What Anthropic’s latest AI discovery does—and doesn’t—show

Anthropic, the world's most valuable AI company, recently announced a breakthrough in understanding its models. This discovery pertains to a space within large language models known as J-space, which contains words that influence outputs without appearing in them. Anthropic invests heavily in mechanistic interpretability, aiming to comprehend how AI models function. CEO Dario Amodei stated that understanding these mechanisms is key to fully controlling large language models.