Commentary on X
@dMxwABXhoVgGr1Y @hyhieu226 The models don’t merely imitate. You can absolutely get superhuman AI capabilities through RL and related methods. See e.g. AlphaZero which didn’t use any human training data at all.
3:00 AM · Jul 15, 2026@hyhieu226 Really? We have open problems falling by the wayside every week, not to mention the absolute domination at the atcoder finals.
3:24 AM · Jul 15, 2026@hyhieu226 5.6 Pro is still struggling against P3/P6 level
2:03 AM · Jul 15, 2026View all
@hyhieu226 Questions is how to find problems that participants can arrive solve while being difficult for AI. More geometry and combinatorial problems? But that doesn't seem to be a weakness anymore
5:03 AM · Jul 15, 2026IMO 2026 happens in the next 2 days. Perhaps the *last* time humans still have an edge against AI at mathematical olympiads.
12:19 AM · Jul 15, 2026