Anthropic just open-sourced a tool to read the minds of AI

Anthropic dropped a circuit tracer that lets you peek inside LLMs and see how the machine actually reasons through a problem. Like watching neurons gossip in real time.

‍

Using it, you can track how models like Llama-3 or Gemma-2b "decide" that the capital of Texas is Austin. This could be a big deal for AI safety. Or at least a handy way to catch your chatbot hallucinating before it claims Napoleon invented the iPhone.