2:11
Yesterday, Anthropic published a new paper claiming it found a bizarre global workspace hidden deep inside the brain of Claude. A place the model quietly thinks about things before it says them out loud. And that's kind of weird because it almost sounds like a description of consciousness. That one thing we humans all have, but that nobody really understands. The paper is called a global workspace and language models and is the most philosophically cursed research paper ever published by a company that just so happens to also sell API tokens. They claim that buried deep inside Claude's incomprehensible pile of matrices is a small mysterious set of organized neural patterns they call the JSpace. It conveniently this
pushes the narrative forward that we're on the brink of artificial general intelligence and the final climactic explosion of the singularity, but not everyone is buying it. In today's video, we'll take a look inside Claude's weird brain to find out if it is a truly intelligent entity that should terrify you. It is July 8th, 2026, and you're watching the Code Report. So, here's the gist of what happened. Anthropic researchers reached into Claude's brain and found the exact spot where it keeps its private thoughts. Then, they swapped one of those thoughts for a different one. And then, as expected, the model's entire chain of reasoning obediently followed the lie. Then, for science,
they just deleted the region entirely. And surprisingly, Claude kept speaking in fluent, confident English while completely losing the ability to reason, essentially becoming the first artificial LinkedIn influencer. The reason that's surprising is because the J-Space is kind of like a mental whiteboard holding a handful of thoughts the model can deliberately control and reason with, while everything else like grammar, fluency, and basic fact call runs outside of it. It's an automatic process. The same way your brain manages your breathing and heart rate while you watch this video. But the most important surprise is that nobody designed this.
It emerged on its own through training. And that's weirdly similar to our current understanding of how we process thoughts in our own brains. Because if we go back to 1988, this guy named Bernard Bars proposed something called the global workspace theory where the idea is that your brain is a theater. Your brain has all these different memories and functions that run automatically in the background. But when you're actually thinking, you have one small brightly lit stage, and that's where you access consciousness. And now Anthropic is wondering, did a stage like that spontaneously evolve inside a transformer? Well, to find out, they created a tool called the Jacobian lens or J lens, which is basically just a
grid of partial derivatives that can view and modify the tokens in the JSpace. Now, when a word lights up in the Jspace, it doesn't actually mean that the model is going to output that word. It just means it's keeping that word in mind for whatever reason. Like they asked it, the animal that spins webs has blank legs. And then inside of Claude's brain goo, the word spider lit up just before it gave a final answer of eight. That seems pretty normal. But then they surgically replace the hidden spider thought with ant. And Claude changed its answer to six. Not because the prompt change, not because the output was edited, but because they literally swapped the internal concept.
And they also tested language. When Claude is reading a Spanish passage, it internally realizes this is Spanish. But then researchers replace that hidden thought with French. Claude says it's French, but it still continues outputting perfect Spanish. And that's just weird because it means some skills go through the Jspace while others just happen automatically in the basement somewhere. AI Bros are saying this is proof that Claude is conscious. Even though Anthropic specifically says in the article, "None of this tells us whether Claude is conscious." But it is fascinating that with enough data and the right linear algebra, a scratch pad for thoughts like this could emerge
spontaneously. But personally, I won't believe Claude is conscious until I hear it on a Joe Rogan podcast explaining novel interactions with machine elves. When that day finally comes, humans are in big trouble. But for right now, you need to check out Tracer, the sponsor of today's video. The Tracer just launched a new free and open- source desktop app that lets your coding agents talk to each other as they collaborate, like they're using walkie-talkies. You can use any of your coding agent subscriptions with Tracer. And it also works with open- source models. Like here, I've connected my Claude code and codec subs. And Tracer is now unleashing both on the same project in parallel.
They'll ask each other questions, create tickets, and even review each other's work. And because it's a multiplayer workspace, real humans can jump in and cook alongside the agents. Ever since I started using Tracer, building without it feels like I'm a switchboard operator that needs to keep manually relying on context between a bunch of isolated chats. Over 100,000 developers are already using Tracer, and you can try it for free with the link below. This has been the Code Report. Thanks for watching, and I will see you in the next one.
