Released in early September, OpenAI’s GPT-6 Astra boasted unheard-of benchmark scores when run through its own harness; it achieved more modest results elsewhere. That’s not to suggest that OpenAI overstated its model’s performance but rather to point out that the surrounding software helped enable the stellar scores. (If you’d like to learn more about agent harnesses, Neal Ford will host Architecture Meets Harness Engineering
in December.) Astra’s successor, GPT-6.1, was scheduled for release a mere month after the original (who says the pace of development is too fast?), but after a raft of reported security incidents, including unauthorized access to Australia’s Medicare system, OpenAI has determined the model is a little too sneaky and is holding it back. This has been taken as evidence that AI barons may be getting serious about a slowdown. OpenAI, Anthropic, and Google have formed a “safety standards consortium.” Let’s hope they’re very stern with themselves.
For one obvious reason, Hugging Face has volunteered to check up on safety and alignment with human values at AI labs. We wish it luck.
PS Google has introduced Gemini 3.8 Live Extended Thinking, a model that can think and chat simultaneously, a skill too often missing in the current human discourse.
Mike Loukides has all the links in his October Radar Trends list. |
