Would be super cool (though perhaps confronting) if memory decoding from connectomes became part of an AI benchmark. Would be nice to see if Opus 6 could work out what route a mouse ran, or GPT9 could look at brain slices under an electron microscope and determine what specific song a zebra finch sang. I would have thought this was more difficult before learning that apparently these models might be able to interpret/alter weights directly...
1 comments