[cube2222]

Hey, I'm curious about the pulumi-ai cli.

Specifically, did you solve the problem of stale API information?

What I mean is that using gpt-4 to generate code is generally very straightforward, but due to the knowledge cutoff it won't know about i.e. new AWS APIs like Lambda URLs.

Is this something you've managed to solve? Or is it just the "even with that knowledge cutoff there's enough value" situation?

[AaronFriel]

Pulumian here - this is something I'm working on and hopefully we'll have more to share soon.

We're in a good position here in that our providers have rich schemas: https://raw.githubusercontent.com/pulumi/pulumi-kubernetes/m...

However our larger providers, primarily cloud platforms, have schemas much larger than the context length of the model. So the trick is scoping that down to the necessary & sufficient amount of data into a prompt, whether via plugin (not yet available via API), preprocessing the prompt, or using a langchain-esque approach.

As Károly Zsolnai-Fehér[1] says, "what a time to be alive!"

[1] of Two Minute Papers fame: https://www.patreon.com/TwoMinutePapers

[cube2222]

Thanks for the explanation and good luck, then! Very curious to see what you come up with.

[verdverm]

There are attempts at solving this issue more generally, by giving the GPTs access to external sources of information, and ofc the right prompts & chaining