> Contrary to what the average HackerNews seems to believe, LLMs don’t “think,” they just predict.
Anecdotally, I can't recall ever seeing someone on HackerNews accuse LLM's of thinking. This site is probably one of the most educated corners of the internet on the topic.
> They shouldn’t be expected to improve in accuracy because of what they are and how they work.
> And there’s nothing in them that will constrain their token prediction in a way that improves accuracy.
These are both incorrect. LLM's are already quite better today than they were in 2022.
Anecdotally, I can't recall ever seeing someone on HackerNews accuse LLM's of thinking. This site is probably one of the most educated corners of the internet on the topic.
> They shouldn’t be expected to improve in accuracy because of what they are and how they work.
> And there’s nothing in them that will constrain their token prediction in a way that improves accuracy.
These are both incorrect. LLM's are already quite better today than they were in 2022.