Gary Lupyan

@glupyan.bsky.social

1847 Followers 173 Following

We leverage these traces to present the first large-scale study (129,134 projects) of the adoption of coding agents on GitHub, finding an estimated adoption rate of 15.85%–22.60%, which is very high for a technology only a few months old–and increasing.
A common view in cognitive science views language as a system for expressing our thoughts with the thoughts themselves being the products of other, nonlinguistic systems. On this view, one might expect that developing a human-like intelligence would require instantiating those nonlinguistic systems. What does it mean then, that most significant advances in developing domain- general artificial intelligence have come from training general-purpose neural networks on language? Exposure to (large) amounts of language appears to be sufficient to instantiate many core aspects of human intelligence. One response is that these developments show us that it is possible to reverse-engineer some aspects of our world from language, but tell us nothing about the role of language in human cognition. I will argue instead that there are compelling reasons to think that human minds are structured by exposure to language in some of the same ways that large language models are. The reason that language proves to be such a powerful type of training is that it provides learners with ready-made abstractions that allow both humans and artificial systems to learn highly compressed representations of the world. Adopting this perspective helps demystify some of the successes of large language models and points to the critical role that language plays in enabling human-like intelligence.
Chart showing the number of hours per day gaming by young men; 2019 - 1.08 hrs. 2022 - 1.82 hours
aerial view of the Madison isthmus - evolang fifteen - Madison, Wisconsin May 18-21, 2024
Aerial view of downtown Madison