DeepSeek has created an algorithm that enables an LLM to bootstrap itself by starting with a small dataset of labeled theorem proofs and create increasingly higher quality example to fine-tune itself.