What Anthropic really found inside Claude's 'brain'

The breathless headlines around Anthropic’s latest research would have you believe we can now read an AI’s mind. Not quite. But what the company actually found is genuinely interesting if you ignore the hype. James O’Donnell sat down with senior editor Will Douglas Heaven, who holds a PhD in computer science, to sift through the findings. Heaven’s take is characteristically sober: this is a clever new window into how Claude reasons, but it’s a narrow view through frosted glass, not a clear pane.

Anthropic has made a habit of publishing quirky, fascinating research that peers into the black box of its models. The latest work traces the internal pathways Claude uses when it answers a question. Think of it like mapping which neurons fire, but for a large language model. The team identified specific circuits that light up during reasoning tasks. The discovery is a meaningful step in mechanistic interpretability, a field that tries to reverse-engineer what these giant models are actually doing under the hood.

But let’s not get carried away. Heaven cautioned that this isn’t a full accounting of Claude’s ‘thoughts.’ The models remain profoundly opaque, and the circuits Anthropic found represent just a tiny fraction of the overall computation. It’s like finding a single gear in a watch and claiming you understand how it tells time. The research is a solid incremental advance, not a breakthrough that suddenly makes AI safe or trustworthy. For anyone who’s been around the block, the real headline is that progress in interpretability is happening at all.

Meanwhile, the push to make AI understand the physical world is heating up on a different front. MIT Technology Review is hosting a LinkedIn Live event today with Heaven and Sam Sinha, head of world models at 1X Technologies. They’ll dig into how these models could finally endow robots with the kind of common-sense physics that current AI disastrously lacks. If you’ve ever watched a robot fumble with a door handle, you know exactly why this matters. The session kicks off at 12:30 PM EDT.