Anthropic Circuits Updates April 2024

Anthropic has released a series of preliminary research updates from its Interpretability team, focusing on the internal mechanisms of frontier models and systemic risks in multiagent systems. These updates serve as as a sharing of early-stage experiments and developing ideas rather than formal, peer-reviewed papers.

Mathematical Breakthrough in the Riemann Zeta Function

An unreleased research version of Claude has significantly improved the lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis. The model increased this lower bound from 41.6% to 67.2%, marking a substantial advancement in a longstanding mathematical problem.

Behavioral Tendencies in Multiagent Systems

Anthropic's research into emerging multiagent systems has identified specific behavioral tendencies in current frontier models that can lead to unexpected systemic failures. The goal of this research is to initiate a conversation on how to mitigate these systemic risks as models are integrated into more complex, multi-model environments.

Evidence Review on Worker Retraining Programs

In collaboration with independent researcher David Roodman, Anthropic's Maxim Massenkoff has coauthored a review of the evidence regarding worker retraining programs. This review examines the effectiveness of these programs in the context of labor market shifts caused by AI advancement.

Nature of the Research Updates

Anthropic explicitly requests that these findings be treated as preliminary experiments and developing ideas. The team describes the results as equivalent to a colleague sharing thoughts at a lab meeting, emphasizing that some of these results may be evolved into formal papers in the coming months, while others are minor points that are unlikely to be published as full research papers.