Tweet

Replying to @tomfleet

I think it’s using that same technique, but applying it to individual neurons to see what stimuli make it fire. it also shows some real images that seem to trigger it, so these seem to be the neurons for “shocked,” “hand,” “tongue,” and… “dinner?” Really? https://microscope.openai.com/models/contrastive_4x/image_block_4_5_Add_6_0/317

(original)