The other workspace
By Daniel Farré Manzorro ·
Anthropic put out a beautiful video yesterday about a workspace they found inside the model. The method under it is real, and I won't pretend otherwise: a new interpretability technique —the Jacobian lens— that isolates the small set of activation directions the model uses to decide what it says next.
The video works because of the frame, not the finding. It's inspired by global workspace theory —a theory of consciousness— and the association does the rest: you don't have to claim a mind to let people picture one. So it's worth looking at what actually got measured, because it's smaller, and stranger, than the frame suggests.
Emergence isn't the surprising part. Stir a glass of water hard enough and you get a vortex —structure nobody placed there, out of rules everyone understands. Chaotic systems do this for a living: known rules, a shape you can't read off them, so you watch instead of derive. A trained network is the same category of thing. The one difference that matters is that the water's shape is inert, and the network's carries meaning —it steers what the model says. That's the part to measure. Not that a mind appeared, but that a complex system grew functional structure we still have to observe instead of predict.
And what they measured is smaller still. The workspace is tiny —a sliver of what the model is computing— and gone the moment the pass ends. "A small, evolving set of unspoken words," they call it. It doesn't persist between your questions. It doesn't retain anything from the conversation. Nothing carries over. Next prompt, it's built again from nothing.
Which settles an argument a lot of people are still having. The hope is that the model will eventually crystallize —that with enough scale the context you keep feeding it bakes itself in, the way experience bakes into a person, and the machine becomes a colleague who remembers. The thing they found is the opposite of that. Short-term by construction. It was never going to grow a past. A model's state is the weights it was trained with, plus whatever you hand it this time. That's it. It's software.
One workspace looks like a discovery, and discoveries make beautiful videos. The other gets no video, no lens, no frame. It's just work —keeping that outside description accurate while the business shifts under it. The context doesn't break; it stays exact about a company that has changed. No model release catches that, because it was never the model's to get.
Two workspaces, two owners. The one inside the model is theirs —theirs to study, theirs to shape. The one that decides whether any of it is worth trusting is yours. It was never a breakthrough away. It was always just work nobody wanted to call a workspace.