Vision IA

On peut maintenant lire dans la tête d'une IA

Anthropic has recently published a method to translate the internal processes of AI models, specifically Claude, into human-readable text. This development, which allows humans to understand what happens inside an AI's "mind," is significant because, until now, the internal workings of AI have been largely a "black box" – incomprehensible to even their creators. When interacting with AI models like Claude or ChatGPT, we exchange words, giving the impression that they "think" in words. However, AIs actually process information as long series of numbers called "activations," similar to how a brain scan shows electrical activity rather than sentences. The problem has been that these numerical processes were unreadable to humans, making the AI's decision-making opaque. This "black box" concept is akin to the flight recorder in aviation; you see what goes in and what comes out, but not what happens in between. AI models have been built this way, trained on vast datasets, becoming increasingly performant without a full understanding of their internal mechanisms.