Hacker News Reader

Comment by mirekrusin | Hacker News Reader

If OpenAI can scrape what humans did then can you create public p2p database of request-responses?

I mean that sounds like a natural evolution; if generating a new response costs money, it makes sense to save them in a cheap database for future lookup / reuse.

But that's quite possibly OpenAI's model as well - cache results and reduce costs while keeping pricing.

Also known as "mainly poem database".

That would really only work on intrusive prompts, before existing state had been established.

It also would not create the non deterministic responses users have come to expect.

> It also would not create the non deterministic responses users have come to expect.

Probably slightly modifying text is less expensive than coming up with brand new text. So the text can be cached, and when retrieved, slightly altered before served. A "cached-but-not-really" cache

Good point, randomly rephrase slightly is probably much easier task resource wise.

> before existing state had been established.

As I understand it, "state" is actually just feeding in the history into the prompt as "context", so realistically this could help build history and context via lookups

eg.

Answer this as an assistant, blah blah blah.

You already said ${QUERY 1}, ${QUERY 2}. Refer to ${QUERY 3}, ${QUERY 4} for ideal formatting.

Prompt: "This is what the user asks chat gpt this round"

I was thinking about that when working on https://eli5.gg. I implemented a caching system internally but it would have been simpler to connect to an OpenAI proxy that handled this automagically.

They can put it in the T&Cs saying you can't do that though. Like what most geocoding APIs do - e.g. Google says you cannot cache/store the geocoding results.

What you do with the results is a legal argument after the fact. It's like saying we give you food, but you can't store the poop. Honestly - there is a house of cards just waiting to be challenged. Then again .. this is Google ($$$)

Do they follow license requirements of data they're scraping for training?

You just need to cache it with AI and you should be good to go.

Reply on news.ycombinator.com