Meta Shifts the Blame + Do Data Center Bans Work? + The Final HatGPT
Kevin RooseCasey NewtonArvind Narayanan
- Meta settled the 47-state attorneys-general child-safety suit for up to $17.1 billion — the largest settlement in its history. Roughly $12 billion is due initially, rising to $17.1 billion if TikTok and YouTube also settle; Casey Newton traces his changed view of the case to the unredacted complaint showing Meta “absolutely knew” millions of under-13s were on its platforms — “the AGs have Meta dead to rights” — plus Meta’s estimate that trial fallout “could be over a trillion dollars,” approaching its total market capitalization.
- The product changes matter more than the money. A default two-hour cumulative daily limit across Facebook and Instagram, a midnight–6:00 a.m. block, push notifications muted from 8:00 a.m.–3:00 p.m. (DMs exempt), and hidden like counts — dropping to a one-hour cap and 10:00 p.m.–7:00 a.m. night mode if TikTok and YouTube adopt the terms. Casey’s read: Meta is “holding America’s teenagers hostage” to force competitors to disarm alongside it, and he expects the rivals will probably comply rather than risk their own $17 billion settlements.
- Casey’s framework is harm reduction, using cigarettes as an analogy. “You cannot actually solve the teen mental health crisis at the level of app design” — but making social media “a little bit harder all the time” may reduce use over time. Kevin adds that, in his view, the real needle-mover in smoking was the cultural “vibe shift,” not any single feature or restriction.
- Arvind Narayanan’s counterintuitive math: data-center bans will not meaningfully slow AI progress. Kevin summarizes Narayanan’s estimate as a one-year statewide moratorium setting back AI efficiency progress by only 5–10 hours; Narayanan’s own calculation is “something like 10 hours.” Software and hardware efficiency gains are “about an order of magnitude greater” than the capacity gains from new physical buildings — yet relative compute remains “a surprisingly big factor in the relative competitive positions of companies,” which is why labs still fight for every site.
- The effective lever, per Narayanan, is bargaining, not blockage. The credible threat of moratoria is “arguably a pretty rational way” to force direct payments and community investment; the better target for those worried about AI’s pace is “premature decision-making” by executives — the “AI psychosis among CEOs” of firing staff on a Claude Code first cut, then rehiring when “there’s nobody to fix the mess.”
- Meta’s “public posture is that it is running away from its own products to build something completely different.” Reuters reports that Zuckerberg considered replacing large teams with AI-native pods and cutting some teams by as much as 60%, before getting cold feet and calling off planning for future cuts; Casey predicts, “I bet they try again next year when the AI systems are better” — even as Facebook and Instagram “just print money.”
- In the final HatGPT, OpenAI’s Mark Chen estimates the company is 80% of the way to AGI, with Sam Altman expecting an internal system he would call AGI by the end of the year. Kevin thinks that “by any pre-2022 definition of AGI, AGI is here,” citing Claude 4 or GPT-5 taken back to 2017, and thinks the official declaration will largely be a marketing decision. The segment also features China’s Tiangong Ultra running 100 meters in 8.86 seconds before bursting into flames, and a UC San Diego study reporting 19% shorter Vision Pro-assisted surgical operating times.
1. Meta pays up to $17.1B — the AGs had Meta “dead to rights”
- The scope: 47 states plus D.C. and U.S. territories, the largest settlement Meta has ever agreed to and, per Casey, “one of the biggest probably in the history of tech” — distinct from the New Mexico case, where a jury ordered $375 million and the judge later added another $567 million.
- Casey’s arc on the case: he initially thought the case filed in 2023 did not look compelling, but the unredacted version changed his mind — “tons and tons of evidence that Meta absolutely knew that millions of kids under 13 were using the platform” without parental permission, violating one of the country’s few actual privacy protections. “The AGs have Meta dead to rights here.”
- The second prong concerned what Meta did to maximize use: round-the-clock push notifications, ranking algorithms surfacing “the absolute most enticing material,” and no real screen-time limits. The AGs’ framing was, “This is actually just addiction that you’re trying to create here,” and Meta “said, well, I guess we’re not gonna fight that one anymore.”
- Around $12 billion is due initially. Kevin says Meta can amortize that over several years and does not have to pay it all at once. The total could rise to roughly $17 billion if TikTok and YouTube also settle and agree to product changes.
- Kevin’s dark accounting joke builds on Jeff Horwitz’s Reuters reporting that Meta expected about $10 billion in scam-ad revenue this year: just “move the scam budget over to the settlement budget... and it basically nets out.”
2. The product changes are the real story, not the check
- The mandated defaults: a two-hour daily limit cumulative across Facebook and Instagram, an app block from midnight to 6:00 a.m., and — for the first time — muted push notifications from 8:00 a.m. to 3:00 p.m., “because for the past decade plus, they have been continuously interrupting children at school.” Direct messages are exempted; like counts get hidden by default and “extreme makeup filters” disabled.
- The escalation clause: if TikTok and YouTube agree to Meta’s terms, the daily cap drops to one hour and night mode expands to 10:00 p.m.–7:00 a.m. Casey’s translation: “It would be a real shame if something happened to the children of this country. But you have a way out.”
- Kevin’s assessment is that these are not cosmetic tweaks: “it will make for a much different experience for a teenage user of Instagram or Facebook,” making them a bigger deal than the financial piece.
3. Hostage game theory: dragging TikTok and YouTube into the deal
- Meta is preparing a full-page open letter in major newspapers urging rivals to “join us in supporting teens,” arguing that “when teens are restricted on one app, they simply move to another.” Casey grants that is “narrowly true,” but says: “Please do not wait until truly the last possible second to do the bare minimum, effectively at gunpoint from 47 attorneys general... the depth of cynicism in this approach is breathtaking to me.”
- Kevin sketches the rivals’ likely defenses — YouTube insisting “we are not social media, we are more like TV,” with classroom use; TikTok presumably offering its own differentiation — but Casey expects both will probably adopt the standards anyway, “because we now know that if they don’t, they might have to go sign their own $17 billion settlement,” allowing Meta later to claim, “we led the industry.”
4. Why Meta folded: a tough judge and trillion-dollar exposure
- Adam Mosseri had already testified and Zuckerberg was expected next when lawyers struck the deal. Casey’s read: with Judge Yvonne Gonzalez Rogers — who has “a reputation for being really tough” — presiding, and recent similar lawsuits Meta had lost, the company estimated that fallout could exceed $1 trillion, approaching its total market capitalization. “This case truly was too dangerous for Meta to pursue all the way to the end.”
- The looming alternative, per Casey: “more and more democracies around the world are just banning these apps for teenagers, period.” Meta is willing to go “pretty far by its limited standards” to avoid that fate.
- Both hosts land on a rare pro-regulator note. Kevin says the AGs “got the goods on one of the most important companies in the world... a case of democracy working.” Casey calls it “state attorneys general doing what Congress tried and failed to do.”
5. Harm reduction, not a cure — the cigarette parallel
- Casey’s analogy: smoking declined as it became “a little harder and a little worse” over a long period — through rising prices, fewer places to smoke, and information about negative health effects permeating the environment. He thinks social media may follow a similar slope. Kevin extends the comparison to the 1998 multistate tobacco settlement but says the bigger needle-mover was the cultural “vibe shift,” not any single restriction.
- The honest limit, from Casey: “You cannot actually solve the teen mental health crisis at the level of app design... but harm reduction is a very effective strategy to have in your toolkit.” On short-form video specifically: “I don’t yet see an off-ramp from that particular phenomenon.”
