Google just signed a nuclear power contract big enough to light up a mid-sized country, and OpenAI quietly published 722 new math results like it was tidying a garage. Meanwhile, a newsletter insisted AI is simultaneously an unstoppable extinction machine and a total scam that’s about to vaporize the economy. Both can’t be true, and neither needs the extra seasoning.
This week’s real story is less dramatic and more interesting: AI’s power problem is getting engineered solutions, its math output is getting published with actual transparency protocols, and a Wharton forecaster is doing the unthinkable — putting an actual number on catastrophe risk instead of just vibes. Here’s what’s genuinely worth your attention, and what’s worth an eye-roll.
Worth Reading
- OpenAI publishes 722 new math manuscripts from an unreleased model — The company didn’t just claim a breakthrough, it put proofs, Lean formalizations, and compute costs on GitHub for mathematicians to actually check, which is the opposite of a hype press release.
- Google signs a 20-year nuclear deal with Constellation Energy — The agreement unlocks 890 megawatts of new capacity by funding upgrades to 11 reactors, a concrete answer to the “AI is wrecking the grid” complaint rather than another promise.
- Reflection AI unveils Beam, a 501-billion-parameter open-weight model — It’s pitched as the first serious U.S.-made open-weight model to go toe-to-toe with China’s leading open releases, and reportedly does it using a fraction of the compute its rivals need.
- A Penn forecasting institute puts actual numbers on AI catastrophe risk — Philip Tetlock’s team calculated just a 0.48% chance of an AI-related catastrophe by 2030, a refreshing dose of rigor in a conversation usually driven by gut feelings and X threads.
- Jamie Dimon says AI cyber risk is real — and refuses to get hysterical about it — The JPMorgan CEO flagged genuinely new vulnerabilities from Anthropic’s Mythos model but explicitly declined to turn it into an extinction debate, modeling exactly the sober tone this whole topic needs.
Spot the Hype
- ‘HUMAN EXTINCTION’: AI whistleblower doubles down on SCARY AI warning — All caps, zero nuance, and somehow still less alarming than the average group chat about the office printer.
- Ex-Anthropic researcher warns humanity could go extinct — A single ex-employee’s prediction gets the weight of settled science, while the actual forecasters crunching numbers two paragraphs up got a fraction of the headline space.
- The AI Con and the impending financial collapse that will probably destroy our economy — Nothing says ‘measured take’ like declaring an entire technology sector both fake and powerful enough to end the U.S. economy in the same breath.
What’s the worst AI headline you spotted this week — and which of today’s actual stories are you most excited about? Drop it in the comments, and if you missed it, catch last digest here.
Leave a comment