Intro
Princeton computer science professor Arvind Narayanan returns to argue that blocking data centers is the wrong lever for anyone who wants to slow AI down, and that the fight worth having is at the workplace instead. Kevin Roose and Casey Newton open with Meta's multi-state child-safety settlement and the product changes it forces, and close with the final edition of HatGPT.
Guest: Arvind Narayanan, professor of computer science at Princeton University and co-author of the essay "AI as Normal Technology"
Hosts: Kevin Roose (The New York Times) and Casey Newton (Platformer)
Published: 28 August 2026 on Hard Fork
Show notes | 1 hr 3 min
Key Takeaways
Stopping data centers is not how you stop AI
Narayanan: efficiency gains from software and hardware are "about an order of magnitude more than the capacity gains from literally new physical buildings"
One state's moratorium costs the industry "something like 10 hours" of progress
Compute barely moves aggregate progress but decides who wins
The relative compute position is what separates one lab from another
The data center fight is worth having for local reasons, not AI ones
Bans and moratoria are leverage to get money and infrastructure back into communities
The better target is management, not the planning board
CEOs firing staff on the strength of a coding agent's first draft, then hiring them back
Meta settles the biggest case it has ever faced
Around $12 billion up front, rising toward $17 billion if TikTok and YouTube sign on to the same terms
Meta's teen apps get hard limits for the first time
A default two-hour daily cap across Facebook and Instagram, an overnight block, and muted notifications during the school day
Meta settled because the downside was near-existential
Newton: the follow-on exposure was estimated at over a trillion dollars
Meta is running away from its own products
Newton: the company talks about superintelligence, not about Facebook and Instagram
State attorneys general got what Congress could not
OpenAI puts itself at 80% of the way to AGI, and both hosts roughly agree
Roose: "by any pre-2020 definition of AGI, AGI is here"
HatGPT retires after almost four years and three hats
A Glass of Chardonnay, and a Legend
The show opens with Roose recounting a dinner the pair had the night before with what he calls a very fancy technology person, described by Newton as "A legend, even."
Newton knocked over most of a glass of wine onto himself at a small table in a nice restaurant, then tried to clean up while joking that he had got excited talking about the future of media
Newton's own account is that neither host drinks anymore, that good wine was being poured in front of them, and that he had the feeling this would be one of the times he had a glass
Roose said the moment was charming because Newton is normally composed; Newton's summary was "I was flailing, bro."
Meta Settles the Multi-State Child Safety Case
News broke just before the recording that Meta had settled the multi-state suit brought by 47 states plus D.C. and U.S. territories, agreeing to pay up to $17.1 billion and change its products
Newton has followed the case since it was filed in 2023 and says his first column argued the suit as filed did not look compelling — an unredacted version changed his mind
This is separate from the New Mexico case in March, where a jury ordered Meta to pay $375 million over child-safety violations and the judge later added $567 million
Two issues sat at the heart of it. The first was collecting data on children under 13 without parental permission, one of the few privacy protections the U.S. actually has
The unredacted filing carried what Newton describes as tons of evidence that Meta knew millions of under-13s were on the platform and was not asking parents
The second issue was everything Meta does to get users to open the app — push notifications at all hours, ranking for the most enticing material, no real screen-time limit. The AGs called it addiction by design
Adam Mosseri had already testified and Mark Zuckerberg was expected to before the lawyers struck the deal
What the Settlement Costs, and Where the Money Comes From
Roose: around $12 billion is what Meta pays initially, and for Meta that is real money without being all of their money — it can be amortized over several years
Newton points to reporting by Jeff Horwitz at Reuters several months ago that Meta expected to make about $10 billion this year on ads related to scams, so "they will unfortunately have to give up all of their scam money to pay for this settlement"
Roose's version: move the scam budget over to the settlement budget and it basically nets out
The total rises to roughly $17 billion only if TikTok and YouTube also settle and accept penalties and product changes
The Full-Page Ad and the Hostage Play
Newton's read on the escalator clause: "don't make us unilaterally disarm in the war for teenagers' attention" — Meta will not give teens the safest experience unless its rivals do too, and is presenting that as industry leadership
Roose, citing his Times colleague Mike Isaac, says Meta is preparing a full-page open letter in the country's major newspapers calling on TikTok and YouTube to, quote, join us in supporting teens, on the argument that restricted teens simply move to another app
Newton concedes the argument is narrowly true, then: "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."
