Each bar in this ring is one signal inside the model. This rebuild has 64. A real model has thousands. Nine out of ten of them run on autopilot — the model can't see that work, or talk about it.
64 signals · mostly autopilotThe rest is different. Anthropic found a way to spot the ideas a model is about to use. The six strongest light up the glowing sphere — the workspace. This is the part of its thinking the model can actually put into words.
6 / 64 · the workspaceA model doesn't think in seconds. It thinks in layers, and your scroll now moves through them. Early layers hold raw pieces. Later layers build the answer. Watch ideas rise into the workspace and drop back out as you go.
scroll · layer 0 → 32Now the big result. Take FRANCE out of the workspace and put CHINA in. Every answer follows: the capital, the language, the continent — all three flip. The workspace doesn't just show the thought. It steers what the model says.
swap · all three answers flipIn Anthropic's tests, a FAKE idea showed up in the workspace while the model was getting ready to lie — before it said a single word. The same moment plays out here around layer 22. That is why being able to read the workspace matters.
deception · visible earlyCareful now. It acts like an inner workspace: you can ask what's in it, you can change it, and changing it changes the answers. But it isn't a brain. It works layer by layer, not moment to moment, and it holds only word-like ideas. A strong likeness — not the same thing.
a likeness · not a brain