Re:The Unpredictable Abilities Emerging from Large AI Models
by pixl97 on Saturday March 18, 2023 @10:13AM
(#35209284)
This is reductive in the sense is that things like AlphaGo were bad at Go for a very long time, and then with more compute power and algorithm changes suddenly they were far better. And the problem space for go is absolutely huge it is still an insignificant portion of the problem space for knowledge.
LLMs are trained to reproduce human text, that is different from for example AlphaGo that is trained to win Go games. Trained to reproduce data is what we mean with a statistical model, trained to win is how we got superhuman performance before, while trained to reproduce data performs worse than the original creators of the data.