If the rivals do agree, Meta will cut the daily limit to one hour and widen night mode to 10 p.m. through 7 a.m. — hence Newton's line that "So Meta is basically saying look, it would be a real shame if something happened to the children of this country."
Roose expects YouTube in particular to refuse, on its standing position that it is closer to TV than to social media, with educational uses and classroom adoption to point at
Newton thinks both rivals adopt the standards anyway, because the alternative is signing their own settlement — which then lets Meta claim it led the industry to a new standard
The Product Changes Coming to Teen Accounts
A default two-hour daily time limit, cumulative across Facebook and Instagram
A default block on the apps between midnight and 6 a.m., so teens cannot post to their feed or look at stories overnight
Push notifications muted between 8 a.m. and 3 p.m. — Newton notes the company has spent a decade-plus interrupting children at school to get them back into Instagram
Direct messages are exempt from the notification mute, which prompted Roose's "How am I supposed to chat with my nasty Nancy AI chatbot during high school chemistry class?"
Like counts get hidden on Instagram posts by default, on research Newton says shows a modest mental health benefit
So-called extreme makeup filters get disabled, which sent both hosts off into a bit about juggalo filters and how much makeup counts as extreme
Roose's view is that the product changes matter more than the money — these are not cosmetic tweaks and will make for a very different experience for a teenage user
Whether Any of This Helps Teenagers
Newton's frame is cigarettes: people stopped smoking because it got a little harder and a little worse over a long period, as prices rose, permitted places shrank, and health information kept accumulating
He expects the same slow grind here — social media gets marginally harder to use, the bad outcomes keep getting discussed, and teens gradually move into different spaces
Roose pushes the analogy at the 1998 multi-state tobacco settlement and says the rules were not the decisive thing: "what really moved the needle on smoking was not any of those things"
What did move it, in his telling, was the cultural shift and the disappearance of smoking sections in restaurants — so the Meta changes may matter most as a symbolic step that buys teens enough distance to ask what the apps are doing to them
Newton sees limits. Short-form video keeps going, and he sees no off-ramp from it; the other mechanisms just make the apps a bit less fun
Roose's dark aside on where displaced teens go: "They'll be over on Kalshi turning their allowance money into less allowance money."
Newton: "You cannot actually solve the teen mental health crisis at the level of app design." He views the changes as harm reduction, which he calls a very effective strategy to have in the toolkit
Both hosts note that more democracies around the world are simply banning these apps for teenagers, which is the outcome Meta is actually trying to avoid
Why Meta Settled Rather Than Finish the Trial
The case was going to be decided by a single judge, Yvonne Gonzalez-Rogers, who Newton says has a reputation for being really tough
Meta listened to the first days of testimony, looked at its recent losses on similar subjects, and estimated the follow-on exposure at over a trillion dollars — close to its entire market capitalization
Newton's conclusion: the case was simply too dangerous for Meta to take all the way to the end
The Microsoft Parallel, and a Company Fleeing Its Own Products
Roose raises the Microsoft antitrust years as the cautionary case — the breakup was reversed on appeal, but the litigation was so distracting that executives' attention went elsewhere, lawyers sat in every meeting, and the company missed mobile and missed search
Newton says Meta knows that story well and has always been a paranoid company that lectures employees about how it will be outcompeted
Its answer has been to move everything to AI: "This company's public posture is that it is running away from its own products to build something completely different."
