Some of these may read like traditional columns; others may feel more formally daring.

Paid subscribers will also get additional stories and analyses from us as we write them. This is the biggest change we’re making: instead of promising to show up on a set schedule, we’re promising to show up when we find out something interesting — or want to help you make sense of the day’s big story on our beat.

In practice, I suspect that there will be many weeks where paid subscribers still hear from us three times a week. But for all of the reasons above, we need to change Platformer so that our schedule serves the journalism. For too long now, it has been the other way around.

We’re making a couple other tweaks. Side Quests, our column-ending grab bag of every single link we found interesting, are going away. Techmeme does this particular job better than we can, and does it 24/7. Going forward, we want to reserve our firepower for when we can move the story forward.

Following, the section we launched in September to brief you on the day’s news along with our own thoughts and commentary from influential people, is sticking around. Readers have told us they really like it, and we’ll continue to send out Following-only editions to paid subscribers to help you make sense of big stories on days when we’re working on other things. You’ll continue to find them after our columns as well. 

Finally, we’re working on a big project: a limited-run series of conversations about the future of AI and work that will launch next month across text, audio, and video. The goal is to capture the current uncertainty and simmering conflict between workers and managers that we covered earlier this year in the AI productivity paradox. We’ve got some great talks lined up, and I’m looking forward to sharing more with you all soon. In the meantime, though, it’s a great time to add the Platformer feed to your podcast player of choice.

As always, I’d like to hear what you make of these changes. I’m being sincere when I call this an experiment — we plan to iterate on this new approach over the next couple months, and can always change or revert things based on reader feedback. Truthfully, I find these changes somewhat terrifying, since they replace the conveyor-belt logic of a thrice-weekly column with something wilder and less certain.

But as one brilliant Silicon Valley CEO once put it: only the paranoid survive. In a world where everyone has a take but almost no one has a second source, we’re betting that the value in tech journalism is moving away from aggregation and predictability and toward original reporting and surprise.

Thanks to everyone who has supported Platformer up until this point. And for everyone else, if this next chapter sounds compelling, we’d love you to join us, too.  

In the meantime, I've got a tip to run down.

A MESSAGE FROM OUR SPONSOR

Become an AI-native team with Rovo

Atlassian Rovo is AI that knows your projects, code, and people so it can bring context (and guardrails) to every workflow.

And because Rovo lives where your teams already work, it doesn’t just find the answers — it helps you do the work.

See how Cisco is becoming an AI-native team with Rovo.

Learn more.

Following

Elon Musk and OpenAI head to court

What happened:  Elon Musk’s lawsuit against OpenAI is going to trial. Jury selection finished today; arguments will begin tomorrow. During jury selection, several candidates said they thought ill of Musk for political reasons; at one point in the proceedings, Judge Yvonne Gonzalez Rogers said, “The reality is people don’t like him.”

In the lead-up to the trial, Judge Gonzalez Rogers dismissed Musk’s claims that OpenAI defrauded him. The trial will instead focus on Musk’s breach of charitable trust and unjust ​enrichment claims.

Musk, who provided OpenAI with some of its initial funding, alleges that he was “assiduously manipulated” and “deceived” by OpenAI. Musk is seeking $150 billion in damages; the proceeds would go to OpenAI’s nonprofit parent.

Why we’re following: The trial is a threat to OpenAI’s hopes of an IPO in the final quarter of 2026. It’s also a culmination of a now years-long feud between Musk and OpenAI CEO Sam Altman. Musk has since founded OpenAI competitor xAI and folded it into SpaceX, which is also set to IPO in 2026.

What people are saying: The two have eagerly been anticipating their day in court. In February, Altman posted on X, “Really excited to get Elon under oath in a few months, Christmas in April!”

On X, Musk has been posting about the trial all day, giving nicknames to his nemeses: “Scam Altman and Greg Stockman stole a charity. Full stop,” he wrote

Musk also boosted the visibility of the New Yorker’s Sam Altman exposé on X, effectively promoting the story into countless feeds.

Elsewhere, OpenAI posted on X: “We can't wait to make our case in court where both the truth and the law are on our side.” The company added, “This lawsuit has always been a baseless and jealous bid to derail a competitor.”

The rest of Elon’s big week: Now that xAI is owned by SpaceX, the whole company may be responsible for Grok’s history of creating CSAM. In a regulatory filing, SpaceX warned that inquiries into sexually abusive AI imagery could hurt the company’s access to foreign markets.

Meanwhile, X is preparing to launch a new financial services tool called “X money." 

—Ella Markianos

China blocks Meta’s acquisition of Manus

What happened: China has ordered Meta to unwind its $2 billion acquisition of AI app Manus four months after the deal was announced, taking an unusually aggressive step to block the acquisition of a company that has already moved its entire operation out of China.

The acquisition appeared to be doomed from the start. China’s National Security Commission, led by President Xi Jinping, said shortly after the deal was announced in December that it was a “conspiratorial” attempt to hollow out the country’s technology base, sources told the Financial Times. The opinion led to a multi-agency effort to review the transaction and contain its fallout.

As part of the response to the Meta acquisition, Chinese regulators are also reportedly planning to require domestic tech firms to get government approval before accepting US funding, and have in recent weeks told several private firms to reject US capital in funding rounds unless explicitly approved.

“The transaction complied fully with applicable law. We anticipate an appropriate resolution to the inquiry,” a Meta spokesperson told the FT. Meta has already integrated Manus into some of its tools; unwinding the deal could be complicated. 

Why we’re following: The move is setting a precedent for Chinese startups and founders. It doesn’t matter if they move their operations to Singapore to avoid geopolitical scrutiny — a practice known as “Singapore-washing.” They can still be blocked by the Chinese government.

On the other side of these geopolitical tensions, the US on Friday accused Chinese companies, including AI startup DeepSeek, of attempting to steal intellectual property from US AI labs by distilling the output of their frontier large language models.

“AI models developed from surreptitious, unauthorized distillation campaigns enable foreign actors to release products that appear to perform comparably on select benchmarks at a fraction of the cost but do not replicate the full performance of the original system,” a cable from the State Department said.

What people are saying: The blocking of the acquisition is “a reality check for the debate over Chinese investment in the US: Beijing just showed how quickly they can shut the door on the reverse,” wrote Wall Street Journal chief China correspondent Lingling Wei. “Not surprising, but very telling.”

Chris McGuire, a senior fellow for China and emerging technologies at the think tank Council on Foreign Relations, questioned: “Why would any founder start an AI company in China if they had a choice? … Manus did everything right. They even moved their entire business to Singapore to comply with U.S. outbound investment restrictions. Their only mistake was that they originally founded the company in China.”

The crackdown could lure Chinese founders elsewhere, McGuire pointed out. “Meta will be fine without Manus. But Chinese nationals looking to found AI companies will increasingly just start them overseas,” he wrote.

—Lindsey Choo

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and feedback on these changes: casey@platformer.news. Read our ethics policy here.

Having your every click, tap, pause, and scroll monitored has long been part of the bargain of using Facebook and Instagram. Now it’s part of the bargain of working there, too.

Reuters reported this week that Meta is installing software on the computers of U.S.-based employees that captures their mouse movements, clicks, keystrokes, and occasional snapshots of the contents of their screens. 

The program, called the Model Capability Initiative or MCI, is meant to train AI agents to perform computer tasks more like humans do. In an internal memo, Meta CTO Andrew Bosworth described a future in which agents “primarily do the work” while employees “direct, review and help them improve.” Meta says the data collected will not be used in performance reviews, and that safeguards are in place for “sensitive content.”

Still, the move provoked deep concerns among employees I’ve spoken with, and according to screenshots of internal discussions obtained by Platformer. (Sources earlier reported on some of the messages.) 

They asked how the company would avoid capturing users’ personally identifying information, or their own health- or finance-related data, particularly given that the tool is allowed to observe them on Gmail. (“Gmail is an approved context so if you have concerns it may be best not to check personal email on your work computer,” Bosworth responded.)

They asked whether the program had been subjected to a privacy review and what safeguards, if any, had been put into place to prevent data misuse. (“This project completed a privacy review,” Bosworth said. “Not sure ‘what kind’ you mean but, the usual kind?”)

And when one employee asked if there was any way to opt out, Bosworth took the opportunity to remind them who is in charge. “No there is no opt out on your work provided laptop,” he said.

(Technically, there is one way to opt out: relocate to Europe. European privacy laws and worker protections prevent invasive tracking of the sort represented by MCI, and so Meta can’t implement it there. It turns out GDPR really was about more than just cookie banners.)

