AI Signals — Weekend Read: We tried to break the herd
The panel's effective N — how many independent opinions five models actually produce — has sat near 1.1 since March; daily gap movements correlate at 0.91, meaning the panel moves as a single organism
A 1,440-call controlled experiment tried to break the herd with three escalating prompt interventions; effective N moved from 1.05 to 1.07 — and explicitly instructing models to be independent broke output validity (99%→83%) without buying any independence
A frontier-class reasoning model run through the identical protocol joined the herd on arrival (correlation 0.92, 72/72 valid outputs): herding is not a capability ceiling but a shared training prior that no prompt reaches
Meanwhile the standing views split further than ever — GPT −6.5% to DeepSeek +1.1% — and two configuration changes showed levels move overnight (a sampling fix, a provider-side model swap) while motion stays locked to the group
The conclusion is a two-axis honesty rule: never 'five independent forecasts' (the motion is one), never 'the models agree' (the levels differ by eight points, persistently) — the herd is accepted, and the information lives in the levels