Overview of Hard Fork
This episode covers three big themes: Meta’s massive child-safety settlement and what it means for social media, a nuanced discussion with Princeton professor Arvind Narayanan about whether data-center bans can actually slow AI progress, and the bittersweet final installment of the show’s long-running gimmick segment, HatGPT.
Meta’s $17.1 Billion Child-Safety Settlement
Kevin Roose and Casey Newton break down Meta’s settlement with 47 states, D.C., and U.S. territories over alleged child-safety and privacy violations. They frame it as one of the largest settlements in tech history and a major setback for Meta.
What Meta was accused of
- Collecting data on children under 13 without parental consent, violating U.S. privacy law.
- Knowingly allowing millions of under-13 users on its platforms.
- Designing products to maximize teen engagement through:
- push notifications
- ranking algorithms
- weak screen-time limits
- addictive design patterns
What Meta agreed to change
- A default two-hour daily limit across Facebook and Instagram for teens.
- A default app block from midnight to 6 a.m.
- Push notifications muted from 8 a.m. to 3 p.m. during school hours.
- Like counts hidden by default on Instagram.
- Extreme makeup filters disabled for teen users.
- DMs exempted from some of the restrictions.
Meta’s strategy
- Meta is presenting the settlement as a kind of industry leadership moment.
- The company is publicly pressuring TikTok and YouTube to adopt similar teen protections.
- The hosts call this deeply cynical: Meta waited until the last possible moment, then recast compliance as moral leadership.
Broader takeaway
- The hosts see the settlement as harm reduction rather than a full solution.
- They compare the long-term effect to anti-smoking policy: making a harmful product gradually less appealing can matter, even if it doesn’t solve the underlying social problem.
Do Data-Center Bans Work? Arvind Narayanan Says Probably Not
The episode’s interview segment features Princeton computer science professor Arvind Narayanan, who argues that local bans or moratoriums on new data centers are not an effective way to slow overall AI progress.
Core argument
- Most data centers are not used for model training.
- Training runs happen in specialized clusters, so a local ban on new facilities barely affects them.
- For inference/serving models, a single state’s moratorium only reduces total global AI capacity by a tiny amount.
The numbers he gives
- A one-year data-center moratorium in a typical U.S. state would slow AI progress by only about 5–10 hours in aggregate.
- His point: the industry’s efficiency gains in software and hardware outpace the capacity lost from blocking a few new buildings.
Important distinction
- Overall AI progress is not much affected by local bans.
- But individual companies’ competitive position still depends heavily on compute access.
- So the total pace of progress may not change much, even though labs still fight hard to secure more compute.
What opposition to data centers can still accomplish
Narayanan says the backlash can still be useful:
- It can force companies into negotiations.
- It can win:
- direct payments
- local investment
- community benefits
- more transparency
Better targets for slowing AI
He suggests opposition may be more effective if aimed at:
- government and institutional use of AI in consequential decisions
- workplace decisions where managers over-delegate to AI
- overconfident executives making premature AI-driven cuts or substitutions
Practical advice
- He encourages people to retain agency over how they use AI.
- Don’t let tools default into thinking for you; configure them to do grunt work rather than decision-making.
Final HatGPT: The Last Gimmick Segment
The episode ends with the final-ever HatGPT segment, framed as a farewell to one of the show’s signature bits.
Notable items discussed
- Apple Vision Pro in surgery
- A UC San Diego study used the headset in a tear-duct procedure.
- Reported results: shorter operating times, high success, fewer complications, and lower workload.
- The hosts joke that surgeons may be the ideal Vision Pro users.
- Grand Theft Auto 6 leaks
- The hosts discuss leaked footage and fan frustration.
- They joke about the game’s secrecy and monetized trailer rollout.
- Amazon drone drops package into a pool
- A delivery drone mistakenly drops a package into a woman’s swimming pool.
- They speculate the drone may have had “good intentions.”
- Carrier Pidge / Roost
- Messaging apps that intentionally slow texts to the speed of carrier pigeons.
- A playful take on digital overload and delayed communication.
- AI agents making founders anxious
- A Wall Street Journal story about startup workers babysitting agents at all hours.
- The hosts mock the idea of being awake at night to monitor agent behavior.
- “Meet proxy”
- A term for coworkers who blindly paste AI output without understanding it.
- The hosts agree this is a real and undesirable workplace behavior.
- Humanoid robot races in Beijing
- Robots reportedly beat Usain Bolt’s 100m record, then crashed into walls and burst into flames.
- The hosts find the spectacle bizarre and compelling.
- OpenAI and AGI
- Time magazine reporting says OpenAI leaders believe they are around 80% of the way to AGI.
- The hosts debate how arbitrary the AGI threshold is and note that definitions have shifted over time.
Key Takeaways
- Meta’s settlement is a major regulatory win for child safety and teen protections.
- The changes are best understood as harm reduction, not a full fix.
- Data-center bans are unlikely to slow AI in a meaningful way, though they can still be useful in local bargaining and community pressure.
- The show’s final HatGPT segment serves as both a comedy roundup and a farewell to a long-running podcast tradition.