Hacker News Reader

What is more interesting for me, is why anyone is surprised that an algorithm based on choosing a statistically probably combination of words, trained with whatever random crap appears on the internet, happens to choose some combination that makes grammatical sense, but doesn't match reality.

View on HN

Because it was not trained with bias in favor of "reality" or facts but grammatical sense. Given that gpt is a statistical model in theory it should be possible to bias towards facts, but that means being able to identify bs from facts. This is a subject on its own.

Do you think it’s even possible? Imagine written facts as a thick thread and bs as long hairs growing from it (both written and hallucinated, there’s no limit for bs). We can ask something very well-known and likely(?) the answer will stay within that thread. But if you ask something non-trivial, i.e. start in the hairy area, which force could drive its “thinking” back into the thread? I believe that there’s no way to create it (in this model). Once far away from the main road, LLMs can wander bs space indefinitely. If my analogy makes sense at all, obviously.

Reply on news.ycombinator.com