I built this after considering whether it is possible to create AI without an LLM, such as coding agents, world models, and so on. My initial goal was to achieve that kind of artificial intelligence. (I hope the past extension code will be gradually cleaned up.) I am also aiming for dynamic derivation (for checkpoints?). It seems to be being discussed in other threads, but initially, there is separate code written in Rust that is compatible with nixpkgs, influenced by tvix. However, since supporting nixpkgs was not the original goal, we will have to wait and see if it is merged back in later for quality testing.
I’m sure you’re a nice person and you’ve got great ideas but what the hell did I just read?
AI-generated config gate — classify AI-produced Nix/K8s/SaaS config as :ok / :held / :reason before it is applied.
What?
Semantic CI — verify not just “tests pass” but that evaluator / machine / lowering / fuzzer all mean the same thing.
What?
Agent capability + witness layer — when an AI agent/plugin touches files, env, or host objects, gate it on a capability and leave a witness/receipt
Wha- Okay no I actually want to engage with this. I’m going to risk wasting my own time and argue that the issue with inserting AI as a buzzword into everything is not source provenance (i.e., receipts) but actually the statistical likeliness of a fuckup. You need to establish a really good sense of “capability” here (which I don’t see explained properly) and even better focus on not letting LLMs touch anything that can fuck up. Because they can, and they will. Neither capability nor receipts fix this.
Quite frankly I’m getting tired of all influx of slop in this forum. First of all, I think you have an idea. I think it is worth developing and if you think it’s a good idea I think you should continue pursuing it. However this README is entirely Claude and it does what Claude does: obfuscates the point . I cannot see what you’re trying to acheive, and I cannot see what you want out of sharing this link besides a few cheap stars. Please do me a favor.—do all of us a favor and write the README in your own words so we can discuss and think about the actual project.
On a matter unrelated to your prose, I think the idea is conceptually flawed in the sense that treating hosts as “limbs” (I have little idea what that’s supposed to mean) and letting only agents engage with them puts so much trust in agents and so little in the user. AI is non-deterministic. Not even formal verification can fix this. A receipt can help you find out which agent fucked up and maybe why (assuming the receipt won’t be forged) so you can kick it out of the loop but I don’t see what this achieves. So it’s IaC… bu not declarative? Why?
4 Likes
Will we ever know whether there was an actual person behind this post? I find it more concerning that there might be. But also less likely.
2 Likes
You either didn’t read my response, or your agent is misbehaving. Either way, good luck in your future endavours.
3 Likes
What a fever dream of a post. I don’t have time to properly dissect every paragraph of yours, but I’d like to reiterate what has already been said - use your own words. Nobody wants to converse with an LLM, and that’s exactly what’s happening here. Also, you clearly did not read Discourse LLM Policy.
3 Likes