DeepSeek
At the start of 2025, DeepSeek AI took the Western AI ecosystem by storm. The team was able to meet or beat state-of-the-art benchmark targets with a smaller, clever model running on non-state-of-the-art GPUs with a bare-metal approach, at a cost orders of magnitude lower than their Wedtern counterparts. Ok, I am simplifying, but basically everyone in the West freaked out, but it felt like the "DeepSeek" moment came and went as a blip on big tech's aggressive timeline. In reality, the DeepSeek moment is still with us, the team continues to produce exceptional work and has inspired a wave of innovation and re-energized the Chinese, and more broadly the non-US, AI ecosystem. And its technical influence is wide-reaching: everyone developing or working on LLMs with reinforcement learning is integrating DeepSeek's innovative approaches.
The response from the US, in addition to a temporary but significant dip in the financial markets, was to dismiss DeepSeek as interlopers who used other companies' outputs, a small outfit that made a small innovation of little consequence, or worse. DeepSeek is none of that. An article in this week's Nature opens the curtain on DeepSeek's model R1 and proves everything the company claimed. In an editorial, the journal Nature thinks it is the first time a major LLM goes through rigorous (and given the stakes, extra super rigorous) peer review process. The editorial itself is a fascinating read, and if you have the time, I encourage anyone to read all the reviewers' comments and the authors' responses, all published as well.
It is not just an exercise in transparency. It sets a precedent and puts others to shame. That R1 has become Hugging Face's number 1 open weight model as measured by downloads (close to 11 million times) may have spurred others to bring open weight models to the world as well. OpenAI's gss open weight models are excellent and a welcome addition.
Kudos to DeepSeek AI for this work, kudos to Nature's reviewers for undertaking a difficult task that is and will stay under the spotlight.