Questions followed by answers
Added 2023-06-07 19:37:24 +0000 UTCgang back
Comments
This “sport moment” made me lol https://www.tiktok.com/t/ZT81qYVW8/
Skyler Johnson
2023-06-13 22:48:01 +0000 UTCre: ChatPPP, you can kind of do this now (not that you should). You don't exactly train anything — what people are doing is indexing the documents (e.g., all of PPP's stuff) with a vector DB (e.g., Pinecone https://www.pinecone.io/learn/vector-database/). Then when someone types in a search, you run a query on the DB to get all the chunks of text the index considers similar to the search, and then you basically dump all that plus the original search in carefully worded prompt to the ChatGPT API, like "answer this question _____ based on these documents ______". GPT-3.5 is not that expensive, it would cost less than a cent per search. GPT-4 is noticeably better but it's about 15x that. Obviously that's all pretty complicated, so search companies are already putting together things that will do it for you, mostly aimed at things like searching documentation: https://www.algolia.com/products/neuralsearch/ This is the best post I've read about this, it's a good overview: https://about.sourcegraph.com/blog/cheating-is-all-you-need
David Crespo
2023-06-09 14:39:49 +0000 UTC
