Until this has been independently verified, I have my doubts. This wouldn’t be the first time for China to vastly exaggerate its technological capabilities.
Deepseek seems to have done a clever thing w.r.t. training data, by having the model train on data that was emitted by other LLMs (as far as I’ve heard). That means there is sort of “quality-pass”, filtering out a lot of the definitely bogus data. That probably leads to a smaller model, and thus less training hours.
Google engineers put out a paper on this technique recently as well.
The article mentions the cost of tokens to end users, not training time.
Ah ok, I didn’t catch that. Other articles were discussing v3’s training using only 2.8M GPU hours.
https://www.ft.com/content/c82933fe-be28-463b-8336-d71a2ff5bbbf
Until this has been independently verified, I have my doubts. This wouldn’t be the first time for China to vastly exaggerate its technological capabilities.
Deepseek seems to have done a clever thing w.r.t. training data, by having the model train on data that was emitted by other LLMs (as far as I’ve heard). That means there is sort of “quality-pass”, filtering out a lot of the definitely bogus data. That probably leads to a smaller model, and thus less training hours.
Google engineers put out a paper on this technique recently as well.