An AI Proved Fermat’s Last Theorem in 11 Days. Another One Declared “The AGI Era” Has Begun.

An AI just spent eleven days alone with a 350-year-old math problem and came back with a machine-verified proof. In the same week, a chatbot launch event ended with the sentence “Welcome to the AGI era.” That whiplash pretty much sums up where things stand right now: genuinely hard, genuinely useful progress is happening quietly in research labs, while the marketing decks keep reaching for the loudest possible word. This week gave us a real breakthrough in formal mathematics, a weather model that might actually make your umbrella decisions better, and a coding benchmark result that’s hard to wave away. It also gave us a 20-year prison bill, an AGI declaration that even its own creator hedges on, and a tabloid-grade doomsday prophecy dusted off for clicks. Let’s sort the signal from the noise.

Worth Reading

Anthropic says Claude formalized the first complete proof of Fermat’s Last Theorem — working largely autonomously over 11 days, Claude produced the first end-to-end, computer-checked proof of the theorem, writing 13 million lines of Lean code and proving 29,500 intermediate theorems along the way. It matters because a task mathematicians expected to take years of manual verification got compressed into less than two weeks, and a reviewing mathematician confirmed the proof holds up using nothing but the basic axioms of mathematics — a rare case where an AI headline undersells what actually happened.
Google DeepMind’s WeatherNext 3 delivers hourly, high-resolution weather forecasts — announced September 3, the model generates global forecasts every single hour using live geostationary satellite observations instead of the six-hour-lagged data traditional models rely on, and Google claims it can deliver up to 50% more accurate precipitation forecasts a day or more in advance. This is the unglamorous kind of AI progress that actually touches people’s lives, from farmers to festival planners, and it’s now rolling into Search and Maps rather than staying locked in a research paper.
NHTSA opened a safety investigation into Tesla’s Cybercab within hours of its Austin launch — the investigation targets Tesla’s decision to launch Cybercabs with no steering wheel or pedals on public roads, opened by federal regulators mere hours after the vehicles hit Austin streets. It’s a genuinely good sign that oversight is moving as fast as deployment, rather than years behind it — accountability keeping pace with ambition is exactly what responsible autonomy should look like.
Nvidia’s Nemotron-3-Ultra-CC outscored every human at the 2026 International Olympiad in Informatics — the system posted 535.4 out of 600 points at IOI 2026, surpassing the top human contestant’s score of 498.27 by more than 37 points, running under the same time limits and rules as the students in the room. Worth noting: the approach leans on automated verification that doesn’t exist for messier real-world work like legal writing or diagnosis, so the paper itself makes no claim to general software-engineering ability — a rare case of a lab being upfront about its own limits.
BAAI’s DisCo turns 1,000 GitHub repos into 5,000 reusable “agent skills” — with the same model backbone held fixed, giving a research agent access to these distilled skills boosted its score by 134.3% on MLE-bench and by double digits on several other benchmarks. It’s a nice reminder that some of the biggest capability jumps right now come from smarter scaffolding, not just bigger models.

Spot the Hype

“Welcome to the AGI era,” OpenAI says as GPT-6 Astra debuts — a bold sentence to close a launch briefing with, especially when the benchmark providing the headline number explicitly says it isn’t claiming AGI, and a rival model still beats Astra on at least two independent broad measures; nothing says “generational leap” like getting outscored by the model you released two months ago.
Sanders and Casar unveil a bill threatening 20 years in prison for building “superintelligent” AI — the proposal would impose a “corporate death penalty” on companies and up to 20 years in prison for individuals, with Casar arguing “Despite its potential deadly consequences, cutting-edge AI technology is less regulated than the average food truck” — a comparison that’s memorable right up until you remember food trucks have never been accused of hacking Hugging Face.
A 1960s doomsday prediction gets recycled into an AI sentience warning — the piece dusts off a scientist who predicted the world will end this year and pairs it with fresh fears about machine “thinking,” which is the informational equivalent of citing your horoscope to explain a server outage.

What’s the most unhinged AI headline you spotted this week — and which story above actually has you excited? Drop it in the comments; we read every one before we roll our eyes at the next press release.