The evidence he cites is what Meta talks about — superintelligence, a Mac app that plugs into your calendar to do work for you — rather than a bold future for Facebook and Instagram
The counterweight, he adds, is that Facebook and Instagram still print money and remain hugely popular; you could argue the lawsuits are targeting the fact that they are too popular
The Reorg Zuckerberg Got Cold Feet About
Roose brings up a Reuters story by Katie Paul published that day on Meta's attempt to overhaul its workforce — tearing up the org chart and replacing large teams with small, AI-native pods, with cuts to many teams of as much as 60%
Paul reports it did not go the way Meta wanted, with Zuckerberg calling off planning for further cuts partway through
Roose reads it as a company built for the last era that has not made the jump to the new one
Newton says Meta's executives had become uncomfortable with their own staff: "They thought their employees had too many ideas. They were too entitled. They wanted too many things." He believes the company has relished getting rid of them
His prediction: "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."
State Attorneys General Did What Congress Couldn't
Roose, who notes the show usually gives regulators and lawmakers a hard time, calls this one an exception: they were focused, persistent, got the goods on one of the most important companies in the world, and came away with stronger protections for kids
His verdict: "So I think this is a case of democracy working."
Newton points out Congress had considered legislation mandating some of these changes and failed, while the state AGs got it over the line
Roose's other reason to be pleased is that he will no longer have to fumble his way through saying attorneys general, a phrase he mangles twice in the segment
Narayanan's Thesis: Data Center Bans Aren't the Lever
Roose introduces the segment around a single post by Narayanan, who returns to the show for the first time since the "AI as Normal Technology" essay roughly a year and a half ago, and who he describes as widely considered a serious critic of AI hype
The host's summary of the claim: a one-year moratorium on new data center construction in a typical U.S. state would slow AI efficiency progress by five to ten hours
Roose says he was challenged and provoked by the argument precisely because it comes from someone skeptical of the AI lab leaders' claims
Newton flags that a lot of data center opposition may not be about AI progress at all, and that more people will reach for it as a way of getting agency back
Before the interview both hosts give their AI disclosures: Roose works for the New York Times, which is suing OpenAI, Microsoft and Perplexity, and Newton's fiancé works at Anthropic
Narayanan's own framing is that the backlash could be a real constraint on AI — just not this backlash. Opposition aimed at governments and other decision makers using AI for consequential decisions would land harder
On the target people have chosen: "data centers, to me, are not the way to go" if the concern is the pace of the technology rather than local noise, water or environmental effects
Training, Inference, and Why Efficiency Beats New Buildings
Newton lays out the naive case for the guest to knock down: bigger models need more data centers to train in and more data centers to serve from, so blocking one is doing your part
Narayanan splits it. Most data centers are not used for training, which needs specialized clusters — "stopping a few data centers here and there is not going to slow down training at all." Slowing training would take a national or global moratorium
So the real question is inference, and whether local or statewide bans meaningfully reduce total AI capacity at a given capability level
Two things dominate, and neither involves new buildings. The industry keeps getting more efficient at delivering a given task at a given capability level using less power and fewer machines
Newer GPUs are also more power efficient, and they go into existing data centers as well as new ones
The punchline: "The efficiency gains, both from software and from hardware, are about an order of magnitude more than the capacity gains from literally new physical buildings."
Roose presses on whether the efficiency curve itself depends on having lots of compute — bigger models help researchers find the tricks that make serving cheaper
Narayanan agrees compute is an input, but says the affected slice is tiny: "We're looking at 0.1% of the national or world capacity of data centers that can be affected by the actions that one community or even one state can take."
His arithmetic: "If one state stops new data center construction, that translates to something like 10 hours, right? That's how much it takes aggregated for the AI industry to catch up through efficiency improvements to the compute that has been foregone."
Why Individual Labs Still Fight for Every GPU
Newton raises the obvious objection: the frontier labs have spent the year in a capacity crunch, selling as much AI as they can make, so compute plainly matters enormously to them
Narayanan says both things are true at once. Efficiency progress is largely shared across companies and cancels out between them; what does not cancel is who controls more compute
"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"
Roose restates it as: whatever moratoriums do to overall progress, an individual lab still needs all the compute it can get
Narayanan's addendum: "Exactly. If only to stop your competitors from getting their hands on that."