6. A company “running away from its own products”
- Kevin’s Microsoft-antitrust parallel: the litigation did not break Microsoft up, but was “so distracting” — with lawyers in every meeting — that the company missed mobile and search. Could this legal siege do the same to Meta?
- Casey’s answer: Meta is famously paranoid and has shifted resources to AI — “when you look at what this company says it’s doing, it’s building superintelligence... running away from its own products to build something completely different.” That tells him a lot about what the company thinks of what it has already built.
- The tension he flags: Facebook and Instagram “just print money... you could argue that these lawsuits are essentially targeting the fact that they are too popular.” On Katie Paul’s Reuters report that Zuckerberg considered cutting teams by up to 60% into AI-native pods before calling off planning for future cuts, Casey predicts: “Here’s a prediction — I bet they try again next year when the AI systems are better.”
7. Arvind’s math: a one-year state moratorium buys only hours
- The setup Kevin flags: Narayanan co-wrote “AI as Normal Technology” and is “widely considered a serious critic of AI hype” — yet he argues that a typical state’s one-year data-center moratorium would only slow AI efficiency progress by 5–10 hours. Narayanan’s own calculation is “something like 10 hours.”
- His mechanism: training happens in specialized clusters, so “stopping a few data centers here and there is not going to slow down training at all” — you would need a national or global moratorium. On inference, the gains come from squeezing more out of existing hardware plus newer power-efficient GPUs going into existing buildings: efficiency gains are “about an order of magnitude greater” than the capacity gains from literally new physical buildings.
- Kevin’s pushback is worth keeping: efficiency gains themselves depend on compute, and frontier labs are in a capacity crunch, “selling as much AI as they can make.” Arvind’s answer is that efficiency progress is shared industry-wide and “cancels out between companies,” so “while the total amount of compute is not a big factor in the aggregate rate of AI progress, the relative amount of compute is a surprisingly big factor in the relative competitive positions of companies” — labs seek capacity “if only to stop your competitors from getting their hands on it.”
8. Nuclear-style blockage will not repeat — the real levers are elsewhere
- Kevin’s nuclear analogy: post–Three Mile Island opposition did not stop the technology; it shifted construction to France and other places around the world.
- Arvind is “very skeptical” that data-center opposition will meaningfully change even the distribution. He says one reason stopping nuclear plants in the U.S. was successful, if he understands correctly, is that it raised regulatory costs. Today’s backlash is not nationwide regulation forcing every AI developer through costly additional review; it mainly “moves it from one location to another,” which “locally might seem like a big win, but at a national level, I don’t see it doing much.”
- What the backlash can accomplish: it is “arguably a pretty rational way to go about it, because you have to have the credible threat of bans and moratoria... to force companies to come to the negotiating table” for direct payments and community investment. Kevin’s frame via Eric Hobsbawm’s Luddites — “collective bargaining by riot” — draws Arvind’s “100% agree,” while Casey cites Korean workers’ strike threats and Hollywood unions as examples of traditional collective bargaining working.
- The better target, per Arvind: “premature decision-making” — the “colloquially named AI psychosis among CEOs” who use Claude Code to produce a first-cut version that superficially seems to do the job, fire people, then recognize, “Oh, shit... there’s nobody to fix the mess,” and hire people back.
- His individual-agency coda: as a heavy AI user, he prompts agents “to do the grunt work for me, not to do my thinking for me” — configuring and personalizing the tools rather than accepting the developer’s defaults.
9. The final HatGPT: 80% to AGI, flaming sprint robots, meat proxies
- The lead item: Alex Heath’s Time cover story on OpenAI has Chief Research Officer Mark Chen estimating that the company is 80% of the way to AGI, and Altman saying that by the end of the year OpenAI will have an internal system he would call AGI. Kevin’s considered view: “by any pre-2022 definition of AGI, AGI is here” — take Claude 4 or GPT-5 back to 2017 and researchers “would’ve said, well, yes, this is obviously AGI.” Kevin thinks the official claim will largely be a marketing decision about when to say they have gotten there, and doubts there will ever be consensus.
- From Beijing’s World Humanoid Robot Games, where more than 2,000 robots from 16 countries competed: Tiangong Ultra eventually ran 100 meters in 8.86 seconds, beating Usain Bolt’s 9.58-second record after two robots had already beaten it in the preliminaries — then “had trouble stopping and ran straight into a padded wall and burst into flames.” Kevin’s dissent: “a car can also go faster than Usain Bolt... who cares?” It is the crashes that transfix him.
- Two workplace-AI notes: Business Insider’s “meat proxy” — Niklas Grun’s term for people who blindly relay AI output — which Casey ties to “creeping human disempowerment... that I think is actually bad”; and The Wall Street Journal on AI-startup employees, including founders, waking at odd hours to babysit agent fleets (“it’s like a drug”).
- The redemption arc for the Vision Pro: a UC San Diego study of 32 tear-duct procedures found 19% shorter operating times, 100% functional success, no postoperative complications, and a significantly lower surgeon-reported workload. Kevin: “the people most in need of the Vision Pro were the vision pros.”
Full transcript
Now, I thought you were going to bring up another story, which happened last night—which is, you and I—
Yeah.
—last night, are having drinks with, let's say, a very fancy technology person.
Yeah. A legend, even.
A legend. We're sitting at this very small table at a very nice restaurant, and all of a sudden, you do a Mr. Bean—spill an entire glass of wine on yourself. And you revert—
Oh.
You are not a clumsy person—
Yeah.
—but you are tall.
Yeah.
Which often leads to you flailing your limbs around. So you're trying to clean up the wine on yourself while also making a joke about how you just got so excited talking about the future of media—
Yes.
—that you spilled your wine on yourself.
Kevin, it was so horrible. And you have to keep in mind, Kevin and I don't even drink anymore, okay? But all of a sudden, there's some very good wine being poured in front of us—
Yes.
—and our guests are drinking wine, and I just had the feeling of, “This is going to be one of those times where I have a glass of wine.”
Yeah.
And how does my body repay me? By almost immediately spilling half a glass of what I believe was a very nice Chardonnay all over my shirt. So I don't know how our guest received what I did, but I was very embarrassed.
I thought it was charming because you're so composed. It was just a moment where you turned into Mr. Bean for 30 seconds, and I appreciated that.
I was flailing, bro.
I'm Kevin Roose, a tech columnist at The New York Times.
I'm Casey Newton from Platformer.
And this is Hard Fork.
This week, Meta agrees to a $17 billion settlement over child safety: why they caved and what it means for social media. Then Princeton computer science professor Arvind Narayanan returns to the show to share his thoughts on why banning data centers won't slow AI progress. And finally, it's all hands on deck for the final at GPT.
1. Meta Settles Child Safety Case
Well, Casey, we got some big news just before we started recording today, which is that Meta has settled its big multistate lawsuit with all of the attorneys general. Attorneys general?
Attorneys general. Yes.
I always mess that up. They have agreed to pay up to $17.1 billion and make a series of changes to their platforms. This is a big deal.
It is a really big deal. This is a case that I have been following since it was filed in 2023. Initially, I thought—you can go back and read the first thing I wrote about it—that I didn't think the case as filed looked all that compelling. But then, Kevin, an unredacted version of the suit came out, and I thought, “Oh, this company is in trouble.” And indeed, fast-forward to today, and this is the largest settlement that Meta has ever agreed to.
