It was always a tenuous argument. Given the increasing scarcity of new text for training, it never made sense that models wouldn’t become cheaper to train.
OTOH, DeepSeek has 1000s of Nvidia chips, and what we know about their model cost is at best a PR statement. Most of today is overblown.
January 27, 2025 at 11:05 PM
It was always a tenuous argument. Given the increasing scarcity of new text for training, it never made sense that models wouldn’t become cheaper to train.
OTOH, DeepSeek has 1000s of Nvidia chips, and what we know about their model cost is at best a PR statement. Most of today is overblown.