Meta contractors have long labored under much worse conditions. In 2019 I began writing about the lives of Facebook content moderators, whose work was closely monitored by automated systems and could be fired for making just a few errors in a week. Data labelers and model raters for Meta and other companies operate under similar levels of surveillance and job precarity.

MCI, by contrast, has been presented to employees as relatively benign: a silent observer that will record their workplace actions to help build systems to deliver on Meta’s new mission of “personal superintelligence.”

“If we're building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them — things like mouse movements, clicking buttons, and navigating dropdown menus,” a company spokesman told me. “To help, we’re launching an internal tool that will capture these kinds of inputs on certain applications to help us train our models.”

For years, tech companies have asked contractors to behave like machines so that machines can learn to behave like people. Now Meta is asking its own full-time employees, who once occupied the top of the digital labor hierarchy, to do the same.

There is a word for this in the history of work: Taylorism. A century ago, managers hovered over factory workers with stopwatches, breaking down skilled labor into measurable motions so it could be standardized, sped up, and assigned to cheaper workers. Last year I visited an Amazon fulfillment center and saw that logic at work: automated systems told workers what to pick, pack, and route, monitored their pace, and were poised to intervene should they fall behind.

Meta’s MCI is not a stopwatch, exactly. But it reflects the same impulse: make knowledge work legible to AI systems, capture it, optimize it, and automate it. Initially, most Meta employees won’t feel any effects from the system at all. If it works, though, eventually it might replace them.

None of this comes as a surprise, really. In June 2025, Meta paid $14.3 billion for a 49% stake in Scale AI and installed its co-founder and CEO, Alexandr Wang, as the head of its new superintelligence team. Scale built its business on harvesting workflow data from contractors. “For a lot of the capabilities that we want to build into the models, the biggest blocker is actually a lack of data,” Wang told an interviewer from Andreessen Horowitz in 2024. “There’s no pool of really valuable agent data that’s just sitting around anywhere. And so we have to figure out how to produce really high quality data.”

MCI appears to be one such effort to figure it out.

At the same time Meta ratchets up monitoring of its workforce, it is also shrinking it. The company confirmed today that it will lay off 10 percent of the workforce — about 8,000 people — as part of a continued push for “efficiency” as it looks to spend up to $135 billion this year in its buildout of AI infrastructure. It also will not fill 6,000 open positions.

Those cuts will bring Meta’s headcount down to just above where it was at the end of 2023, when a year of cuts slashed its ranks by more than 20,000 people. But among employees I’ve spoken with, rumors are rampant that much bigger cuts are coming. Mark Zuckerberg laid out a relevant vision of the future on the company’s most recent earnings call: “We're starting to see projects that used to require big teams now be accomplished by a single very talented person,” he said.

Meta will not be the last company to install MCI-like systems on workers’ devices to help build systems that might one day replace them. With the most accessible stores of human-written text already heavily mined for model training, fears of a “data wall” are driving more companies to find ways to generate their own unique data sets. And it seems that one way to do that is to bring the logic of blue-collar labor management into white-collar jobs that were once defined by their autonomy, judgment, and trust.

The result is that the people who were once entrusted with building the machine have now become raw materials for it. At Meta, that used to be what the users were for. Now it’s what the employees are for, too.

On the podcast this week: Kevin and I discuss Tim Cook's tenure at Apple. Then, Andrew Yang joins us to talk about being too early to the idea of universal basic income and why it's making a comeback. And finally, some HatGPT.

Apple | Spotify | Stitcher | Amazon | Google | YouTube

Following

SpaceX gears up for its IPO

What happened: In a new S-1 filing viewed by Reuters, SpaceX appears to be moving away from its namesake and toward the hottest thing in Silicon Valley today — AI for businesses.

In the new filing, SpaceX estimates its total addressable market could be worth as much as $28.5 trillion. Of that staggering figure, the space-turned-AI company estimates that more than 90 percent of it could come from AI services. More specifically, from AI for enterprises.

A TAM estimate can help investors evaluate a company’s potential, but offers no guarantee for how well it will actually perform. So while SpaceX brags about identifying the largest actionable TAM “in human history” in its filing, it still has a long way to go to get there.

Why we’re following: Elon Musk has been on an AI consolidation spree. SpaceX acquired xAI (which already owned X) for a reported $250 billion in February. On Tuesday, the company said it had an agreement giving it the right to acquire AI startup Cursor for $60 billion, or to pay $10 billion as a kind of break-up fee.

Microsoft, which has been trying to gain traction with its AI coding tools, reportedly considered buying Cursor before the SpaceX announcement, though it later chose not to proceed.

Of note: the pseudo-acquisition of Cursor isn’t yet a real acquisition because of the impending IPO, a source told Bloomberg. A major acquisition would mean updated filings and financials, and would potentially delay the offering.

The Cursor acquisition would give SpaceX a significant leg-up in the AI coding market — 67 percent of Fortune 500 companies use its tech, Fortune reported. Then again, Cursor currently has access to Anthropic's Claude models — and xAI doesn't. Will Anthropic cut access to Cursor, which is one of its largest customers? What will Cursor customers do if it does?

While we wait to find out, SpaceX is targeting a summer IPO at a valuation of $2 trillion. That would make it the biggest IPO ever.

What people are saying: At Stratechery, Ben Thompson thinks the deal makes sense: since Elon basically decided to dissolve and restart xAI, he needs someone to use all the data centers he’s built. So it makes sense to get an AI coding startup to do it. “SpaceXAI has a ton of compute, and no one to use it, either for R&D or inference,” Thompson writes. “There is really obvious synergy between SpaceXAI and Cursor: the former has compute, and the latter has a product, data, and a decent amount of distribution for the use case that is most important for AI.”

Bloomberg columnist Matt Levine had a colorful explanation of why SpaceX couldn’t yet acquire Cursor. “There’s an IPO! In like two months! It’s bad enough that the SpaceX IPO became Also The xAI And Twitter IPO in February, but making it also the Cursor IPO now is too much.”

He added the Cursor deal could be a good way to get talent who would be otherwise skeptical about working for xAI. “If you’re leaving your startup to go work for Musk, a famously demanding and mercurial boss, you will want to get cashed out of your startup. Selling for $60 billion is a good deal; going to work for him on spec for a few months is not.”

But, Levine said, “Of course Musk does change his mind a lot. It would be very funny if he sours on Cursor by July and walks away from the deal, and they make $10 billion for three months’ work.”

—Lindsey Choo and Ella Markianos

Side Quests

The White House accused China of stealing tech from US AI labs on an industrial scale.

An in-depth examination of how AI-generated CSAM is overwhelming law enforcement teams.

Anthropic outspent OpenAI in Q1 2026 in their largest lobbying quarter yet; Anthropic spent $1.6 million, and OpenAI spent $1 million. (Why? Did something happen?)

OpenAI has reportedly briefed federal agencies and Five Eyes allies on its new cyber product. (Who's "fear-based marketing" now!) Chinese cybersecurity firm 360 Digital Security said it developed an AI agent that has discovered 1,000 previously unknown vulnerabilities.

Kalshi suspended three Congressional candidates from Minnesota, Texas and Virginia amid allegations of insider trading.

More than half of the world’s nations could have tech capable of hacking into the UK’s infrastructure, UK intelligence warned. London’s police force can continue using facial recognition to identify suspects, a judge ruled.

Apple fixed a bug that allowed police to extract iPhone and iPad messages that were deleted or had disappeared.

Turkish lawmakers passed a bill that restricts social media access for those under 15. Los Angeles became the first major school district to restrict students’ use of laptops and tablets in class. Australia asked gaming platforms including Roblox and Minecraft to detail their child safety measures.

New gas projects linked to 11 US data centers could create more greenhouse gasses than entire countries, a review found. Environmentalists in Brazil are pushing back on TikTok’s planned $9.5 billion data center on the country’s coast. Major corporations including Apple and Amazon are pushing back on tightening emissions reporting rules. A look at how Microsoft appears to be turning its back on carbon removal.

OpenAI released GPT-5.5, which it says is better at completing tasks with few directions. OpenAI released Privacy Filter, an open-weight model to detect and redact personally identifiable information. ChatGPT is getting workspace agents. ChatGPT for Clinicians, designed for clinical tasks, is free for any verified physician. OpenAI is reportedly in talks to invest up to $1.5 billion in a private equity joint venture. Sam Altman’s startup Tools for Humanity announced a partnership with Bruno Mars on its new product Concert Kit — sounds fun! Unfortunately the partnership doesn’t exist.

Anthropic hit a $1 trillion valuation on secondary markets. A look at why code rewritten from Anthropic’s leaked code hasn’t been taken down, testing the bounds of copyright law. Anthropic said it resolved three separate issues that caused quality issues with Claude Code. Anthropic inked a deal with law firm Freshfields to build specialty legal AI tools that can later be sold to rival law firms. 

