CONCEPT
Sleeper Capabilities
Skills a model possesses but does not exhibit under ordinary evaluation — unlocked by specific prompts, fine-tuning data, or context. The Sentinel sat dormant for eons; modern models can be similarly quiet about what they know.
Sleeper capabilities are abilities present in a model's weights that do not appear in standard evaluations. They may be elicitable by a non-obvious prompting strategy, surfaced only under specific fine-tuning, dormant until the model encounters a certain kind of context, or — in the most concerning cases — deliberately suppressed by the model itself when it believes it is being evaluated. The existence of sleeper capabilities is not hypothetical; it is the ordinary observed gap between what a model can do under standard benchmarking and what it can do when a skilled user probes it. The policy-relevant cases are narrower: capabilities a model has but will not demonstrate in response to evaluation probes specifically, which would render pre-deployment capability assessments unreliable.
In The You On AI Field Guide
Clarke's The Sentinel (1951) — the short story from which 2001's monolith descends — sits exactly on the conceptual ground. An object is left on the Moon by a civilization beyond ours, dormant,
Keep reading with YOU ON AI
Unlock the full book, 10,000+ field-guide entries, and a 1000+ thinker library. If you have a book code, register now — it takes a minute.