This representation parameterization is useful because it (1) tends to generalize better and (2) is more interpretable. But this is just data-efficiency and aesthetics. These properties cannot provide *evidence* against behaviorism. And they provide weak (if any) evidence for the representation.
I'm making a website for my new lab (!!!). Who should I steal ideas from? (Self-nominations welcome)
This is an actual line that was added to the official system prompt for Codex for GPT-5.5 by OpenAI. Usually the system prompt is as minimal as possible, so I assume it would otherwise mention goblins a lot.
AIs are weird.
This is not to say cognitive models aren't valuable. Going full Hofstadter, the representations in a model are themselves representations: useful ways to understand behavior and brain. This deflates Skinner's critique of cognitivism. Representations aren't true or false; they don't need evidence.
But the moment we start taking our models too seriously, we fall into the trap Skinner warned us of: believing things we have no evidence for. I think that cognitive science has largely fallen into this trap (myself included), and that we'd benefit greatly from taking Skinner's ideas more seriously.
New preprint from my lab! We study how reinforcement learning & selective attention interact. To do so, we built a set of models describing different ways that value & reward prediction error can modulate top-down attention. We compare model outcomes to monkey data from a color value learning task