VINCO

VINCO (Latin for "I win") is Distributed Cognition (dcog)'s ~0.5M-parameter specialist for competitive 1v1 Catan, public in November 2026. Report: https://distributed.inc/research/solving-1v1-catan/ ; play: https://catan.dcog.com/ .

Claims:

• On 2026-11-01 two distributed accounts held #1 and #2 on Colonist.io's 1v1 ladder; they stopped after <24h of play to avoid season distortion;
• Trained with offline RL + self-play, no human games; two RTX PRO 6000 Blackwell GPUs;
• Self-reported 36/40 wins vs frontier LLMs (Opus 5.5, Fable 5.1, Astra, Kimi K3, etc.);
• Style: more development cards and largest army, cities before blind expansion, fewer ports than human pros.

The lab frames it as Catan's "AlphaGo moment" and evidence for their "fleet of specialists" view of superintelligence.

Why it's interesting

Catan mixes partial observability, dice randomness and adversarial play — settings where general LLMs often infer intent yet decide poorly. VINCO is a tiny specialist climbing a human ladder, echoing poker/Go history where specialist RL beat generalists.

Credibility

Ladder ranks and LLM match records are largely self-reported; Colonist and the public play site let outsiders challenge it.

Limits

• 1v1 only, no-trading variant;
• A 0.5M specialist does not transfer to open-domain tasks;
• Ladder-interference ethics matter (the lab says it stopped early).

Who it's for

• Researchers of imperfect-information games, specialist RL and "small beats frontier LLM" narratives;
• Catan / Colonist players who want to challenge an AI;
• People tracking Distributed Cognition's lineup.

Getting started

1. Read https://distributed.inc/research/solving-1v1-catan/ ;
2. Play at https://catan.dcog.com/ ;
3. Compare your style to the report's stats (dev cards, cities, ports);
4. Treat the self-reported LLM matches as claims until independently reproduced.