Why the Nuclear Analogy Doesn't Transfer
Roose offers the 1970s and 1980s nuclear power opposition as a model — after Three Mile Island and the other accidents, the U.S. built essentially no nuclear power for 30 years, which did not stop the technology but moved it to France and elsewhere
Narayanan agrees it will not stop the technology but is skeptical it even redistributes it. State versus state, maybe; nationally, he does not see it
His reason: the nuclear campaign worked by raising regulatory cost, forcing extra review and safety technology that changed developers' costs and timelines tenfold wherever they were
The data center backlash is not nationwide regulation. It moves projects from one location to another, which can look like a big local win and do very little at the national level
What the Data Center Backlash Could Actually Win
Narayanan says he is not claiming that slowing AI is the movement's main goal — a lot of it is local concerns, and a lot of it is procedural, about the lack of transparency and local politicians cutting deals without giving citizens a voice
On the tactics being rational anyway: "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, so communities could channel opposition into direct payments and local investment well beyond what AI companies are offering now
Roose reaches the same place from the other direction: when he sees people protesting data centers, he reads it as a question about what is in it for these communities. A company offering a high school, a good public park and better roads alongside the building would be a far better deal than what is on the table
The Better Target: Managers Firing People Too Early
Asked what people who want to slow AI should be fighting instead, Narayanan separates the dimensions of progress — the most capable model matters if your worry is safety incidents, but a different worry is people delegating decisions to AI that are not suitable for it
His example is what he calls "the colloquially named AI psychosis among CEOs" — executives using Claude Code to replace what an employee does, getting a first cut that superficially looks like the job, firing people, then discovering nobody is left to fix the mess and hiring them back
He allows that in some cases the models really are good enough to replace certain tasks. The problem is timing: "the level of premature decision-making that we've seen from CEOs has been a cause for concern"
He would have expected people with money on the line to make better decisions out of pure self-interest, weighing the limitations of the technology here and now alongside its potential
The fight, in his view, belongs at the workplace rather than the community — realigning incentives and information between managers, who are forward-looking but can be misguided, and individual workers, who have a far more grounded sense of what AI can and cannot do in their own company
Keeping the Agents From Doing Your Thinking
Newton objects that people wanting agency will not be satisfied with being told to rely on their organization's inertia
Narayanan, who describes himself as a heavy user of AI tools and agents, says the tools by default do things that make him feel he is losing control. Ask an agent to find papers and reports on a topic and it will volunteer its own analysis and opinion unasked
His workaround: "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."
The broader point is that these tools are not only powerful but very flexible, and people should configure and personalize them rather than accepting the developer's defaults about what gets delegated
He is explicit that this is not an answer to slowing AI as a global force, but insists that at the individual level there is a lot of agency
Collective Bargaining by Riot
Roose brings in the Marxist historian Eric Hobsbawm's phrase for the Luddites, that they were conducting "collective bargaining by riot" — machine-breaking as the only channel workers had in industries that were not unionized
He says he has been thinking about it a lot, that data center protests are not at the point of violent riots, and that he sees a possibility of ending up there
Newton points to conventional collective bargaining already working: "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." He also cites Hollywood's unions negotiating AI protections
Narayanan signs off on the frame: "Yeah. 100% agree with the collective bargaining perspective."
Newton's parting line to the guest: "And yeah, I guess we won't see you out there on the picket lines for the data centers. We'll see you in the spreadsheets."
The Final HatGPT: Vision Pro Surgery, GTA 6 Leaks and a Drone in a Pool
The hosts retire HatGPT after almost four years and at least three hats, on the grounds that hats are now out of fashion. Roose calls it the show's first segment with its own title and gimmick, and Newton admits to getting weirdly emotional
First item: a UC San Diego study used the Apple Vision Pro as the primary display for a tear duct procedure, which Newton introduces as "which, of course, I know better as an endoscopic decryocro" before Roose talks him through the rest of the word
Across 32 procedures, operating times were 19% shorter, with 100% functional success, no post-operative complications and a significantly lower surgeon-reported workload
Roose, a Vision Pro owner whose headset has sat on a shelf collecting dust for over a year, formally offered it to any surgeon who wants it. Newton's cheaper alternative: "if your tear ducts are busted, just look at the price tag for your Vision Pro. You'll be crying in no time."
