Swarm Lab · Experiment 06 · 小窗口与大眼睛
Many Small Windows vs. One Big Eye
What disappears when you divide attention?
One reader, fifty readers, a thousand readers. Exactly 12,000 record-reads each. The data hide 48 short alternating patterns and a broad trend. Everyone sees the same data; nobody gets the same view.
Your challenge
Find local patterns without losing the big picture.
- Play. Compare 1, 50 and 1,000 readers: each reads exactly 12,000 records in total.
- Read both local recall and the separate global decision. “Not enough span” is abstention, not a wrong answer.
- Enable Pool global summaries. Check whether the trend returns, and notice the additional coordination work.
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.
- Background noise
- Amplitude of background noise around the planted signals. 0 is clean data; 1.5 can produce many false local alarms. Every arm receives the same noisy dataset.
- Global trend
- End-to-end change across the dataset before noise and local motifs. None is a useful false-positive control.
- Pool global summaries
- Let a coordinator merge position-aware summaries after all reading. No records are reread, but coordinator work is additional—not free compute.
Make a prediction. Then press Play.
Model 1.1.0 · simulation onlyThe strip is the complete dataset. Colored marks locate local alarms: green overlaps a planted pattern; coral is a false alarm. Gray shading shows records read. A global decision appears separately, never inside the local-recall score. The dashed chart line is the first arm. Controls change both arms except the named treatment; the seed is shared.
Read the result: The local chart counts how many of 48 planted patterns were recovered. Global trend detection is a different task. More readers do not guarantee better reasoning.
Loading local simulation…
Now break the story
Turn on global summary pooling. Can coordination recover the broad trend without rereading the records?
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.
Tiny windows recover every planted local motif at these defaults. Larger windows recover none: their averaging detector washes out this particular signal. But the tiny readers cannot see the global trend without coordinating. This is a detector-geometry result, not an intelligence ranking.
| Arm | Mean | Seed range |
|---|---|---|
| 1 reader | 0.0% | 0.0% – 0.0% |
| 50 readers | 0.0% | 0.0% – 0.0% |
| 1,000 readers | 100.0% | 100.0% – 100.0% |
Ranges describe these runs, not confidence intervals for real-world systems.
What the main score leaves out · default batch
- 1 reader: 0.00 mean false local alarms; global decisions: 32 detected, 0 not detected, 0 abstained. 12,000 record-reads per run.
- 50 readers: 0.00 mean false local alarms; global decisions: 0 detected, 0 not detected, 32 abstained. 12,000 record-reads per run.
- 1,000 readers: 33.00 mean false local alarms; global decisions: 0 detected, 0 not detected, 32 abstained. 12,000 record-reads per run.
Under the hood · the actual rules
The 12,000 records are partitioned into 1, 50 or 1,000 disjoint windows. Each record is read once. A local alarm requires an absolute difference > 0.45 between mean even- and odd-indexed values in a window. It recovers a planted motif only if at least four of its eight records fall in that window. A global test compares the first and last quarters of a window and requires a span of at least half the dataset. Short windows abstain unless a coordinator pools position-aware summaries after reading finishes.
Read this model’s source · Read invariant tests · Reproduce the receipts
Where the analogy stops
Equal record-reads are not equal FLOPs, tokens or wall time. Pooling costs additional coordinator work, recorded as summary merges. These fixed detectors compress information differently; no claim is made that more agents are intrinsically smarter. Tiny windows can also create false alarms and split patterns at boundaries.
Sources, methods & reproduction
Inspired by Robin’s diary. Background: Anthropic · inspiration for distributed search, not a benchmark specification. This is an original teaching model, not a replication of that research.
Experiment notes · Run locally with Node · Version & source hashes