I think the secret of DeepSeek is basically using RL to train a model that will generate high quality synthetic data. You then use the synthetic dataset to fine-tune a pretrained model and the result is just amazing: https://open.substack.com/pub/transitions/p/the-laymans-intr...