ashbys-law-of-experience ADG-0002

Ashby's law of experience

“Identical deterministic machines given the same inputs converge on the same state; variety does not grow.”

Origin

W. Ross Ashby, 1956. Information put in through a shared input tends to destroy and replace information about each machine's initial state, so a set of machines synchronises. About deterministic systems in general.

Why it should, or should not, apply to models

N instances of one model fed the same context are Ashby's identical machines. The law predicts they converge: whatever diversity they started with (different samples, different partial views) is destroyed by a shared input. That is the mechanism behind two findings the catalog cares about — why broadcasting everything between agents removes the benefit of having several, and why output diversity collapses when many samples share a prompt. It also says where diversity has to come from: different inputs, different models, or a non-shared source of variety.

Standing: Untested

Standing counts claims a person has reviewed. Claims reviewed by AI alone are listed and marked, and do not move it.

Where it holds

Claims showing the heuristic applying to models.

In multi-agent LLM systems where one agent starts with an erroneous shared belief, broadcasting every agent's full context at every step raises the rate of false assertions above doing no synchronization at all, because the error propagates to agents that were previously correct — and the harm appears only in tasks where one wrong fact cascades across semantically linked dimensions (destination to airport to weather), not where agent contexts are largely orthogonal.
Stating false facts confidentlyReviewed by AI
Broadcasting full context synchronised the agents onto one belief, a wrong one. Shared input destroyed the variety that would have caught it.

Add a case where it holds or breaks

A break is worth more than a hold. If you have seen this fail for a model, with the setup written down, that is the most useful thing you can add here.

Notes

Output diversity is filed under fine-tuning-retention and unmeasured by any claim in this catalog; a claim there would be the second piece of evidence for this law.