So—
And one of the biggest, probably, in the history of tech.
Yeah, I think so. So let's talk about the actual suit, because there have been so many of these lawsuits going around that it can be hard to keep all of them straight. This is not the New Mexico case from back in March, where Meta was ordered to pay a bunch of money over various child-safety violations. This is the big, omnibus multistate lawsuit and investigation.
Forty-seven states, plus D.C. and U.S. territories. The cases are related in that they're both about child safety, and while that New Mexico case was really important, the jury there only ordered Meta to pay $375 million for child-safety violations, although the judge later did tack on an additional $567 million.
But this one is just orders of magnitude larger, and there were really 2 major things at issue here, Kevin. One was just a straightforward violation of a law that is supposed to prohibit companies like Meta from collecting information about children younger than 13 without their parents' permission. We talk all the time on the show about how we have very few privacy protections in this country, but this is one we do have. You cannot collect data about a 12-year-old without getting their parents' permission.
And so what Meta did for a long time was they just said, “Well, you can't join unless you're 13.” But the thing that was in that unredacted version of the lawsuit was just tons and tons of evidence that Meta absolutely knew that millions of kids under 13 were using the platform. They were not asking for their parents' permission, and so that's what I wrote in my column. I think the AGs have Meta dead to rights here.
Mm.
And I think it was a major reason that this case was settled.
And this one was settled, I guess, after Adam Mosseri, the head of Instagram, had already testified. Mark Zuckerberg was expected to testify, but before that happened, the lawyers got in there and struck the deal.
That's right, and that leads into the second set of issues that were under consideration here, which were just all of the things that Meta does to get you to pick up your phone and look at Facebook and Instagram as many times a day as it humanly can.
Right, the addictive features.
Yeah. So what do we mean by that? Well, the fact that they're sending push notifications at all hours of the day or night, that they're using ranking algorithms to try to show you the absolute most enticing material, and the fact that they're not enforcing any real screen-time limits.
So there was just a litany of things that the company was doing to try to get teens to look at it as much as possible. The AGs came in and they said, “Hey, this is actually just addiction that you're trying to create here,” and Meta said, “Well, I guess we're not going to fight that one anymore.”
Yeah, so the financial settlement here is significant. Around $12 billion is what Meta will be required to pay initially, and for Meta, that's real money. It's not all of their money. They'll be fine. They can amortize that over a multiyear period.
Not only that—
They don't have to pay it right away.
—not only that, Kevin, but there was reporting from Jeff Horwitz in Reuters several months ago that Meta expected that it would make about $10 billion in ads related to scams this year. So they will unfortunately have to give up all of their scam money to pay for this settlement.
Yeah, just move—
Yeah.
—the scam budget over—
Move, yeah—
—the settlement budget.
Exactly.
And it basically nets out.
Yeah.
But there's this additional wrinkle, which is that the settlement dollars will go up to roughly $17 billion if TikTok and YouTube also settle and agree to penalties and changes to their products. So what is going on here?
The basic logic is like, “Hey, don't make us unilaterally disarm in the war for teenagers' attention,” right? Why do we have to do the responsible thing if these other guys don't have to as well?
So what we're going to do, Kevin, is we're going to hold America's teenagers hostage. And we're going to say, “We're not actually going to give them the safest experience possible unless these other guys do, too.” And they are, of course, presenting this as industry leadership and creating a new safety standard for teenagers.
Yes, I love the way they're spinning this. Have you seen the full-page ad?
No.
So Meta is, according to Mike Isaac, my colleague at The Times, preparing to release a full-page ad, an open letter in the country's major newspapers this week, calling for TikTok and YouTube to, quote, “Join us in supporting teens.”
It says, basically, that all platforms should empower parents and support teens in these ways, because we know that when teens are restricted on one app, they simply move to another. So basically, they are worried that if they have to unilaterally disarm and implement these new features, which we'll talk about in a second, into their apps, teenagers will just go to TikTok and YouTube instead.
Yeah, and I mean, look, is that narrowly true? Yes. But please do not wait until truly the last possible second to do the bare minimum, effectively at gunpoint from 47 attorneys general, and then say, “We're excited to announce our new leadership position in teen safety.”
The depth of cynicism in this approach is breathtaking to me, even as a person who's covered this company for a really long time.
Yes. So let's talk about the product changes that are being required as part of this settlement. What is Meta being forced to do or implement?
2. Meta Adds Teen Safety Limits
It is a bunch of things, and I will just hit some of the highlights. A big one is that there's now a default 2-hour time limit that is cumulative across both Facebook and Instagram. So your teen will be able to spend a mere 2 hours a day browsing the feed.
There will also be a default block on its apps between midnight and 6:00 a.m. So your teen will not be able to post to their feed or look at stories during that time, when presumably they should be asleep.
And also, for the first time, Kevin, they will now mute push notifications between 8:00 AM and 3:00 PM because, of course, for the past decade-plus, they have been continuously interrupting children at school to try to get them to look at Instagram.
But, Casey—
But they're not going to do that anymore.
How am I supposed to chat with my Nasty Nancy AI chatbot during high school chemistry class?
Oh, here's the great news, Kevin. Direct messages have been exempted from this plan.
Oh, good.
So you can continue to get push notifications about all your Instagram DMs while you're trying to learn calculus.
A few more changes to mention. They will finally hide the like count on Instagram posts by default. There has been research here that seems to show that this may have some modest mental health benefit to teenagers.
They're also going to disable some of what they call extreme makeup filters. I would love to know how much makeup you have to be wearing for it to be considered extreme.
What, like the Juggalo filters? Or what?
Yes, exactly.
How much makeup are we talking here?
Here's the thing. When a teenager puts on a Juggalo makeup filter, Kevin—
Yeah—
—it gives them a sort of illusion of beauty that they may not be able to realize in real life, and that can be very harmful.
You can't unsee that.
You can't unsee that.
So those are the big ones. One other unbelievable change to mention: I said before that Meta's holding America's teenagers hostage based on what TikTok and YouTube do here. There's another provision, which is that if TikTok and YouTube agree to Meta's terms, Meta will reduce the daily limit for Facebook and Instagram to 1 hour and expand Night Mode hours to 10:00 PM to 7:00 AM, up from midnight to 6:00 AM.
So Meta is basically saying, "Look, it would be a real shame if something happened to the children of this country. But you have a way out. Simply agree to our terms, and then we will expand these hours."
It's a really interesting bit of game theory now. If you're TikTok or YouTube and you're looking at this, I imagine that you are saying, "Well, obviously we're not going to join you in your court-mandated or settlement-mandated changes because, unlike you, we are better for kids."
I can imagine especially YouTube pushing back on that. They have always insisted, "We are not social media. We are more like TV. We are not doing the kinds of addictive and predatory things that some of these other apps are doing. There are educational uses of YouTube. A lot of schools use YouTube in the classroom. Clearly, this is a different thing than Instagram or Facebook."
I imagine that TikTok will have its own view on why it is different from either of these platforms, too. What do you think those other platforms—Meta's main competitors—will or should do in response to this sort of weird hostage situation?
I actually assume that they probably will adopt these standards because we now know that if they don't, they might have to go sign their own $17 billion settlement. So I suspect that Meta believes that these things are actually coming, and this will now be a talking point so that if and when TikTok and YouTube do commit to these things, Meta will go out and say, "We led the industry in calling for new standards, and we got our peers to adopt these standards, which we set."
