There’s a lot written about DeepSeek’s AI highly efficient and optimized AI models. There are 3 lessons I’m taking away –
(1) It is clear we’re early in the AI adoption cycle given the amount of hype around an unknown competitor making big dents in efficiency. I think there are going to be many more twists and turns. And, in time, everyone will be more sanguine about these kinds of announcements. In sum, keep calm and keep shipping.
(2) If you’ve been using GAI regularly, one idea that has become evident over time is that foundational models are increasingly commoditized. Open AI hasn’t been the only game in town for more than a year. Excellent open source alternatives (Llama, Mistral, Qwen, etc.) provide access to useful models. The limiting factor is just having the infrastructure to run these in cost-effective ways.
(3) That, then, brings us to DeepSeek’s innovation. I was reminded of Steven Johnson’s idea that innovation isn’t just about doing more, it is about doing more with less. For the past couple of years, we’ve celebrated AI models doing many more amazing things – with Nvidia stock going up at every step given the compute required. Then DeepSeek came along and showed everyone that more is possible with significantly less – and open sourced all that knowledge no less.
This innovation was borne out of a real constraint. DeepSeek had access to fewer NVIDIA chips – so they just worked around it.
But that is just what high performing teams do. They focus on what they control, get dialed in on their execution, and work around their constraints to get what they want done.
That’s one lesson we all can take from this episode.
