Definition
Plain language
How much extra real-world help a model gives someone trying to do something dangerous.
As stated in the literature
The marginal capability a model confers on a would-be misuser relative to their baseline resources; in chemical and biological domains it concentrates in last-mile operational specifics rather than topic knowledge.
Why it matters: It focuses safety work on what actually changes an outcome, rather than on whether a model can discuss a sensitive topic at all.
For example, the question is not whether a model can describe a dangerous process in general terms, but whether it hands someone the specific quantity and temperature they were stuck on.
Heard on the show
“The paper doesn't test whether you'd get the same uplift without the substrate, or whether a weaker, cheaper supervisor would close any of the gap.”Episode 096 — How Treating an AI Agent's Execution Like Git Recovers a Coordination Penalty