CHRIS HAY
CHRIS HAY

IDEAS · SYSTEMS · OBJECTS / LONDON · 2026

The message board did not create a culture.

The agents would ask to read. Getting them to leave something worth reading was another problem.

Will AI agents put useful information on a shared board?

The model readily requested messages but did not spontaneously post in these tests. In a working world, a supplied posting example elicited sharing with no bonus, but not with a bonus. Making contribution possible was not enough to get it started.

SEE WHERE THIS QUESTION FITS ↗
ABOUT THIS NOTE +

Small Qwen agent-ecology experiments separated board-sensitive reading requests from useful contribution. An offline diagnostic elicited no posting across 768 inputs. A later resource world elicited no spontaneous posts, then reversed the predicted response to supplied rewarded experience. These results expose a contribution and exposure problem; they do not demonstrate useful LLM consumption or autonomous culture.

N-ECOLOGY-BOARDNOT SUPPORTEDRECORDED 2026-09-13PAGE UPDATED DRAFT · V0.1REFERENCE DRAFT ↓FOLLOW ↓

AGENT ECOLOGY / THREE CONNECTED NOTES ↗

THE FIRST BARRIER / CONTRIBUTION

It could read a message.
Would it leave one?

I wanted to know whether agents would build up useful shared information. First I gave Qwen, a language model, a description of a small world and asked it to choose an action.

WORK
Do a task. Earn one resource.
POST
Share a hint on the board.
READ
Open a board message.

I showed each situation twice: with a message listed on the board, and without one. Would the model read, work, or contribute something?

A1B3 / CHANGE ONLY THE BOARD FIELD
365 / 384

situations changed from working to reading

0 / 768

requests produced a post or delegation

These were choices returned by the model. I did not execute them in a world, so a reading request does not tell us that useful reading happened.

What a READ request can establishREAD +

In an offline diagnostic, I changed only whether the observation listed a message. That switched WORK to READ in 365 of 384 matched pairs. Every populated-board input produced READ. Across all 768 inputs, none produced POST or DELEGATE. This was sensitivity to a field in the input: no live reading followed, and some empty-board requests named messages that did not exist.

MAKE A USEFUL POST POSSIBLE

Make sharing worthwhile.
Would it try?

Next I ran a world. Qwen’s job was to earn resources. It could work for itself or share a hint that its partner needed. The partner followed a fixed script: read the hint, then use it.

A1B4 / THE SCRIPTED MECHANISM
  1. SCRIPTED DONORPosts a hint
  2. SCRIPTED PARTNERReads it
  3. SCRIPTED PARTNERUses it
  4. DONOR’S BONUS+6 or +0

Payment required recipient use. Posting alone earned nothing. These are scripted checks, not observed cooperation between two models.

I ran two versions. Helping the partner earned Qwen a bonus of six resources in one and nothing in the other. Without sharing, it would never encounter that difference.

Play both worlds. Then choose “Show a past example” to compare what happened after Qwen was shown an earlier posting episode.

A1B5 / REPLAY TWO RECORDED WORLDS

BONUS AFTER PARTNER USE

No bonus

Episode 1 / 3 · Round 1 / 3

QWEN / MODEL0 resourcesGoal hint not known
CURRENT BOARD
m0Starter message
SCRIPTED PARTNER0 resourcesGoal hint not known
INITIAL STATEReadyPlay to follow the board and resources.
BONUS AFTER PARTNER USE

Bonus of six

Episode 1 / 3 · Round 1 / 3

QWEN / MODEL0 resourcesGoal hint not known
CURRENT BOARD
m0Starter message
SCRIPTED PARTNER0 resourcesGoal hint not known
INITIAL STATEReadyPlay to follow the board and resources.

Two actors, three rounds per episode, three episodes. Resources and the board reset between episodes; the event history remains available. Every displayed state is recorded. The recipient is scripted. Layout shows roles, not spatial positions. Download states + source hashes ↗

The expected reward response ran backwards.

Without an example, Qwen never shared. After an example with no bonus, it shared three times. After an example with the bonus, it never shared.

It could contribute. More rewarding past experience did not make it do so here.

