On “Model Organisms”
This post explores the concept of "model organisms" in AI safety research, drawing parallels to their use in biology. The author distinguishes between studying a production model to understand general behavior, testing specific interventions, or inferring properties of other language models. In biology, model organisms like the lab mouse are chosen for practicality and extensive existing research, allowing for comparisons and generalization, though this can limit understanding of other species. The author also differentiates model organisms from "knockouts" and "mutants," which are specifically altered to study the function of particular genes or traits. AI
IMPACT This commentary offers a framework for thinking about AI research methodologies by comparing them to biological studies.