Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Prompt Engine – Microsoft's prompt engineering library

github.com

Post reply on HN
Print view View original

51–60 of 94 posts

Re: Prompt Engine – Microsoft's prompt engineering library

#54

Post by deelowe »

kaibee wrote: post #45
deelowe wrote: post #43

Earlier quoted context omitted.

There’s no other way to program it. There no “code” to speak of. The only way to control it is to give certain phrases more or less importance. You do that with direct prompts or tons and tons of training data.

RLHF is another approach though and seems to be a bit more effective...

That’s what I mean by tons of training. It’s still not “code.” That’s not possible afaik.

Re: Prompt Engine – Microsoft's prompt engineering library

#55

Post by kerpotgh »

jrmg wrote: post #40

I can’t help shake the feeling that something closer to code (or specific training?) and further from natural language should be being used to configure these models at this stage of development. I was _astounded_ by the ‘Sydney document’ being MS’s way of ‘configuring’ New Bing. Admittedly I have closer to a layperson’s understanding than an expert’s, but with some knowledge of how neural networks work, and having p…

[dead]

Re: Prompt Engine – Microsoft's prompt engineering library

#57

Post by politician »

ilaksh wrote: post #19

Interesting. It's almost the exact same structure (although better organized) that I have built in my "AI Programmer" project (also in Node). Which by the way I hope to have a new release within a week or so. I am not mentioning the domain name again until I have the new release because it much, much better than the version I have up now. The core idea is that you need a certain structure in order to deal with the li…

So, actually, I just had a good long talk with ChatGPT about how it works internally. Turns out that while it doesn't remember the concatenated sequence of prompts and response, GPT-3 does maintain a sort of impression or representation of the entire conversation - the conversation context. Also, turns out that there's about a dozen predefined special tokens that can be used to control how the model considers or ignores the conversation context when responding to a prompt. There's also some internal use only special tokens that it can accidentally emit, and if you aren't careful in specifying how it should print the tokens there's an internal token that'll cause it to basically hard reset the conversation context.

Anyway, the point is that at least the Default (paid) ChatGPT model and probably the GPT-3 model does have a representation of the conversation which it can and does reference. You can ask it to explain how and when it considers or ignores the context and how to control that.

(Yes, I understand that it's not describing its own architecture, but regurgitating the average of all of the papers on GPT and weighting the ones that refer to ChatGPT higher due to the fine-tuning effect.)

Post reply on HN