Anthropic Circuits Updates July 2024

Anthropic has released a series of preliminary research updates for July 2024, providing a window into the ongoing work of their interpretability team. These updates cover a range of topics from systemic failures in multiagent systems to significant mathematical progress on the Riemann zeta function.

Mathematical Breakthroughs in Claude Research Version

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%.

Behavioral Tendencies in Multiagent Systems

Anthropic's research into emerging multiagent systems has identified specific behavioral tendencies in current frontier models. The team reports that these tendencies can lead to unexpected systemic failures, highlighting a need for mitigation strategies to reduce these risks.

Worker Retraining Program Evidence Review

In collaboration with independent researcher David Roodman, Anthropic's Maxim Massenkoff has coauthored a review of the evidence regarding worker retraining programs. This research focuses on evaluating the evidence base for programs designed to assist workers in transitioning to new roles.

Nature of the Research Updates

Anthropic explicitly states that these July 2024 updates should be treated as preliminary experiments and developing ideas rather than mature, peer-reviewed papers. The updates include emerging strands of research that may lead to future publications, as well as minor points that the team intends to share with the community without the intention of writing formal papers on them.