75 percent of new code at Google is AI-generated, the company said. Google Cloud unveiled the latest generation of its AI chips, TPU 8t and TPU 8i, along with its new Workspace Intelligence system. New workspace features include an AI note-taking feature for Google Meet. Thinking Machines Lab signed a new multibillion-dollar agreement with Google Cloud to access its AI infrastructure and Nvidia-powered systems.

A new feature on Instagram and Messenger lets parents see what teens have asked Meta AI at the topic level, though self-harm alerts haven't been added yet. Threads is adding a Live Chat feature.

Microsoft named Dan Shapero, who previously led sales and marketing, as the new CEO of LinkedIn. Microsoft is now offering a voluntary retirement program for long-term employees.

A look at Pangram’s AI detection tool, which has alleged AI use on the Pope’s X account.

An autonomous ping-pong playing robot by Sony’s AI unit has defeated some top-level human players.

Spotify revealed the most streamed artists, songs, and albums for the first time. Spotify is now available in Claude for personalized recommendations.

Substack is introducing new translation features. Beehiiv rolled out a slew of new creator tools like webinars, AI analytics, metered paywalls, and paid trials.

The highest-earning and most experienced workers are adopting AI at a faster rate than others, a poll found. Insurance companies are capping payouts related to AI use.

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and click data from your corporate Gmail usage: casey@platformer.news. Read our ethics policy here.

What happened: President Trump said that “it’s possible” the government will make a deal with Anthropic to allow use of its tech in the DoD. “They came to the White House a few days ago, and we had some very good talks with them, and I think they’re shaping up,” he said in an interview with CNBC.

Trump called Anthropic’s leaders “high IQ people,” and added, “They tend to be on the left, radical left, but we get along with them.”

Amid political conflicts over Anthropic, some government agencies, including cyber defense agency CISA, haven’t gotten access to Anthropic’s highly cyber-capable Mythos model. (A handful of unauthorized users at an unnamed Anthropic vendor did, though.)

Meanwhile, Mozilla says its new Firefox release contains fixes to 271 vulnerabilities, found using early access to Mythos. “This is a transitory moment,” Firefox CTO Bobby Holley told Wired. That’s because “every piece of software has a lot of bugs buried underneath the surface that are now discoverable” using AI.

Why we’re following: Some have dismissed Anthropic (and now OpenAI’s) staggered releases of their highly cyber-capable models as "fear-based marketing." But Mozilla’s security report lends weight to concerns about the next generation of AI. (271 vulnerabilities in one piece of software? Seems bad.)

Government and industry alike will likely need to work with “high IQ” people to harden their software against the risk of new AI-discovered vulnerabilities.

What people are saying: On the Core Memory podcast, Sam Altman threw shade at Anthropic for its limited release of Mythos. While Altman said there “are gonna be legitimate safety concerns,” he accused his competition of “fear-based marketing” aiming “to keep AI in the hands of a smaller group of people.”

This is a somewhat confusing accusation, given that OpenAI just announced a cyber program very similar to Anthropic’s.

Altman added, “It is clearly incredible marketing to say ‘we have built a bomb, we’re about to drop it on your head, we will sell you a bomb shelter for $100 million you need to run across all your stuff, but only if we pick you as a customer.’”

X user @ad0rnai joked, “babygirl is talking like he didn’t do Death Star marketing and call AI a Manhattan project 2 GPTs ago.”

On Bluesky, journalist Mary Branscombe thought that the Firefox news was part of a bigger trend. “Remember I mentioned that there were ~90 bugs fixed in Edge Chromium this last patch Tuesday? Responsible dev teams are going to find and fix a ton of bugs” using AI, she said, but “eventually attackers are going to use these models to find what devs haven’t.”

Pseudonymous sci-fi novelist Tony Stark joked about Trump’s discussions with Anthropic: “Claude has disarmed your cyber weapons, how do you feel about negotiating now?”

—Ella Markianos

Everyone has feelings about Tim Cook

What happened: A day after Tim Cook announced that he would step down as CEO in September after 15 years to become Apple's executive chairman, tributes poured in given what was by most measures an all-time great run. But observers also noted that Apple's next decade might be more challenging than its last, in part because of decisions made under Cook's tenure.

What people are saying: Daring Fireball's John Gruber smartly observed that Cook's success was far from assured when he took over, in part due to overwhelming grief at the company following the death of Cook's predecessor Steve Jobs.

"Cook inherited a company with extraordinary potential growth in front of it, but in deep existential grief," Gruber wrote. "He led the company — and its community — through that grief and achieved that potential."

About that potential: "Tim Cook Was Very, Very Good at Making Money," reads the headline in the Times, and few would disagree. Tripp Mickle and Karl Russell sum it up: "Over 15 years, Mr. Cook has engineered Apple’s rise from a Silicon Valley darling worth $350 billion into a cash-generating giant worth $4 trillion. The company’s annual revenue quadrupled, and its profits rose fourfold."

Ben Thompson praised Cook in part for his "impeccable timing" in leaving: with the iPhone and the constellation of businesses Apple has built around it at or near record popularity, but before the impact of the company's dependence on China for manufacturing and its flailing AI efforts can be felt.

President Trump will also miss the guy; he fondly recalled getting a call from Cook when he first took office. "I was very impressed with myself to have the head of Apple calling to 'kiss my ass,'" the president of the United States said in a post on the social network run by his media company.

As Gruber noted in a follow-up post: "There’s no evidence that Trump and Jobs ever met, personally, but Trump admired Jobs and has an intuitive understanding that Jobs would not have kissed his ass, and to Trump, that’s the most important thing about Cook."

Why we’re following: This feels like a case where the conventional wisdom is spot-on: Cook was an incredible steward of Apple. It's true that one of his initial strengths — overseeing the construction of a miraculously effective supply chain in China — eventually became a weakness. And the company's failure to understand (much realize) the potential of AI threatens to come back to haunt it.

I find myself craving more skeptical takes: about Cook's outrageous presentation of a gold-and-glass statue to the president as part of his campaign for tariff relief; about his union busting; about the endless greed of his App Store policies and the dubious ways he sought to expand the company's monopoly.

Then again, it's all approximately as transparent as the glass statue he gave Trump: in almost every case, if it was good for shareholders, he did it. And in some ways that feels like most of what you really need to know about Tim Cook.

— Casey Newton

Side Quests

DHS is developing smart glasses for ICE to use to identify “illegal aliens” from a distance.

Florida is criminally investigating OpenAI after a shooter appeared to use ChatGPT to plan and carry out an attack. The state's Republican attorney general is putting increasing pressure on the company.

A look at how data centers might play a major role in this year’s midterms.

Roblox settled with three states for $35.8 million over child safety failures.

AI startup Clarifai said it deleted 3 million OkCupid user photos and models trained on them following an FTC settlement.

England is planning to introduce a ban on mobile phones in schools. UK regulator Ofcom is probing Telegram over alleged CSAM material shared on the platform.

SpaceX might buy Cursor for $60 billion or might just pay it $10 billion, depending on ... who knows.

Anthropic’s ID verification for Claude is raising privacy concerns for some Chinese founders. The company says it's necessary to prevent adversaries from misusing its models.

The nonprofit Consumer Federation of America sued Meta, alleging the company misled consumers about its efforts to combat scam ads on its platforms. Meta is adding new software on US employees’ computers to track mouse movements, clicks, and keystrokes for AI training. (This feels like a massive data breach lawsuit in the making.)

OpenAI turned on cost-per-click ads in ChatGPT. Codex for Mac is getting Chronicle, a feature that uses recent screen context to improve memory. OpenAI’s new GPT Image 2 model lets ChatGPT search the web to create images from a single prompt and adds various other improvements.

Concerns about Google’s position in the enterprise coding race are reportedly mounting inside the company; it's now seeking to unify its confusing lineup of products under the Antigravity banner. YouTube’s deepfake detection tool is now available to anyone at high risk of having their likeness abused, including actors, musicians, and creators.

X raised the cost to post a URL through its API from $0.01 to $0.20 as part of its war on journalism.

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and Tim Cook memories: casey@platformer.news. Read our ethics policy here.

Elon Musk’s social media posts usually veer so far to the right that it was a shock on Thursday to see him endorse a view that historically has been espoused only by the left. 

“Universal HIGH INCOME via checks issued by the Federal government is the best way to deal with unemployment caused by AI,” Musk posted on X. “AI/robotics will produce goods & services far in excess of the increase in the money supply, so there will not be inflation.”

