Comment by Philpax

That's not really true - the current generation, as in "of the last three months", uses reinforcement learning to synthesize new training data for themselves: https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero