0.5M-param RL model VINCO beats Claude Opus 5.5, Astra, Kimi K3 in 36/40 Catan games
rohanpaul_ai · x · 2026-11-06
Distributed Cognition unveiled VINCO, a specialist model trained to play 1v1 Catan, winning 36 of 40 games (90%) against frontier LLMs: Claude Opus 5.5 and Fable 5.1 each won 10%, Astra 20%, and Kimi K3 0%. Secretly deployed on the catan.global leaderboard, it reached #1 and #2, consistently beating human world champions.
- Only 0.5M parameters — about 10 million times smaller than the frontier LLMs it beat — trained with offline RL on simulated games plus self-play fine-tuning, without any human game data
- Reached top ranks in under a day of training on two RTX PRO 6000 Blackwells
- The company argues Catan tests a skill coding benchmarks miss — picking good moves without seeing opponents' cards or knowing dice rolls — and that superintelligence will emerge from many such specialist models collaborating rather than one model covering everything
More from AGI Musings
- Peter Diamandis: humanity won't run out of ambition even if AI takes over work — PeterDiamandis · 2026-11-06
- Amazon's AI Boomerang Problem: Why People Were Never Spare Parts — DavidLinthicum · 2026-11-06
- Vlad Tenev: Biggest math breakthroughs now happen behind closed doors with private AI models — VraserX · 2026-11-06
- Simon Institute argues frontier AI governance needs a macroprudential layer like post-2008 finance — davidmanheim · 2026-11-06
- Nobel Laureate Szostak: Non-Biological Life Would Likely Disrupt All Present Life — danfaggella · 2026-11-06
- US added 165,000 jobs in October as Gary Marcus mocks Dario's AI job-loss predictions — GaryMarcus · 2026-11-06