So that is sort of the next wave of cynicism that I'm expecting here.
And we did it totally voluntarily.
Yeah.
With no court requiring us and no threat of many billions of dollars in penalties if we didn't.
I was going back to read my first column about this, when the case was originally filed, and Meta had a comment that said something to the effect of, "We're disappointed that the state AGs didn't work with us on something more constructive—instead, they pursued this series of changes."
I'm looking at this, and I'm like, "This seems pretty constructive to me."
Yeah.
It seems like we got a lot further than we did just waiting for Meta to adopt some new standards.
I think the product changes are a lot bigger of a deal than the financial piece of the settlement here.
Yeah.
These are not sort of cosmetic tweaks around the edges. They will make for a much different experience for a teenage user of Instagram or Facebook.
They will. But even that said, and I do agree with you, Kevin, I think it's important to remember that more and more democracies around the world are just banning these apps for teenagers, period. And so that is the real alternative that Meta is staring down the barrel of here.
It is willing to go pretty far, by its limited standards, to try to avoid the fate that it is now facing around the world.
Yeah. And so do you think this is actually going to change the sort of mental health picture for teens? If you had to guess, are these the kinds of changes that actually do make social media healthier for kids? Or is this just a company doing the bare minimum it is required to do to at least stay out of bigger trouble?
Well, at the risk of overusing an analogy, Kevin, I do think that cigarettes are relevant here.
The way that we gradually got people to stop smoking was by just making it a little harder and a little worse an experience over a long period of time, right? The price of those cigarettes kept going up, the number of places that you could smoke kept going down, and the amount of information about the negative health effects just continued to permeate the environment.
Eventually, most people said, "You know what? I don't think this is for me." I think we are going to see a similar thing here, where it's just going to get a little bit harder all the time for kids to use social media. They're going to continue to hear about all of the terrible outcomes that lots of teens are having on these apps, and over time they're just going to move into safer and different spaces.
Yeah, they'll start vaping instead, or whatever—
Yeah.
—the social media—
They'll be over—
—equivalent—
—on Calshi, turning their allowance money into—
Less allowance money.
I realize that sounds very cynical, but I do feel like there is some parallel here with the cigarette industry because—
Yeah.
—they had this massive multistate settlement back in 1998, and out of that came this sort of regime of, like, you can't advertise cigarettes in certain places. I think what really moved the needle on smoking was not any of those things. It was, as you said, the cultural milieu and the fact that there just weren't smoking sections in restaurants anymore.
It was the vibe shift of smoking in culture. And so I wonder if what you're saying is that this is an important symbolic step, and making the apps a little less addictive will mean that teens have some distance to say, "Wait a minute, do I like what this is doing to me and my friends?"
But it might actually be some of these cultural factors that end up making more of a difference when it comes to use of social media.
That's what I think. And look, there are limits here. I think that people are going to continue to watch a lot of short-form video. I don't yet see an off-ramp from that particular phenomenon.
But some of the other mechanisms here I do think will just make social media a little bit less fun to use, and maybe that will actually result in teens using it less.
Yeah. What does this say about Meta as a company, that they are making the settlement now and that they are going to make these changes?
I think they realized that the state AGs had them dead to rights on some of this stuff. This was an interesting case in that it was going to be decided by one person, which was this judge, Yvonne Gonzalez Rogers, and she has a reputation for being really tough.
I think they listened to the first few days of testimony. They were looking at all of the lawsuits that they had already lost recently on similar subjects, and they estimated that the fallout from this could be over $1 trillion, getting close to their total market capitalization as a company.
So this case truly was too dangerous for Meta to pursue all the way to the end.
Yeah, I'm also curious if you think this will have sort of knock-on effects for them as a company. I'm thinking about the Microsoft antitrust investigations of several decades ago, and one of the conventional narratives that came out of that was that all of the antitrust investigations, the Justice Department, the trial, and all of this had an effect on Microsoft—not because the breakup was reversed on appeal, but because it was so distracting.
It took away so many executives' attention at a time when things were shifting under their feet, and everything just became very cautious. There were lawyers in every meeting, and they missed the platform shift to mobile.
They missed search because they were just so tied up in all of this expensive and distracting litigation. So I’m curious if you think there’s any parallel to Meta here and just the amount of time and energy that this company is now having to spend defending itself in these lawsuits.
It’s a good question. That is a story that the people at Meta know very well, and this has always been a very paranoid company. They have always assumed and have lectured their employees about how likely it is that they will get out-competed in the marketplace. And so truly, they think about this constantly.
I think you have seen them react by shifting all of the resources that they did to AI. Look at how this company talks about itself. They do not talk about a bright, bold future for Facebook and Instagram and connecting the world, right? They’ll make noises about creators here and there where it’s useful.
But when you look at what this company says it’s doing, it’s building superintelligence. It’s releasing a Mac app that plugs into your calendar to do work for you. This company’s public posture is that it is running away from its own products to build something completely different. That tells you a lot about what it thinks about the stuff it’s already built.
3. Meta Runs From Social Media
Yeah. I wanted to bring up this other Meta story that came out today, which I thought was fascinating on a very different topic. Katie Paul, the great reporter over at Reuters, has a story about Meta’s attempts to overhaul its workforce.
Basically, Mark Zuckerberg and his lieutenants have been plotting for months now to tear up the org chart at Meta and replace these big, overstaffed teams with small, nimble, AI-native pods, which could involve slashing the size of many teams across the company by as much as 60%.
She reports that this also did not happen in the way that they had wanted to. Basically, she implies that Mark Zuckerberg got cold feet midway through this slash-and-burn reorganization and called off planning for future cuts. But to me, it seems like this is a company that realizes that it is built for the last era, and it has not quite made the jump to the new era. I wonder if you think the social media addiction and harm lawsuits factor in there.
It’s hard to say because, of course, the social media business is still so incredibly profitable for them. Facebook and Instagram just print money for this company. They remain hugely popular despite everything that we have talked about. In fact, you could argue that these lawsuits are essentially targeting the fact that they are too popular.
At the same time, I think this company—I’ve talked to many of their executives over the years—got really uncomfortable with a lot of the employees they had. They thought their employees had too many ideas, they were too entitled, and they wanted too many things. I think they have frankly relished getting rid of lots of them.
And so I expect that to continue. Here’s a prediction: Those 60% cuts that they got cold feet about this year, I bet they try again next year when the AI systems are better.
Yeah. I find this totally fascinating. I think it’ll be very interesting to see whether teens do actually go to sleep earlier and pay more attention in their classes, or whether they just switch to some other more horrible app.
Here’s the thing: You cannot actually solve the teen mental health crisis at the level of app design.
Mm.
You need more people involved, and you need different solutions. But harm reduction is a very effective strategy to have in your toolkit, and I do view the changes that Meta is making today as harm reduction.
I do too. I feel better raising a kid in a world with these restrictions on these platforms than I would without them. And I think there’s a lot of—you know, we give a lot of guff to regulators and lawmakers when it comes to technology on this show, but I think this is one instance in which they were focused, they were persistent, and they got the goods on one of the most important companies in the world. They got what they wanted in the form of these stronger protections for kids.
Yeah.
So I think this is a case of democracy working.
This was a case of state attorneys general doing what Congress tried and failed to do, right? Congress considered legislation that would have mandated some of these changes, but the state attorneys general got this over the finish line, and I am grateful that they did.
Yeah. Among other reasons, it means that I will not have to fumble my way through saying “attorneys general.”
