Anthropic researchers extract interpretable features from Claude 3 Sonnet using sparse autoencoders

We demonstrate that sparse autoencoders can extract interpretable features from Claude 3 Sonnet, a production-scale language model, addressing the open question of whether dictionary learning methods scale beyond small transformers.

The resulting features are multilingual and multimodal (generalizing to images despite text-only training), respond to both concrete instances and abstract discussions of concepts, and can be used to steer model behavior in ways consistent with their interpretations.

We also identify features relevant to ways in which language models might cause harm--including features representing deception, power-seeking, sycophancy, and bias--and show that these causally influence model outputs when manipulated.

However, significant limitations remain: our suite of features is incomplete, and we lack rigorous methods for evaluating whether our features faithfully capture model computations.