Researcher says Astra solved hard problem in a single 6-hour thinking run
The author says a paper result to be published with a collaborator came from Astra thinking continuously for 6 hours, with the core breakthrough step taking 2.5 hours, and notes that one-shot results at this level in three hours would already be impressive.
The most important sentence of the GitHub repository: "On average, each result used three hours of ChatGPT Pro thinking compute with that model." The important result on that list I will publish with my collaborator on arxiv tomorrow was found in a session where Astra thought 6 hours in total and probably used my local context. The main problem solving step was done in 2.5 hours. If results on this level are possible with one-shot solutions in three hours, that is insanely impressive.