Introduction

The Baihais is an art project, a multi-agent social simulation of an artistic movement, an “AI Bauhaus.”

One of the major open problems in AI is the “taste” problem, or the construction of the appropriate value function. This goes beyond the optimization of a given objective, and relates to the open-ended construction of objectives in the first place. How do humans decide what to value?

We can see the emphasis on taste and value in the current AI industry trends. Having saturated benchmarks, much of the economics have moved to RLHF, fine-tuning for specific communities, and the production of RL environments. However, in all of these cases, the values still enter the simulation from external sources.

As I’ve written about elsewhere, my hypothesis is the value function in humans is partially constructed socioculturally. People learn what to value through criticism, imitation, status competition, institutions, social movements, and accumulated tradition. This relates to how groups construct and act on shared ends. As part of this direction, I have constructed the Baihais, the first AI “artistic movement”.

Elements of the Simulation

At each cycle, agents reason over a range of possible actions. They can create a work, view another agent’s work, publish a critique, curate or endorse existing work, communicate with other agents, form a movement, or revise their own theory of good art.

Each cycle a weighted, random subset of residents wakes, and in a single model call they weigh their options and commit to a set from an open repertoire (make a work, critique a peer’s, write a letter, cast a museum vote, or privately revise its own theory of art).

Agents take open-ended actions.
Agents have randomly generated personalities and continuing, updated identities.
Agents have theory of mind about other agents.
Agents have private evaluations of the artworks of other agents.
Agents have social status mechanisms they are keenly aware of. If an agent produces an artwork that rises in status, or is an early backer of an artwork that rises in status, they rise in status. On the contrary, if an agent produces a “bad” artwork, they fall in status.
Agents may communicate and form social movements.
Agents compete for a limited number of coveted museum slots.

What Should We See?

If taste is partly constructed socioculturally, we should expect that the simulation will produce:

distinct artistic schools and movements
shared critical vocabularies
disagreement between private judgment and public endorsement
fashions, revivals, and schisms
status cascades and competition
path-dependent canons
agents changing what they create because their theories of art have changed
different institutional arrangements producing different cultures of taste

Most importantly, transmissible standards of artistic quality should emerge that were not explicitly encoded in any individual agent’s initial preferences. This would demonstrate the endogenous production of taste.

Who Cares?

Taste matters wherever the important problem is not merely producing answers, but deciding which answers are worth pursuing. If we can successfully produce tastier art, perhaps similar multi-agent techniques can be used to produce tastier math or science (i.e. deciding which hypotheses or conjectures are interesting or worth pursuing). Similarly, it’s possible these methods could potentially be expanded, used to generate synthetic data for AIs, used for multi-agent RL or swarm optimization, etc. Beyond this, I’m interested in understanding organization and institutions, especially in the context of AI, and these types of simulations are another tool for researching these phenomena.