[gremlinsinc]

I'm probably not google-level, but I have an idea to recreate ask jeeves, where at first it brings back all the results and people check the box next to the ones most likely to fit their need, using reinforcement learning it'll eventually know the best ones, and then have gpt3 summarize the top 10, and eventually just answer the damn question. The goal being chatGPT+Google+wikipedia(citations) to combat chatGPT's misinformation issues.

Another idea is to basically create LLM's for every city's laws/legislation/codes etc and basically be like an ai lexus nexus.

[runnerup]

> at first it brings back all the results and people check the box next to the ones most likely to fit their need

My friends+colleagues would 100% write bots for that in less than a week. You'd be completely swamped with fake feedback to SEO/game the results. The bots will beat every CAPTCHA you have and have completely normal human-like client fingerprints.

[tensility]

Unfortunately, having watched so many mass-input rating systems fail by being gamed, I suspect that this strategy might only really work well if the reinforcement model was built and maintained independently for each user...which seems like an approach that could be very effective but which would be more difficult and likely costly to scale.

[_pjee]

I'd imagine the search engine of the future will be like that. Let a summarization AI do the search and then report back on results, with citations.

[freediver]

Kagi has a technology demo that does exactly that, with varying rate of success:

https://labs.kagi.com/ai/contextai