←back to thread

No clicks, no content: The unsustainable future of AI search

(bradt.ca)
137 points bradt | 2 comments | 31 Aug 25 15:37 UTC | HN request time: 0s | source
Show context
visarga ◴[31 Aug 25 17:04 UTC] No.45084809[source]▶
The content is now created in private chats with AI, probably a trillion tokens per day flow between humans and LLMs. When new AI models are made they will incorporate some experience from past usage. This dataset might be the most valuable source for training AI and solving our tasks. So in case humans decide to abandon publishing, there is a new experience flywheel spinning up.
replies(3): >>45084967 #>>45085233 #>>45089835 #
8organicbits ◴[31 Aug 25 17:22 UTC] No.45084967[source]▶
I assume that AI chat is a lot of questions being asked by people unfamiliar with the subject matter. Would that training dataset have any information from experts?
replies(2): >>45090461 #>>45093353 #
1. visarga ◴[01 Sep 25 15:15 UTC] No.45093353[source]▶
Question answering and learning are just a corner of LLM usage, but they have learning signals for the AI. Say a user asks about Pythagoras, the LLM provides an explanation, the user doesn't get it. The LLM tries again.

Repeat this loop a million times with diverse students and you get a distribution of what kind of explanations work. The model gets better at explaining through its own experience.

replies(1): >>45093811 #
2. 8organicbits ◴[01 Sep 25 15:59 UTC] No.45093811[source]▶
Sounds like you'd end up with pop science. The loop stops when the explanation is satisfying, not when it's correct. Vibe science isn't based in reality.