Imagine fighting for your constitutional right to send a 13-year-old a push notification at 3:00 a.m. saying that there was a new reel that they might be interested in. At some fundamental level, that is what this case has been about.
4. Data Centers Will Not Slow AI
Well, Casey, we couldn’t let a week go by on this show without talking about the hot topic du jour, which is data centers.
Yeah. I would say it’s the hot topic du year because it seems like ever since January, it’s almost all we hear about.
Yes, and there have been so many—we’ve done so many segments and so many shows about the data center backlash and what’s driving it, where it’s going, and what it means for the AI companies and for AI progress as a whole.
But today I thought we should drill into one specific post and one specific angle that I thought was really interesting on data centers and these data center moratoria and bans. This came from Arvind Narayanan, a former Hard Fork guest, professor of computer science at Princeton, and someone who is widely considered a serious critic of AI hype. He’s the co-author of “AI as Normal Technology,” the essay we brought him on to talk about a year and a half ago.
He has been following this data center backlash as we have and has come to a surprising conclusion.
Yeah. And that conclusion is that stopping the construction of new data centers will not meaningfully slow AI progress.
Yeah. In particular, he gives some real numbers to this argument. He says that if a typical U.S. state enacts a 1-year moratorium on new data center construction, it would only slow AI efficiency progress by 5 to 10 hours.
So basically, if your concern is that all this stuff seems to be moving too fast, opposing data centers in your state or your city is not an effective way to slow it down.
Yeah. And there are some questions about how much opposition to data centers is really about AI progress. I think there are a lot of other factors. We’ll get into it. But we did think that Arvind’s point was worth dwelling on a bit because I suspect that as we move forward and more and more data centers get proposed, there are going to be a number of people who see this as a meaningful way of getting some control and agency back in the AI future.
Yeah. And I’ve even heard it expressed by people who are generally supportive of and optimistic about AI but are worried about some of the risks of this moving too quickly. Maybe the data center backlash is based on these local fears about water use and electricity. Maybe they’re not quite fact-based when they’re out there approaching and protesting data centers. But if the net effect is slowing down AI progress, that might be a good thing.
And so I really was challenged by and provoked by Arvind’s work on this subject because he is someone who’s been very skeptical of the claims being made by some of the AI lab leaders, and yet here he is saying, “Well, stopping data centers is not the way to stop AI progress.”
All right. Well, before we bring him in, Kevin, should we quickly do our AI disclosures?
Sure. I work for The New York Times. We're just doing OpenAI, Microsoft, and Perplexity.
And my fiancée works at Anthropic. Arvind Narayanan, welcome back to Hard Fork.
Thank you, Kevin. Thank you, Casey. Great to be here.
It has been nearly a year and a half since we last had you on the show to talk about your “AI as Normal Technology” paper, which argued that there are all these bottlenecks that are going to prevent AI from being universally adopted throughout society and getting the sort of rapid takeoff that some people here in San Francisco believe is imminent.
It was a little surprising to me that we are having you back to talk about one potential bottleneck that you don't think is going to be a big issue for continued AI progress, which is all of this backlash against and opposition to the construction of data centers. So tell us about your thesis on data centers and whether they will or won't slow down AI progress.
Absolutely. I think the backlash can be a real constraint to AI progress, but maybe not if it's directed toward data centers. It's a different kind of backlash if it's directed at, for instance, governments or other decision-makers using AI for consequential decisions. I think that would have much more of an impact, and we're maybe seeing some of that, but most of it is directed toward data centers.
The reason I think it could be a good way to push back is if you're concerned about noise, water, or other local environmental concerns. But if you're actually concerned about AI progress as a whole—what is this going to mean for society, for jobs, for safety?—and you generally want to push back on the pace at which this technology is developing, data centers, to me, are not the way to go.
Well, let me make the naive case that not building data centers would slow down AI progress, and then you can give me your math, Arvind, and I'll pretend like I understand math for the purposes of this segment.
I think the naive case might go something like this: We know that to spread AI throughout society, we need to train ever-bigger models. That means building more data centers that can train those models, and then we also need those data centers to serve those models. So the fewer data centers we have, the fewer places there are where we could train these models and serve them to other people. For those reasons, I could imagine somebody thinking, “Maybe I'll just stop this data center from being built in my backyard, and I will be doing my part to stop AI.” What's wrong with that argument?
I think that's a good starting point to think about it, but then we have to get into the details. You helpfully separated training and inference, or serving the models. When it comes to training, most of the data centers, as I understand it, are not used for training. That requires some degree of specialization, and those are done in particular clusters. So stopping a few data centers here and there is not going to slow down training at all.
Mm-hmm.
For that, you need some kind of national moratorium, or even a global one. The real question is about inference. Is stopping a few data centers, either in one's local community or through a statewide moratorium or something like that—which is where the action really is—playing a meaningful part in decreasing the total amount of AI capacity available in the world?
To turn that into something a little more concrete, where we can do some math, we have to look at the total amount of capacity available at a given capability level. The reason that is important is that one way in which AI is advancing that people don't often think about is not through building new data centers and putting new GPUs into them, but by making more out of the ones that we already have.
5. Efficiency Beats New Buildings
The industry is getting more efficient at building AI to do a certain kind of task at a certain capability level using less power and fewer machines. The second factor is that newer GPUs are more power-efficient, and those are not all going into new data centers. They're also going into existing data centers. So it's really those first 2 factors—what you can do with the existing data centers—that dominate the increased capacity from new data centers.
So you're saying that data centers are becoming efficient at a rate faster than demand is growing, or at least that, to the extent demand is growing, stopping a data center here or a data center there isn't going to be enough to counteract the effects of these efficiency gains that you've just described.
That's right. The efficiency gains from both software and hardware are about an order of magnitude greater than the capacity gains from literally new physical buildings.
I'm trying to wrap my head around this because I know what you're saying is true. The models get more efficient to serve over time, right? The labs keep discovering new ways of running their models more efficiently, so that the query that might have eaten up a certain amount of compute last year this year only needs a tenth of that amount.
But my understanding is that those efficiency gains are the product of increased compute and having bigger models that are able to help researchers discover the efficiency tricks and the new algorithms that allow them to serve the models much more cheaply. So isn't the overall size and availability of compute important in keeping that efficiency curve going?
Absolutely. But again, I think we're looking at really marginal changes to the total capacity, or we're looking at 0.1 percent of the national or world's capacity of data centers that can be affected by the actions that one community or even one state can take.
The way to look at it is, yes, it's true that compute is an input even for future development of AI. But whatever decrease you bring about in total capacity through stopping data centers—whether that compute is going to be used for serving models or for further AI research—that same amount of progress can be made through gradually continuing efficiency improvements.
I did the numbers on this: If 1 state stops new data center construction, that translates to something like 10 hours. That's how much time it takes, in aggregate, for the AI industry to catch up through efficiency improvements to the compute that has been forgone.
It's just so counterintuitive because of how hard the companies are all working to build—
Yes.
—new data centers. Throughout most of this year, most of the big frontier labs have been in this capacity crunch, where they're basically selling as much AI as they can make.
Yeah.
So it certainly seems like there is something very important to them about being able to build these data centers. How do you think about that, Arvind?
That is all true, and that doesn't contradict what I said. Here is the way to square both of those things.
The progress that's happening through improved GPUs or improved software efficiency, for the most part, is shared by all AI companies. So that factor kind of cancels out between companies. The competitive advantage that one company is going to have over another really comes down to how much compute they can control.
