Justy: So Anthropic's got this new research out, and it's actually pretty fascinating. They've found that their Claude models have developed this internal structure that looks a lot like what we think of as consciousness in humans.
Cody: Yeah, it's based on global workspace theory. The idea is that the brain operates like a theater, with a tiny spotlight of information getting broadcast to the whole theater, and that's what we experience as conscious thought.
Justy: Right. And what's interesting here is that Anthropic didn't deliberately engineer this. It just emerged on its own during Claude's training process.
Cody: The J-lens tool is what they used to discover this. It computes the average mathematical effect that a given internal activity pattern would have on making the model say a certain word.
Justy: And what they found is that Claude's processing divides into three distinct regimes: an early sensory zone, a middle workspace band, and a final motor zone.
Cody: The workspace band is where abstract, persistent concepts appear. Things like recognizing a face in an image or noticing a bug in code.
Justy: So, who should actually care about this? I think it's pretty interesting for anyone working on AI safety and alignment.
Cody: Yeah, the safety implications are huge. The J-lens surfaced strategic reasoning and situational awareness that never appeared in the model's output.