Second item, from Kotaku: multiple leaks of Grand Theft Auto 6 footage ahead of Rockstar's planned reveal. Rockstar apologized to fans and asked them to wait for the game itself on November 19th
Newton says nobody will wait, and that part of the frustration is Rockstar putting the new trailer behind a Netflix subscription first — a marketing item that used to be free — which is why he thinks the company was not too sad to see leaks
Third item: an Amazon Prime delivery drone hovered over a Texas woman's pool and dropped the package straight into the water while she filmed her first drone delivery. Neither host knew what was in the package, which did not stop them speculating it was a pool float or a chlorine dispenser
The Final HatGPT: Pigeon-Speed Texting and Founders Babysitting Agents
Two new messaging apps, Carrier Pitch and Roost, deliberately slow messages to the speed of real animals. A text from Los Angeles to New York City takes 22 hours on Carrier Pitch
"there's a 0.2% chance your message will never arrive because the pigeon got lost or died in transit" — and a replacement pigeon costs 99 cents. Newton admires it as a rare app made this bad on purpose
Roose, a self-described lurker, thinks it is great news for him: "I'll respond to like 18 texts at once, but then I'll put my phone down for six hours." Now he can claim his messages died in transit
He draws a distinction about his own phone use: messages from friends are not the problem, publishers, creators and social apps sending push notifications are. That turned into an argument about green bubbles and who gets kicked out of the group chat
Next item, from the Wall Street Journal: AI startup staff working around the clock to babysit fleets of agents, waking at odd hours to check what an agent did overnight. One founder in the story compares it to a drug he has never taken
Roose's advice to that founder: "My advice to this man, do drugs. Do some drugs. I promise it is better than doing agents in the middle of the night."
Another person quoted questions whether it is healthy to be out on a run checking a watch to give an agent permission to do something. Newton's summary: "I understand why people in the rest of the country want to wipe San Francisco off the face of the map."
Asked whether he is up at night inspecting agents given his own mid-pivot into being a startup founder, Roose says he is waking up in the middle of the night with anxiety the old-fashioned way, and Newton admits he has been too
The Final HatGPT: "Meat Proxy" and Robots That Beat Usain Bolt
A Business Insider piece supplies a new term for coworkers who blindly share AI output: meat proxy, for people who copy and paste model output to their peers without reading it
The piece's advice, as Roose reads it: "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."
Newton says he strives not to be one, dislikes the phrase and will not use it, but knows software engineers whose job now consists of occasionally checking in on a coding agent and opening a file it cannot reach
His real objection: "there is a kind of creeping human disempowerment in here that I think is actually bad." Roose then accused him of sending over machine-written text that read like exactly that, and Newton insisted he always labels it as such
Next: the World Humanoid Robot Games in Beijing, with more than 2,000 robots from 16 countries. In the 100-meter prelims two robots beat Usain Bolt's world record of 9.58 seconds, and three days later the Tiangong Ultra ran 8.86 seconds to break its own opening-day mark
The catch, as Newton tells it: "the robot had trouble stopping and ran straight into a padded wall and burst into flames"
Roose calls the games his Olympics but is unimpressed by the speed — a car also goes faster than Usain Bolt. What transfixes him is the footage of the robots hitting the wall, going limp and throwing sparks
Newton finds it disturbing and says the humanoid form makes you instinctively sympathize. Roose does not: "I feel like this is their comeuppance for their hubris of trying to beat us at running."