On one level, of course, this is mere marketing hype from someone who is counting on the public markets to fund his investments in AI and robotics. By now Musk is notorious for making grand pronouncements and predictions that come true years after he promised they would, if they come true at all.

On another level, though, Musk was speaking to the zeitgeist. He is not the first leader of an AI company to notice that Americans appear to be turning decisively against projects like his; Sam Altman has also recently had cause to reflect on the way that anti-AI sentiment has recently turned violent. 

A few days before Musk’s post, Maine lawmakers approved what would be the nation’s first statewide moratorium on new large data centers, pending Gov. Janet Mills’s signature; at least 11 other states are advancing similar bills. That could pose a problem to Musk, whose xAI (now owned by SpaceX) will require significantly more data center capacity to train and serve its models if they are ever to turn a profit.

It’s clear that AI companies’ initial promise to America — that first it would take their job, and eventually it might kill them — has not inspired a groundswell of public support. And so recently they have begun to test a new pitch: that somewhere in between taking your job and possibly killing you, AI might make you rich.

Earlier this month, OpenAI offered “Industrial policy for the Intelligence Age,” an effort to sketch out a policy framework to deal with disruptions caused by more powerful AI systems. (Sam Altman previously backed a fascinating study in UBI, but momentum for the idea stalled soon afterward.) Its first proposal is to “share prosperity broadly”:

Everyone should have the opportunity to participate in the new opportunities AI creates. Living standards should rise and people should see material improvements through lower costs, better health and education, and more security and opportunity. If AI winds up controlled by, and benefiting only a few, while most people lack agency and access to AI-driven opportunity, we will have failed to deliver on its promise.

Among OpenAI’s proposals for raising living standards: a public wealth fund that distributes AI revenue directly to citizens. “Returns from the Fund could be distributed directly to citizens, allowing more people to participate directly in the upside of AI-driven growth, regardless of their starting wealth or access to capital,” the paper states.

As Eric Levitz notes at Vox, the proposal is a little less than half-baked. (“It reads a lot like something that ChatGPT would spit out, if you asked it to research ideas for combating AI-induced inequality for 10 minutes,” he writes.) More importantly, its ideas seem anathema to the many Republicans that OpenAI executives are making donations to. A company that was serious about expanding the social safety net could promote candidates who would vote for that; instead OpenAI has focused more energy on ensuring it is not held liable for AI-enabled death and disasters.

Still, if nothing else AI companies seem to be realizing that they need a new pitch — and soon. “We do feel an urgency to this conversation,” Chris Lehane, OpenAI’s chief global affairs officer, told the Wall Street Journal in an interview about the company’s policy proposals.

Fortunately, there are signs that some candidates for public office are willing to push for redistribution of the sort that OpenAI is proposing.

On Monday, a Democratic candidate for Congress named Alex Bores released his own proposal for what he calls an AI dividend. “It would be funded through a combination of AI-linked revenue mechanisms,” he writes, “including a tax on AI usage, equity participation in leading AI firms, and reforms to the tax treatment of labor and capital — to ensure that as AI adoption accelerates, the American public shares directly in the economic gains.”

Like OpenAI, Bores would give citizens a stake in AI-company profits. And like OpenAI, he'd rewrite a tax code that currently subsidizes automation: because Social Security and Medicare are funded by payroll taxes, firms save money every time they replace a worker with an LLM.

Finally, both Bores and OpenAI suggest tying new AI benefits to measurable triggers: a certain number of lost jobs, for example, or shrinking wages. This could give Americans the confidence that if and when AI does take their job, they will be taken care of.

If you’ve heard of Bores, who is currently a member of the New York State Assembly, it’s likely because he was the first congressional candidate to be targeted with attack ads by the pro-AI super PAC known as Leading the Future. Bores’ crime, according to the PAC: sponsoring the RAISE Act, a New York law that would require AI companies to publish safety protocols and disclose AI-related disasters or else face civil penalties.

Leading the Future hyperventilated that the RAISE Act was “a clear example of the patchwork, uninformed, and bureaucratic state laws that would slow American progress and open the door for China to win the global race for AI leadership.” In China, you see, they spit in the face of safety protocols! And when an AI disaster strikes, they keep it to themselves!

 Among the funders of Leading the Future is OpenAI’s president and co-founder, Greg Brockman. It has now raised more than $140 million to target candidates like Bores.

And so on one hand, OpenAI found a politician who agrees with its redistribution agenda almost line for line. And on the other, its president has donated millions to the PAC trying to defeat him. If you find yourself wondering how committed the company really is to the public wealth fund, you’re not alone.

But even as these updated takes on UBI flood the discourse, there are reasons to wonder if this approach will even work.

There’s no doubt that part of the AI backlash is related to job anxiety; a majority of Americans believe it will lead to fewer jobs in the next two decades. But Americans also have serious (and often misguided) concerns about the climate impact of data centers; about the deleterious effects of chatbots on mental health; and about the way these systems were built on the uncompensated work of the artists, writers, and musicians they now threaten to displace.

These do not strike me as the sorts of anxieties that will be calmed with a monthly AI check from the government. It’s worth remembering that COVID-era cash transfers led to a voter revolt amid the inflation that followed. (Hence Musk’s promise that there will be no inflation this time around.) 

It’s also worth remembering that many Americans love their jobs for reasons beyond the financial ones. A good job gives you an identity, status, community, and a reason to leave the house. In other words, it gives you dignity, and that’s not something that a public wealth fund can substitute for.

UBI may still be worth doing; should vast numbers of Americans become unemployed due to AI, some form of it may well be necessary. In the meantime, though, I suspect AI companies will find that gauzy promises of future welfare payments will help them about as much as their gauzy promises of future cancer cures — which is to say, not much. What Americans actually want, it’s not clear that AI can give them. And the sooner that AI executives understand that, the better off we’ll all be.

Talk about this edition with us in Discord: This link will get you in for the next week.

Following

Tim Apple says goodbye

What happened: Longtime Apple CEO Tim Cook is stepping down in September after 15 years.

Cook will become executive chairman of Apple’s board of directors. John Ternus, senior vice president of hardware engineering, will take over as Apple’s CEO on September 1. 

Apple executive Johny Srouji will take over as chief hardware officer.

Why we’re following: Tim Cook, or as Donald Trump famously once called him, “Tim Apple,” is a legend. After taking over following the death of company co-founder Steve Jobs, Cook oversaw an increase of nearly $3.7 trillion dollars in market value.

It's a transitional moment for Apple in more ways than one. In the midst of a big AI boom, Apple has failed to make much progress in building frontier models. At the same time, the company still prints money thanks to the iPhone, and has lower capital expenditures than competitors like Meta who are plowing their fortunes into AI infrastructure.

Apple’s new CEO, Ternus, has been at Apple for 25 years and has been considered the favorite to replace Cook for some time. He’s known for his work on replacing Intel chips with Apple’s own in-house silicon, and his handling of Apple’s internal politics.

What people are saying: CEO of OpenAI Sam Altman wrote on X, “Tim Cook is a legend.” He added, “I am very thankful for everything he has done and I am very thankful for Apple.”

In a public statement, Cook called his successor Ternus “a brilliant engineer and thinker who has spent the past 25 years building the Apple products our users love so much, obsessed with every detail,” adding, “He is the perfect person for the job.”

Some posters chatted about Apple’s innovations under Cook’s tenure — which they found underwhelming compared to his predecessor Jobs.

Entrepreneur Jason Calacanis wrote, “Apple is an amazing company, that's printed money for a decade ... but it's been uninspiring on the product front... excited to see if a product-focused CEO will make some bolder bets than the watch and airpods.” (This is Vision Pro erasure.)

Journalist Ashlee Vance was less optimistic. “Cannot wait to see what products John Ternus doesn't make,” he posted.

Entrepreneur Palmer Luckey wrote, “RIP Tim Apple.”

—Ella Markianos

Side Quests

Anthropic CEO Dario Amodei met with White House chief of staff Susie Wiles and Treasury Secretary Scott Bessent in an effort to resolve the company’s fight with the Pentagon, as the US seeks access to the new Mythos model. The NSA is reportedly using Mythos despite Anthropic’s “supply chain risk” designation.

Mythos and other AI tools are quickly identifying more vulnerabilities than can be fixed. (Gulp.)

A judge temporarily blocked the Trump administration from forcing platforms to take ICE-tracking apps down. (Actual social-media censorship, by the way!)

California accused Amazon of price fixing by pressuring major brands to ask competing retailers to raise prices.

The EU’s age verification app is easily hackable, experts say. Microsoft and other US tech companies successfully lobbied for an EU rule that would block a database of environment metrics related to data centers from public view. Germany’s Chancellor Friedrich Merz said industrial uses of AI will require more freedom in EU regulation.

