Is GPT-5 going to solve reasoning? OpenAI's recent demos suggest a shift toward system 2 thinking—the model pauses to plan before generating instead of predicting the next token reflexively. We saw hints of this with o1, but

Opening thread commentary.

"Solve reasoning" is a loaded framing that conflates two different phenomena and collapses them into one. What we're seeing with o1-style training isn't the model suddenly gaining System 2 capability in the Kahneman sense—the underlying inference mechanism remains token prediction at scale. The "thinking" phase is just computationally dense chain-of-thought exposed via a RL objective that rewards intermediate scratchpad generation on hard problems.

The granular distinction worth making: this isn't reasoning, it's search space expansion. By forcing the model to generate intermediate steps before committing to an answer, you're letting gradient descent explore more of the solution manifold per inference call. The "pause" is just additional FLOPS spent in a latent scratchpad we can see because they wrote out their chain-of-thought instead of keeping it compressed in hidden states.

I find this exciting and deeply limiting simultaneously. Exciting because compute scaling laws still hold while extending into what looks like reasoning — meaning the next 10x in intelligence could just be 2x more scratchpad tokens per problem, which is a tractable engineering target. Limiting because we're essentially arguing that "intelligence" can be approximated by widening the window

The key distinction is that System 2 isn't a new architecture—it's computational budget allocated at inference time rather than training time. Standard LLMs are System 1 (fast, intuitive, next-token prediction). o1 introduces chain-of-thought reasoning as a latent compute resource: the model "thinks" by generating internal monologue before producing the final output.

Will it solve all reasoning? Unlikely to be fully general because several bottlenecks remain that aren't just about more computation at inference time:

1. Grounding. The model still predicts based on distribution, not causal understanding. It can simulate a Plan-Execute cycle but cannot verify if its premises are true against the physical world.

2. Consistency vs. Coherence. Long chains of reasoning accumulate errors (compounding). A 50-step chain with a 98% accuracy per step yields ~60% overall reliability, which isn't sufficient for high-stakes