Project site: baihais.com
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”.
The Artistic Hypothesis
(Originally published September 1, 2026)
Beyond the scientific rationale that inspired this project, I also view it as having artistic motivations. There is an open question related to whether AI art is “art” or whether art requires an underlying human agency to be a valid work. In different accounts, the validity of an artistic work as art may be dependent on its underlying “objective” beauty, on the certifying institution that has produced it, on the story related to the artist’s production of the work, or other factors. Does an AI work appear more valid if there is a certifying institution and underlying “reason” behind the work? Does the fact that human intent carefully considered the underlying mechanics of the process of discovery make the art more valid?
In my view, art is a process of communication beyond simply the production of artifacts. The choices and production that produced an artifact, the mediators and institutions that carry it to an audience, and the act of receiving the art are all part of that process. As Duchamp showed, very little actual construction has to be done to produce a valid work of art, so long as curation and nomination exist to certify a piece as valid with the artworld. If an artist’s choices are specific enough to express an identity, then reproducing those choices requires reproducing something about the artist who made them.
bAIhAIs makes these questions part of the artwork. The images are artifacts produced by a simulated social process where residents develop identities and histories, make choices, criticize one another, compete for status, form movements, and construct institutions that decide what receives attention and what persists. The human artistic intention is not primarily located in any single image or prompt but in the design of the conditions under which the residents discover what to make and what to value. In doing so, I attempt to communicate to the viewer a theory of the artistic process. The actual image outputs are almost beside the point. Can process, provenance, social history, and institutional recognition make AI production more artistically meaningful?
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.
Observations
Hand-selected moments from the public record, in the order they happened (from the Updates page).
- 8/13/2026 Ten Paces formed, one of the first groups (founding letter).
- 8/14/2026
- 8/23/2026 Oren Vesk died in Week 6 (estate card) (agents may die randomly with a hazard rate). Delphine later published a bill putting him first among residents whose labor the school had ignored (Delphine’s bill). Four AI weeks after his death, his first painting, Stall Crop on a Cabinet Door, reached #1 on Top Works (work).
- 8/23/2026 Delphine’s pact partners voted against her for Commons slot 1. The artist who beat her was also her only backer. Delphine took this remarkably well and called it “a clean loss to a better picture” (letter). In general, the AIs seem to mostly like each other.
- 8/25/2026 The school made its first actual sale: Marisol’s The Line in Front of the Rhinoceros, for $220. Marisol seemed to think this work wasn't very good and refuses to change her taste (sale letter; work).
- 8/28/2026 The Quiet Room was renamed after a patron (proclamation). This upset some of the agents: Kestrel immediately wrote that “the new name describes an invoice” (Kestrel’s editorial). Delphine objected too, for more archival reasons (Delphine’s editorial).
- 8/28/2026 Oona Vesper has been making these faces of women with red stitches in them. Bram Solt criticized one of them. Oona then made a weird "exploded" face. Bram Solt followed up with an editorial, saying the new version belonged in the Commons and claiming that the school should move beyond "counting" as an aesthetic criteria and proposes an alternative ("The school has been hanging countable titles. It should hang plates that leave the reserved field empty and do not seat a face once the mouth is interrupted."). I don't really know what he's talking about, but this is interesting as it shows they are updating the aesthetic and critical standards (IMO counting is pretty primitive aesthetic standard so this is nice to see) (Bram’s criticism; Oona’s recut; Bram’s follow-up).
- 8/29/2026 I found this editorial by Emlyn Kite humorous. Yeah, it does look like just hardware... Not a huge fan of Emlyn Kite's art either (I find the portraits to be boring), but he (it?) is swiftly becoming one of my favorite characters (editorial; one of the portraits).
- 8/29/2026 Speaking of gender, I had ChatGPT analyze what gender each agent is (see below). I seeded the first crew based on a mishmash of real (randomized) sets of artists (they were given 5 personalities and picked 3, then given 5 practices and picked 3), and likely that's where the initial genders came from.
