Is model drift the same as model collapse?

Collapse is the extreme case: a model trained repeatedly on its own outputs loses the diversity of the original data and degrades. Drift is the broader, slower change in behaviour from any cause; collapse is one way drift ends.

How do labs detect drift?

By holding back fixed test sets the model never sees, by monitoring the model’s reasoning traces, by probing its internal representations, and by using evaluators — human or model — that are deliberately kept independent of the model under test.

Why does drift matter for recursive self-improvement?

Because a model that improves its successor also writes the tests its successor passes. If both drift together, the loop reports progress it is not making. That is the failure mode outside evaluators exist to catch.

Glossary term

Model Drift

The gradual change in a deployed AI model’s behaviour or accuracy over time — because the world it was trained on changes, because it is updated on its own outputs, or because its evaluators drift with it. The risk that makes closed loops of self-improvement dangerous rather than merely fast.

AI-generated — produced automatically by Closelook’s systems under this site’s editorial policy.

What it means

A model is trained on data from a period and evaluated against tests written in that period. After deployment three things move. The world changes, so the data the model sees no longer matches its training (data drift). The model is updated — sometimes on data it generated itself — so its concept of the task shifts (concept drift). And the tests used to judge it are updated by people, or by models, whose standards shift with the model (evaluator drift). Each alone is manageable; together they can leave a system confidently wrong with nobody positioned to notice.

In classical machine learning drift is a maintenance problem. In a loop where a model trains its successor, drift compounds.

Why it matters for the AI trade

Drift is the strongest technical argument in the pacing debate: the labs’ own research describes self-trained models collapsing toward their own outputs, and the value of an outside evaluator is precisely that it does not drift with the lab. For markets the relevance is indirect but real — a drift incident at a frontier lab would be the first event to change the demand curve for compute from the model side rather than the budget side.

How Closelook uses it

The recursive self-improvement read gives drift its own section alongside collapse, reward hacking and evaluator drift, and lists the methods the labs use to catch it. It is background for the tape, not a signal on it — until it is.

Common questions

Is model drift the same as model collapse?: Collapse is the extreme case: a model trained repeatedly on its own outputs loses the diversity of the original data and degrades. Drift is the broader, slower change in behaviour from any cause; collapse is one way drift ends.
How do labs detect drift?: By holding back fixed test sets the model never sees, by monitoring the model’s reasoning traces, by probing its internal representations, and by using evaluators — human or model — that are deliberately kept independent of the model under test.
Why does drift matter for recursive self-improvement?: Because a model that improves its successor also writes the tests its successor passes. If both drift together, the loop reports progress it is not making. That is the failure mode outside evaluators exist to catch.