Overview
- The UN independent scientific panel on AI warned Tuesday that governments should regulate powerful AI agents before their full risks are known, citing the precautionary principle and urging international cooperation on safety and accountability.
- OpenAI asked the United States to lead creation of shared technical standards for frontier models and announced an independent Mathematics and AI advisory group after reporting that a recently trained model solved more than 100 long‑standing math problems, a claim the company says it will vet with experts.
- Reporting shows OpenAI and Anthropic have discussed a legally binding pact to pressure‑test each other’s models, a tentative industry proposal that would let competitors hunt for vulnerabilities before public release.
- Alibaba’s Pingtouge unveiled the TrueWu V900 training‑and‑inference chip and detailed a Qwen model roadmap that plans 5–10 trillion‑parameter versions plus new supernode servers set for early 2027, an explicit push to expand compute capacity and model scale.
- Banks and creators are already sounding alarms: lenders warn AI shopping agents could raise fraud and privacy risks, and companies are lobbying governments—including Australia—over copyright rules that would affect how models are trained and how creators are paid.