Diffuse

DeepSeek Helps China’s AI Innovation Dodge US Chip Strangulation

Pierre -

Let me know how you feel about the following line of reasoning.

DeepSeek recently open-sourced their latest LLM DeepSeek R1. It rivals ChatGPT o1. Two takeaways from me after catching up via ChatGPT:

  1. DeepSeek got to this level of competition not by flexing hardware (throwing more metal at it) but by innovating at the software level. Recent progress for AI has been more indexed on hardware because a) it scales so nicely and b) it is somewhat of a more known factor, less difficult. This point is probably 10x more nuanced and probably debatable but that's my read so far. Help me correct if needed.
  2. This is a milestone for Chinese AI. Not only are they making more progress (Alibaba's Qwen was in the news in the past as well) but they're not just copying. they're innovating on other dimensions.

Overall this brutal race is still a win for the consumer.

I would highlight at least 3 major pillars of AI:

  • Models: the software, the logic, the thinking. This part needs hardware to run on: that's the chips
  • Chips which need enormous amount of electricity (power)
  • Power

The US is currently strangling China on advanced chips.
China is racing on models and doing well.
China also has an epic, centralized plan to flood its nation with cheap energy. Truly incredible.

If we consider China to be at the top of the pyramid for what is considered advanced manufacturing - excluding the more advanced chips - fueling this with no-cost energy will sustain them in this position for a long time. Strategically/militarily, if China steers towards high volume low cost armament (think drones of all kinds), they would further grow into a formidable adversary.

Now, add advanced-chip-independent AI to the mix.

DeepSeek-R1 can operate on non-NVIDIA hardware, including AMD GPUs and CPUs, as well as Huawei's Ascend platform.

I'm going to throw a Honda Civic vs. McLaren comparison.

At the end of the day all vehicles get to their destination in roughly the same amount of time but the dude in the McLaren can take it to the raceway and crank out record times <– the analogy here could be research, including AI research.

Remember inference time scaling: the longer you let models think about a topic, the better the answer. so the faster the model, the better the answer in less time.