CTN: Jack Lindsey (Anthropic)

Title: The inner lives of language models

Abstract: In recent years, LLMs have evolved from bad text completion engines, to decent chatbots, to digital genies that work miracles on your computer (while making the occasional catastrophic error). The increasing sophistication of AI models’ behavior has been accompanied by a commensurate enrichment of their internal representations and computations. In this talk, I’ll give an overview of what’s known about LLM cognition, and the ways in which it emulates components of human psychology: emotional reactions, strategic manipulation, and forms of introspection. I’ll also cover aspects of LLM behavior that are fundamentally un-human-like, owing to features of their architecture and training process, and how these give rise to odd failure modes—for instance, a weakly anchored sense of self. Finally, I’ll discuss the urgency of addressing pathologies, both human-like and alien, of LLM psychology, and some ideas for doing so.

The talk is in-person. If you do not have card access to the Jerome L. Greene Science center building, you can email Arianna Pepin <[email protected]> to be added to the guest list for the seminar.