I have many questions to ask after reading this blog post. I will keep them until I read all DeepSeek papers to understand their techniques better. One high-level intuition is that engineers always have innovative optimizations given restricted compute resources, like in the time of computers with 64KB memory. However, such optimization doesn't change the ever-growing demand for computing. The door to a new world is open when resource constraints are improved.
Therefore, on top of mind to understand DeepSeek papers, I want to reason whether the above intuition is true at some degree in the context of building better AI. In my previous comment, I asked for the moat of OpenAI (or any closed models), which I just realized that my question is about whether DeepSeek's method will lead to better AI (not only models better than o1) with much smaller compute demand.
I have many questions to ask after reading this blog post. I will keep them until I read all DeepSeek papers to understand their techniques better. One high-level intuition is that engineers always have innovative optimizations given restricted compute resources, like in the time of computers with 64KB memory. However, such optimization doesn't change the ever-growing demand for computing. The door to a new world is open when resource constraints are improved.
Therefore, on top of mind to understand DeepSeek papers, I want to reason whether the above intuition is true at some degree in the context of building better AI. In my previous comment, I asked for the moat of OpenAI (or any closed models), which I just realized that my question is about whether DeepSeek's method will lead to better AI (not only models better than o1) with much smaller compute demand.
You have the right intuition. More AI is coming.