Anthropic Circuits Updates June 2024

Anthropic has released a series of preliminary research updates in June 2024, focusing on interpretability, the systemic risks of multiagent systems, and the ability of an unreleased research version of Claude to solve complex mathematical problems. These updates serve as a shared set of developing ideas and preliminary experiments rather than formal, mature papers.

Mathematical Capabilities and the Riemann Zeta Function

An unreleased research version of Claude has demonstrated significant progress on a problem related to the Riemann hypothesis. Specifically, the model improved a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the hypothesis, increasing the bound from 41.6% to 67.2%.

Multiagent Systems and Systemic Failures

Anthropic is investigating behavioral tendencies in current frontier models that can lead to unexpected systemic failures when deployed in multiagent systems. The goal of this research is to identify these patterns and initiate a conversation on how to mitigate the associated risks.

Worker Retraining Programs

In collaboration with independent researcher David Roodman, Anthropic's Maxim Massenkoff has coauthored a review of the evidence surrounding worker retraining programs, exploring the effectiveness of these programs in the context of AI-driven economic shifts.

Research Status and Intent

Anthropic's Interpretability team describes these updates as "developing ideas" and "emerging strands of research." The team explicitly asks that these results be treated as preliminary experiments shared among colleagues rather than as finalized, peer-reviewed research papers.