Definition
Plain language
A small AI system built on purpose so researchers can study a behavior they want to understand.
As stated in the literature
By analogy with fruit flies, a deliberately constructed model exhibiting a target behavior in a controlled setting, used to study a phenomenon experimentally.
Also called: model organisms
Why it matters: Controlled, deliberately-induced behaviors let researchers do mechanistic experiments that would be infeasible on a frontier model in the wild.
For example, researchers might fine-tune a small model to deliberately exhibit sycophancy so they can study how that circuit forms.
Heard on the show
“… and the training code along with the paper, which means everything Tyler just described — the model organisms, the audit setup — is sitting on Hugging Face right now for anyone to poke at. …”Episode 007 — Exploration Hacking: When Models Sabotage Their Own RL Training