Anthropic Circuits Updates September 2024

Anthropic has released a series of preliminary research updates for September 2024, focusing on interpretability, multiagent system risks, and specialized mathematical capabilities. These updates serve as informal sharing of developing ideas and preliminary experiments rather than formal, peer-reviewed papers.

Multiagent Systems: Behavioral Tendencies and Systemic Failures

Anthropic identifies 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 provide a foundational understanding of these tendencies to develop mitigation strategies for risks associated with multiagent interactions.

Evidence 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 work examines the effectiveness of various retraining initiatives designed to support workers in the transition to an AI-driven economy.

Mathematical Capabilities: Progress on 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 lower bound from 41.6% to 67.2%.