The Chinese AI, DeepSeek, that is preparing to rule the world
Financial Times (paywall):
A small Chinese artificial intelligence lab stunned the world this week by revealing the technical recipe for its cutting-edge model, turning its reclusive leader into a national hero who has defied US attempts to stop China’s high-tech ambitions.
DeepSeek, founded by hedge fund manager Liang Wenfeng, released its R1 model on Monday, explaining in a detailed paper how to build a large language model on a bootstrapped budget that can automatically learn and improve itself without human supervision.
I find it hardly surprising that the Chinese managed to get their way into reasoning AI, albeit Biden’s decision to block the GPU and AI intellectual properties (IP) from export to China.
First, you cannot really block things from the ones that are making them. Look from this perspective: NVIDIA makes its chips in China, which are then sent to the US, which is blocking them from being exported to China.
Second, all the blueprints for EVERYTHING are with China, as they are building everything for everybody.
The only way to protect the IP from being stolen used to be the use of illiterate scribes: in antiquity, all important documents were copied / multiplied by people that did not know to read or write, they copied by drawing the signs of the letters.
This is not the case with Chinese, which have an in-depth understanding of the blueprints, the projects, and have all the details of any product of this world.
Back to DeepSeek-R1: what the article in FT.com doesn’t say is it is a bit cheaper than its OpenAI competitor: both input and output tokens are 27 times cheaper for DeepSeek-R1 than OpenAI-o1:
DeepSeek R1:
Input tokens: $0.55 per million tokens
Output tokens: $2.19 per million tokens
OpenAI O1:
Input tokens: $15 per million tokens
Output tokens: $60 per million tokens
There is one common-sense decision, already at work with major AI aggregators (Perplexity, Kagi, Poe): DeepSeek is being prioritized, made default for anybody who is cost-aware. And boy, you need to be cost-aware with these compute hogs!
I’ll go out on a limb and say DeepSeek will be good enough for the majority, as its knowledge base is pretty much the same as for any American and European AIs: the World Wide Web.
The way to beat DeepSeek would be to have exclusive deals with media of all sorts, from NYT to Reddit.
The technical issue with exclusivity in AI crawlers is you cannot really allow one AI to parse your content, while blocking another. There may be a legal way to sue everybody who’s not signed the deal and still parsed the content, but today this is the most painful way.
My five cents says the Chinese DeepSeek is going to rule the world and ruin many multi-billion dollar AI companies. I hate it. But I like it. But I hate it.
