Live data from Hacker News

Pulumi Insights – AI generated IaC programs

pulumi.com

1–10 of 55 posts

Re: Pulumi Insights – AI generated IaC programs

#2
ohhhh this is really smart; every terraform user has at some point run up against the limits of it not being a language and flirted with pulumi. natural excuse now to get started, esp if the auto import works

targeting translation between declarative-ish languages is a natural day-1 application for LLMs

declarative languages generally good, bc 1) it's relatively easy to check them, 2) they are low-boilerplate (less to get wrong + better support for small token memories)

wonder if this backfires bc pulumi is more powerful, giving the LLM more chances to get in trouble? like there are cases where in order to check a pulumi program you would need something like TLA+ with a good model of the system you are targeting

Re: Pulumi Insights – AI generated IaC programs

#5

CDK/Pulumi are going to be used by LLMs to spin up website infra exactly to spec. Devs will fill in the business logic interfaces. Going to be wild

but how will it be maintained? "IAC" seems like a bit of a misnomer here, if I'm understanding correctly. Now, if Pulumi generated terraform for me to maintain, that'd be a different matter, but this article seems to just produce cloud resources based on LLM prompts, which are not deterministic at all.

Re: Pulumi Insights – AI generated IaC programs

#6

CDK/Pulumi are going to be used by LLMs to spin up website infra exactly to spec. Devs will fill in the business logic interfaces. Going to be wild

Yeah but ChatGPT can already write Terraform and Cloudformation really well. So this doesn't seem that special besides it's already baked into the pulumi eco system.

Re: Pulumi Insights – AI generated IaC programs

#7
Sorry to be offtopic, but I've been using Pulumi at work for the past 6 months and I'm really not impressed. It's basically just Terraform but worse, with a million ways to declare your infrastructure instead of just one. Infrastructure people tend not to write the best code and from my observation the extra freedom of an imperative language just makes stuff even more complex and harder to maintain. It's also much harder to automate than Terraform, I am not aware of any equivalent to Atlantis.

Also, Pulumi previews (equivalent to plans) are complete bullshit. If you don't write your code carefully, resources can be created and removed and you won't know it's going to happen until you start applying... it's an engineer's worst nightmare when a tool lies.

Re: Pulumi Insights – AI generated IaC programs

#8
Pulumi employee here!

Worth noting that while lots of insights are leveraging LLMs/AI, it's only one pillar of the launch today.

The cross cloud search and analytics are really valuable additions to the product. On day 1, being able to export all your resource definitions to BI intelligence tools and drill down into specific areas of interest is something that has traditionally been very difficult.

On a personal note, being able to simply search for "who spun up this EC2 instance" without traversing through cloudtrail is a godsend, too.

Re: Pulumi Insights – AI generated IaC programs

#9

CDK/Pulumi are going to be used by LLMs to spin up website infra exactly to spec. Devs will fill in the business logic interfaces. Going to be wild

but how will it be maintained? "IAC" seems like a bit of a misnomer here, if I'm understanding correctly. Now, if Pulumi generated terraform for me to maintain, that'd be a different matter, but this article seems to just produce cloud resources based on LLM prompts, which are not deterministic at all.

It produces pulumi programs which are the equivalent of a terraform module or main.tf file. Whether or not said programs are deterministic is another matter.

Re: Pulumi Insights – AI generated IaC programs

#10

CDK/Pulumi are going to be used by LLMs to spin up website infra exactly to spec. Devs will fill in the business logic interfaces. Going to be wild

but how will it be maintained? "IAC" seems like a bit of a misnomer here, if I'm understanding correctly. Now, if Pulumi generated terraform for me to maintain, that'd be a different matter, but this article seems to just produce cloud resources based on LLM prompts, which are not deterministic at all.

Pulumi is an IAC tool similar to Terraform (it actually usually calls terraform providers over gRPC under the hood), just written in languages like Typescript/Python. I think the intention would be to use this AI prompt thing to generate pulumi and then to insert it into your codebase, just like you'd do with terraform, and then it becomes deterministic.

I've used ChatGPT and other tools to generate Pulumi before, so I'm not totally sure how this specific tool from Pulumi is different, but I'd guess they've somehow trained it more specifically on their sdks/docs or something

Post reply on HN