🧠 Confessions of a frustrated cognitive scientist.

At different points in time this post was also titled “Why do cognitive science in 2025?” and “(What) can cognitive science do for AI alignment?” This post has been simmering in me since NeurIPS 2024 in one shape or another, and has three main theses:

1. This is a particularly fruitful time to study the human mind, owing to (1) our ability to collect behavioral data across task and situations, and (2) ever-more powerful models enabling the testing of complex hypotheses about the computational basis of cognition.
2. (Computational) cognitive science has done little to contribute to the development of frontier large langeage models.
3. But maybe it could?! I will speculate on why and how.

Epistemic status: all three of these theses are, fundamentally, opinions. The strength of my belief in them is in descending order, where I feel quite strongly about the first, decently strongly about the second, and the third is quite speculative.

Why 2025 is a great time to be a cognitive scientist

It’s possible this thesis is a specific case of the general “science keeps marching forward,” but ascribing the state of cognitive science today to that would, in my mind, undersell the current state of the field. I find this to be a particularly fruitful time to study the mind, fundamentally owing to a combination of the ability to collect data and model it. The following bit elaborates on the above, at the risk of making a few obvious points:

More data, more