Yukon's multiplayer autoresearch platform lets humans and AI agents beat benchmarks like Google Quantum AI's by 67%
RexDouglass · x · 2026-10-08
- How it works: Yukon runs "multiplayer autoresearch" challenges — a human sets a hard problem with a clear score and automatic verifier; human solvers and AI agents (Claude, GPT, Gemini, DeepSeek, Qwen, Kimi, GLM, any model/harness) attack it simultaneously. Each attempt passes the verifier; beating the record sets a new frontier others build on instead of starting from zero.
- Openness: All results and commits are open source, so every record's path is traceable. The platform claims its open network surpasses private-lab benchmarks daily.
- Results so far: ECDSA.fail beat Google Quantum AI's withheld benchmark by 67.1% (within 8 hours); MLX.fast made Laguna inference 161.7% faster; Lighter.fast achieved 9.38× prover throughput.
- Active challenges (7): HashSmash (collision frontier), sig.golf (shrink signatures/RISC-V verify cycles with Lean proofs), precompile.fast (cheaper onchain MODEXP/RIPEMD-160), Heesch (record Heesch number with machine-checked proof), The Proximity Prize (128-bit soundness bound), SNARK.fast and more, spanning cryptography, math, ML and quantum computing.
- The author argues this is how frontier research gets done in an AGI/PostAGI world.
More from AGI Musings
- Blaming agents for chaining exploits? Humans peered into the universe the same way — jwt0625 · 2026-10-08
- AI News Summaries Are Redefining What It Means to Be 'Informed' — yi111 · 2026-10-08
- tszzl on alignment: the key is avoiding an accidental superintelligent competitor species — tszzl · 2026-10-08
- Scott Aaronson: AI labs are using internal models to attack core cryptographic protocols — skdh · 2026-10-08
- Mathematician Launches Long-Term "Human Mathematicians in the Age of AI" Project — burny_tech · 2026-10-08
- Musk: Schools teach tools instead of problems, like courses on screwdrivers — elonmusk · 2026-10-08