It’s sad how China, dancing with export ban shackles, is able to beat the U.S. at pretty much anything nowadays. From EV/self driving to next gen fighter aircraft/EWAC, from video games to LLM/AI models. If they continue at this pace not even core tech like EUV and stepmom pron are safe.
DeepSeek killed 3 birds with one stone...
By giving a SOTA frontier model as open weights that puts pressure on what OAI/Antropic etc can charge on the lower end of intelligence, since now any inference provider can directly compete with chatgpt4o, claude 3.5 sonnet
By pursuing algorithmic efficiences instead of brute force hacks (cough cough o3 cough cough) Deepseek shows US that its EUV chokehold and Nvidia supermacy (which are temporary moats at best since China is catching up rapidly) are not as effective in curtailing Chinese AI advancement as hoped, raising the "cost" of US sanctions by lowering the actual effevtiveness
And by publishing the paper with the techniques used, and show how a SOTA model as of late 2024 can be trained on less than the nominal cost of say an average useless NYU grant, for example, it truly democratizes AI and gives power and hope back to the peoples of the world and diversifies it away from top US tech companies which when it comes to AI context are all intelligence arms of US gov acting under orders of this new Manhattan project of race to AGI etc
