Abstract: This will be a very informal talk! We'll start with a high-level overview of how modern LLM systems like Codex and ChatGPT work -- sequence models, transformers, pre-training, post-training, and agentic tool-use harnesses which let today's models take actions, write code, and otherwise interact with their environments.

Then I'll share some of my personal perspective working at an AI lab alongside increasingly-capable AI systems. I'm not a mathematician (anymore!), but I work in a highly-technical domain, and hopefully some of the things I say will resonate or be interesting for you in your own work.

Contact: Karen Smith