Balls Deep
(Posted 14:31:30 on 1st February 2025 by Rag)
Balls to the US and DeepSeek to China. It's been well over a year since I've posted to this blog, but we've had a rather staggering change this week with the release of DeepSeek an AI from China. It's being likened to the Sputnik moment when the US realized they were not as far ahead of Russia in the space race, this is China's moment to catch America with its pants down, only in the development of AI versus space travel.
I have to laugh if you read the article I posted 18 months ago (either scroll down to the title “David(s) and Goliath(s)” or
click here) where I talked about the key to success being able to train models cheaply and focus on dspecifc areas. I will admit that it may not always be beneficial to train a model on limited data as it's been proven that the ubiquitous models like ChatGPT or Gemini perform better on specific topics than models trained on specific data, however, that may change.
So what is special about DeepSeek? Nothing if you were to just compare it to the models of the major players. It performs pretty similarly. The shock, if true, is that the model was trained for just $6m instead of the $bns that its cost the major players. Additionally, it was done using out of date technology as the US has banned the export of the latest chips to China (ironically to prevent against them getting a leg up in the AI development race). I guess they never heard the saying “necessity is the mother of invention.”
So this model is cheap, lightweight and portable with pretty much the same performance as its much more expensive competitors. (Reminds me of a car nicknamed Godzilla, but that's a whole other story). At a bear minimum it raises a lot of questions and hopefully opens the door to more competition. This may just see the start of a lot of advancement which may, in turn, result in more blog posts as it's definitely piqued my interest.