Hacker News Reader

Comment by elaus | Hacker News Reader

Any idea how this is done? I.e. is it just priming the underlying GPT model with plugin information additionally to the user input ("you can ask Wolfram Alpha by replying 'hey wolfram: ...' ") and performing API calls when the GPT model returns certain keywords ('hey wolfram: $1')?

View on HN

Yup, basically.

Edit: see here: https://github.com/openai/chatgpt-retrieval-plugin/blob/main...

I did this a while back with ARKit: https://github.com/trzy/ChatARKit/blob/17fca768ce8abd39fb27d...

Thanks, very interesting! Weird that it never occurred to me before reading OpenAI's announcement (and missing all the cool projects like yours beforehand).

On the bright side, now you’re basically caught up. This stuff is shockingly easy :)

I like to think that to get a sense of how this might be done, one way maybe to extrapolate from this experiment at https://til.simonwillison.net/llms/python-react-pattern .

I wonder, can these instructions be revealed with prompt injection?

Everything that is in the context window can be potentially revealed with prompt engineering.

(In this case, there's no prompt injection to speak of because letting the user input an arbitrary request is part of the UI. I think it's more accurate to call it "injection" only when that's not anticipated, like when Bing picks up instructions from the webpage you tell it to summarize.)

I guess that's technically true, but then what would you call tricking the model to reveal its instructions via anticipated input?

Reply on news.ycombinator.com