Experiments

Small studies I run for fun: borrowing methods from cognitive science to probe how machines see, think, and search.

Guided search but for vision-language models June 2026

Three follow-up searches suggested by Jeremy Wolfe, each one with no color to guide attention. GPT-5.5 searches a T-among-L display serially while Claude treats it as a pop-out; both count accurately; and both land on the same hard case in the classic search asymmetry, one in effort and the other in accuracy.

Vision-language models Visual search Guided search
Do vision-language models search like humans? May 2026

I ran a classic feature-vs-conjunction visual search experiment on Claude and GPT models, then lined their results up against 80,000 human trials from the Wolfe lab. The models reproduce the human signature, with one striking exception.

Vision-language models Visual search Evals