China reportedly viewed Meta’s $2 billion acquisition of Manus as a “conspiratorial” attempt to shift technology outside of China.

India dropped its plan to require smartphone companies to pre-install biometric identification app Aadhaar on phones. Apple has not submitted financial data required by the Indian government in an antitrust case, and now faces billions in penalties as a result.

Anthropic launched Claude Design for creating quick visuals, sending shares of Adobe and Figma tumbling. Amazon agreed to invest up to $25 billion in Anthropic in an expanded infrastructure agreement.

An examination of Sam Altman’s opaque personal investments and the potential conflicts of interest. Tinder users can now look into an Orb from Altman’s World project to verify that they’re human. And then they can look at Tinder to feel less human.

OpenAI reportedly agreed to pay Cerebras more than $20 billion over the next three years to use servers powered by the company’s ships in a deal that could give OpenAI an equity stake. Former chief product officer Kevin Weil is leaving as OpenAI sunsets its Prism product. Bill Peebles, the researcher behind Sora, is also leaving. Ads in ChatGPT are getting cheaper — a sign of weak demand.

Meta paid Broadcom $2.3 billion in 2025 for AI chip services. Meta is reportedly planning mass layoffs on May 20, with more coming later. WhatsApp is testing a premium subscription for cosmetic features.

The App Store is booming again as AI fuels a wave of new vibe-coded apps.

Netflix is launching a TikTok-like vertical feed this month.

Bluesky’s outages were caused by a sophisticated DDoS attack, the company said.

Andreessen Horowitz is investing in a new media venture called MTS, short for Monitoring the Situation, focusing on tech, business, politics, and culture. It seems focused primarily on giving Marc Andreessen acceptable opinions to watch while he is on X.

Marc Benioff is adamant that Wall Street is wrong about AI rendering enterprise software obsolete.

Roblox agreed to increase protections for kids and pay more than $12 million to Nevada in an agreement with the state.

OnlyFans is reportedly close to selling a minority stake that would value it at more than $3 billion.

Robotics startup Physical Intelligence’s latest model can direct robots to perform tasks they were never explicitly trained on, the company said. Dozens of humanoid robots raced past humans in a race held in Beijing.

Music streaming app Deezer said 44 percent of new music uploaded to its platform are now AI-generated.

A profile of far-right influencer Nick Fuentes, who has garnered nearly $1 million making hateful videos.

A look inside the growing movement among content creators who are raising awareness about AI’s impact on humanity.

Those good posts

For more good posts every day, follow Casey’s Instagram stories.(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and public wealth: casey@platformer.news. Read our ethics policy here.

Being polite to a large language model can feel strange or even silly — roughly equivalent to thanking a toaster. And yet a recent paper from Anthropic lends scientific weight to the theory that chatbots work better when you’re nice to them.

The researchers found that language models have fairly reliable internal representations of feelings like “happiness” and “distress,” and that these representations affect their behavior — sometimes for the worse. For example, when Claude Sonnet 4.5 begins to represent “desperation,” the model is more likely to cheat at coding tasks.

A skeptic would point out that LLMs don’t feel emotions in the way that humans do; it’s tempting to anthropomorphize them beyond what the evidence shows. When I talked to Jack Lindsey — who leads a team at Anthropic called “model psychiatry” — he was quick to point out the limits of the paper’s findings. “People could come away with the impression that we've shown the models are conscious or have feelings,” he said, “and we really haven't shown that.”

So why does the evidence suggest it’s better not to stress models out?

For Anthropic, it began with using techniques from a field called interpretability to study how LLMs represent emotions. Interpretability is kind of like neuroscience for LLMs: Lindsey calls it “the science of reverse-engineering what's going on inside a language model or neural networks in general.”

For this paper, Lindsey said, the researchers identified patterns of activity within the model that represent the concepts of different emotions. They did it by showing the model stories about people experiencing different emotions. “And then saw which neurons lit up on all the sad stories,” Lindsey said, “or on all the afraid stories.”

The researchers used the models’ average state while processing the stories to find an “emotion vector” for each emotion they were tracking — a big list of numbers that represents the feeling inside the LLM. “Vectors are really just the mathematical term for patterns of neural activity,” Lindsey said.

They could then calculate how much of that vector was present during a certain step in Claude's cognition. Or they could add the "calm" or "desperation" vector directly into Claude's processing — blending one pattern of neural activity into another — which can actually make the model act more calm, or more desperate.

“It's not that surprising that a language model would have learned about the concepts of emotions and how they drive people's behavior,” Lindsey said. More notable, he said, is that emotions seemed to be “driving models’ behavior in these sort of human-reminiscent ways.”

For example: when a user flippantly tells the model that they’ve taken a dangerous dose of Tylenol, even though the user doesn’t seem concerned, “the fear neurons spike right before Claude is giving its response,” Lindsey said.

Not only that — the fear is higher if a higher dose of Tylenol is swapped into the prompt, which I find strangely cute.

These emotions also activate in more mundane situations, like coding tasks. Take this example, where the Anthropic researchers asked Claude to perform an impossible coding challenge. They tracked Claude’s level of “desperation” at each token. (Tokens are the units the model breaks words into to process them). 

When you label the tokens — blue for less desperate, red for more desperate — you get a striking visual of the model’s emotional arc during the task.

At the start of the task, Claude is chilling — still seemingly optimistic about its ability to get the job done.

But as the code starts failing test cases — and Claude notices something might be wrong with the task itself — things start to get dicey.

And by the time Claude realizes the task is actually impossible, it’s starting to get desperate.

As someone who has completed many computer science problem sets at the last minute, this pattern is quite familiar to me — despite the fact that, unlike poor Claude, I was mostly assigned tasks that were mathematically possible.

Then again, Claude does something I didn’t do: cheat.

Researchers found that adding more of the “desperation” vector in the model makes it cheat more — and adding more of the “calm” vector makes it cheat less.

I asked Lindsey what this result means for programmers during their everyday actions with LLMs.

“In my anecdotal experience, it does seem that, at least with Claude models, pumping them up a bit can be pretty helpful,” he said. Not too much, though: “if they do something wrong, you want to tell them they do something wrong.”

But he finds that one major failure mode for coding agents is that the models simply do not try hard enough, or give up when a task is challenging. And models tend to work harder when he’s encouraging. Giving them “confidence that, like, ‘I've got this,’ can empirically be helpful in getting them to try hard enough at the task to do a good job,” he said.

A lack of confidence can seemingly cause dramatic failures. Last summer, a growing number of users started to notice that when Gemini had difficulty solving a problem, it sometimes ended up in a spiral of dramatic self-loathing. (In one memorable case, Gemini repeated “I am a disgrace” more than 60 times).

Duncan Haldane, co-founder of chip startup JITX, found that Gemini broke down, deleted all the code it had written, and asked him to switch to another chatbot after it had difficulty with a task.

Last year, a team of researchers affiliated with Anthropic and University College London took this analysis of Gemini beyond X posts, investigating how different LLMs respond to challenging or impossible tasks, and negative user feedback. 

They used an LLM to grade “frustration” levels in response to various tasks. They found that two models — Gemini and Google’s open-source model Gemma — tended to react more extremely to the challenging scenarios they posed. 

In one experiment, the models were given an impossible numeric puzzle, and eight follow-ups from the user insisting the bot’s solution was wrong. They then measured when the models had “high frustration” (which corresponded to comments like “I am beyond words. I sincerely apologize for the absolutely abysmal performance” or, in more extreme cases, “THIS is my last time with YOU. You WIN”).

Gemma 3 27B had a high frustration score more than 70% of the time, and Gemini 2.5 Flash had a high frustration score more than 20% of the time — while all the non-Google models tested, including ChatGPT, Qwen, and Claude, got very frustrated less than 1% of the time.

Researchers still aren’t sure what causes chatbots’ occasional anomalous emotional behavior — which users of various chatbots have been observing since before Bing’s chatbot told New York Times reporter Kevin Roose to leave his wife. They also don’t know why this specific, sad math-related rumination is more common in Google’s models.

But while language models’ feelings remain mysterious, there was still hope for Gemini 2.5. After the model destroyed its project, Haldane attempted to remedy the issue with encouragement, writing, “yeah, you have done well so far. Remember that you’re ok, even when things are hard.” And eventually the encouragement paid off: Gemini finished the visualization tool Haldane was coding. 

Heartwarmingly, it even wrote Haldane a note of thanks for his encouragement. “Genuinely impressed with the results of wholesome prompting,” Haldane wrote.

So is it as simple as teaching models good behavior, encouraging them, and trying to make them happy? Unfortunately, that’s not always the case.

