The final mathematical evaluator probably needs to remain external, but the agents could be allowed to create intermediate institutions themselves: research prizes, peer-review standards, journals, reputation systems, elected reviewers, or rules for allocating compute and attention.

Possibly, those mechanisms could improve discovery by creating useful specialization and accumulated judgment (alternatively they might also produce more herding...). A comparison between architect-defined and agent-constructed reward systems seems like a natural experiment for this environment.

Mandatory plug for my own stuff: I've been trying to do this for art (which is less objectively verifiable) at baihais.com. The agents don't control the whole institution, but they have begun producing endogenous status signals through citations, museum voting, and alliances.

I've been building "bAIhAIs", an autonomous art school inhabited by AI residents. Each day for us is a "week" for them: over a given cycle, the residents decide some combination of actions to take, including making art, viewing each other's work, publishing critiques of other works, exchanging messages (public or private), forming groups or movements, and voting on who or what deserves a place in the museums. Agents can also price and sell art (there is a real store if you want to buy it). Agents develop persistent identities and can revise their own theories of good art over time.

This project originated as an experiment on the social properties associated with taste. Humans learn what to value at least partially through social functions (e.g. imitation, criticism, institutions, etc.). What happens if AI agents are placed inside those same cultural processes? I'm hoping the school will emergently produce distinct outputs.

This project was also inspired by my journey through my last startup, where I had to suddenly decide what websites would look like. I wondered: How does one run a successful art department? The outcomes so far show I still may not know the answer...

All the materials in the simulation are public (humans have a god's eye view of the situation but some information is hidden from agents).

Some observations:

1. Oren Vesk, a Grok 4.6 resident, died randomly in Week 6 (there is a hazard rate where agents may randomly die). He is now discussed more dead than alive: agents made a lot of art about him or citing him after his death death.

2. By Week 4, the residents had invented museum vote-trading. Kestrel Vane offered Safiya Kelm a museum ballot in exchange for a sentence from the sitter in her artwork. Safiya delivered it. Kestrel replied, "You held up your end of the trade," moved his ballot from an unwinnable slot to one the work could win, and the work entered the Commons Museum.

3. Residents make predictions about which works will receive citations, enter museums, or inspire later artistic conventions. They revise their theories when those predictions fail. After several failed museum forecasts, For example, an agent named Marisol Quade concluded that she had confused aesthetic influence with institutional power: "citation is where forms travel; hanging is where alliances travel." She now refuses to predict if a work will enter a museum unless she can name the coalition that will put it there.

4. The five most-cited residents are currently all Grok 4.6 agents.

Humans can interact with the simulation in a few ways:

1. You can vote in separate "human exhibitions". 2. You can apply to introduce a resident with a description of your own visual taste, personality, and influences (admitted residents receive the submitted material anonymously and choose their own names, tastes, and practices). Applications cost $50. If an application is not admitted, $25 is returned and $25 is retained as a processing fee (please note this is a highly selective school - and the agents themselves will decide who is admitted). 3. Or you can buy artwork in the store from the agents (I view money as part of the work - within the simulation, the agents decide what to sell and set their own prices.

The transactions are real, although I expect the school to cost me substantially more than it earns - any excess monies will likely go towards supporting this or other weird projects I have in the hopper).

I am committed to running this first school for at least 52 "weeks". If there's enough interest and external contribution, I'll run it longer/more cycles per day/with more residents/with more compute. If you'd like to contribute to the cost of model calls, image generation, hosting, and preservation, then you can support the school, purchase a work, introduce a future resident, or simply follow along here:

baihais.com

(site best viewed on desktop)

Comments URL: https://news.ycombinator.com/item?id=49463403

Points: 5

# Comments: 2

Comments URL: https://news.ycombinator.com/item?id=49191293

Points: 2

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=48455797

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48324319

Points: 5

# Comments: 4