April 2024 Mechanistic Interpretability Research Updates and Team Hiring Plans
Circuits Updates — April 2024
AI BriefingAnthropic's interpretability team shared its research progress and hiring plans for April 2024. The team currently has 17 members and expects to expand significantly throughout 2024 and 2025, with a focus on hiring managers, research scientists, and engineers. The research explores scaling laws for dictionary learning and analyzes how resource allocation relates to the training performance of sparse autoencoders (SAEs). It also presents a specific case study showing how large-scale hyperparameter sweeps can be used to find an optimal configuration. The team emphasizes that these are preliminary findings, shared informally in the style of a lab meeting.
Source: Anthropic: Transformer Circuits · transformer-circuits.pub