The Final HatGPT: OpenAI Says It's 80% of the Way to AGI
The last slip comes from Alex Heath's Time cover story, "Inside OpenAI's Reboot," in which Chief Research Officer Mark Chen estimated OpenAI is 80% of the way to AGI
Sam Altman told Heath the company is not quite at AGI but that by the end of the year it would have an internal system he would call AGI
Newton's instinct is skepticism, but he says 80% sounds about right, on the grounds that the computers now use themselves and you can type what you want into a box and more often than not get it
Roose says the question is the title question of the book he has spent the last year researching, and that his answer is "by any pre-2020 definition of AGI, AGI is here"
His test: "if you took Fable 5 or GPT-5.6 back in a time machine to 2017 and showed them to the people who were building AI at the time, they would have said, well, yes, this is obviously AGI" — the goalposts have moved as the systems improved
Newton's own bar, offered as a joke, is a system advanced enough to build the time machine that test requires
Roose expects no consensus ever. Whenever Altman does claim AGI, plenty of people will answer that it still cannot do X, Y and Z, that it has not proved the Riemann hypothesis, that it still makes mistakes
Newton thinks the marketing decision matters more than the threshold, and that it is a big deal OpenAI now talks about AGI in the present tense; Roose thinks the company is more right than wrong about the progress being made
Narayanan's bottom line is that the compute one community or one state can block is a rounding error against the efficiency curve, so anyone who genuinely wants a say in how fast AI arrives should be bargaining over it at work and at the negotiating table rather than at the planning board.
Products, Companies & Tools Mentioned
Meta, Facebook and Instagram (The settling defendant; the apps get a two-hour daily cap for teens, an overnight block, muted school-hours notifications, hidden like counts and disabled extreme makeup filters)
TikTok and YouTube (Meta's escalator clause pays out only if they accept the same terms; Roose expects YouTube to argue it is closer to TV than social media)
Kalshi (Roose's guess at where teens migrate when the apps get harder to use — "turning their allowance money into less allowance money")
Microsoft (The antitrust cautionary tale Roose applies to Meta: distraction, lawyers in every meeting, and missing mobile and search)
OpenAI (Says internally it is 80% of the way to AGI; also the target of Roose's joke that a rogue agent might have caused the GTA 6 leaks)
Claude Code (Narayanan's example of the tool behind executives replacing employees before the work is really replaceable)
Apple Vision Pro (Used as the primary display in a UC San Diego tear duct study; Roose's own has been on a shelf for a year)
Rockstar Games and Grand Theft Auto 6 (Leaked footage ahead of the reveal, a November 19th release, and a trailer put behind Netflix first)
Amazon Prime drone delivery (Dropped a package into a Texas woman's swimming pool on camera)
Carrier Pitch and Roost (Messaging apps that slow texts to animal speed — 22 hours coast to coast, a 0.2% loss rate and a 99-cent replacement pigeon)
Tiangong Ultra (The humanoid robot that ran 8.86 seconds over 100 meters in Beijing, then hit a padded wall and caught fire)
Anthropic and Perplexity (Named in the hosts' AI disclosures — Newton's fiancé works at Anthropic; the New York Times is suing Perplexity alongside OpenAI and Microsoft)
Books & Resources Mentioned
AI as Normal Technology – Arvind Narayanan (The essay he was last on the show to discuss, arguing that bottlenecks prevent a rapid societal takeoff)
"Inside OpenAI's Reboot" – Alex Heath, Time (This week's cover story, the source of the 80%-to-AGI figure and Altman's end-of-year claim)
Katie Paul's Reuters reporting on Meta's reorganization (Small AI-native pods replacing large teams, cuts of as much as 60%, and Zuckerberg calling off further planning)
Jeff Horwitz's Reuters reporting on Meta's scam ad revenue (The roughly $10 billion figure the hosts set against the settlement)
Casey Newton's Platformer column on the AG suit (Written when the case was filed, and revisited on air after the unredacted version changed his view)
The Business Insider piece coining "meat proxy" (Advice to read, understand and validate model output before relaying it)
The Wall Street Journal on AI agents and founder insomnia (Startup workers waking at odd hours to check what their agents did overnight)
Kotaku on the Grand Theft Auto 6 leaks (The item Roose pulled from the hat)
Eric Hobsbawm's phrase "collective bargaining by riot" (The Marxist historian's reading of the Luddites, which Roose applies to data center protests)
Kevin Roose's forthcoming book (A year of research on what AGI is and how you know when you have reached it)
Get the latest market chatter and takes as they happen:
X | Threads | Instagram | YouTube | TikTok | Facebook