Source Posts on X
oh, do you mean it's not capable of solving an erdos problem that was solved by another model (in this case i think this is just that people don't try, but i agree no data) or solving a new problem (in which case this is just obvious since ppl use more closed model to do this + no one is arguing open model are better than frontier closed model for pure intelligence)
7:51 AM · Jul 15, 2026@eliebakouch also one concern: - models are good on similar things - it's possible that GPT-5.2 and successor models have already picked all the low hanging fruits so of course we don't see open models solving anything new
7:49 AM · Jul 15, 2026@scaling01 lol ok you just answered me imo would be interesting to have a public benchmark no internet access for already solved problem
7:53 AM · Jul 15, 2026@eliebakouch yeah i mean that's kind of what FrontierMath T4 is there open-models also haven't caught up with GPT-5.2-xhihg
7:55 AM · Jul 15, 2026View all
@scaling01 lol but it's private :( will take a look and see if it's easy to cook a quick benchmark for this
7:58 AM · Jul 15, 2026