3 ms·
You're getting downvoted, but you're completely right. There are very few cases in which narrowing a model down is buying you anything worthwhile. It seems lik
by ACCount37 3mo ago
You're getting downvoted, but you're completely right. There are very few cases in which narrowing a model down is buying you anything worthwhile.
It seems like for LLMs, "general intelligence" is expensive, but "one more domain" is fairly cheap.
- Danox 3mo agoThe demand for smaller models and single purpose AI agents will only get bigger in the future. It’s no different than going for mainframe computers back in the day to personal computers. It’s inevitable…..
- andy99 3mo agoHas nothing to do with demand, it’s a question of the most efficient way to train a good model. Narrower models are in general worse than general purpose ones, so you’re better using the largest general purpose model that fits in your compute budget vs trying to somehow remove capability or knowledge and assume the model stays as capable for the task you want.