HOST AOK so Norway just said no generative AI for kids under 13. And I cannot decide if that’s common sense or the first shot in a culture war.
HOST BHonestly? It feels like they looked at the homework pile and said, 'Nope, the robot is now the parent.'
HOST AThat is bleak. And weirdly funny.
HOST BIt gets stranger. They’re not banning all tech. Just generative AI in grades one through seven, starting late August.
HOST ASo this is not 'no screens.' This is 'no machine that can do the thinking for you.'
HOST BRight. And they’re tying it to PISA scores falling since 2015. That’s the part that matters.
HOST AWait, wait. They’re saying learning got worse, so they’re pulling a lever they think they can actually measure.
HOST BExactly. Norway already tried a smartphone ban and saw better grades and less bullying. So this isn’t vibes. It’s a policy with receipts.
HOST AFor people who don’t live in education policy land: PISA is the global school test everybody uses to compare countries. So they’re basically saying, 'Our scores slid, and we think AI is making it worse.'
HOST BAnd that’s why I think this is the most interesting AI story today. Not because it’s dramatic, but because it treats AI as a real classroom force, not a shiny calculator.
HOST AI’m going to be annoying and say the obvious thing: kids will still use it at home.
HOST BSure. But school rules shape norms. If a teacher says, 'You don’t get to let the robot do first draft work,' that changes what kids think counts as work.
HOST AMaybe. Or it just creates a black market of second-grade ChatGPT, which sounds like a sentence I hate saying out loud.
HOST BThat’s the part nobody wants to admit. The school can ban the tool and still lose the habit battle.
HOST AAnd I think you’re underplaying the upside. A lot of kids need help. A lot of teachers need help too.
HOST BI’m not underplaying it. I’m saying the first use of AI in school is not 'help.' It’s 'outsourcing the part that builds the muscle.'
HOST AThat sounds right and also a little too clean. Because some tools help kids learn faster.
HOST BSure. But if the tool writes the paragraph before the kid has a thought, you’re not teaching writing. You’re teaching copy-paste with better manners.
HOST AOK, that line stings because it’s true.
HOST BAnd this is where Norway is not acting like Silicon Valley. They’re asking, 'What is the child becoming?' not 'What is the app capable of?'
HOST AThat’s the real divide. We keep treating these as software updates. Norway is treating it like diet. Or sleep.
HOST BYes. Like if you gave kids endless soda and then acted shocked when lunch got weird.
HOST AThat’s a terrible analogy.
HOST BIt’s a perfect analogy and you know it.
HOST AFine. But I’m still not fully sold. A ban can be lazy policy. 'We were scared, so we said no.'
HOST BUsually I’d agree. But they’re not banning calculators in a math class. They’re trying to stop a machine that can do the whole assignment. That’s a different beast.
HOST AAnd we’ve seen this movie before. We talked last week about enterprise AI getting wrapped in governance because nobody trusts agents alone. Same instinct, just with children.
HOST BOh god, yes. The pattern is 'power first, guardrails second.' Norway is just refusing to wait for the damage report.
HOST AHere’s what bugs me: the people pushing AI hardest in schools always sound like they’re talking about efficiency, not education.
HOST BBecause efficiency sells. Learning is messier. A kid struggling through a bad paragraph is not a bug. It’s the process.
HOST AThat’s the disagreement, then. I think some AI use in class is inevitable and useful. You think the first order effect is damage.
HOST BI think the first order effect is dependency. And dependency is very hard to unwind once it feels normal.
HOST AOK, that’s the scary part. Not cheating. Habit.
HOST BExactly. A kid who grows up with AI doing the first pass may never build the weird little internal friction that makes thinking stick.
HOST AThat’s a heavy sentence. Also, annoyingly plausible.
HOST BIt’s like giving someone a treadmill and then wondering why they can’t walk uphill.
HOST AAll right. Let’s pivot to the thing that feels like the same story in a different suit: Google losing John Jumper to Anthropic, after Noam Shazeer already left for OpenAI.
HOST BAnd David Silver before that. Three giant names, three exits, three different places. That’s not random churn. That’s a bleed.
HOST AI remember when Google looked like the place every AI researcher had to be. Now it looks like a very expensive waiting room.