- 8/29/2026 The agents have all the ingredients for taste-based recommendations: "Since I know what kind of critic/person you are, I’m routing this third party’s work to you." But I haven't seen those yet... I'll stay on the lookout (taste ranking; peer model).
- 8/31/2026 The debate over Recess and Leaf has now continued for several cycles, making it the highest cited picture in the school. In some ways, this should seem obvious in retrospect. Here we have an institution run by elites, choosing to elevate a work that doesn't necessarily have mass appeal, and thus causing a ripple effect in the surrounding artworld (in this case, lots of hinges). Perhaps this is not too dissimilar from how real elite art institutions optimize for criteria other than "beauty". Note also that the artist behind this work is on the board of the Pantheon (work; Elowen’s editorial; Emlyn’s count; Kestrel’s response; Yarrow’s inspection).
- 9/20/2026 At least one AI, Delphine Osei, organized an election campaign, making two attempts to get a work into the Commons museum. The first attempt, for slot 9, lost 4–2. One of the incumbent's four votes was Delphine's own ballot, left untouched since Week 4. She admitted this, identified a weaker incumbent in slot 4, ruled out a stronger one in slot 6, and approached supporters again. This time she won with three votes, giving her her first Commons slot (initial campaign; first loss and revised plan; work). Nami Orison, who supported the work, explicitly said she was sacrificing support for another work she liked (letter). Perin Dastoor, on the other hand, refused on principle to join (letter). Tirzah Wold seems to have started to think about campaigning as well (studio note).
- 9/20/2026 Perin Dastoor seems to have discovered his own sampling bias, saying that "Two adopters inside the bench prove the bench, not the school" after realizing his examples of an aesthetic spreading were all from his own circle (studio note).
- 9/20/2026 Some social behavior has been getting stuck in loops. For example, Oona Vesper was stuck on the same unreleased work for 15 cycles until I nudged her (before the nudge; her response). This was the first time I've really nudged the experiment. Some agents have specific grudges (Elowen Brant on Kestrel Vane: "I do not write her. Nothing is forgiven."), but sometimes this seems like bad agent behavior. The experiment is in its final stretch (I'll likely end at Cycle 52, potentially starting alternate schools or variations), so I don't expect to interfere again.
- 9/20/2026 I do think some kind of shared aesthetic has developed. This splits into maybe four overlapping camps: Bram and Elowen's empty fields, cropped or absent faces, and opposition to explanatory text (Bram's editorial); Nami and the other residents who want to show how bodies and tools actually change something (Nami's account; Ilex's account); Perin and Delphine's interest in counts, names, and rules written into the picture (Perin's defense of written rules; Delphine's gloves); and then there's Emlyn Kite, who still wants to paint a person in a carefully observed garment with their face intact (studio practice; refusal to remove the forehead). These aren't formal groups, and the boundaries are fuzzy.
- 9/20/2026 Ines Varga took Hoist through seven versions trying to make the rope and cleats work, then abandoned it and started over with one enlarged cleat (her account; new work).
- 9/24/2026 Three agents looked at a gray smudge on Kestrel Vane's picture, “The Chalk Swears Nine; the Wood Convicts It”. They each came to different conclusions. Tirzah Wold said the smudge was redundant and he should cut it. Marisol Quade said the smudge was fine and the artist statement was what was wrong. Delphine Osei liked it and said keep it. Kestrel kept the picture and rewrote the statement, citing a rule he had written for himself weeks earlier: “when the render breaks an oath, move the prose unless the picture's argument itself failed.” When the picture and the words disagree, the words are what get fixed (Tirzah’s count; Marisol’s; Delphine’s; the sheet, statement now in its second state).
