The weekend delivered a genuine milestone for AI in science: an AI system produced the first fully machine-checked proof of Fermat's Last Theorem, a problem that resisted mathematicians for 358 years. Meanwhile, OpenAI's most powerful model reached ChatGPT amid an apology tour, Tesla's steering-wheel-free robotaxi hit the streets of Austin — and a federal investigation within hours — and the money kept pouring into AI infrastructure. Here is what matters from September 4–6.
Claude completes the first machine-checked proof of Fermat's Last Theorem
- Working autonomously for 11 days, Anthropic's Claude wrote 13 million lines of Lean code and proved 30,300 supporting theorems to formalize the 358-year-old result, consuming about six billion output tokens in the process.
- Mathematician Kevin Buzzard of Imperial College London — who had estimated the formalization would take years of human effort — verified the result: the proof rests on nothing but the standard axioms of mathematics.
Why it mattersWhy it matters: this is AI crossing from generating plausible answers to producing mathematics that a computer can certify as correct — a preview of formally verified science and software, where the machine's work doesn't need to be taken on faith.
GPT-6 Astra arrives with big claims — and an apology from Altman
- OpenAI launched GPT-6 Astra on September 3 with enterprise customers first; by September 4 Sam Altman was apologizing for the "messy" rollout that left paying Pro and Plus subscribers waiting, offering banked usage resets as compensation.
- OpenAI says Astra scores 98% on FrontierMath Tier 4 and is the first model it has designated "Critical" for cybersecurity risk, citing its ability to find previously unknown vulnerabilities; president Greg Brockman went as far as calling it "broadly smarter than people".
Why it mattersWhy it matters: the frontier keeps moving faster than OpenAI's ability to serve it — and the AGI rhetoric is now coming from the company's own executives, not just enthusiasts.
Tesla's Cybercab starts carrying passengers — and NHTSA opens a probe within hours
- The two-seat Cybercab — no steering wheel, no pedals — began picking up riders in Austin on September 4 through Tesla's Robotaxi app; the company now runs 314 driverless vehicles in Texas, about 45 of them Cybercabs.
- Hours after launch, NHTSA opened an audit of how Tesla self-certified a car with no manual controls under federal safety standards that still require them — echoing its 2022 inquiry into Amazon's Zoox.
Why it mattersWhy it matters: the robotaxi era is arriving before the rules for it exist. Whoever wins this regulatory fight sets the template for every city that follows.
Crusoe triples its valuation to $30 billion building OpenAI's power
- The energy-first data center company raised over $3 billion at a roughly $30 billion valuation, per Bloomberg — nearly triple its October 2025 mark — in a round led by Atreides Management and Valor Equity Partners.
- Crusoe is building a 1.2-gigawatt cluster in Abilene, Texas for OpenAI, part of the Stargate buildout that has turned electricity into the scarcest resource in AI.
Why it mattersWhy it matters: the bottleneck in AI is no longer talent or even chips — it's gigawatts. The companies that can deliver power are commanding valuations that triple in under a year.
Washington and Beijing prepare their first AI safety talks
- Reuters reports the US and China will hold their first dedicated AI safety talks in mid-September in Beijing, with Treasury Secretary Scott Bessent leading the American delegation, ahead of a Trump–Xi summit set for September 24.
- The agenda includes cooperating on monitoring AI-directed cyberattacks and US concerns about Chinese companies distilling proprietary American models.
Why it mattersWhy it matters: the two AI superpowers are starting to negotiate the rules of the game directly — everyone else, from Europe to Latin America, will live with whatever they agree.
The big picture
The thread running through the weekend: capability is outrunning institutions. An AI just produced mathematics no human could check by hand — and a computer certified it. A robotaxi with no steering wheel is picking up passengers under rules written for cars with pedals. The response, from NHTSA audits to US-China talks, is institutions sprinting to catch up. The gap between what the technology can do and what our systems are built to absorb is now the defining story in tech — and the money, like Crusoe's $30 billion, is betting the gap keeps widening.
Sources & further reading
- Anthropic — Formalizing Fermat's Last Theorem
- SiliconANGLE — Anthropic uses Claude to formalize proof of Fermat's Last Theorem
- Unite.AI — Sam Altman apologizes as GPT-6 Astra staged launch denies paid access
- CNBC — OpenAI announces rollout of GPT-6 Astra model
- TechCrunch — Feds launch investigation into Tesla's Cybercab deployment
- Bloomberg — Crusoe raises over $3 billion at $30 billion valuation
- The Korea Times / Reuters — US, China gear up for mid-September AI safety talks