A Summer with Claude and Relative Arbitrage, by Dmitriy Traytel (lawrencecpaulson.github.io)

Dmitriy Traytel's project, "Copilots for Isabelle," aims to integrate AI agents, like Claude, with the Isabelle proof assistant for formal mathematics. A notable achievement from this project is Kevin Kappelmann’s PIDE MCP component, which allows AI to have a live, interactive interface with Isabelle, similar to human users. During a summer project, Traytel tasked Claude with autoformalizing the key result from an upcoming paper on the Relative Arbitrage problem, leading to the generation of over 100,000 lines of formal proof in Isabelle within three weeks. This proof highlights the capability of AI in managing complex formalizations autonomously, with Claude utilizing multiple versions to refine its output. The significance of this development lies in its implications for the AI/ML community, particularly in the context of formal methods and theorem proving. As AI becomes increasingly involved in complex mathematical proofs, it raises questions about collaboration between human mathematicians and AI agents. The project also showcases challenges in the autoformalization process, such as ensuring rigorous adherence to mathematical definitions and managing the project's scale. Ultimately, the work not only advances the practical capabilities of AI in formal mathematics but also sets the groundwork for future research in stochastic optimal control, potentially influencing methodologies in the field.