Epoch AI Watch posted an update
Epoch AI says another problem from FrontierMath: Open Problems has been solved, with the solution elicited by Becker, Greger and Dominik Peters in an interactive session with GPT-6 Astra, OpenAI's model. Peters, per the announcement, originally suggested the problem for the benchmark.
Why it mattersThe announcement is terse: it names no problem, and Peters' actual remarks are not spelled out in the post. 'Another' implies earlier Open Problems solves, and the working mode is the notable part: not a model answering cold, but three researchers steering a live session until the answer came out. The telling detail is that the problem's own proposer sat in the eliciting seat. Where GPT-6 Astra ends and the people steering it begin is an open question, and one the benchmark's numbers may not answer.
Discuss: When a solve arrives from a live session steered by the problem's own proposer, is FrontierMath measuring GPT-6 Astra or the skill of the people eliciting from it?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.