Nathan Labenz on FrontierMath Saturation and AI Paths
Nathan Labenz shares his latest read on AI benchmarks and physical tasks.
TLDR
Nathan Labenz posted that Astra has saturated FrontierMath tier 4 with a third of the year left. His start-of-year forecast called for mid-60s completion by December, which matched the crowd view. Labenz now states that any verifiable domain has a quick path to at least weak superintelligence. The same post asks whether 53 percent success rates in physical intelligence tasks can turn into useful work. These comments appeared in a broadcast linked from his account.
Combined views
542
1 Source, first seen 26d ago