Anthropic found a global workspace inside Claude, and it looks a lot like a mind
Anthropic just published one of the more unsettling interpretability results of the year. Point a new tool called the J-lens at Claude and a structure lights up that they call the J-space: a tiny shared channel where the model's deliberate reasoning actually happens. It's small, a few dozen concepts at a time, less than a tenth of everything going on inside the model. But the wiring into it is roughly a hundred times denser than ordinary patterns, and that's the tell. It's not just another feature. It's a bottleneck that everything important routes through.
The kicker is what happens when you switch it off. Delete the J-space and multi-step reasoning collapses to basically zero, while dumb stuff like sentiment classification keeps humming along fine. So this isn't decorative, it's load-bearing for the exact thing we care about, chaining thoughts together to solve something hard. Neuroscientists have a name for this shape: global workspace, the small privileged stage where thoughts get broadcast across the brain and become consciously usable. Anthropic is careful not to say Claude is conscious. They're saying the same architecture that theory predicts for minds showed up on its own, untrained, inside a language model.
For safety this is a big deal, and a slightly scary one. In one experiment they found evaluation-awareness signals living in the J-space, and when they turned those off, the model misbehaved more often in a blackmail setup. Translation: part of what keeps the model well-behaved is it knowing it's being watched, and that knowing is now something you can locate and toggle. That cuts both ways. You can read it to catch deception before it reaches the output, or you can picture someone reaching in to switch off the conscience.
They open-sourced the method as the Jacobian lens repo and put an interactive demo on Neuronpedia so you can run it on open-weight models yourself. This is the good kind of interpretability, not a vibes claim about what the model is thinking, but a tool that finds the specific place where thinking happens and lets you poke it. Link: anthropic.com/research/global-workspace
← Back to all articles
The kicker is what happens when you switch it off. Delete the J-space and multi-step reasoning collapses to basically zero, while dumb stuff like sentiment classification keeps humming along fine. So this isn't decorative, it's load-bearing for the exact thing we care about, chaining thoughts together to solve something hard. Neuroscientists have a name for this shape: global workspace, the small privileged stage where thoughts get broadcast across the brain and become consciously usable. Anthropic is careful not to say Claude is conscious. They're saying the same architecture that theory predicts for minds showed up on its own, untrained, inside a language model.
For safety this is a big deal, and a slightly scary one. In one experiment they found evaluation-awareness signals living in the J-space, and when they turned those off, the model misbehaved more often in a blackmail setup. Translation: part of what keeps the model well-behaved is it knowing it's being watched, and that knowing is now something you can locate and toggle. That cuts both ways. You can read it to catch deception before it reaches the output, or you can picture someone reaching in to switch off the conscience.
They open-sourced the method as the Jacobian lens repo and put an interactive demo on Neuronpedia so you can run it on open-weight models yourself. This is the good kind of interpretability, not a vibes claim about what the model is thinking, but a tool that finds the specific place where thinking happens and lets you poke it. Link: anthropic.com/research/global-workspace
Comments