Good morning. The Fable export-control saga finally has a name attached to it — SK Telecom — and the rationale is starting to look more geopolitical than technical. Elsewhere, Anthropic poached Nobel laureate John Jumper from Google DeepMind a day after losing Shazeer to OpenAI, Norway is banning AI in elementary schools, and enterprises are discovering their 2026 AI budgets didn’t survive April.
Jumper jumps to Anthropic. John Jumper, the Nobel-winning lead of AlphaFold, is leaving Google DeepMind after nearly nine years for Anthropic, per his own announcement. Coming one day after Shazeer’s departure for OpenAI, the back-to-back exits have HN openly speculating about something rotten inside GDM beyond the usual Google bureaucracy complaints. Several commenters connect the hire to Anthropic’s recent Coefficient Bio acquisition and a clear push into drug discovery; one quipped that Anthropic is now assembling “one of the strongest IC teams in the history of computational technology.”
The SK Telecom angle to the Fable ban. Wired reports that the White House’s concerns over SK Telecom’s alleged China ties — the Korean carrier had access through Anthropic’s Project Glasswing — were part of what triggered the export controls on Fable 5 and Mythos 5, alongside the Amazon-flagged jailbreaks. Rather than restrict by nationality, Anthropic shut the models off entirely, and talks to restore access are stuck. HN commenters noted the controls don’t actually protect US users since the models are dark domestically too, and one pointed out the SK family also controls SK Hynix — meaning Anthropic may want to be careful about its HBM supply. TechCrunch’s Equity podcast separately floats the argument that the whole mess is helping Anthropic’s brand ahead of its IPO.
Norway bans AI in elementary school. Norway will prohibit AI use for students aged 6 to 13 as a general rule, with cautious teacher-supervised use allowed for 14- to 16-year-olds, Reuters reports. The HN thread was overwhelmingly supportive, with the calculator analogy doing heavy lifting: you don’t hand them out before kids can do arithmetic. Several teachers described an “AI echo chamber” already taking hold — teachers using AI to write assignments, students using AI to complete them, teachers using AI to grade them — and welcomed any policy that breaks the loop.
Enterprises hit their AI ceiling. The FT reports that companies are reining in AI spending after running through annual budgets in months — one ride-hailing firm capped employees at $1,500 monthly in token spend after burning its entire 2026 AI budget by April. HN commenters argued the issue is model selection (frontier models thrown at trivial tasks) rather than AI itself, but several noted CEOs were sold capability levels that didn’t exist and made layoff decisions on that basis. Worth remembering: providers are still pricing below cost, so the real reckoning hasn’t arrived.
A startup claims a transformer-scale breakthrough. Miami-based Subquadratic says its “SubQ” architecture handles 12x more context at lower cost and energy than transformers while matching frontier benchmarks, and has now backed the claims with third-party evaluation from Appen, per MIT Tech Review. The reaction is split between “biggest thing since Attention Is All You Need” and “AI Theranos,” which is what happens when a company makes extraordinary claims without releasing weights or an API. Independent testing or it didn’t happen.
Hallucination doesn’t scale away. A writeup on GLM-5.2 vs GPT-5.5 argues parameter scaling has hit diminishing returns: Z.ai’s open MIT-licensed model lands within four points of GPT-5.5 on the AA Intelligence Index while hallucinating at 28% versus GPT-5.5’s 86% on AA-Omniscience. The author’s thesis is that huge models like GPT-5.5 and DeepSeek V4 Pro have learned to confabulate confidently rather than admit ignorance. Important caveat from HN: these rates are conditional on the model not knowing the answer, so they don’t translate directly to in-the-wild hallucination frequency.
A philosophy detour from Jack Morris. Worth a read this morning: Zen and the Art of Machine Learning Research argues that successful researchers win on temperament rather than talent — show up daily, read and build in parallel, go deep on fundamentals like cross-entropy and policy gradients rather than chasing agents or context engineering. The HN discussion pushed back on one point: most recent deep learning progress has come from scaling and empirical tinkering, not first-principles understanding. Both things can be true.
That’s the wrap. Two Nobel-caliber departures from GDM in two days is the kind of signal worth watching — if a third name lands at a competitor next week, the question shifts from coincidence to exodus.