I heard about something called J-space earlier today, so I went and read into it. So basically, LLMs probably aren't doing pure "next-token prediction" (autocompleting) in the flat way people describe it. Anthropic's new interpretability paper shows Claude has developed a small internal workspace with a handful of neural patterns, which they call the J-space, where certain concepts get held, reported on, and reasoned with before generating a response, somewhat similar to how our own neural activity works before we speak. This wasn't designed into the model. It was discovered when researchers looked into how Claude actually generates its answers. Most of what Claude does still runs on autopilot and never touches the J-space. It's only the harder tasks, such as requiring multitasking or complicated reasoning, that seemed to pull information from this shared space first. Thus, "It's just autocomplete" isn't wrong exactly. Some of what the model does really is close to that. But some of it looks like something else: information written once into a shared space, then read by many different parts of the network at once. None of that means Claude is conscious the way we humans are. Anthropic is careful about this. What they found speaks to what philosophers call "access consciousness," the ability to report a thought, reason with it, direct it on request, as opposed to "phenomenal consciousness," which is whether anything is being subjectively experienced at all. And that second question is exactly where David Chalmers' hard problem still lies; We know a ridiculous amount about our own neurons, how they fire, how they wire up, how different regions of the brain hand information to each other. We can point to the actual circuits behind vision, memory, language. And we still can't explain why and how any of that produces a first-person experience instead of just information moving around in the dark. So finding a workspace-like structure in Claude doesn't prove anything about experience, one way or the other. But it does make me wonder if we've been asking the wrong question the whole time. "It's just next-token prediction" is about as complete an explanation as "it's just neurons firing." Both are technically true, and neither tells you why there's something it's like to be the thing doing it, if there's anything at all. We're getting better at how intelligence works mechanically. Whether that produces experience is a different question, and getting better at the mechanics might never answer it. So how the heck do we know anyone else has an inner life to begin with? Not just Claude, other people. We infer it from behavior, from what they report, and from the fact that they're built roughly like we are. That question is hundreds of years old, and this research doesn't answer it.