His reasoning is quite fresh and interesting:

> What LLMs in Debian development will do, I fear, is eliminate any incentive to scrap boilerplate or reform policies that require a lot of other senseless human effort. If I had had access to LLMs 30 years ago, I might have just had them generate the rules files, replate with complexity. So they will make Debian even more firmly what it is, and ever less likely to explore what it could become.

I use and like LLMs, but it does feel like structural social changes are harder to do. For example, creation of new protocols. I hope I'm wrong, and LLMs are used to build collaborative, community projects - not just everyone vibe coding their own incompatible software.

That's perfectly wrong. Since strong coding AI, people venture into huge rewrites and other big changes that automatically make sense but otherwise would not.

Yeah it's a bizarre claim. Even better, I've literally taken some of the crappy experimental projects from the pre-Opus 4.6 era, and turned them into real software as models have become smarter.

It's crazy picking up a project that took me days of frustration pre-AI and just adding features like it's nothing.

Or being able to get down something that you've had in your mind for 20+ years but could never quite get down into a real working thing.

I would say it's a mixed bag. On the one hand, yes, you can experiment with big changes much more easily, but on the other hand, you can have cruft build up in a big way if you're not actively looking for it and trying to cut it back.

In Debian, often the barrier to change is one or two people actively disagreeing with a change who are in a position to block it through their own action or inaction, so just being able to do more work on the change yourself is not necessarily going to make it easier to change things.

That's a very narrow view. If anyone could vibe code their printer drivers, maybe free software wouldn't have existed, because it would not have been needed in the first place.

Yeah. We shouldn't have standardized languages either. Everyone was fine with their own regional dialect for hundreds of thousands of years.

Certainly, but if everyone can vibe code their printer drivers then we have free software in the sense of free speech and so users would be developers and the entire need is met. That is wonderful, not awful.

If it had happened early enough, code copyright might never have even been possible and no one can deny you your machine because you can just vibe up stuff.

Yeah I totally agree. It's wonderful, even though it may cause weird localized effects.

I use LLM's to make massive structural changes I never would have done by hand all the time. It's intractably tedious for me but easy for an LLM.

Yes, I experiment much more. It's easier to justify, and, perhaps more importantly, just WAY psychologically easier to throw away a 20 minute experiment that your robot wrote for you than the same when it took you a week of tedious manual coding.

Yeah experimentation is probably the best use case of LLMs

And at the end of the experiment I usually have my solution

Yes, because LLMs have no online learning, it's a mistake to innovative in programming languages and tooling. Use what exists instead, the power of having training data is often underestimated.

His point is using LLMs to build software will lead to code so complex that it requires an Agent. Or that it will cause avoiding cleaning up the boilerplate because we can tolerate it now because of agents.

And thus this will all bias towards complexity and fragility..

Aside from Bun, which is owned by Anthropic, do we know of open source projects that allow AI where this has happened?

OpenClaw IMO, It's a such big mess that only agent can operate it.