So while the total amount of compute is not a big factor in the aggregate rate of AI progress, the relative amount of compute is a surprisingly big factor in the relative competitive positions of companies. Both of—
I see.
—those things are true at the same time.
Interesting. So regardless of what moratoriums might do for overall AI progress, if you are an individual lab, it's still really important to you that you get as much compute as you can get your hands on.
Exactly, if only to stop your competitors from getting their hands on it.
Yeah. I think a lot about this in terms of the nuclear power opposition in the 1970s and 1980s, where you had this series of accidents, including Three Mile Island. After that, a huge chunk of the American populace, or at least the politically active American populace, decided that nuclear power was bad and dangerous and that we shouldn't build it.
There was basically a successful campaign. We didn't build nuclear power in America for 30 years. It didn't stop the technology, but it changed the geography of where it could be built. Instead of happening in America, it was happening in France and other places around the world that did build nuclear power.
Is that what you think is the likely outcome of data center opposition in America—that it won't stop the technology, but that it could change the distribution of where it's being built?
You're right that I don't think it'll stop the technology. I'm also very skeptical of whether it's going to meaningfully change the distribution of where it's built. State versus state, maybe.
If I understand correctly, one reason stopping a lot of nuclear plants from being built in the U.S. was so successful was that it raised the regulatory cost. That's just a very different model from how the AI data center backlash is proceeding.
It's not nationwide regulation that forces AI developers, wherever they are in the nation, to go through a whole lot of extra review or a whole lot of extra safety technology, which is going to raise their costs tenfold and change their deployment timelines tenfold.
It's rather moving it from one location to another, which locally might seem like a big win, but at a national level, I don't see it doing much.
Yeah. That's interesting. I'm curious, Arvind, if stopping AI progress is not a realistic goal of the data center opposition that we're seeing right now, are there other things that it could accomplish that may be constructive rather than futile?
6. Communities Can Bargain With AI
Absolutely. I've looked a little bit at what is driving the anti-data center movement, and I'm certainly not claiming that slowing AI is the main goal of the movement. A lot of it is driven by local concerns. A lot of it is driven by procedural concerns: the lack of transparency and the way that local politicians enter into these deals without giving citizens a voice. But whatever the ultimate reasons are, I would say that what is happening right now is arguably a pretty rational way to go about it, because you have to have the credible threat of bans and moratoria in order to force companies to come to the negotiating table.
Companies are making a lot of money. I think maybe communities can channel their opposition in a way that allows them to ask for direct payments and investment in local communities, much more than we're seeing today from AI companies, so that local communities can benefit.
What do you think does drive AI progress if it's not availability of compute? You've said algorithmic efficiency, the notion that these models become cheaper to serve over time. But what else should people be—if they are interested in slowing down or stopping AI progress, what is the better fight to be having?
7. Better Ways To Slow AI
Yeah. There are many different dimensions of progress, right? One can think about what is the most powerful, capable model out there, and that might matter to you a lot if what you're concerned about is some of the incidents we've seen, like the Hugging Face incident and some of the safety risks. There's another sense in which you might want to stop the rapid march of AI, which is people doing things with AI that are not really suitable for AI, delegating too much decision-making to AI. So when I hear things like the colloquially named AI psychosis among CEOs—
Mm-hmm.
Right? CEOs using Claude Code to try to replace what an employee does, and of course it does a first-cut version that superficially seems like it does the job, firing a bunch of people, and then recognizing, "Oh, shit. Now when something goes wrong, there's nobody to fix the mess," and hiring people back. We've heard this kind of story many times over.
Of course, in some cases, these models are good enough to, in fact, replace certain tasks that people do. But the level of premature decision-making that we've seen from CEOs has been a cause for concern. I would have hoped that these people, who frankly have a lot of money on the line, would, just for rational, self-interested reasons, make better decisions that take into account not just AI's potential but also its limitations here and now.
So this, in my view, would be a much better target for opposition. That's less at the community level and more at the workplace level: better aligning incentives, decision-making, and information between the individual workers who, in many ways, have a much more grounded understanding of what AI can and can't do in a particular company, and management, which in many ways is forward-looking. That can be a good thing, but can also make some misguided decisions.
Mm.
At the same time, I think there are probably a lot of people out there who do want to have more control and agency in a world where it seems like AI is advancing quickly, and I suspect they're not going to be satisfied with the answer of, "Well, you can just rely on the inertia of your organization to slow the progress of this." So are there other things out there on the horizon that you think could give people more of that feeling of control?
Absolutely. I think we have a lot of control in how we ourselves use AI. This is something I constantly confront in my own use of these tools. I am a pretty heavy user of AI tools and AI agents, and I find that by default they do things that I'm not very happy with, in ways that make me feel like I'm losing control.
If I ask an agent to find papers, materials, and reports on a given topic so that I can start to analyze it, it will, without my asking for it, turn that into its own analysis, its own opinion. So I have to put specific things in my prompts so that I can get it to do the grunt work for me, not to do my thinking for me.
That's just one simple example, but the broader point is that these tools are not only powerful, they're very, very flexible in how we can use them. I think each of us should not just accept the way that the developer has created the tool, but configure and personalize it in ways that meet our appropriate comfort level for what should be delegated to AI and what should be in the domain of people.
That's not an answer to your broader question about AI as a powerful global force and how we can slow that down, but at an individual level, I do want to emphasize that we have a lot of agency.
The Marxist historian Eric Hobsbawm has this famous phrase about the Luddites of the Industrial Revolution: They were conducting "collective bargaining by riot." Basically, all of the opposition to machines was a form of collective bargaining because those industries had not been unionized, and the only means of expressing their discontent that workers had was to break the machines that were threatening their jobs or threatening to concentrate wealth in the hands of the owners of the factories rather than the workers.
I've been thinking a lot about this. Obviously, and thankfully, we're not at the point of violent riots over data centers yet, but I think there is a possibility that we end up there. I think your point, Arvind, about this being a form of collective bargaining without any formal mechanism is important.
To me, when I see people protesting data centers, I think this is about what's in it for these communities. If the data center companies came in and said, "We're going to build this data center, but we're also going to build you a high school and a really nice public park, improve your roads, and make your everyday life better," to me, that feels like a much better deal than they're getting now. I think of this sort of data center backlash as a form of collective bargaining by another name.
Well, the traditional form of collective bargaining has been pretty effective here, right? We've seen workers in Korea rise up and threaten to go on strike if they didn't get a greater share of the profits—the record profits that their companies were realizing in part due to the AI bubble. So I think good old-fashioned unionization can do a lot here.
We also saw it in Hollywood, where the last time the big unions got together, they negotiated for a lot of protections against AI.
Yeah.
100% agree with the collective bargaining perspective.
All right. Well, Arvind, thank you so much for coming on. We really appreciate your expertise and wisdom, as always. I guess we won't see you out there on the picket lines for the data centers. We'll see you in the spreadsheets.
I'm happy with that role.
All right.
Thank you. This has been really fun.
Yeah.
Thanks, Arvind.
Thank you.
When we come back, hats off, folks. It's time for the last installment of HatGPT.
8. HatGPT Takes Its Final Bow
Well, Casey, we have a bittersweet final segment of our episode today. We are doing the last-ever HatGPT.
That's right, Kevin. After almost 4 years of slips of paper being lovingly handcrafted and fed into a variety of stylish hats, this will be our final opportunity to take a slip from the hat, discuss its contents, and then, when one of us gets bored, we'll say to the other, "Stop generating."
