Anthropic has formalized Fermat's Last Theorem. Claude did the work over 11 days, mostly autonomously, producing what Anthropic calls the first complete computer-verified formalization of Fermat's Last Theorem in Lean. It is not a new proof; Andrew Wiles already provided one. The work expresses that proof in a form whose logical steps can be checked mechanically. It deserves more attention than another mathematics benchmark would get.1
Can AI now design circuit boards? EEBench is trying to build the electronics version of executable code evaluations. Instead of judging whether a generated schematic looks plausible, it builds a circuit graph, runs SPICE simulations and checks measurable criteria such as voltages, thresholds, ripple and cost. Claude Opus 5 leads its published September 1 results with 61.6%. I like this approach: once models do engineering work, the useful benchmark is whether the device works.2
Google has patched an actively exploited Chrome zero-day. CVE-2026-85046 is a type-confusion bug in V8 that allows arbitrary code execution inside the Chrome sandbox through a crafted page. Browser security matters more for agent infrastructure now, since so many computer-use systems effectively hand a model control of a Chromium instance.3
GPT-6 Astra is already available via OpenRouter. It took about a day for yesterday's frontier launch to become one more model you can pick through a standard API. Models are getting more portable at the same time as model-specific harnesses matter more, which is an interesting tension.
Who is using MCP in production? started a useful discussion on Hacker News today.4 I care more about reports of real operational problems, authentication decisions, tool boundaries and deployment patterns than about another explanation of what MCP is supposed to become.
The Rust React Compiler is now native in Vite. Another small example of Rust becoming part of the tools most JavaScript developers use without caring what language they are written in.
Google AI Mode shows products priced 21.6% higher than traditional search does. This connects to my post from two days ago on retrieval systems and purchasing decisions. If AI search systematically shows a different set of products from conventional search, how retrieval and recommendation work is more than an SEO question.5
References
Footnotes
-
Anthropic · “Formalizing Fermat's Last Theorem,” 4 September 2026. This is the original source citing Anthropic's assertion that Claude achieved the first fully computer-verified formalization of Fermat's Last Theorem following 11 days of mostly autonomous work in Lean. ↩
-
EEBench, “Can AI design circuit boards yet?”, 4 September 2026. This article describes EEBench's deterministic circuit-evaluation method and shows results from its current model. ↩
-
BleepingComputer · "Google warns of a new Chrome zero-day vulnerability being exploited in attacks", 4 September 2026. Report on
CVE-2026-85046and Google's Chrome 152 security update. ↩ -
Hacker News · “Ask HN: Who is using MCP in production?” A discussion by practitioners regarding actual MCP deployments. ↩
-
ProductRise · Study into Google AI Mode's pricing model. A study that compares products found using Google AI Mode with those obtained through traditional search. ↩