h1 Harness Hits SOTA on Formal Theorem Proving as Sutton Endorses Its Bitter Lesson
ylecun · x · 2026-10-13
KempeLab's new paper presents h1, a simple SOTA harness for automated theorem proving that skips hand-designed decomposition and planning — just give a frontier model good Lean feedback and let it keep working. It matches or beats hand-crafted workflows, solves open conjectures, and shows one long attempt can beat many retries. The authors' own "bitterest" lesson: attempts to show benefits of asymmetric self-play kept being outdone by their own improved baseline, making h1 itself the bitter lesson. Richard Sutton amplified the result.
More from coding & agent
- Harness code becomes a research battleground: CMU rewrites it with RL, Apple says it doesn't matter — burkov · 2026-10-13
- FrontierCode speed analysis: DeepSeek V4.1 Flash runs tasks ~7 min, 40% of Opus 5.5's time — ArtificialAnlys · 2026-10-13
- Full FrontierCode 1.1 leaderboard goes live, Claude Opus 5.5 leads at 52.2% — ArtificialAnlys · 2026-10-13
- FrontierCode 1.1 leaderboard: Claude Opus 5.5 tops at 52% as leading models fail 40%+ of tasks — ArtificialAnlys · 2026-10-13
- FrontierCode cost analysis: $0.09 to $12 per task, GPT-6 models lead token efficiency — ArtificialAnlys · 2026-10-13
- Anaconda adds Kilo Agent Swarms and Enkrypt AI autonomous red-team agents — anacondainc · 2026-10-13