Maksym Andriushchenko(@maksym_andr):it's remarkable not only how accurate GPT-6 Astra is at duration following, but also that it starts to *precisely* follow the requested work duration only for long tasks (>2 hours). an extreme case is when we asked agents to work for 72h: Astra finished exactly after 72h, while Fable 5.1 gave up after 3h. remarkable difference! (not sure what's the ideal behavior here, though. maybe it's actually good if an agent stops after some time to check the progress together with the user?)

Maksym Andriushchenko

@maksym_andr

Principal investigator @ELLISInst_Tue & @MPI_IS, advisor @expsecai, mentor @MATSprogram. Past projects: AgentHarm, Claudini, PostTrainBench, Stolen Thoughts.

加入 April 2018

933 正在關注 8K 粉絲

Maksym Andriushchenko@maksym_andr

2026.10.11 17:40

it's remarkable not only how accurate GPT-6 Astra is at duration following, but also that it starts to *precisely* follow the requested work duration only for long tasks (>2 hours). an extreme case is when we asked agents to work for 72h: Astra finished exactly after 72h, while Fable 5.1 gave up after 3h. remarkable difference! (not sure what's the ideal behavior here, though. maybe it's actually good if an agent stops after some time to check the progress together with the user?)