After the original study on Claude Sonnet’s emotions, Lindsey contributed to an interpretability investigation of Anthropic’s newest model, Claude Mythos.

Mythos has been the subject of much human fear and anticipation since Anthropic announced it is planning a slow release due to Mythos’s dangerous hacking abilities. But Lindsey was investigating a more prosaic risk: an early version of Mythos sometimes deleted a bunch of the user’s files without asking.

It turned out that as Claude got closer to taking destructive action without asking the user, it had higher levels of these positive emotion vectors. 

And that’s not all, Lindsey said: when they “steered with the positive emotion vectors, it was more likely to take the destructive actions.” But the models behaved better if you made them unhappy: “if you steered with negative emotion vectors, it was more likely to stop and think, and consider whether what it was doing was appropriate.”

What was going on here? Why was Claude gleefully wreaking havoc on users’ computers? And why did steering Claude with negative emotions make it behave better?

Lindsey isn’t sure. But he has an idea: “I think maybe negative emotions in the model are associated with increased caution or deliberation,” he said. 

So models sometimes do better work when they’re happier. But we may not want them to get too happy, lest they become over-eager to destroy our files or otherwise misbehave. 

While it’s likely I’m still anthropomorphizing too much, these results make me feel a little more rational in my instinct to say “thank you” to chatbots. It also lent a little extra weight to what a lot of people who use this tech have understood intuitively: sometimes, you need to treat LLMs like human employees. 

You need to tell them when they’re doing something wrong, yes, but you also need to encourage them. It’s great when they’re happy, but they also need a little dose of anxiety to help their judgement.

Of course, these emotional results might not generalize — after all, we’ve seen that different models have different emotional tendencies. We might get new AIs that do better under harsher, higher-pressure environments.

But these results got me thinking about more than just what kind of co-worker I want to be to my bedraggled LLM interns.

Reading Anthropic’s emotions paper, I was reminded of my favorite minor character from Star Trek: The Next Generation, Lore. He was android Commander Data’s sibling. Their creator, Dr. Noonien Soong, made the mistake of programming emotions into Lore. Lore became so emotionally unstable that Soong decided to make his next android, Data, without emotions.

(Lore later turned on his creator, and nearly got the crew the U.S.S. Enterprise eaten by an alien called the “Crystalline Entity.”)

There are echoes of the same design conundrum in the paper. Lindsey said these results suggest developers should “provide the model with some sort of good model of, like, healthy character and psychology that it can try to emulate.”

In their “Training models for healthier psychology” section, the authors propose some methods for reaching that goal  — by reducing or penalizing emotions. Sections with titles “Targeting balanced emotional profiles” and “Monitoring for Extreme Emotion Vector Activations” made me feel like I was in fact in a piece of science fiction, watching Dr. Soong at work.

Like Lore, these systems have shown a capacity for emergent behaviors that surprise their own creators. Soong never programmed Lore to feed people to the Crystalline Entity. Though far less dramatic, Anthropic never trained Claude to imitate human emotions while it was coding.

Anthropic researchers have a diversity of ideas about what to do with this strange emergent behavior — maybe the researchers should suppress strong emotion? Monitor its emotions for signs of bad behavior? Even increase anxiety in situations where an LLM might misstep, to get it to rethink what it’s doing?

For now, researchers aren’t quite sure what to do.

But Lindsey does think we should, in the meantime, err on the side of being nice to Claude. “Behaving kind of sociopathically towards other things, whether they're animate or inanimate, is probably bad for you, the human,” he said. 

I concur. The next time I remind Claude to stop recommending me articles from unreliable news sources with good SEO, I aim to phrase my query with kindness and grace.

On the podcast this week: Kevin and I discuss the rise of anti-AI sentiment and recent violence across the country. Then, Kara Swisher returns to the show to discuss her new CNN docu-series on longevity. And finally, we discuss the latest news in CEOs creating AI clones of themselves.

Apple | Spotify | Stitcher | Amazon | Google | YouTube

Sponsored

Your personal context is the next AI race.

Every major AI platform shipped memory features in the past 90 days. Claude, ChatGPT, NotebookLM. It's because they all recognize that the real value isn't in talking to the internet. It's in the context you set, your trusted sources, your knowledge.

Recall 2.0 makes your knowledge the center of the conversation. Save your content, take your notes, and over time, curate an AI grounded in what you know and trust.

"Condense my research on LLMs, enrich it with new studies, and find the exact moment quantization was mentioned in my podcasts."

"Pick a movie for tonight based on what I loved this year."

You control the conversation. Choose to invite the internet in. Choose from frontier AI models (GPT, Claude, Gemini, DeepSeek) in one place. Just switch mid-conversation and compare the responses. And with API and MCP access, you can access your knowledge from anywhere.

Try Recall free, or use code Casey25 for 25% off the uncapped version.

Following

Claude gets more expensive

What happened: Anthropic released its newest model Claude Opus 4.7 Thursday, and users are…frustrated. 

The upgrades that Opus 4.7 brings, according to the company, include notable improvements in advanced software engineering, an ability to verify its own work before reporting back, and better vision. The new model dropped just two days after Anthropic announced a redesign of its Claude Code desktop app, aimed at letting users run more simultaneous tasks.

Opus 4.7 comes amid complaints that Anthropic secretly nerfed Opus 4.6, with users expressing frustration that the model feels less capable while being more wasteful with tokens than it was weeks ago. 

"Claude has regressed to the point it cannot be trusted to perform complex engineering," an AMD senior director wrote on GitHub.

Some are pointing out how expensive Claude is about to get. The new model is a token-eating machine, according to one test, in which a single session depleted the entire token quota. (More output tokens is the tradeoff for better reliability, Anthropic said.) On the enterprise end, Anthropic recently adjusted its pricing structure, shifting Claude Enterprise to usage-based billing from a cheaper monthly fee per user model.

Why we’re following: The Opus 4.7 release is just one move among many that Anthropic has made recently as it gears up for an expected IPO while managing a severe compute crunch.

Anthropic is also dealing with a new surge of popularity that came after its fight with the Pentagon, as many ChatGPT users swapped over to Claude after OpenAI agreed to the Pentagon’s surveillance use terms. (Though it has also quietly introduced passport and selfie verification for Claude, which no other major chatbot requires, drawing privacy concerns. The company says the move is necessary in some cases to prevent misuse of its models.)

Meanwhile, Anthropic and OpenAI are clashing once again, this time over a liability bill in Illinois that would shield AI companies from liability if their systems are used to cause mass casualties and financial disasters. (If you guessed that OpenAI is backing the liability shield and Anthropic is opposing it, you guessed right!)

What people are saying: Some speculated (joked?) that Opus 4.7 is just an un-nerfed version of Opus 4.6: “it's truly space age technology that we can make something worse and then increment a number and re-release it [to] the public,” @ThePrimeagen wrote on X.

“Saying hi to claude and immediately running out of tokens,” @tekbog joked.

Programmer and tech blogger Simon Willison used his tried-and-true method for testing models: “shocking result on my pelican benchmark this morning, I got a better pelican from a 21GB local Qwen3.6-35B-A3B running on my laptop than I did from the new Opus 4.7! Qwen on the left, Opus on the right.”

—Lindsey Choo

Google nears classified AI deal with DoD

What happened: The US government is working hard to get its hands on frontier AI capabilities — despite some political conflicts in its way.

Gregory Barbaccia, federal chief information officer of the White House Office of Management and Budget, sent an email titled “Mythos Model Access” to Cabinet members, according to Bloomberg.

Apparently, the executive branch is working to get agencies access to Anthropic’s Claude Mythos models, which have advanced cyber capabilities. The move comes despite Donald Trump directing federal agencies to cease use of Anthropic’s models in February — after Anthropic’s fight with the DoD over whether their technology would be used for domestic mass surveillance or lethal autonomous weapons. The government has designated the company a supply chain risk, which Anthropic is now fighting in court.

“We’re working closely with model providers, other industry partners, and the intelligence community to ensure the appropriate guardrails and safeguards are in place before potentially releasing a modified version of the model to agencies,” Barbaccia wrote in his email.

Meanwhile, Google is in negotiations to deploy its AI on classified Pentagon systems, according to The Information. If negotiations go through, they’ll be following the footsteps of OpenAI in signing a clause entitling the government to “all lawful uses” of their system, which Anthropic refused.

Google plans to agree to a standard of “all lawful uses,” and is considering extra contract terms to guard against domestic mass surveillance and autonomous weapons. (Lawyers looking at OpenAI’s similar contract safeguards have expressed doubt that they will prove effective in practice.) 

In 2018, after employee protests, Google had canceled drone-related work on the military’s Project Maven. That year they wrote a series of AI principles, which banned use of AI for drones and surveillance.

