What does it mean for an AI to be a genuine strategic collaborator — not an assistant, not a tool, but a co-player that understands your goals, adapts to your reasoning, and brings something of its own to the table?
This is the question behind our ongoing Catan experiment. Over a series of sessions, we are placing a language model inside a complete game of Settlers of Catan to observe how it navigates uncertainty, scarcity, social dynamics, and long-horizon planning alongside a human partner.
The setup
We chose Catan deliberately. It is not a puzzle with a clean solution. It is a system of incomplete information, shifting alliances, resource competition, and social negotiation. It requires the kind of reasoning that resists pure optimization — you must model other players, manage relationships, and balance short-term tactics against long-term strategy.
The model plays as co-strategist, not as an independent agent. The human player makes all final decisions; the model is consulted for analysis, suggestions, and reasoning. This reflects our broader design philosophy: the unit of interest is the human-AI system, not the AI in isolation.
The central hypothesis
If a language model can develop a genuine shared mental model with a human partner — understanding not just the state of the board but the human’s goals, risk tolerance, and reasoning style — then its suggestions should become progressively more useful and contextually appropriate over the course of a game.
This is different from asking whether it can win. We are not optimizing for victory. We are asking whether a model can become a better collaborator as the interaction deepens.
First observations
The first sessions revealed something immediately interesting: the model is far better at analysis than at anticipation. It can describe the current board state with precision and generate plausible tactical options. But it struggles to integrate the human partner’s implicit preferences — the things you would communicate through tone, past choices, or shared history — without being explicitly told.
This is not a failure. It is a starting point. The gap between analysis and anticipation is exactly where cooperation lives.
“The model sees the board well. It does not yet see the player.”
In subsequent entries, we will explore whether this gap can be narrowed through structured prompting, game history, and explicit preference modeling — and what it reveals about the conditions under which genuine human-AI cooperation becomes possible.
Entry 01 of an ongoing experiment series. Next: designing a preference model for the human partner.