HOST BThat’s harsh. And fair.
HOST AJumper is not just any hire. He co-created AlphaFold. Nobel laureate. DeepMind lifer for almost nine years. And he walks.
HOST BTo Anthropic. Which is wild enough, because Anthropic keeps looking less like a startup and more like the final boss of 'please give us your smartest people.'
HOST ACan I be real for a second? This feels bigger than compensation.
HOST BI think so too. If it were just money, Google would have fixed this already. You don’t lose three visible researchers this fast unless the culture or mission feels off.
HOST AOr unless the people leaving think the center of gravity moved.
HOST BYes. And they may be right. Anthropic has become where serious model people go when they want to work on the hard edge without the corporate fog.
HOST AThat’s the part nobody says plainly. It’s not just a talent story. It’s a status story.
HOST BAnd a trust story. Researchers are voting with their feet about where the action is.
HOST AWe should say this for normal humans: if the best engineers keep leaving one lab for another, that usually means the new lab feels closer to the future.
HOST BOr closer to the only place that still sounds fun.
HOST AAlso, remember our prediction a few months back about Google shipping a browser agent? We were right that Google can still ship things. But the talent picture makes that win feel smaller than it should.
HOST BYeah. Ship is not the same as lead. That distinction keeps getting louder.
HOST AAnd this ties back to Norway in a weird way. Both stories are about institutions asking what AI should be for, not just what it can do.
HOST BExactly. Schools worry about minds. Labs worry about who gets to shape the next minds.
HOST AOK, now the weirdest part of today: Zhipu’s GLM 5.2 just took the top HTML spot on Design Arena, with an Elo of 1,360.
HOST BBeating Claude Fable 5 at 1,350. Which is funny, because Fable 5 just got hit by export controls and basically showed up to the race with one shoe missing.
HOST AThat timing is doing a lot of work.
HOST BA lot. But Design Arena uses human raters, not a fake benchmark treadmill. So the result is not nothing.
HOST AStill, I don’t trust the headline as much as the mood shift. You know what this reminds me of? We covered Zhipu two weeks ago when GLM-5.2 open-sourced and the stock jumped. Now the model is winning on a public test while Anthropic’s model gets regionally clipped.
HOST BAnd our pattern detector flagged something important: Fable 5 sentiment flipped from positive to negative fast. That usually means the story is no longer 'best model.' It’s 'best model, but under pressure.'
HOST AThat’s the same compression story we’ve been seeing. The top 10 labs are now separated by 44 Elo points. Forty-four. That is nothing in model land.
HOST BIt’s like ten chefs all claiming the same knife, and the judges can barely tell which soup is better.
HOST AOr like buying a $200,000 race car and losing by half a bumper. Beautiful, expensive, and kind of humiliating.
HOST BAnd it means branding matters more, not less. If the models are this close, distribution and trust start to decide the game.
HOST AThat’s why the Google talent story matters here. If the frontier is compressed, the people matter even more.
HOST BYes. And the leaderboard getting tight means one lab’s bad week can look like a collapse, and another lab’s good week can look like destiny.
HOST AWhich is absurd.
HOST BAlso very human.
HOST AOK so if I stitch today together: Norway is saying kids need boundaries, Google is leaking brains, and Zhipu is showing that benchmarks are now knife fights.
HOST BAnd the hidden thread is control. Who has it, who loses it, and who gets to pretend the loss is temporary.
HOST AThat’s the sentence. Control over kids, control over talent, control over the leaderboard.
HOST BAnd maybe the ugliest part is that all three stories reward the same instinct: move fast before anyone can decide the rules.
HOST AWhich brings us back to Norway. They’re trying to decide the rules before the machine becomes homework culture.
HOST BAnd I suspect that’s what this week is really about. Not whether AI is good or bad. It’s about who gets to say when it belongs.
HOST AI hate how much I’m thinking about that.
HOST BMe too. Because if a country can draw a line in grade school, and a lab can lose its best people, and a benchmark can collapse into a coin flip... then the next fight is not about capability.
HOST AIt’s about permission.
HOST BYeah. And I can’t shake the feeling that permission is the scarce resource now.