3 ms·
They shouldn’t be expected to improve in accuracy because of what they are and how they work. Contrary to what the average HackerNews seems to believe, LLMs don
by eschaton 2y ago
They shouldn’t be expected to improve in accuracy because of what they are and how they work. Contrary to what the average HackerNews seems to believe, LLMs don’t “think,” they just predict. And there’s nothing in them that will constrain their token prediction in a way that improves accuracy.
- rybosworld 2y ago> 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.
- reshlo 2y ago> I can’t recall ever seeing someone on HackerNews accuse LLMs of thinking. There are definitely people here who think LLMs are conscious. https://news.ycombinator.com/item?id=41959163 https://news.ycombinator.com/item?id=41959163
- BlackSwanMan 2y ago[dead]
- cozzyd 2y agoIf anything, they may regress due to being trained on lower-quality input.