Hacker News Reader

Comment by dmd | Hacker News Reader

"I need to build a skyscraper. I haven't tried using steel, but using wet spaghetti has been garbage."

The vast majority of humans on the planet haven't been able to afford to work with ChatGPT-4 (and I don't consider M$'s offerings of it to have nearly the same range of use). Perhaps there are good reasons our interlocutor hasn't tried ChatGPT-4. I also admire that someone has been testing with local models; we benefit from their experience as well.

True, but we're not talking about "the vast majority of humans on the planet", we're talking about the sort of person who plays around with Nix - who can probably afford gpt4's $0.03 / 1K tokens.

But I absolutely agree with your 2nd point!

I understand that claim. I'll agree the vast majority of Nix users I've encountered could likely afford to use ChatGPT-4 if they wanted.

As to our domain of discourse, I'm here to speak about what we owe to all people, and I'm not here to make that assumption about our interlocutor. Maybe you know our interlocutor, but I don't think your standalone quote is a morally appropriate argument with a stranger. Setting aside those who might be the sort to possibly play with nix who haven't, I've also met people who have played with nix but cannot afford ChatGPT-4, or don't have access to pay for it. There really may be good reasons our interlocutor hasn't tried ChatGPT-4, and I think your joke can be easily misinterpreted as a flippant response or a putdown.

I appreciate the meme though. You're correct about the vast difference in ability between ChatGPT-4 and every other model the public can access.

You keep saying “chatgpt-4”; I wonder if you’re aware that you don’t need to pay $20 a month (which is what that costs) to use gpt4. If you use it through the API instead, even a fairly heavy user comes out to a dollar or two a month.

I understand one can pay to use APIs to various models offered by OpenAI. API use with maximum context does add up for significant users. It does cost something, and not everyone can pay that cost.

Even GPT-4 has been hit or miss for me. I think it's less about the model and more about the data. I'm intentionally aiming to write my home-manager stuff as modern as possible (e.g. flakes, recent home-manager syntax) but most of the content out there uses older paradigms, so it's often hard or impossible to get chatGPT to produce quite what I want.

Still helpful, just not as helpful as it is for like, python 3.11.

I probably get more mileage out of it because I stick to one giant skiddieable configuration.nix file (~2k lines so far) for a base on a single machine, with some scripting to get some stuff done after that. I don't use so much of what makes Nix radically powerful, `/nod`. Like, I'm runnin' a noob's GNU/Systemd/Loonix up on my Rube-Goldberg Machine for the 'tism points, my sibling in Christ; it's spaghetti, and I wish I could share it with Saint Terry. So, yeah, I can see how for those seeking modern nix specific affordances, it's a much worse experience. =/. One can hope that the next model will make it rain.

There are so many facets of the nix ecosystem that have so little data in the training corpus, and hopefully that will be fleshed out further in future generations. Probably no way around RLHF from some experts, and whatever magic sauce they can cook up after that either, to get the right sort of data.

Reply on news.ycombinator.com