Today’s models are astonishing: their thinking states are passed on only through sequences of text. This means they can stop halfway through a thought, go to sleep, and perfectly restore their thinking state. Humans are different. It is fascinating when a thought comes together and takes written form, but when I later reread what I wrote, it can feel difficult to understand, dry, and lifeless. To me, thinking is a dense, high-dimensional structure. When it is compressed into text and then reconstructed, the patterns collapse. For a model, then, reading each token amounts to reconstructing the previous token’s thinking state with perfect consistency and continuity. I find this continuity astonishing and unlike human thinking; it is also an overlooked, counterintuitive aspect—that this can work at all. This question points toward model architectures with hidden states.[1] [2]
The consistency of the models constructed during thinking may extend further, imposing negative constraints on the space of thought. We like “new brains” that bring different ways of modeling knowledge. Logical correctness may resemble a fractal, with convergence at the level of knowledge. I think this points to a direct relationship between the performance of architectures with hidden states and pretraining–post-training.[3]
For this reason, I also think it is difficult for two-faced models with hidden plans to emerge in architectures without hidden-state transmission. The model’s state is constrained by the state of the input text, and breaking those constraints requires additional effort. Findings from limited evaluations can therefore be extrapolated to some extent.