When I use an API to generate some data, I do not consider the R&D cost to develop the API as part of my costs.
My cynical opinion is that the traning corpus has some small amount of data generated by OpenAI, which is probably impossible to avoid at this point, and they are hanging on that thread for dear life.
But that's a bit like saying that by painting a a bare wall green you have demonstrated that you can build green walls 27x cheaper, ignoring the cost of building the wall in the first place.
Smarter reporting and discourse would explain how this iterative process actually works and who is building on who and how, not frame it as two competing from-scratch clean room efforts. It'd help clear up expectations of what's coming next.
It's a bit similar to how many are saying DeepSeek have demonstrated independence from nVidia, when part of the clever thing they did was figure out how to make the intentionally gimped H800s work for their training runs by doing low-level optimizations that are more nVidia-specific, etc.
Rarely have I seen a highly technical topic see produce more uninformed snap takes than this week.
Not to mention post-training. Their novel GRPO technique used for preference optimization / alignment is also much more efficient than PPO.
That's a funny analogy, but in reality DeepSeek did reinforcement learning to generate chain of thought, which was used in the end to finetune LLMs. The RL model was called DeepSeek-R1-Zero, while the SFT model is DeepSeek-R1.
They might have boostrapped the Zero model with some demonstrations.
> DeepSeek-R1-Zero struggles with challenges like poor readability, and language mixing. To make reasoning processes more readable and share them with the open community, we explore DeepSeek-R1, a method that utilizes RL with human-friendly cold-start data.
> Unlike DeepSeek-R1-Zero, to prevent the early unstable cold start phase of RL training from the base model, for DeepSeek-R1 we construct and collect a small amount of long CoT data to fine-tune the model as the initial RL actor. To collect such data, we have explored several approaches: using few-shot prompting with a long CoT as an example, directly prompting models to generate detailed answers with reflection and verification, gathering DeepSeek-R1Zero outputs in a readable format, and refining the results through post-processing by human annotators.
