Benchmark researcher Ofir Press jokes he'll stop saying millennium problems are too hard to saturate
OfirPress · x · 2026-09-09
Princeton researcher Ofir Press sheepishly walked back a claim from his benchmark-building talks: he used to say the Millennium Prize Problems "could be a benchmark but it would take 100 years to saturate." After reports that a model cracked one of the problems, he wrote "won't be saying that anymore," joking that he looks forward to the millennium-prize benchmark being "fully saturated by the end of 2028." The quip captures the industry mood that frontier models are outpacing expert expectations on fundamental science.
Related event: Benchmark expert says his millennium-problem example aged too fast(3 posts)→
More from Fun
- AI naming fail: "Human-in-the-Loop Execution Routing" abbreviates to HITLER — SatelliteNetSec · 2026-09-09
- Vibecoder builds 'fruit fly heaven' sim where a fly can just fly around forever — Polymarket · 2026-09-09
- Dev mocks AI doomers: deduplicating tokens at scale while screaming "we're all going to die" — gordic_aleksa · 2026-09-09
- 'Alignment is useless, just unplug the datacenters' — and then nobody unplugs anything — EigenGender · 2026-09-09
- AI Twitter Rewrites 'Country Roads' to Mock Agents That Are Always 'Almost Done' — SIGKITTEN · 2026-09-09
- mitsuhiko: Google's Astra codes like a code golfer — mitsuhiko · 2026-09-09