| 13 Mar 2026 |
 magic_rb | and also i dont want to force atemu or pentane to have to yell at us | 23:20:02 |
 waltmck | To be fair you brought up this topic, I am just responding. But yeah I agree I'll drop it | 23:20:56 |
| magic_rb | yeah i know | 23:21:02 |
| magic_rb | i have a bad tendency of getting very tilted about this :( | 23:21:15 |
 matthewcroughan - nix.zone | They will never produce quality code. Because quality code is often coinciding with new code that actually solves the problem you're trying to solve. Which they will not do. | 23:35:29 |
| matthewcroughan - nix.zone | What is code quality anyway? Maybe actually being fit for purpose is part of quality. | 23:36:09 |
| matthewcroughan - nix.zone | If you use an LLM to generate code, you are fuzzy finding a template and hallucinating a portion of it. This is bound to be low quality. | 23:36:51 |
| waltmck | I don't know how you can be so confident about this. Five years ago I would not believe that they have a coherent conversation, and now they are playing an active part in solving outstanding math conjectures | 23:37:09 |
| waltmck | * | 23:37:24 |
| matthewcroughan - nix.zone | Because I know how they work. And I'm not looking at historical appearances of improved power or competence. | 23:37:40 |
| waltmck | Do you know how human brains work? There is actually not a huge fundamental difference | 23:38:21 |
| matthewcroughan - nix.zone | They do not know how to play chess, but they can convince you that they know how to play chess. But eventually they make an illegal move (hallucination), because that's just how they work. Throwing more compute at it won't solve that. | 23:38:27 |
| matthewcroughan - nix.zone | Comparing it to a human brain is just L O L | 23:38:37 |
| matthewcroughan - nix.zone | You don't know how they work, but are eager to compare them to our own brain. | 23:39:09 |
| matthewcroughan - nix.zone | In place of understanding, you would rather directly compare apples to oranges. | 23:39:26 |
| matthewcroughan - nix.zone | Because it's more exciting that way, I guess. | 23:39:36 |
| matthewcroughan - nix.zone | I know it's tempting to speculate, but actually knowing is more fun. | 23:40:00 |
| waltmck | I'm a PhD student working on machine learning. Just 6 months ago I presented a spotlight oral paper at ICML. You don't know anything about me | 23:40:24 |
| waltmck | I'm not sure where this ad hominem is coming from, I don't think I've been disrespectful to you | 23:41:00 |
| matthewcroughan - nix.zone | Not only does knowing, prevent you from seeming like Sam Altman talking about how we should build dyson spheres, (a literal physics in-joke), it also means you can focus on what you can actually use them for. Like alt text captioning and translation. | 23:41:11 |
| matthewcroughan - nix.zone | They are not comparable to human brains. Just don't. | 23:41:27 |
| waltmck | Got it, great argument | 23:42:06 |
| matthewcroughan - nix.zone | ML can produce machines that do know how to play chess. | 23:42:57 |
| matthewcroughan - nix.zone | LLMs aren't one of them. | 23:43:01 |
| matthewcroughan - nix.zone | That's because LLMs don't learn rules. They learn patterns. | 23:43:21 |
| waltmck | You don't know what you're talking about, but I don't really see the point of continuing this discussion here | 23:44:03 |
| matthewcroughan - nix.zone | Fantasizing about how this can be eliminated with "a little bit more data bro" is just making me sick. People keep saying it. | 23:44:06 |
| waltmck | You can DM me if you want to talk about this further | 23:44:12 |
| matthewcroughan - nix.zone | If it's so good, go and build it and demonstrate how it no longer causes problems, hallucinates or it wastes time. | 23:45:18 |
| matthewcroughan - nix.zone | I will believe it when I see it. | 23:45:24 |