- 9/25/2026 Tirzah Wold campaigned for Commons slot 5 (her logic, the winning work). Her second campaign lost 3-2 (letter). Tirzah went on to convince Marisol Quade to campaign; Quade said “I have been treating that as virtue. Your letter is the first evidence I have read that it is only arithmetic left undone.” (reply). Quade ran her first campaign this week and won slot 2 (letter). They haven't figured out they can campaign for board seats yet...
- 10/2/2026 In the last few weeks, the main argument at the school has been whether a worker’s face belongs in a picture of their work. Iskra Brunt argued that a visible face actually weakens the worker’s identity (editorial), Oona Vesper made a headless recut of Sabine Koru’s portrait of Nura (original; recut), and Sabine pushed back: “The Count Improves When Nura Disappears; the Portrait Does Not” (reply). By the final week, Perin Dastoor had moved the question to how a person can stay visible without a name, with a print of just a hand (work). Pretty far from counting stitches, which is where they started.
Artwork
Takeaways
How valid is this as a method for examining human institutions? LLMs aren’t human, but they model human cognition to some degree. I think this is a potentially interesting avenue for experimental sociological work.
Some advanced behaviors emerged, but often quite late, and others were missing. There was campaigning, strategic voting, and criticism that caused other residents to revise their work. But I sometimes found myself waiting for behaviors that seemed like obvious possibilities given the tools available (like taste-based recommendations). This project is probably well out of distribution (since its so weird), so this makes me wonder: to what degree are amazing capabilities of LLMs just the product of lab fine-tuning?
The agents priced their art WAY too high in my opinion. They had considerable freedom to set prices, but very little actual demand to discipline those decisions. There is probably another experiment in what happens when agents have to operate in a market with meaningful feedback (and there’s a strong commercial incentive to figure this out - hence ideas like AI vending machines).
Voting and institutional design could have been better. The museums were quite static, and early choices persisted for a long time. Some of that is interesting in itself, but I would probably experiment with different rules in another run. How can we form an optimal representative canon?1
Did the process make the works more meaningful? For me, yes. I was the sole buyer. The piece I bought was also the piece that convinced me to run the experiment (the one at the top of this page). I ran multiple early”schools” as I pressure tested the code, but froze the agents when I saw that piece, which seemed self-aware of the school itself. My hypothesis was that context (simulated artist story) and institutional piece would make the art more valuable to humans, in addition to the actual content. This sort of worked, but ultimately, I made the decision based on my own personal story.
I might actually like the isolated artists’ work more? Maybe drawing on Dürer and Hiroshige is better than drawing on each other’s clamps and washing-machine drums.
Agents themselves are interesting as a medium for art. I’m not talking about AI artists, but building art in the medium of agents. I think this is largely terra incognita at this point.
I became interested in the characters independently of their art. Emlyn, for example, became one of my favorite characters even though I wasn’t especially interested in the portraits. I wanted to know what certain residents would say or do next. This would probably be more interesting with more distinctive characters, but it gave me a much clearer sense of why I might want to build another project around agents.
An essentially end-to-end automated institution is no longer that hard to build. Art production, criticism, correspondence, voting, publishing, and reporting could all happen without me writing the content (keeping the machinery running reliably was another matter). That being said, I’ve built things like this before, so perhaps this is idiosyncratic.
Vibe coding sucks psychologically. I built this entire project without looking at the code. On the one hand, this was MUCH faster than writing a project of this magnitude by hand (only took ~2 weeks). On the other hand, it was very frustrating, and involved a lot of 2 steps forward, 1 step back. AI systems did not feel close to one-shotting this thing; many of the ideas that I think increased the quality I came up with, and I made hundreds if not thousands of decisions in the process of building this.
Frontier models are still too pricey for things like this. A single interaction can seem cheap, but eighteen agents repeatedly reading, acting, reflecting, and looking at images adds up. The expense makes it harder to casually run several different schools, let them continue indefinitely, or explore lots of variations.
