The author questions the practice of fine-tuning models on summarized or censored state-of-the-art chain-of-thought (CoT) traces, arguing that it is not a magic solution for improving output quality.
- Distillation does not raise output quality above what the base model is capable of.
- Fine-tunes based on Fable and Anthropic's models are criticized because their reasoning traces differ significantly from actual model outputs.
- This mismatch guarantees that the resulting fine-tuned models will perform worse than before.
The author suggests that relying on these summarized traces is counterproductive for improving reasoning capabilities.