Anthropic interpretability team publishes January 2024 research progress

Anthropic's interpretability team published its January 2024 Circuits update, summarizing several research ideas still in progress and stating that over the coming months it will focus on three directions: scaling dictionary learning, using dictionary learning to attack superposition in real models, and understanding circuits on that basis. The team applied dictionary learning to the MLP activations of an 8-layer Transformer and found that features in the final layer resemble those of a single-layer Transformer, while features in the middle layers (roughly layers 4 to 6) are more abstract, and gave several examples of safety-related features. On larger-scale models, the team said it has preliminarily extracted multilingual features, for example the same concept triggering the same feature across multiple languages.