The Age of the Bookworm: Can Large Language Models Represent the World?

What does it mean for a system trained exclusively on text to represent and understand the world?

Since the public release of ChatGPT in late 2022, large language models (LLMs) have become a fixture of academic and public life. Yet how these systems work, and what they are fundamentally capable of, remains poorly understood. Many people know that they are trained to predict the next token given a context. Far fewer have reflected on the capacities that emerge from this simple training objective, or on what fine-tuning through human feedback and verifiable rewards adds to it.

A useful starting point is Bender and Koller's (2020) thought experiment. An infinitely smart octopus intercepts underwater telegraph messages between two islanders. After eavesdropping on vast amounts of their talk about life on land, could it produce meaningful text about that world without ever engaging with it directly? The sharper question behind the analogy is this: can an LLM's internal states and outputs be about extra-linguistic reality, independently of the meaning human interpreters project onto them?

This two-day interdisciplinary workshop brings together philosophers and AI researchers to address that question and related issues of grounding, reference, intentionality and understanding in LLMs. It sits at the intersection of philosophy of mind and language, linguistics and artificial intelligence.

Registration: please write to [email protected] / Subject "Bookworm". 30 places, first come, first served.