@mayintoronto I'm just absolutely astounded that there's this many professional coders who *aren't* required to use it in some form for work yet.

The enterprise-grade/enterprise-cost tools are far better than the basic stuff.

We have a monthly per-dev credit budget so literally on a prompt by prompt basis I have to decide which model to send it to, based on what I'm doing and how much budget I have left.

Claude Opus 4 is definitely the best. If I get all the context loaded right and give an essay-length prompt full of requirements, it will usually get something I can send out for code review with little corrections. It is also the most expensive by far.

Claude Sonnet and Claude Haiku are not worth using.

GPT-5 Codex High is next best and gets you 90% of what Claude does but at 1/3 the cost. I usually reach for it as my primary model.

GPT-5 Codex Medium is half the cost of High and I use it for simpler tasks or fixing up other models minor mistakes.

The whole gemini family is infuriating. It often does the right thing on the first prompt but when it gets things wrong it does it in the most infuriating, non-obvious way and once you see it, it absolutely refuses to take correction.

@mayintoronto It's mixed. I like writing code, and developing. It's the fun part of my job. So I'm coming at it from the angle of it making the fun part tedious and boring. I'm also specifically using Copilot in VS Code since that's what we're allowed to use currently.

So my work is converting whole Chef Cookbooks to Ansible Roles. When I've asked it to convert the whole thing, it doesn't finish entirely. I kind of think its context is getting filled up, and moves on to the next file before it finishes the current one. I get a lot of half complete playbooks.

When I ask it to convert individual recipes, it does a pretty good job. It's using some older Ansible formatting, but that's not a big issue to me. For the most part, the output was correct and how I expected it to be.

@mayintoronto @enobacon same, I'm in the it's complicated camp, at the same time I'm using it for work as it's required and I have a feeling this will be part of the performance review at some point

I never don't bring up the environmental and human impact of our companies' LLM use

@mayintoronto It’s indeed weird.

It comes from the reality that as an operator, you need to understand what your tool does and what results it provides.

And a user should also have ecological and ethical accountability for its use.

@mayintoronto @EdwinG I tend to agree with Edwin. The commit/PR is under a single person’s name.

So if you push garbage, generated by Ai or not, it’s on you.

If you push code and lines you don’t understand, it’s on you. And that happened with slack overflow copy pasting before too.

@mayintoronto I think that, looking more broadly, "centralization of production power" doesn't adequately capture the main issue with surveillance capitalism.

(Frankly, I think that Zuboff's book is one of the most important books of the XXI century)

People don't use LLMs only (or even primarily) for coding. They often share very personal information about themselves (and people around them), because they don't feel judged.

As a result, those who control these models have the capability of making "Google Street View" of peoples' minds, but accessible only to the "congnoscienti".

As a result, corporations have more and more power over people (in every aspect of their life), and - because corporations slip away from democratic control - consequently people have less and less power over themselves (which is why the term "technofeudalism" is probably more adequate than "capitalism").

As to purpose-built tools, I'm not sure I entirely agree. I think that LLMs owe a lot of their capabilities to their "general intelligence", and that it generally shows that the more advanced models are more capable than the less advanced ones.

But it also turns out that LLMs themselves are eager users of human-made frameworks, as that those frameworks often save their time and allow them to make less mistakes.