0.5M-param specialist VINCO beats frontier LLMs at Catan with 90% win rate
mhdfaran · x · 2026-11-06
Distributed Cognition unveiled VINCO, a 0.5M-parameter model trained solely for 1v1 Catan, secretly deployed on the catan.dcog.com global leaderboard where it holds the #1 and #2 spots, consistently beating human world champions.
Key numbers: VINCO wins 90% of games versus 20% for the best general LLM (Astra) — a parameter gap of roughly 10 million times. The takeaway: specialist models that know one thing extremely well can crush general models at narrow tasks. You can play against VINCO on the site.
More from Fun
- OpenAI rumored to drop Astra-line models Friday as Qwen teases the same day — skeole · 2026-11-07
- Dev builds sailing sim and makes frontier models race each other — abhi3188 · 2026-11-07
- Fuzzy moss-covered 3D spheres star in a striking generative demo — thespite · 2026-11-07
- Henge 0.1 Lets You Run macOS Directly on iPhone and iPad, No Mac Needed — Scobleizer · 2026-11-07
- Nintendo Uses Rare Fluid-Sim Fog Most Games Skip as Too Expensive, Analyst Says — banteg · 2026-11-07
- AI Video Meme Goes Viral: 'He Doesn't Gallop—He Dynamically Fractures Space' — LudovicCreator · 2026-11-07