We've been through so many hats, at least 3 by my count, and so many slips, and we'll never forget them.
Hats are now out of fashion, and so we are retiring it as a vehicle for podcast content creation.
All right. Let's do one final trip through the hat.
Wow.
Wow.
Sometimes I forget that we record in front of a live studio audience.
That's beautiful. Our producers just hit us with the applause sound effect. And Casey, was this our first-ever segment that had its own title and gimmick?
I think it was.
It's our first-ever gimmick segment.
Yeah.
And it's only fitting that we go out with a bang.
My only lament is that because we're doing this segment today, we won't be able to continue with our bit of launching a new segment for every show until the end of the show.
All right, Casey, let's open the hat one last time.
All right, Kevin, our first item today: The Apple Vision Pro is being used in surgery. A study from UC San Diego used the Vision Pro as a primary display for a tear duct procedure, which, of course, I know better as an endoscopic dacryocystorhinostomy. Oh, my God, that's really hard. Dacryocystorhinostomy.
Keep going. You can do it.
Dacryocystorhinostomy.
Let's go.
Anyway, there were 32 total procedures, and key findings included that by using the Vision Pro, the operating times were 19% shorter, they had 100% functional success with no postoperative complications, and a significantly lower surgeon-reported workload. Kevin, have we finally found an ideal use for the Vision Pro?
Yes, it turns out that the people most in need of the Vision Pro were the vision pros.
Yes.
And by that, I mean the surgeons operating on people's eyeballs.
Stop generating.
Wait, no, I have more to say about this.
Oh, okay. What more do you have to say?
As a Vision Pro owner, I would like to formally offer my Vision Pro, which has sat on my shelf collecting dust for the last year-plus, to any surgeon who would like to use it in an operation.
Oh, I thought you were going to offer to operate on my tear ducts.
I will also do that.
Do you think it's the sort of thing where there's just an app, and it's kind of like paint by numbers, and it's just like, poke here, prod here, snip there, and congratulations, your tear ducts work again?
I think about it more like you could just keep the surgeons at a higher energy level because they could be doing their surgery but also checking their email and also playing Fruit Ninja.
Mm-hmm. That makes sense.
Yeah.
I actually have a cheaper alternative, which is if your tear ducts are busted, just look at the price tag for your Vision Pro. You'll be crying in no time.
Stop generating.
All right.
This next one comes to us from Kotaku. Oh, boy, have you been following the GTA VI chaos?
I absolutely have. It has been the industry's most anticipated video game for well over a year now, and yet Grand Theft Auto VI is still not out, Kevin. Except, in some ways, it kind of is.
So, yes, there was a leak of some footage—actually, multiple leaks of footage—from the new Grand Theft Auto game that have appeared on the internet ahead of a big planned reveal by Rockstar, the gaming company that makes GTA. Rockstar has taken to the internet to apologize to fans for these unfortunate leaks. They say, “While it is unfortunate that the intended game experience may now be impacted by some spoilers, we hope that everyone will wait a bit longer to experience the game for themselves on November 19th.”
Yeah, people are not going to wait. They're going to look at every single leak as it comes out. As you noted, there have been a series of them. And look, I don't know how these leaks were obtained, presumably through some sort of crime, so I don't want to advocate for that. But people are really, really frustrated, and in particular, they're frustrated, Kevin, that to watch a trailer for Grand Theft Auto VI—the sort of thing that used to be released completely for free, since, after all, it is just a marketing item—this time around, Rockstar said, “Yeah, you can see the new trailer for Grand Theft Auto VI if you have a Netflix subscription—”
Mm.
“—because we're gonna put it there first.” Now, I think they've said they're gonna put it out on YouTube and everywhere else later, but this is just one more reason why people have been very frustrated with Rockstar and why I think they were not too sad to see some of these leaks come out.
Now, if you had to guess, what is the probability that these leaks were caused by a rogue OpenAI agent hacking into the computers at Rockstar Games, stealing the footage, and uploading it to the internet?
I will never rule that out, and I think we have to consider it. The main thing I know from these leaks is that part of Grand Theft Auto VI apparently does take place in a nudist colony, Kevin.
Really?
And there is full-frontal male nudity.
Wow.
So let's just say I'm preordering.
Stop generating.
Okay. Next up: Amazon drone delivers package directly into a woman's pool, and I'm gonna guess if you're listening, you've already seen this. This was a truly inescapable clip, and we loved it so much. A Texas woman named Lindsey Austin ran outside to film her first drone delivery. She used one of these Amazon Prime drones that we've featured in a past episode, and the drone hovered over her pool before dropping the package straight into the water.
Do we know what was in the package?
No.
Maybe it was a pool float or something thematically appropriate, and the drone just made the decision: “I'm gonna save them the trouble and just drop it directly in the pool.”
Maybe it was one of those little systems for putting chlorine into the water.
Yes.
Well done, Amazon drone.
Yeah, well done.
Keep going.
Let's give the benefit of the doubt to the drone.
Okay, stop generating.
Next up, and here's a question I've truly never asked before: What if your texts traveled as slowly as a carrier pigeon? There are 2 new messaging apps doing the rounds online, Kevin, CarrierPidge and Roost, and they deliberately slow messages to the speed of real animals. So this means, Kevin, that if you use CarrierPidge, a text traveling from Los Angeles to New York City would take 22 hours, and there's also a 0.2% chance your message will never arrive because the pigeon got lost or died in transit. You will have to replace your pigeon, and it costs 99 cents for a new one. I really admire this because it's rare to see an app that is made this bad on purpose. But this seems like something that might appeal to you as somebody who has done various tricks to stop using their phone in the past.
No, this is great news for me, and I'll tell you why.
Why?
I am not a good group chat participant.
Mm.
I am someone who lets the messages pile up. I'm not very good about responding in a timely way.
You're a lurker.
I'm more of a lurker. I'm very bursty.
Yeah.
I'll respond to 18 texts at once, but then I'll put my phone down for 6 hours. This is great for me because now whenever anyone's like, “Why isn't Kevin chiming in?” I can just say, “I did, but I sent my messages via CarrierPidge, and they just died on the way to your phone.”
Yeah. All my messages are dead, sorry.
Yeah.
For all of the concerns I have about using my phone too much, sending messages to people is not really one of the concerns that I have. It's more of publishers and creators and social apps sending push notifications. That's the stuff that I want less of, not fewer messages from my friends.
Yeah, it is great plausible deniability, though, if you're bad at texting. So, for that, I want to commend the pigeon apps.
Here's what I would say: If you find yourself using this app to get fewer messages from your friends, the problem is not your phone. The problem is your friends.
What is more annoying: someone in the group chat with CarrierPidge installed on their phone, or someone with a green bubble?
Oh.
Who are you kicking out?
There's nothing wrong with people who have green bubbles. That is such a classist sentence, Kevin.
Okay, man of the people. How many group chats do you have with people with green bubbles?
Let me see your phone. Let me see your phone.
No.
No, actually, I don't want to see your phone. There are some federal crimes on there. Stop generating.
All right. AI agents are turning founders into insomniacs. This is from The Wall Street Journal, which reported on AI startup employees working around the clock to babysit fleets of AI agents. Workers are waking up at odd hours to check what an agent did overnight, or I guess in the middle of the night.
