Re:Yann LeCun on GPT-3
by arugulum on Tuesday October 27, 2020 @02:28PM
(#24910292)
>However, even if they are, the point stands: currently, there are teams of people at companies all over the world tuning models for these shallow and limited use-cases. GPT-3 can replace them all, without OpenAI needing to invest another cent in training for a particular customer's use-case. That is in fact game-changing for the ML/DL world and current applications thereof.
The counterpoint is that it would be significantly cheaper AND have better performance to fine-tune models to each customer's use case than it is to just run GPT-3 at inference.
The counterpoint is that it would be significantly cheaper AND have better performance to fine-tune models to each customer's use case than it is to just run GPT-3 at inference.