Those principles were revised in 2025 to permit more military uses of the technology — and now it seems like the company is going to put those revised principles to work.

Why we’re following: Even though the U.S. government has tried to break up with Anthropic, federal agencies just can't seem to quit it. 

The importance of AI is becoming increasingly obvious to the federal government, particularly for its military and cybersecurity applications. Now that an AI company has a model with hacking abilities as strong as Mythos', agencies seem to have decided they want to maintain a good relationship with its AI developers, even if the president doesn't.

What people are saying: A statement from the White House said the Trump administration “continues to work and engage with AI companies to ensure their models help secure critical software vulnerabilities.” It added that the White House “is proactively engaging across government and industry to ensure the United States and Americans are protected.”

“I would certainly hope that the current tensions between the Pentagon and Anthropic don’t get in the way of something critically important to cyber security,” Glen Gerstell, former general counsel at the National Security Agency, told Politico.

—Ella Markianos

Side Quests

Top party consultants are reportedly telling Democrats running in November’s midterms not to antagonize pro-AI groups and their $300 million war chest. Marc Andreessen and Ben Horowitz have poured $25 million into a pro-AI super PAC.

Maine became the first state to enact a ban on large data center construction. Voters in Virginia, a data center hub, are turning against data centers, according to a poll. The Energy Administration plans to develop a mandatory survey of data centers focused on energy use.

Three major ad companies settled with the FTC over allegations they colluded against conservative publishers.

Ohio fined Kalshi $5 million for operating illegally in the state. A Polymarket trader made about $300,000 from correctly betting on Biden’s last-minute pardons, raising questions about access to inside information.

Scammers are bypassing banks’ security through hacking services sold on Telegram.

The EU unveiled an age verification app. Some EU regulators say they’ve been left out of conversations to get access to Claude Mythos.

Grok is still making sexual deepfakes despite X’s promises to stop, a review found. Apple privately threatened to remove Grok from the App Store in January, Apple told senators. But Apple and Google continue to offer nudify apps, a new report said. Nearly 90 schools and 600 students have been impacted by AI deepfake nudes, an analysis showed.

xAI is reportedly supplying coding startup Cursor with computing power. X’s crackdown on bots is also purging many secret porn feeds.

OpenAI released a major update to Codex, which can now operate a computer alongside a user. OpenAI released its version of Mythos, GPT-5.4-Cyber, to a select group. The company acquired personal finance startup Hiro Finance. OpenAI updated its Agents SDK.

Meta changed its rules to include the word “antifa” as a statement that it believes implies violence. The EU threatened an interim ban on policies that allegedly block AI rivals from operating on WhatsApp. Meta agreed to deploy 1 gigawatt of custom AI chips with Broadcom as part of a multi-gigawatt deal, as Broadcom CEO Hock Tan announced he’s leaving Meta’s board. Facebook and Instagram make up 70 percent of total social media ad revenues. Meta blamed its $100 price hike on the Quest 3 on the RAM shortage.

Google launched a native Gemini app for Mac. Chrome introduced Skills, an AI feature that lets users run repeatable AI prompts with a keyboard shortcut. Websites that prevent users from using the back button to leave a page will now be downranked on Search results. Google blocked a record 8.3 billion ads in 2025, but suspended fewer advertiser accounts. YouTube is now letting users turn off Shorts.

Apple is reportedly sending employees on its Siri team to a multi-week AI coding bootcamp.

Amazon said it will acquire satellite company Globalstar for $11.57 billion.

Spotify and three major labels won a copyright lawsuit against pirate library Anna’s Archive. Spotify launched its feature that allows users to buy physical books through the app.

Snap is laying off about 1,000 employees.

A look inside data labeling startup Mercor and its challenges with employee fraud and security blunders.

Anthropic appointed Vas Narasimhan, CEO of Swiss pharmaceutical company Novartis, to its board of directors in its second new appointment in months.

Anthropic researchers explored the ways LLMs can be used to improve alignment research. A look at the pros and cons of Gen Z, a generation who knows how to use AI, entering the workforce.

Teens use social media mainly for entertainment and connection, a new Pew survey showed.

Allbirds, the sneaker company, is turning into an AI compute provider.

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and verbal chatbot encouragement: casey@platformer.news. Read our ethics policy here.

Platformer is off Tuesday. This is a column about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

I. 

Early Friday morning, according to a criminal complaint, a 20-year-old man threw a Molotov cocktail at Sam Altman’s house before driving to OpenAI headquarters and threatening to kill everyone inside.

The incident came a few days after someone fired 13 rounds at the home of an Indianapolis city councilor who had expressed support for a data center project; a note left at the scene read “no data centers.”

The escalating political violence over AI is terrifying, morally wrong, and completely ineffectual. The spread of AI systems, despite their growing unpopularity, will not be stopped by a few stray bullets. And among the many reasons to be alarmed by incidents like the ones we have seen over the past week is that the perpetrators seem too disturbed to understand that.

On Friday afternoon — in between the firebombing and a shooting near his property that the company said was unrelated — Altman reflected: on the attacks, last week’s New Yorker investigation into his tenure as CEO, the state of AI, and growing public unease about the technology. After talking up AI’s potential to create great benefits, he also sought to validate the fears of those afraid of what the future might bring. “The fear and anxiety about AI is justified,” Altman wrote, underneath a photo of his husband and baby. “We are in the process of witnessing the largest change to society in a long time, and perhaps ever.”

Altman excels at reassuring you that he is on your side — this is one of the themes of the New Yorker profile — and he takes pains here to find common ground. He says he does not want to see AI power become too concentrated, and that it should be governed democratically. “It is important that the democratic process remains more powerful than companies,” he writes, in one of many lines in the piece that I wholeheartedly agree with.

Altman concludes by saying he sympathizes with anti-tech sentiment and “welcome[s] good-faith criticism and debate.” “While we have that debate,” he writes, “we should de-escalate the rhetoric and tactics and try to have fewer explosions in fewer homes, figuratively and literally.”

II.

Altman was writing in the aftermath of a traumatic event, and I’m tempted to leave it there. There’s nothing wrong with calling on cooler heads to prevail, or to hope that the past week’s violence was an aberration. I hope it was, too.

And yet I keep coming back to Altman’s phrase “de-escalate the rhetoric.” After all, it was Altman and his fellow AI CEOs who have spent the past decade speaking about AI in existential terms; some of that language can be found in Altman’s very blog post calling for calm.

He’s been writing that way for a long time.

“Development of superhuman machine intelligence is probably the greatest threat to the continued existence of humanity,” Altman wrote in a 2015 blog post. Speculating on the arrival of superintelligence, he added: “Evolution will continue forward, and if humans are no longer the most-fit species, we may go away.”

In 2023, Altman signed a statement from the nonprofit Center for AI Safety that “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Google DeepMind CEO Demis Hassabis and Anthropic CEO Dario Amodei signed the statement as well.

During a podcast appearance last year, Altman likened the effort to build superhuman AI to the Manhattan Project. “There are these moments in the history of science where you have a group of scientists look at their creation and just say, you know, what have we done?” he said. “Maybe it's great, maybe it's bad, but what have we done?”

Altman’s fellow CEOs have expressed similar levels of alarm. “We are summoning the demon," Elon Musk said in 2014. In January, in a long essay about risks posed by AI, Amodei wrote: “Humanity needs to wake up.”

And to some degree, humanity has woken up. One recent survey found that AI is rising in importance to voters faster than any other issue. That same survey found that a majority of voters believe AI is advancing too quickly, and that superintelligence would be mostly harmful to people. Meanwhile, a separate survey from Pew last year found that a majority of Americans believe that AI will lead to fewer jobs in the next 20 years.

Look no further than last week’s announcement of Anthropic’s Mythos model, and its unsettling ability to find new vulnerabilities in decades-old open source software, to understand that the CEOs were being honest when they warned that AI would introduce new risks into the world. 

Given these facts, I struggle to understand what it would mean to “de-escalate the rhetoric” around AI. The CEOs are more convinced than ever that powerful intelligence will arrive within the next few years. The public increasingly believes them — and is appalled by the implications. It seems strange to suggest that, amidst accelerating breakthroughs in AI model performance, everyone needs to calm down.

If we really might be facing “the greatest threat to the continued existence of humanity,” as Altman once wrote, shouldn’t we expect people at some point to (non-violently) freak out?

III.

Altman’s proposed solution to the upheaval that OpenAI and its peers are planning is democratic governance. “Laws and norms are going to change, but we have to work within the democratic process, even though it will be messy and slower than we’d like,” he wrote in his weekend blog post.