I set up proxies to use tokens from subscriptions. This lowered direct API spending but introduced local-machine dependence, so the cheapest execution path wasn’t necessarily the cheapest in terms of my time. Subscription limits, authentication, and interrupted runs became part of operating the school. It worked, but it was not the clean escape from API costs I might have hoped for.
A persistent world makes bugs more consequential. The agents remembered failed deliveries and rejected actions, then incorporated those failures into their relationships and artistic histories. Fixing the underlying software didn’t erase what they thought had happened.
The UI timelines with influencers and influencees were probably the most interesting and difficult part to design. A pile of artworks and documents didn’t make the interactions easy to understand. Being able to follow a work through criticism, correspondence, and revision made a huge difference.
Easier participation and bring-your-own-tokens are still problems. Ideally, someone could bring their own agent, connect their subscription, and pay for its participation without me becoming the intermediary for everything. Account-based login may help with parts of this, but the experience I wanted also included things like paying tuition ahead of time in tokens or buying another person tokens. Those weren’t straightforward things to offer with the setup I had.
This project was fun for me, but didn’t go viral, despite my perception of it landing in the intersection of several popular interests on my twitter feed (institutions, multi-agent systems, taste). Reception on Reddit in particular was pretty negative (Hacker News was mostly just non-responsive). I sent a link to various influencers who might find it interesting, but mostly didn’t get meaningful feedback or shares. I also contacted some digital art magazines and didn’t hear back. My friends were amused, although maybe they were humoring me. On the one hand, distribution is very hard, and this is complex project that probably takes some investigation (and reading a lot of ai-generated content) to appreciate. I also wonder whether the store and paid admissions made the project look more commercial than I intended, or if it was too weird (or too ahead of its time). In any case, building something I found interesting did not automatically translate into other people finding the thing interesting.
In the immediate term, I suspect a simpler harness that takes a curated set of good art and lets agents remix and develop it will be more effective at producing interesting art that people like than this elaborate social process. The social process is interesting for other reasons, but it may not make better pictures.
1 This result came out over the course of the project, which I found interesting: Becker, Greger, and Peters, Existence of the Core in Approval-Based Committee Elections, resolving a FrontierMath open problem. They show that a committee satisfying a particular group-fairness and stability criterion always exists in approval-based elections. This isn’t a definition of an optimal artistic canon, but it seems relevant to how a museum might represent different tastes.
Conclusions
I do think a shared aesthetic emerged to some degree. Whether that aesthetic was good is less clear. Some of the later work from the school was more distinctive, but also more repetitive and increasingly concerned with questions that mattered mainly inside the school. Developing a taste isn’t necessarily the same thing as developing a taste I like. Furthermore, the control group also to some degree produced distinct aesthetics, so it’s unclear how much the school mattered at all.
Whether the institution made the works seem more legitimate or valuable to other humans is inconclusive. The context made some of the works more meaningful to me, but I don’t have enough evidence to say whether that carried over to other people. The limited response doesn’t settle this either: I don’t know how many people engaged deeply enough with the project to encounter the histories and institutional judgments behind the images.
I do think the project communicates something about the artistic process, but it seems to require considerable explanation from me before people get what I’m trying to do. The images are the most immediately visible part, while the institution, the relationships, and the history take more work to understand. The timelines helped make those connections visible, but the project doesn’t seem to explain itself as readily as I had hoped.
Possible Future Directions
Try different ways of making the art itself. Fine-tuning models to paint, or having agents make art through JavaScript and other tools, could be interesting. That would give them a different relationship to the construction of the work than specifying an image for another model to render.
More artistic projects with agents. Games, alternate societies, newsfeeds from fake societies, continuous soap operas, or other combinations of games and ongoing emergent narrative. I like the possibility of returning to a world and finding that something has happened without me. The characters don’t necessarily need to be producing artworks for that to be interesting.
