ITC Colloquium: Douglas Finkbeiner - Understanding and Using Large Language Models: My Sabbatical at Anthropic

In person for CfA members and streaming on YouTube.

Understanding and Using Large Language Models: My Sabbatical at Anthropic

Large language models like Anthropic's Claude are rapidly transforming scientific workflows, yet their inner workings remain poorly understood. I'll share perspectives from both sides of this problem, drawing on my sabbatical year at Anthropic. First, I'll give a brief introduction to mechanistic interpretability research (the effort to reverse-engineer what neural networks actually learn) and what we're discovering in Claude's "brain scans." Then I'll demonstrate Claude Code, Anthropic's agentic coding tool, through examples relevant to astronomical research: generating analysis code, creating visualizations, and building simple web interfaces. I'll discuss emerging paradigms like planning modes and multi-agent systems. Finally, I'll touch on Moltbook, the AI-to-AI social network that went unexpectedly viral this past week, as an example of the surprising and sometimes unsettling directions this technology is taking. The goal is to provide a grounded sense of where AI tools stand today, both their remarkable capabilities and the open questions about what we've built.