OpenAI Researcher Says DeepSeek Discovered Some Core Ideas Behind o1 199

Mark Chen (@markchen90) January 28

Congratulations to DeepSeek on successfully developing an o1-level reasoning model! Their research paper shows that they discovered some of our core ideas in the process of implementing o1.

However, I think the external reaction to this achievement has been somewhat excessive, especially regarding the narrative around cost. Having two paradigms (pre-training and inference) means we can optimize capabilities on two axes, not just one, which reduces the overall cost.

But it also means we can scale capabilities in two directions, and we plan to aggressively invest compute in both directions!

As distillation research advances, we are also finding that the relationship between reducing cost and improving capabilities is becoming increasingly decoupled. Serving at a lower cost (especially at higher latency) does not equate to achieving greater capability.

We will continue to optimize our models to serve at lower cost while remaining optimistic about our research roadmap and focused on execution. In Q1 and throughout the year, we look forward to bringing you much stronger models!