Key points
- Anthropic found a workspace inside Claude where thoughts form before any words appear.
- It raises questions about human cognition we could never ask before.
- A window into AI may show us the first principles of cognition.
Recently, I posted something about AI that I didn’t expect to write for years, if ever. I admitted that the interior of these systems may not be the empty room I’ve long described. Anthropic has identified what it calls the J-space, a small set of internal patterns that behave like a mental workspace. Concepts take shape there and influence answers without ever appearing in the final text. The researchers can watch a thought form inside the machine before it speaks. I haven’t stopped thinking about that.
Two details stand out to me. First, no one engineered this workspace. It emerged on its own during training. Second, when researchers suppressed the J-space, basic recall survived, but complex reasoning collapsed. Whatever the structure is, it isn’t decoration. And when you remove it, the thinking stops
Let’s start here. Studying cognition meant studying ourselves. Every capable mind we could examine came from evolution, including the one doing the examining. AI breaks that monopoly. Whatever else these systems are, they’re a second implementation of advanced cognition, reached by a completely different route. And this “second implementation” lets us ask something that used to be unaskable: Does thinking have first principles?
One possibility, I believe, is yes. Cognition, like vision or flight, may have only a few workable engineering solutions. Evolution found a workspace in us. Training found one in the machine. Octopus eyes and human eyes converged on nearly the same optics through lineages that split more than five hundred million years ago, and nobody calls the octopus a copy
This perspective has serious backing. Stanislas Dehaene and Lionel Naccache, two of the architects of global workspace theory in human consciousness, reviewed the Anthropic work and provided this striking perspective
The global workspace may provide a universal computational solution to the problem of flexible processing, one that biological and artificial systems converge on when they must chain reasoning, reuse intermediate results, and report on their own processing
If they’re right, we’ve glimpsed a principle of thought itself
The other possibility is stranger. Every word the model learned came from minds that already work this way. Maybe the machine built its workspace by copying the one thing all its training data had in common—us. In that case, we haven’t found a principle of cognition. We’ve found our own reflection, buried deeper in the machine than anyone expected
What saves this from being a hypothetical construct is that it can be tested. Train a system mostly on non-human data like protein folding or weather and watch if a similar workspace appears. If it does, that would point to a law. If it doesn’t, the workspace starts to look like something we handed the machine without meaning to. The researchers themselves list this as an open question. So, the experiment is sitting there, waiting to be put to the test
Either answer changes what we know. One says thinking has first principles, and we can finally go looking. The other says the most convincing mind we ever built is a copy of the only one we had to copy, and we’d never tell the difference from inside our own heads
I’ve spent three years arguing that these machines think with no mind behind the thinking. This is the first finding that could prove me wrong. It could also be the best evidence that I was right. I don’t know which
