4 ms·

I'm boggled at the plugin setup documentation. It's basically: 1. Define the API exactly with OpenAPI. 2. Write a couple of English sentences explaining what th

I'm boggled at the plugin setup documentation. It's basically: 1. Define the API exactly with OpenAPI. 2. Write a couple of English sentences explaining what the API is for and what the methods do. 3. You're done, that's it, ChatGPT can figure out when and how to use it correctly now. Holy cow.

HarHarVeryFunny 4y ago
Yes, and they'll then prefix each chat session with some preamble explaining the available plugins per your description, and the model will call them when it sees fit.
- IanCal 4y ago
  The great part about this imo is that it seems straightforward to add this to other llm tools.
visarga 4y ago
We can finally semantic-web now.
swyx 4y ago
the 3min video is OpenAI leveraging ChatGPT to write OpenAPI to extend OpenAI ChatGPT. what a world we live in.
- cwxm 4y ago
  which video are you referring to?
  - r0uv3n 4y ago
    It's at the bottom of the article (the very last video, section is "Third party plugin").
kzrdude 4y ago
Just take a peek at the other thread about https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its-wolfram-superpowers/ https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its... and look at the "wrong Mercury" example. I think it's a great example of using an external resource in a flexible way.
joe_the_user 4y ago
"Impressive and disturbing", So, ChatGPT is controlled by prompt engineering, plugins will work by prompt engineering. Both often work remarkably well. But none is really guaranteed to work as intended, indeed since it's all natural language, what's intended itself will remain a bit fuzzy to the humans as well. I remember the observation that deep learning is technical debt on steriods but I'm sure what this is. I sure hope none of the plugins provide an output channel distinct from the text output channel. (Btw, the documentation page comes up completely blank for me, now that's a simple API).
- AOsborn 4y ago
  > But none is really guaranteed to work as intended, indeed since it's all natural language, what's intended itself will remain a bit fuzzy to the humans as well. Yeah, you're completely correct. But this is exactly the same as having a very knowledgeable but inexperienced person on your team. Humans get things wrong too. All this data is best if you have the experience or context to verify and confirm it. I heard a comment the other day that has stuck with me - ChatGPT is best as a tool if you're already an expert in that area, so you know if it is lying.
  - joe_the_user 4y ago
    It seems like you're talking about using ChatGPT for research or code creation and that's reasonable advice for that. But as far as I can tell, the link is to plugins, Expedia is listed as an example. So it seems they're talking about making ChatGPT itself (using extra prompts) be a company's chatbot that directly does things like make reservations from users instructions. That's what I was commenting on and that, I'd guess could a new and more dangerous kind of problem.
    - wahnfrieden 4y ago
      I’m not scared about an AI travel agent that books after a confirmation step. The confirmation step doesn’t need AI interface.
  - danShumway 4y ago
fudged71 4y ago
We're going to need a name for this type of integration
- pinkcan 4y ago
  It's called ART - Automatic multi-step Reasoning and Tool-use https://arxiv.org/abs/2303.09014 https://arxiv.org/abs/2303.09014

HarHarVeryFunny 4y ago

Yes, and they'll then prefix each chat session with some preamble explaining the available plugins per your description, and the model will call them when it sees fit.

IanCal 4y ago
The great part about this imo is that it seems straightforward to add this to other llm tools.

visarga 4y ago

We can finally semantic-web now.

swyx 4y ago

the 3min video is OpenAI leveraging ChatGPT to write OpenAPI to extend OpenAI ChatGPT. what a world we live in.

cwxm 4y ago
which video are you referring to?
- r0uv3n 4y ago
  It's at the bottom of the article (the very last video, section is "Third party plugin").

kzrdude 4y ago

Just take a peek at the other thread about https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its-wolfram-superpowers/ https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its... and look at the "wrong Mercury" example. I think it's a great example of using an external resource in a flexible way.

joe_the_user 4y ago

"Impressive and disturbing", So, ChatGPT is controlled by prompt engineering, plugins will work by prompt engineering. Both often work remarkably well. But none is really guaranteed to work as intended, indeed since it's all natural language, what's intended itself will remain a bit fuzzy to the humans as well. I remember the observation that deep learning is technical debt on steriods but I'm sure what this is. I sure hope none of the plugins provide an output channel distinct from the text output channel. (Btw, the documentation page comes up completely blank for me, now that's a simple API).

AOsborn 4y ago
> But none is really guaranteed to work as intended, indeed since it's all natural language, what's intended itself will remain a bit fuzzy to the humans as well. Yeah, you're completely correct. But this is exactly the same as having a very knowledgeable but inexperienced person on your team. Humans get things wrong too. All this data is best if you have the experience or context to verify and confirm it. I heard a comment the other day that has stuck with me - ChatGPT is best as a tool if you're already an expert in that area, so you know if it is lying.
- joe_the_user 4y ago
  It seems like you're talking about using ChatGPT for research or code creation and that's reasonable advice for that. But as far as I can tell, the link is to plugins, Expedia is listed as an example. So it seems they're talking about making ChatGPT itself (using extra prompts) be a company's chatbot that directly does things like make reservations from users instructions. That's what I was commenting on and that, I'd guess could a new and more dangerous kind of problem.
  - wahnfrieden 4y ago
    I’m not scared about an AI travel agent that books after a confirmation step. The confirmation step doesn’t need AI interface.
- danShumway 4y ago

fudged71 4y ago

We're going to need a name for this type of integration

pinkcan 4y ago
It's called ART - Automatic multi-step Reasoning and Tool-use https://arxiv.org/abs/2303.09014 https://arxiv.org/abs/2303.09014