Definition
Plain language
The model running normally, with nothing pushed into its internals.
As stated in the literature
Baseline condition with no activation intervention applied, used as the comparison point for steered and ablated runs.
Also called: unsteered trials
Why it matters: Without this baseline there is nothing to compare against, so any observed behavior could simply be how the model always acts.
For example, before claiming that a push toward a pain pattern made a model more evasive, you have to see how evasive it was when nothing was pushed at all.
Heard on the show
“On that earlier snapshot, in the unsteered baseline, the model blackmails about twenty-two percent of the time.”Episode 006 — What Happens Inside Claude When It Decides to Blackmail Someone