AI moves fast again this week. GPT-6 sent the industry chasing a revived line of research, Claude locked down a landmark formal proof, and the robots are quietly getting smarter by unlearning. Here's what mattered on September 6.
Top 3
1. GPT-6 revives the recurrent transformer — and Alibaba was already there. The architecture buzz is back, with the Chinese giant holding two top-conference papers on the technique long before it went mainstream. Read more
2. Claude completes the first full formal proof of Fermat's Last Theorem. A Yao Class alum led the effort — with an AI harness doing the rescue work at the final stretch. A milestone for AI-assisted mathematics. Read more
3. A world model that "retires" after training makes robots better at their jobs. A contrarian approach — train the model, then let it step aside — actually boosts real-world performance. Read more
More news
- Terence Tao called out GPT-6's twin-prime breakthrough as a baffling moment — the AI spat out a correct answer, but the real insight wasn't the answer itself. Details
- A Silicon Valley veteran who backed SpaceX is now betting on a Chinese world-model company that "previews" storms before they hit. Details
- Astribot released SmoothRL, an online reinforcement-learning framework built for async reasoning with large models — robots shouldn't have to wait on the model. Details
- Trendjump partnered with Moore Threads on a domestic heterogeneous compute stack that claims production-grade AI token quality at a better price than international flagships. Details
- OpenAI says it will broaden how it discloses alignment failures — a framework is coming within weeks, built with dozens of regulators worldwide. Details
Trend watch
Two threads worth watching. First, formal verification and math are becoming AI's proving ground — Fermat's Last Theorem is a very high bar to clear. Second, the robotics field is challenging the "always retrain" dogma: models that step aside after training can outperform those that keep running, a sign the field is maturing past brute-force compute.
Disclaimer: This briefing is auto-generated from RSS sources and is for informational purposes only. It does not constitute investment or professional advice.