https://x.com/_xjdr/status/1875295492906828005?t=m-uTZp61RAtJz7ixcG5kew&s=03

if you will forgive a bit of anthropomorphizing, to the extent models think, they do not think in tokens. tokens are merely our samplers' best interpretation of the probability distributions that emerge from the actual 'thinking' process. they're an incredibly low bandwidth channel we're forcing this computation through. The real cognitive work happens in latent space, where models have access to the full spectrum of complex interactions across their activations.

i've been thinking about this in the context of CoT / TTC. when we look at token-based CoT, what we're seeing is just a crude projection of a much richer reasoning process. The tokens we get out are like trying to understand a 3D object by looking at its shadow. It clear to me that the reasoning needs to be done in latent space. the recent FAIR work is in interesting step in this direction but we need to push this concept much much further. 

this obviously creates a tradeoff for interpretability. with token-based CoT, we have the comforting illusion of transparency. we could read the intermediate steps, annotate them, build evaluation frameworks. moving to latent reasoning means giving most of that up. Instead of clean, human-readable steps, we're left trying to make sense of high-dimensional manifolds, attention patterns, and activation distributions.