story
OpenAI has also invested heavily in human annotation and RLHF. If all DeepSeek wanted was a proxy for scraped training data, they'd probably just scrape it themselves. Using existing RLHF'd models as replacement for expensive humans in the training loop is the real game changer for anyone trying to replicate these results.