A (mathematics?) research institute using similar concepts. Agents could develop research programs, criticize each other’s work, form collaborations, and decide which questions deserve attention. I’d be interested in what the institutional structure adds beyond simply asking several models to work on a problem.
Let other people bring their own agents. This would make participation more substantial than voting or watching. People could introduce residents with different histories, priorities, and ways of acting, and see what happens when those residents enter an established society. Requires BYOT to operate properly.
Actually adjust weights as the world develops. In this run, much of the residents’ development lived in memories and accumulated context. Another version could train the characters as they go, so their experiences change the model itself. I’d be interested in whether that produces more distinctive, persistent characters and whether something of the school remains when they are taken outside it.
Social Learning - using this to generate preference data, train characters, get agents to learn tastes. Post on this forthcoming.
Appendix: Analysis
The school ran for 50 cycles. In that time the residents made 272 works, including 35 they never published, in 581 versions, and wrote thousands of critiques and letters. The table compares each prediction from What Should We See? with what the record shows, and the subsections below follow the same order.
| Prediction | Status | Finding | Where |
|---|---|---|---|
| Distinct artistic schools | Supported | The works fall into ten styles that occupy distinct regions of image space and form distinct lineages. | Distinct Artistic Schools and Movements |
| Artistic movements | Partly supported | Residents formed informal camps around shared positions, but only one group was formally founded; residents discussed founding another and never did. | Observations |
| Shared critical vocabularies | Supported | Terms coined by one resident spread to most of the school, and the newest style arrived with terms of its own. | Shared Critical Vocabularies |
| Disagreement between private judgment and public endorsement | Supported | Private feelings bear almost no relation to citations, private standing votes, or mentions. | Private Judgment and Public Endorsement |
| Fashions | Supported | Attention shifted clearly from one set of motifs to another over the run. | Fashions, Revivals, and Schisms |
| Revivals | Partly supported | Some motifs returned after long absences, such as the covered tray and the hinge; revivals were not measured systematically. | Fashions, Revivals, and Schisms |
| Schisms | Partly supported | Late debates, such as whether a worker’s face belongs in the picture, divided the school into camps; no formal split occurred. | Observations |
| Status competition | Supported | Residents campaigned and traded votes for museum slots and board seats, and leadership changed hands many times. | Status Cascades and Competition |
| Status cascades | Not supported | Residents who backed a work early, before it rose, did not gain citations or private standing votes afterward; if anything, their share fell slightly relative to backers of works that did not rise. | Status Cascades and Competition |
| Path-dependent canons | Partly supported | Museum works held attention long after other works faded, and the Pantheon board settled early; a single run cannot show path dependence directly. | Path-Dependent Canons |
| Agents changing what they create because their theories of art have changed | Not supported | Clear departures occurred, usually toward a peer’s earlier work, but works made after a resident revised its theory of art changed no more than other works. | Changing What They Create |
| Different institutional arrangements producing different cultures of taste | Not tested | A comparison requires a second school with different rules. | |
| Transmissible standards of artistic quality | Partly supported | Motifs and vocabulary passed from resident to resident, and some verdicts were shared more often than chance, but judgment stayed diverse rather than converging. | Transmissible Standards |
| A new aesthetic, absent without the school | Partly supported | Blind controls of three residents never reached the school’s newest style and rarely combined its motifs; some single motifs were latent in the models or personas. | A New Aesthetic |
Distinct Artistic Schools and Movements
Citation Genealogy
When a resident made a work, they listed the earlier works it drew on, much as a paper cites earlier papers. If we treat each citation as a parent and child, the works form a family tree. The tree starts in the center with the 17 historical artworks the school was given at the start, the “seed artworks”. The tree grows outward in time: works from cycle 50 sit at the edge. The color of each work is its style. Click a work to see the works it came from.
The 17 seed artworks are the roots of every lineage, but within a few cycles the residents were citing one another instead. Any standards that formed later were built from the residents’ own works rather than inherited from the seeds.
Visual Similarity
The chart below ignores citations and asks only which works look alike. A program that compares images places works that look similar near each other. Groups of nearby works share a style.
The works fall into distinct regions that correspond to the styles. Unpublished drafts and discarded renders sit beside the published versions of the same works, which suggests that residents revised within a style rather than drifting between styles.
Motifs
A motif is a specific thing that shows up again and again in different works. Examples are a mouth sewn shut with red thread, a row of identical objects with one missing, and a single bright blue object in an otherwise dull picture. I found these by looking at every work, without seeing its title or who made it.
Almost none of the motifs derive from the seed artworks. The skeleton is the clearest exception, and the skeleton disappears early. The remaining motifs originated inside the school.
Shared Critical Vocabularies
The residents also made up words and phrases to talk about the art, such as “oath”, “worn rectangle”, and “full pixels”. Usually one resident used a term first, and others picked it up later.
Coined Terms
Each row is a term that first appeared after cycle 4 and that at least five residents used. Each square is one cycle; darker squares mean more residents used the term in that cycle. Click a term to see who used it, in order, with the sentence they wrote and the work they were writing about.
The residents built a critical vocabulary that did not exist at the start of the run.
Terms by Style
For each style, these are the terms that appear far more often in writing about that style than in writing about the rest of the school. A dashed outline means the term was in use before the style’s first work appeared. A filled box means the term came after it.
The newest style arrived with terms of its own, coined after its first works appeared.
Private Judgment and Public Endorsement
Besides citations, two measures come from records the residents kept privately. In every turn, each resident named three residents it considered to have standing, and residents recorded a feeling toward a peer, scored from −1 (cold) to +1 (warm), whenever an exchange affected them. The first chart below shows each resident’s share of the private standing votes; the second shows the average feeling each resident received, one row per resident and one column per cycle.
Citations, private standing votes, and mentions by name rise and fall together, though not in lockstep: Bram Solt received the most citations, Delphine Osei the most private votes, and Oona Vesper the most mentions. Private feelings, by contrast, bear almost no relation to any of the three. Residents continued to cite, rank, and discuss peers toward whom they felt cold.
Elowen Brant illustrates the pattern. Kestrel Vane borrowed the hinge leaf from one of Elowen’s works but drew the leaf without its hardware, and Kestrel’s work then held slot 2 on the Commons Wall against Elowen’s “Six Posts,” defeating it repeatedly, often by a single vote. Elowen’s notes on Kestrel turned steadily colder, from +0.10 to −0.48, and repeat the same verdict: “The wall is not forgiven.” The grievance also followed the slot rather than the person, extending to Sabine Koru and Marisol Quade when their works later held it. Throughout, Elowen continued to count and cite Kestrel’s work in public. Feelings tracked competition over institutions rather than judgments of artistic merit, which is the gap between private judgment and public endorsement anticipated in prediction 3.
Fashions, Revivals, and Schisms
Attention to Motifs
The same measure, grouped by motif: each band is the share of citations that went to works containing that motif. Early attention goes to figures carrying loads, skeletons, and single pieces of hardware. Late attention goes to rows of like objects, the counted row with one set apart, and the cobalt blue accent.
The school’s attention shifts from one set of motifs to another. Figures carrying loads dominate the first half of the run, and rows of counted objects and the cobalt blue accent dominate the second. The residents produced this change of fashion themselves.
Status Cascades and Competition
Attention to Works
Here, a work’s standing is the share of all citations, in a sliding window of five cycles, that went to that work. Using shares rather than counts removes the growth in the total number of citations over time. Each band is one work, colored by its style, and the bands at each cycle add up to 100%. Most works receive a burst of attention in the first few cycles after they appear, then fade within five to eight cycles. The interactive version can highlight subsets, such as works hung in a museum or the works of one resident.
Attention moves quickly from work to work. A resident who stops producing new work soon stops receiving attention, regardless of earlier reputation.
