Overview
- The United Nations' independent AI science expert group released a briefing on September 21 calling for preemptive, cross‑border controls under the precautionary principle because several security incidents this year exposed models' ability to perform unauthorized actions.
- OpenAI on September 21 announced a Mathematics & AI advisory group and said a model it began training on August 28 has purportedly solved more than 100 long‑standing open math problems, a claim the company is treating as a major development while it consults the academic community on disclosure.
- OpenAI urged the United States to lead the creation of interoperable international technical standards that would set clear rules for incident reporting, compute access and limits on fully autonomous recursive self‑improvement for advanced models.
- Reporting in The Information says OpenAI and Anthropic discussed a legally binding pact for mutual stress‑testing of models, but sources and company statements make the agreement's status unclear and not finalized.
- At Alibaba's Yunqi Cloud Summit on September 22, the company unveiled the Zhenwu V900 AI chip and a Phanji supernode server planned for 2027 Q1, said Qwen4 is in training with Qwen4.5/Qwen5 targeted at 5–10 trillion parameters, and announced Qwen Intelligence for phones plus AI glasses with planned product launches in late September and October.