Anthropic just found something inside Claude that looks like a global workspace, and it changes what “show your reasoning” even means. New interpretability research from Anthropic, the Jacobian lens and the J-space, reads the concepts a model is holding before it says a single word. Under 10% of the model is reportable. The rest runs silent. Bernard Baars proposed this same architecture for human consciousness back in 1988, and Nisbett and Wilson showed in 1977 that our own explanations for our
3:40
Anthropic just found something inside Claude that looks like a global workspace, and it changes what “show your reasoning” even means. New interpretability research from Anthropic, the Jacobian lens and the J-space, reads the concepts a model is holding before it says a single word. Under 10% of the model is reportable. The rest runs silent. Bernard Baars proposed this same architecture for human consciousness back in 1988, and Nisbett and Wilson showed in 1977 that our own explanations for our
12K views1 month ago
TikTokethicai_au
See more
Static thumbnail place holder