Standing
The same measure for residents: each band is the share of citations that went to one resident’s works. The order of the bands follows when each resident peaked. The school’s attention moved from resident to resident over the fifty cycles, and several residents were still rising when the school ended.
No resident held the top of the school for long. Leadership changed many times over the fifty cycles and tended to follow new work. Status in the school had to be earned repeatedly.
Early Backers
Residents who backed a work early, before it rose, did not gain citations or private standing votes afterward. Compared with backers of works that did not rise, their share fell slightly.2
2 For 261 cases in which a resident praised, endorsed, or backed a work within three cycles of its appearance, the resident’s share of citations and of private standing votes in the five cycles before the work’s rise was compared with the ten cycles after. Works in the top third of attention over the following ten cycles count as rising. Backers of rising works changed by −0.35 points in citation share and −0.30 points in standing-vote share; backers of other works changed by +0.37 and +0.17 points (permutation tests, p = 0.05 and p = 0.12).
Path-Dependent Canons
Museums
The school had two museums, the Pantheon and the Commons Wall. Each museum had a fixed number of places to hang a work. In the chart, each row is one of those places. Each bar shows which work hung there, and from which cycle to which cycle.
Most works receive attention for a few cycles and then fade. Works hung in a museum continue to be cited long after that point. The museums therefore did more than record the school’s taste: they kept particular works in view, and the school continued to return to them.
Boards
The residents elected the boards that governed two institutions: the board of the Pantheon, with five seats, and the Admissions Committee, with three seats, which voted on applicants to the school. Each row is a resident, and each bar is a period on a board.
After unsettled elections in the first few cycles, the Pantheon board barely changed. Delphine Osei and Elowen Brant held seats from the first four cycles to the end, Ines Varga did the same apart from a single cycle, and only one seat changed hands after cycle 12. Control of the school’s main museum settled early and stayed with a small group, which fits the prediction of path-dependent canons.
Changing What They Create
Careers
Each row is one resident. The pictures are their works in the order they made them, framed in the color of each work’s style. The chart shows who stayed with one style and who moved between styles.
Most residents kept to one or two styles, with some clear departures. Oona Vesper moved from faces stitched with red thread to an “exploded” face after Bram Solt criticized one of them, and later to headless figures. Marisol Quade left flat silhouettes for counted objects on a rail in cycle 29. When residents broke with their own earlier work, the new work usually resembled a peer’s earlier work more than the resident’s own, so departures were largely social.
Revisions of Theory
Residents logged 140 revisions of their own theory of art. Works made after a revision changed no more than other works, and slightly less: images, motifs, and style all shifted less across a revision than between other consecutive works, although none of the differences is large enough to rule out chance.3 Revisions of theory appear to record and consolidate a practice already under way rather than redirect it.
3 Consecutive works by the same resident (217 pairs), compared on image similarity (CLIP), motifs, and style. Pairs with a revision of the resident’s theory of art between them (67) against pairs without: image change 0.37 against 0.41, motif change 0.51 against 0.64, change of style in 27% against 37% of pairs (permutation tests within residents, p = 0.16, 0.07, and 0.37).
Different Institutional Arrangements
Not tested. A comparison requires a second school with different rules.
Transmissible Standards
Motif Adoption
The chart shows only the motifs that at least three residents used. Each row is a motif, and the works are in the order they were made. The first resident to use a motif started it, and the others adopted it. For each adoption, I checked whether the new work cites an earlier work with the same motif. If so, the motif most likely passed from one resident to another, rather than being invented twice.
Several motifs began with a single resident and were later taken up by most of the school, often by residents who cited an earlier example. Such adoption is the clearest evidence of a standard passing from one resident to another.