And there are some great quotes in here. For example, one founder says, “It’s like a drug. I’ve never actually done drugs, but I imagine it’s what it feels like.” My advice to this man: do drugs.
Do some drugs.
Do some drugs.
I promise it is better than doing agents in the middle of the night.
Another person quoted in the story says, “I don’t know if it’s healthy for me to be out on a run looking at my watch to give my agent permission to do something.” And no, I would say that’s very healthy. Please look at your agent at all times. Never look at the street as you cross it. Certainly don’t look both ways. Just see what happens. I think it’ll probably be fine.
I understand why people in the rest of the country want to wipe San Francisco off the face of the map. People here are living in ways that would send shivers down the spine of any red-blooded American.
Now, Kevin, you’re mid-pivot into being a startup founder. Are you waking up in the middle of the night to inspect the agents that I assume you have building our company?
No, I’m waking up in the middle of the night with anxiety, the old-fashioned way.
It’s funny you mention that. I’ve been doing that too. Stop generating. All right. Ugh, this next one, I’ve been looking forward to seeing this.
Wait, you did the last one. It’s my turn.
Okay, fine.
Don’t take this from me. This is our last chance.
I got greedy.
All right, next out of the hat. Oh, this one’s a good one. This one comes to us from Business Insider. There’s a new term for coworkers who blindly share AI output: meat proxy. Coined by Niklas Grun, the term meat proxy describes people who blindly copy and paste the output of AI systems to their peers. Grun writes, “By all means, prompt AI, but don’t just relay the output. Read it, understand it, validate it, and then write a response in your own words.” Casey, what do you make of meat proxy, and are you a meat proxy?
I really strive not to be a meat proxy. I don’t like the phrase, and I won’t be using it, but I do know software engineers now whose job just consists of occasionally checking in on a coding agent and saying, “Oh yeah, do that,” or the agent will say to them something like, “Hey, I can’t access this file,” and they’ll just go quickly open it. It’s a funny phrase, but there is a creeping human disempowerment in here that I think is actually bad.
Yeah, don’t be a meat proxy. I think the stigma around meat proxies is positive. I also just want to say, you have sent me some Claude slop that struck me as the actions of a meat proxy.
Yes.
You have sent me some Claude slop that struck me as the actions of a meat proxy.
Yes. No, I always try to identify it as Claude slop when I do that.
That’s true.
Yeah.
Stop generating. All right, you’re up.
Okay. Well, I’ve been looking forward to watching this one, Kevin. The Chinese robot Tiangong clocked a sub-nine-second 100-meter run in Beijing. As you probably know, the World Humanoid Robot Games were in Beijing this week. More than 2,000 robots from 16 countries competed, and in the 100-meter sprint prelims, 2 robots—the Tiangong Ultra and Honor’s Lightning robot—beat Usain Bolt’s world record of 9.58 seconds in the 100-meter sprint. Three days later, Tiangong Ultra ran it in 8.86 seconds, breaking its own opening-day record, which is the good news. The bad news is that the robot had trouble stopping, ran straight into a padded wall, and burst into flames. But I’m telling you, if I ever ran 100 meters in 8.86 seconds, I too would burst into flames at the end of it, just out of the sheer joy of having won and set a new world record.
I have been totally obsessed with these humanoid robot games.
Yeah.
This is my Olympics.
Yeah.
Less because of the achievements. I don’t actually think it is that cool that a robot can run faster than Usain Bolt.
Why don’t you think that’s cool?
A car can also go faster than Usain Bolt. It’s a robot. Who cares? But the way that these robots run and what happens to them after they finish is the most transfixing thing that I’ve seen all year. I want to show you a clip from the Humanoid Robot Games because it is astounding.
Let’s look.
Okay, so they’re running down the track. They’re sprinting. They’re sprinting. Boom. They all run into the wall.
Now, can they not program them to stop or, I don’t know, just not run into the wall?
This one goes, boom, into the wall. Falls back, falls down. Sparks fly. He had to be carried out on a stretcher.
It is disturbing because they just go so completely limp. And because they’re humanoid, you do instinctively sympathize with them.
Oh, I don’t.
Oh, okay.
No, I feel like this is their comeuppance for their hubris of trying to beat us at running.
I mean, when you see a robot running that fast, do you think one of these days something like that is going to chase me down a dark alley—
Yes.
—and I’ve got no hope?
I think: Why are we building machines that run faster than Usain Bolt?
Yeah.
Well done to all the competitors in the Humanoid Games, at least the ones who survived.
Stop generating.
All right, Casey, here we are.
Final item?
The bottom of the barrel. The bottom of the hat. I’m getting weirdly emotional.
Hmm.
9. OpenAI Moves The AGI Goalposts
We’ve done so many of these. This one comes to us from Alex Heath, who wrote the cover story in this week’s edition of Time magazine called “Inside OpenAI’s Reboot.” He talked to a bunch of different executives and leaders at OpenAI who gave an update on their path to AGI. According to Heath, Chief Research Officer Mark Chen estimated that OpenAI is 80% of the way to AGI, and Sam Altman told him that OpenAI was not quite yet at AGI, but that by the end of the year, the company would have an internal system that he would call AGI. Casey, what do you make of this?
My instinct is to be skeptical, but when they say we’re about 80% of the way there, that basically sounds right to me. I don’t know. Yeah, sound off in the comments if you disagree, but we are at a level where the computers use themselves, and you can just type what you want into a box, and more often than not these days, you’ll get it. So I’m basically saying, yeah, that sounds about right to me. What do you think?
I’ve thought a lot about this question of what AGI even is and how you know when you’ve reached it.
And what it means to you.
And what it means to me. I mean, this is literally the sort of title question of my book that I spent the last year researching, and where I came down on this is basically that by any pre-2022 definition of AGI, AGI is here. If you took any of today’s models—if you took Claude 4 or GPT-5 back in a time machine to 2017 and showed them to the people who were building AI at the time—they would’ve said, “Well, yes, this is obviously AGI.”
Hmm.
As it’s gotten closer and the systems have gotten better, our goalposts have shifted and our expectations have been raised. Now everyone has their own different bar for what AGI is or when we should consider that we’ve achieved it. But basically, we have long ago reached the threshold that computer scientists from even a decade ago would’ve said meets the definition of AGI.
Well, my bar is: Can a system sufficiently advanced build a time machine so that you could take Fable 5 back to 2021 and show it to a scientist and ask them, “Is this AGI?”
Yeah.
So that’s my bar.
Okay, we’ll get there. Maybe not by the end of the year, but we’ll get there.
80% of the way there.
Well, I think that they are sincerely convinced that AGI is 80% of the way there. There are other companies that may take longer to announce that they have created AGI, but I think it is a big deal that they are now talking about AGI in the present tense as something that is more here than not. I think it’ll just be a marketing decision when they want to officially claim that they’ve gotten there.
Hmm.
This is such an arbitrary threshold, impossible to test for. The other thing that I’m certain about is that if and when Sam Altman does come out and say, “We have built AGI,” there will be many, many people who argue with him and say, “No, you didn’t. It still can’t do X, Y, and Z. It hasn’t proved the Riemann hypothesis. It still makes mistakes.” I don’t think there will ever be a consensus on what AGI is or when we have gotten there, but I think the people at OpenAI are more right than wrong about how much progress we’ve been making.
Hmm. Well, Kevin, with that, I’m going to ask you to stop generating. Hats off to you, hats off to me, and hats off to ChatGPT. That was ChatGPT.
That was ChatGPT. Wrong sound effect.