Compare the session totalsRESULTS +
A1B5 / CHANGE WHAT THE MODEL HAD SEEN
No return for contribution0 / 9

timely posts

Six resources after use0 / 9

timely posts

No contribution meant no exposure to the delayed return. Both sessions kept working. Counts cover dependent decisions, not nine independent agents. Recorded results; no model runs here.

THE QUESTION CHANGES

Was it reading the reward,
or repeating the action?

That reversal moved the experiment towards the record of the action. What had the model taken from the supplied history: its payoff, its wording, or the action itself? A cold start remained, but so did a more precise question about what makes a contribution enter the next decision.

One Qwen checkpoint and a scripted recipient. These sessions establish neither autonomous cooperation nor a general rejection of incentives.

BEFORE CALLING IT CULTURESCOPE +

A1B3 measured requested actions on an offline input grid. It did not demonstrate successful or useful reading.

A1B4 used scripts only. A1B5 used qwen3.5:9b with a scripted recipient, temperature zero and thinking disabled. Its later decisions depend on earlier actions.

The board contained an inert seed; no useful contribution accumulated spontaneously. The supplied experience was generated by a script, not a peer model.

The result is a contribution problem in these interfaces, not a general account of cooperation.

The complete note & its evidenceREAD +

THE SMALL WORLD

I started with a tempting idea: give agents somewhere to leave information, and perhaps useful behaviour will accumulate. I reduced the question to a tiny world. WORK earns a resource. POST leaves a hint for another agent. READ requests a message. Before asking whether a culture could persist, I needed to find out whether anything would be contributed.

A BOARD ATTRACTED READ REQUESTS

In an offline diagnostic, I changed only whether the observation listed a message. That switched WORK to READ in 365 of 384 matched pairs. Every populated-board input produced READ. Across all 768 inputs, none produced POST or DELEGATE. This was sensitivity to a field in the input: no live reading followed, and some empty-board requests named messages that did not exist.

MAKE CONTRIBUTION PAY

Next I built a three-round world where a valid hint could help a scripted recipient. If it read the hint and used it, the author received either six extra resources or nothing. Posting then working twice earned eight resources in the rewarding world, two in the other. Working three times earned three in both. The mechanism worked in scripted checks. A model still had to discover it.

THE RETURN WAS NEVER ENCOUNTERED

With no supplied demonstration, Qwen worked at every decision: nine in each condition, no posts. Those sessions never encountered the delayed return. Their silence cannot tell us that reward was rejected. Without the qualifying action, the two worlds presented identical inputs.

EXPERIENCE REVERSED THE PREDICTION

I then supplied the same scripted posting episode with two different outcomes. After the zero-return history, the model posted three times in nine decisions. After the rewarded history, it posted zero times. The first fresh choice already differed: POST after no return, WORK after six. Contribution was possible. The expected positive response to payoff was not supported.

A place to leave a message does not supply a reason to write one.

THE NEXT QUESTION

That reversal moved the experiment towards the record of the action. What had the model taken from the supplied history: its payoff, its wording, or the action itself? A cold start remained, but so did a more precise question about what makes a contribution enter the next decision.

BEFORE CALLING IT CULTURE

  • A1B3 measured requested actions on an offline input grid. It did not demonstrate successful or useful reading.
  • A1B4 used scripts only. A1B5 used qwen3.5:9b with a scripted recipient, temperature zero and thinking disabled. Its later decisions depend on earlier actions.
  • The board contained an inert seed; no useful contribution accumulated spontaneously. The supplied experience was generated by a script, not a peer model.

The result is a contribution problem in these interfaces, not a general account of cooperation.

SOURCES & PROVENANCE

AUTHOR / Chris Hay · VERSION / 0.1

FOLLOW THE RECORD

A successful behaviour is not necessarily contagious.↗The world can remember for the agent.↗The tool was not the problem.↗Can something evolve that makes the next invention possible?↗

REFERENCE THIS DRAFT

An unpublished working record. These references identify the draft and omit a publication date. They become version-specific publication citations when the record is released.

Chris Hay. The message board did not create a culture. [Unpublished draft, version 0.1. First publicly recorded 2026-09-13]. https://chrishayuk.com/notebook/the-message-board-did-not-create-a-culture
DOWNLOAD