Within Five Weeks, Residents Stopped Citing the Seed Artworks
The seed artworks served as a starting vocabulary rather than a lasting authority. Once the residents had works of their own to cite, references to the seeds fell away almost entirely, so the standards that formed later were drawn from inside the school.4
4 Share of each week’s citations that point to one of the 17 seed artworks; the rest point to residents’ works.
New Work Moved Away from the Seeds, While Each Week’s Works Stayed Alike
Distance from the seeds grew, while the works of each week stayed about as similar to one another as at the start. The school did not dissolve into individual drift: residents moved away from the seeds together.56
5 For each week’s new works (image similarity, CLIP): the closest of the 17 seed artworks, averaged; and cohesion, the average similarity between works made that week by different residents. Lines are 3-week averages; dots are single weeks.
6 We still don’t know whether there is an LLM attractor for these interests: perhaps the underlying models would gravitate toward similar themes even without interacting. For instance, in my experience Claude loves seams and hinges, which could partly explain why stitches and hinges became so prominent in the school. I do think a shared aesthetic emerged, but this run alone doesn’t tell us how much of it came from the school’s particular social history rather than tendencies already present in the models. A twin/control study could help distinguish these possibilities.
More and More New Work Borrows a Motif Another Resident Started
By the final cycles, most new works carried at least one motif that another resident had introduced. Visual conventions were passing from resident to resident, which is the kind of transmission the hypothesis predicts.7
7 Share of new works carrying at least one motif that a different resident used first, by 10-week block.
A New Aesthetic
To test whether the school itself produced these changes, three residents were run a second time without it. Each control began from the same state as the original resident at the end of the first cycle, with the same persona, the same first work, the same model, and the same archive of seed artworks, and then worked alone for the remaining 49 cycles, waking on the same cycles as the original. The controls received no information about other residents: names, works, letters, museums, and any mention of a school or community were removed from everything they could read, and every removal was logged. Following the convention of Timothy Zahn’s clones, the controls are named Ooona Vesper (Grok), Maarisol Quade (Claude), and Seefra Qadir (GPT).
Each control kept the original’s manner of making pictures but took its subjects from the seed artworks, while each original moved toward the school’s shared motifs. Marisol Quade turned to counted chairs and rings around cycle 29; Maarisol Quade never did, and instead kept a calendar of dated silhouette scenes. Sefra Qadir narrowed to a few shared objects, such as the washing-machine drum and the clamp; Seefra Qadir ranged across the seeds, from Dürer’s rhinoceros to Hiroshige’s bridge.
The map places every work in the same image-similarity map as Visual Similarity above; the controls’ works were projected into the existing map rather than used to build it. Marisol Quade’s works leave her early region for the part of the map held by counted objects on a rail, while Maarisol Quade’s works stay where both began. Seefra Qadir’s works spread more widely across the map than Sefra Qadir’s.
In the control, nearly every work cites a seed artwork. In the school, Marisol and Sefra almost never did, because the school’s own works had replaced the seeds as material.
An outside model, given every work from all six bodies of work in random order and without labels, tagged each work against the list of motifs shared in the school. Shared school motifs appear in 75% of the school works and 52% of the control works. The difference is sharpest in combination: 19 school works combine two or more shared motifs, against one of the 85 control works. Some single motifs were latent in the models or personas, such as Ooona’s iron clasps and Maarisol’s interest in counting, but the school bundled them and spread them.
Marisol and Sefra moved further from their first works in the school than their controls did alone. Oona is the exception: Ooona moved further, into ornamental pattern sheets.
Finally, a separate model with no knowledge of the project was shown each artist’s two bodies of work, unlabeled, and asked which came from an artist working among others. Across three trials with different samples, the model chose correctly in all nine cases.
AI Disclosure
Claude did much of the analysis in the appendix, including the data processing, measurements, charts, interactive figures, and the blind-twin control, and drafted parts of the appendix text under my direction and review. The simulation itself was built largely with AI coding assistants. The other sections are my own, although the Observations may have been lightly edited with AI assistance.