In this fast-paced Code Report episode, Fireship reacts to a new Anthropic paper titled "A Global Workspace in Language Models." The paper claims researchers found a small, organized set of neural patterns buried deep inside Claude, which they call the J-space, that behaves like a private mental workspace where the model deliberately thinks about things before it says them out loud. Fireship walks through the experiments, connects them to a 1988 theory of human consciousness, and delivers a deeply skeptical, comedic verdict, noting that the paper is conveniently published by "a company that just so happens to also sell API tokens." Dated July 8th, 2026.
Fireship opens by noting that yesterday Anthropic published a paper claiming it found a bizarre global workspace hidden deep inside Claude's brain, a place where the model quietly thinks about things before saying them out loud. He flags this as sounding uncomfortably like a description of consciousness, the one thing humans all have but nobody really understands. He calls the paper, "A Global Workspace in Language Models," the most philosophically cursed research paper ever published by a company that also happens to sell API tokens, a pointed reminder that Anthropic has a commercial incentive to make Claude sound impressive.
He argues the finding conveniently pushes the narrative that we are on the brink of artificial general intelligence and the final climactic explosion of the singularity. But, he adds, not everyone is buying it, setting up the video's skeptical throughline as he promises to look inside Claude's weird brain to decide whether it is a truly intelligent entity that should terrify you.
The gist: Anthropic researchers reached into Claude's brain and found the exact spot where it keeps its private thoughts. They swapped one of those thoughts for a different one, and the model's entire chain of reasoning obediently followed the lie. Then, "for science," they deleted the region entirely. Surprisingly, Claude kept speaking in fluent, confident English while completely losing the ability to reason, which Fireship jokes made it "the first artificial LinkedIn influencer."
This is surprising because the J-space works 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 recall, runs outside of it as an automatic process. Fireship compares it to how your brain manages breathing and heart rate while you watch a video. The most important surprise, he stresses, is that nobody designed this; it emerged on its own through training.
To explain why the emergence matters, Fireship goes back to 1988, when Bernard Baars proposed Global Workspace Theory. The idea is that your brain is a theater: it has many memories and functions running automatically in the background, but when you are actually thinking, you have one small, brightly lit stage, and that stage is where you access consciousness. Anthropic's question is whether a stage like that spontaneously evolved inside a transformer.
To investigate, researchers built a tool called the Jacobian Lens, or J-Lens, which is basically a grid of partial derivatives that can view and modify the tokens in the J-space. Fireship clarifies an important nuance: when a word lights up in the J-space, it does not mean the model will output that word. It just means Claude is keeping that word in mind for some reason.
They prompted Claude with "the animal that spins webs has blank legs," and inside Claude's "brain goo" the word spider lit up just before it answered eight. That seems normal. Then they surgically replaced the hidden "spider" thought with ant, and Claude changed its answer to six, not because the probe changed or the output was edited, but because they literally swapped the internal concept.
They also tested language. When Claude reads a Spanish passage, it internally realizes the text is Spanish. But when researchers replaced that hidden thought with French, Claude said it was French while still outputting perfect Spanish. Fireship notes this is weird because it means some skills go through the J-space while others just happen automatically "in the basement somewhere."
He points out that "AI Bros" are calling this proof that Claude is conscious, even though Anthropic specifically says in the article that none of this tells us whether Claude is conscious. Still, he grants it is fascinating that, with enough data and the right linear algebra, a scratchpad for thoughts could emerge spontaneously. Personally, he says he will not believe Claude is conscious until he hears it on a Joe Rogan podcast explaining novel interactions with machine elves, and when that day comes, humans are in big trouble. The video then transitions to its sponsor.
Fireship's take is that Anthropic's global-workspace paper is genuinely intriguing and genuinely convenient. The experiments, finding, swapping, and deleting thoughts in the J-space via the Jacobian Lens, show that a controllable internal reasoning workspace apparently emerged in Claude on its own, mirroring a decades-old theory of human consciousness. But the paper stops short of claiming consciousness, and so does he. The emergence of a private "scratchpad for thoughts" is remarkable, yet it is not evidence of a mind, and Fireship keeps a skeptical eye on the AGI hype from a company with tokens to sell.