Swarm Lab · Experiment 08 · 共享频道
Shared Channel
A shared lead—or a very crowded rabbit hole?
A thousand searchers have 128 directions to explore. Someone posts an exciting lead. Open the message board, turn up the volume, and watch whether the room becomes a conversation or an echo.
Your challenge
Share useful leads without turning the whole swarm into one searcher.
- Play isolated and shared-board search together. Watch effective diversity and true/false flagged directions.
- Raise Messages per decision from 0 to 1, 2, 4, 8 and 16. Compare with the saved sweep below.
- Uncheck First lead is true. Does attention stay trapped around a noisy mistake?
First time here? The buttons explained
Play starts or pauses. Step advances one frame; the time slider rewinds. Changing a control prepares a new run; press Play again. Reset replays the current settings. Seed chooses a random world, not difficulty; keep it while changing one rule, then try Next seed. Share saves your setup in a link. The downloads let you inspect the data.
What every parameter means
- Seed
- A repeatable random world, from 0 to 4294967295. The compared arms use the same seed.
- Messages per decision
- Number of modeled message exposures before a choice. With trust p and volume m, following probability is 1 − (1 − p)^m. More messages can reinforce the same lead, not add new evidence.
- Trust per message
- Chance one exposure persuades an agent to follow the shared board. 0 makes the two arms identical.
- Evidence threshold
- Positive observations required to flag a direction. Even false directions occasionally yield positives, so flagged does not mean proven.
- First lead is true
- Whether the initially promoted direction contains a planted anomaly. There are always 16 true directions overall.
Make a prediction. Then press Play.
Model 1.1.0 · simulation onlyEach dot is a research direction. Size reflects the number of searchers there. Green directions contain a planted anomaly; coral ones do not. A gold ring means accumulated positive observations reached the evidence threshold—not that the direction is proven true. The dashed chart line is the first arm. Controls change both arms except the named treatment; the seed is shared.
Read the result: Diversity is effective directions ÷ 128. It accounts for crowding; 70 barely visited directions can coexist with nearly everyone in one. Below 25% is this model’s declared collapse line, not a universal law.
Loading local simulation…
Now break the story
Make the first lead false. Does the board correct itself—or reward the loudest mistake?
Move one control, keep the seed, and compare again. The batch below is a fixed baseline—not a live summary of your current settings.
Observed in this model · 32 matched seeds · all runs retained
A result, not just an animation.
Messages herd attention long before they necessarily destroy useful coverage. At two exposures per decision, diversity crosses our line but all sixteen true directions are still flagged. At eight, only one is. Inspect both measures.
| Arm | Mean | Seed range |
|---|---|---|
| Isolated search | 93.8% | 92.2% – 95.4% |
| Shared board | 1.6% | 1.5% – 1.8% |
Ranges describe these runs, not confidence intervals for real-world systems.
What the main score leaves out · default batch
- Isolated search: 16.00 true directions flagged; 0.00 false directions flagged. 30,000 search attempts per run.
- Shared board: 1.00 true directions flagged; 0.00 false directions flagged. 30,000 search attempts per run.
| Messages / decision | Final diversity | True directions flagged / 16 | Runs crossing 25% diversity |
|---|---|---|---|
| 0 | 93.8% | 16.00 | 0 / 32 |
| 1 | 48.4% | 16.00 | 0 / 32 |
| 2 | 21.9% | 16.00 | 32 / 32 |
| 4 | 6.2% | 15.97 | 32 / 32 |
| 8 | 1.6% | 1.00 | 32 / 32 |
| 16 | 0.8% | 1.00 | 32 / 32 |
Two is the first tested volume crossing our declared line—not a universal threshold. The isolated control stays unchanged. Change trust or evidence requirements and test again.
Under the hood · the actual rules
1,000 searchers make 30 choices among 128 directions, 16 of which contain planted anomalies. The board initially promotes direction zero. A searcher follows it with probability 1 − (1 − trust)^message volume, choosing proportional to squared previous positive-vote counts plus a smoothing term. Otherwise it explores uniformly. A true direction yields positive evidence with probability 0.8; a false one with probability 0.03. Diversity is exp(Shannon entropy)/128; collapse means this modeled measure falls below 25%. Both arms receive identical draw keys and 30,000 attempts.
Read this model’s source · Read invariant tests · Reproduce the receipts
Where the analogy stops
Popularity amplification, trust and evidence noise are chosen rules, not measured swarm behavior. The 25% collapse line is a declared convention, not a universal threshold. Evidence counts can flag false directions. No Hugging Face incident, actual message platform or live AI swarm is recreated here.
Sources, methods & reproduction
Inspired by Robin’s diary. Background: Origin · Robin’s diary questions about swarms. This is an original teaching model, not a replication of that research.
Experiment notes · Run locally with Node · Version & source hashes