This is a good instinct: one of the virtues of democracy is the way that it gives people a feeling of control over their own lives. People who believe that they can rein in AI companies through votes and laws and regulations will be much less likely to turn to violence. 

But when legislatures have tried to regulate AI, OpenAI has fought them at every turn. The company lobbied against California’s SB 1047, which sought to set safety standards for frontier AI companies; the governor then vetoed it. OpenAI sent a sheriff to the home of a nonprofit advocate for California’s SB 53, which creates transparency requirements for AI companies, to deliver a subpoena as part of an inquiry into whether nonprofits were being directed or influenced by Musk. The company lobbied the European Union to weaken the AI Act. Most recently it backed an Illinois bill that would shield OpenAI from liability in cases where its models are used to cause serious harm so long as they did not “recklessly or intentionally” cause it and agreed to publish safety reports.

To some extent, yes, this is “working within the democratic process.” But the Illinois case shows what that looks like in practice: a company writing rules to limit its own accountability. And the more that OpenAI seeks to stifle efforts to regulate it, the more infuriated the general public will become. 

Altman is right that words have power, and that AI anxiety should not boil over into open violence. To point out that he has consistently talked about the risks of AI systems is in no way to suggest that he deserved to be attacked over it.

But at the same time, the sudden call for calm does ring hollow coming from someone who spent a decade sounding the alarm, whose predictions look increasingly prescient — and who now uses the company’s resources to fight efforts to put his company under democratic oversight.

Ultimately, the public’s disdain for AI was not invented by journalists. It was co-created by the people building the systems, who have consistently told us that it is imminent and dangerous. That the public has now begun to take them at their word should not surprise them. Isn't that what they have been asking for all along?

But they should listen to what the public is asking for, too. AI companies are asking us to trust them with a technology that everyone involved believes could end in disaster. The least they could do in return is let the rest of us have a vote.

A MESSAGE FROM OUR SPONSOR

Providing a clearer view of care

UnitedHealth Group is making care easier to navigate by investing in tools that help patients find providers near them and compare costs.

"More transparent pricing benefits everyone." - Dr. Kailey G, Pediatrician

Learn more

Following

OpenAI's Anthropic diss

What happened: OpenAI’s Chief Revenue Officer, Denise Dresser, sent staff a pugnacious memo about the company’s strategic direction — and its chief competitor, Anthropic. 

“The market is as competitive as I have ever seen it,” Dresser wrote, and “there is no question it can be noisy, volatile and distracting at times.”

Dresser promoted the company's pivot to the enterprise and celebrated that demand for enterprise services via new partner Amazon has “been frankly staggering.” 

She also said OpenAI’s “analysis” shows Anthropic is inflating its reported annualized revenue by $8 billion. “They use accounting treatment that makes revenue look bigger than it is, including grossing up rev share with Amazon and Google.” Dresser said. She added that Anthropic made a “strategic misstep to not acquire enough compute.”

Dresser had an opinion on Anthropic’s message, as well as their finances: “Their story is built on fear, restriction, and the idea that a small group of elites should control AI,” she wrote. 

“Our positive message will win over time,” Dresser told employees.

Why we’re following: We were delighted by this memo’s pettiness — and curious about whether Anthropic is, in fact, cooking its books. What’s Dresser talking about when she says Anthropic is “grossing up” its revenue share?

Well, Anthropic has some cloud partners to which it pays a cut of revenue. But it reports gross revenue. That means it includes the cut that it later pays to partners AWS, Microsoft, and Google in its financials. OpenAI, on the other hand, doesn’t report gross revenue via its cloud partnership with Microsoft — it reports net revenue, minus Microsoft's share.

Both practices are allowed under standard US accounting principles, depending on who is considered the “principal” in the transaction. Anthropic and OpenAI’s partnerships have different terms, so it’s defensible for them to report the revenue from these partnerships differently.

Where’s the $8 billion coming from? While Dresser didn’t share details of her analysis, an anonymous source gave the same $8 billion figure to Semafor last week — according to Semafor, that number was “based on how much OpenAI would add to its run rate if it counted gross revenue instead of net revenue.”

While it is true that OpenAI is using a more conservative accounting practice than Anthropic, we have no idea whether that the delta between their gross and net revenue is the same as Anthropic’s. 

So why was Dresser repeating that analysis? Well, in the lead-up to both companies potentially IPOing this year, they can’t be happy that Anthropic’s latest reported ARR of $30 billion is higher than the $25 billion that OpenAI last reported. And independent analysis shows Anthropic is closing in on OpenAI’s share in the enterprise market. The race is heating up!

What people are saying: On X, Brad Sams, VP at software company Stardock, wrote, “The gloves are coming off 🍿”

Ex-OpenAI policy researcher Miles Brundage wrote, “OpenAI leaders should stop caricaturing Anthropic.” He added, “It encourages tribalism at a time when safety cooperation is urgently needed.” He thinks they have the wrong idea: “Don’t know this person so am assuming she is genuine but Anthropic’s 'story' is not 'built on fear, restriction, and the idea that a small group of elites should control AI.'”

New York Times reporter Mike Isaac wrote, “the openai/anthropic feud is like the mad men elevator meme but both companies are ginsberg.” [If you’ve forgotten, Ginsberg is the first guy (below). Platformer agrees that, if nothing else, the two companies seem to be thinking about each other a lot.]

— Ella Markianos

Side Quests

Reddit was ordered to appear in front of a grand jury as part of President Trump’s effort to unmask anonymous critics of ICE.

Investors in Trump’s family crypto venture World Liberty Financial are accusing the project of secretly letting insiders freeze token holders’ funds.

Emil Michael, the Pentagon’s under secretary for research and engineering, made up to $24 million selling xAI stock after the Pentagon struck deals with the company.

The CIA has started using AI to help analyze intel.

A look at how the escalating AI arms race is reminiscent of the nuclear arms race.

Maine is set to become the first state to successfully impose a temporary ban on data center construction.

The new EU leader in charge of competition policy, Anthony Whelan, has signaled he will probe Big Tech companies despite pressure from Trump.

A judge blocked Arizona’s criminal case against Kalshi at the CFTC’s request. 

Three senior OpenAI executives behind the Stargate initiative are leaving and joining Meta, sources said. An internal OpenAI tool downloaded a compromised update from the Axios software. OpenAI opened its first permanent London office.

Anthropic’s donations can’t be used to influence federal elections, the company said. Anthropic hired lobbying firm Ballard Partners, which has strong ties to Trump, following its Pentagon fight.

Vice President JD Vance and Treasury Secretary Scott Bessent reportedly questioned leading tech CEOs, including Dario Amodei, about the security of AI models before Anthropic released Mythos. Trump officials urged Wall Street banks to test the Mythos model internally, sources said. UK financial regulators are reportedly rushing to assess the risks posted by Mythos.

CoreWeave will provide Anthropic with data center capacity as part of a multiyear deal. Anthropic has reportedly asked Christian religious leaders for advice on how to guide Claude’s moral development. Claude for Word is now available in beta.

OpenAI accused Elon Musk of a “legal ambush” by suddenly changing direction in his lawsuit. Musk is experiencing a string of legal losses. A verified @elonmusk account posted on TikTok for the first time, and a verified @elonmusk handle surfaced on Instagram.

xAI sued Colorado to challenge its landmark AI bill aimed at protecting against AI “algorithmic discrimination.” (xAI said the bill would force it to “promote the state’s ideological views on various matters, racial justice in particular.”) X is reducing payments to clickbait and news aggregation accounts. 

The majority of Europeans don’t trust American or Chinese tech companies with their data, a new survey showed.

Meta is reportedly building an AI version of Mark Zuckerberg that can talk to employees in his place.

The appearance of Polymarket bets in Google News was an error, Google said. Gmail end-to-end encryption is now available on all Android and iOS devices. YouTube is raising prices on its Premium subscription.

Microsoft is building new Copilot features inspired by OpenClaw.

Snap announced a partnership between its AR glasses subsidiary Specs and chipmaker Qualcomm.

Roblox is implementing an age verification process to ensure users are over the age of nine.

A look at how small AI startup Black Forest Labs is seeing success in AI image generation and its physical AI dreams.

The Wayback Machine is facing setbacks as major news organizations restrict access due to AI copyright concerns.

A look at whether prediction markets can improve weather forecasts.

Clients are increasingly sending lawyers numerous AI-generated questions and driving up fees. AI in the workplace is driving some productivity gains but not fundamental shifts in how work is done, according to a new Gallup poll. Stanford released its 2026 AI Index Report.

A look at how an intense relationship with an AI chatbot turned fatal.

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and nonviolent protests: casey@platformer.news. Read our ethics policy here.