I first encountered the following concept in one of Oxide's publications; good chance it didn't originate there though:
There is an implicit social contract with writing that the writer has put more effort into writing than the reader will need to read something. Sure you get crackpots still, but there are only so many Gene Rays in this world, so the volume is limited.
I think the same applies to PRs. Pre-AI , it was usually obvious when a PR was either completely terrible or very half-baked, and the required effort to create even a shitty PR was usually more than that required to reject it.
AI makes it trivial to make a completely terrible PR, and much easier to make a not-immediately-obviously-bad PR.
aidenn0
> Back in pre-LLM days, receiving an unexpected pull request (PR) from a fellow coder was a source of excitement and pride.
As a maintainer of a few FLOSS projects, this tracks.
The Pavlovian PR notification response has gone from, "Oh! What do we have here?" to "Groan. Do we have _anything_ here?"
I won't get specific but I just had to remove a contributor from a project after multiple submissions of either cutesy, fluffy bullshit (add ASCI animations!) or "rewrite entire project in other language". Not only did the PRs result in wasted time and energy but they also resulted in conversations about how to deal with this sort of spam. (Probably good to get out of the way and set policy but still...) So, this person probably spent fifteen minutes prompting together these stupid PRs and multiple maintainers had to spend hours agonizing over what to do about them.
ethagnawl
The thing is I totally, 100% get this. The other thing I can't help but see though is how excited my non-programmer friends are to finally be able to make software. The sense of pride and accomplishment from non-coders who are finally able to make something work the way they wanted to.
We almost need like ... noncanonical software? Not so much forks, but like ... Maybe software as like a cluster? an ecosystem? On-demand app store where features / forks are shared/upvoted/evolved by the community where the maintainers don't have to get burnt out, and when it inevitably becomes a ball of mud oh well it does the job? I really don't know!
I hope we can think about some answers and not get tribal though because this is really a huge problem and also a huge opportunity and so a minor reminder that there is a baby in that bathwater?
ctoth
I - and many, many others - learned flask from his mega-guide that he obviously spent a lot of time working on.
I feel bad for people like him who get the brunt of dilettantes who can "code" polluting his time and focus. Reminds me of that mitch hedberg joke: "When someone hands you a flyer, it's like they're saying here you throw this away." but for PRs
fantasizr
I recently built a very large test bench for System Verilog.
I ran a bunch of different compilers on it, including some open source ones.
Some of them failed some tests, and it was natural to have my LLM (Claude Fable 5) root-cause the issues, and to double-check my test bench wasn't to blame.
But now I stood with all these patches that I couldn't just throw at the upstream maintainers all at once. I ended up just filing a few issues and moved on to other things.
It felt weird to just file issues when my LLM had already spent a lot of time root-causing and fixing the issues. But then, maybe they could just have their LLMs do the same.
Still not sure if it was the right call?
thomasahle
> My perception is that there is less interest in open source, and in coding in general. The main reason I love coding is that it is a challenge, and I think this is actually the same reason why a lot of people prefer to give money to an AI lab and get a machine to spit out code for them, even with the risk of the code being subpar.
I maintain the hope that those technically minded who are really interested in coding and care about doing things properly using their own reasoning on all levels of detail will find each other and maybe become less diluted as a community by the coding-just-for-money crowd than in the past decade or two.
layer8
Even if this guy were not anti-AI, as the primary maintainer of OS projects, it sounds like he's dealing with a genuine problem.
> My initial task when a new unexpected PR arrives is to determine if there is a person behind it or not, and luckily this is easy to figure out in just a few seconds.
OK. How? That would have been an interesting explanation to me.
stantaylor
Asking contributors to first make sure there is an approved requirement before creating a PR sounds like a great idea regardless of the use of LLM.
But another issue is - AI disclosure (agent, model etc). I'm sure others tried similar approaches, but in case this is not common knowledge - I tried to see what happens if you ask agents to disclose themselves in the PR description / comments in a rule file.
It seems to work pretty well as most AI "assisted" PRs will be opened by agents using the gh cli or MCP on behalf of the user. (Of course this can be bypassed, but for someone who doesn't mind disclosing or doesn't care, this is a good step forward)
(so far worked on PRs from both Claude Code and Codex - both got a footer disclosure of the agent name and model)
eranation
We had a process at one company where you had to create an issue before filing a PR. I found it most non-sensical and introducing friction for no good reason. Very surprised to see the author suggesting it in the article.
Review is indeed the main bottleneck now for open source, and we need to solve it. Introducing more friction is hardly helping.
kvark
> Generative AI notice: I do not use LLMs, agents or any other generative AI tools for help with writing, coding, image creation or any other tasks related to this blog or my open source work.
This reminds me way back when you would see "Website written in Notepad" on websites.
comments (10)
There is an implicit social contract with writing that the writer has put more effort into writing than the reader will need to read something. Sure you get crackpots still, but there are only so many Gene Rays in this world, so the volume is limited.
I think the same applies to PRs. Pre-AI , it was usually obvious when a PR was either completely terrible or very half-baked, and the required effort to create even a shitty PR was usually more than that required to reject it.
AI makes it trivial to make a completely terrible PR, and much easier to make a not-immediately-obviously-bad PR.
aidenn0
As a maintainer of a few FLOSS projects, this tracks.
The Pavlovian PR notification response has gone from, "Oh! What do we have here?" to "Groan. Do we have _anything_ here?"
I won't get specific but I just had to remove a contributor from a project after multiple submissions of either cutesy, fluffy bullshit (add ASCI animations!) or "rewrite entire project in other language". Not only did the PRs result in wasted time and energy but they also resulted in conversations about how to deal with this sort of spam. (Probably good to get out of the way and set policy but still...) So, this person probably spent fifteen minutes prompting together these stupid PRs and multiple maintainers had to spend hours agonizing over what to do about them.
ethagnawl
We almost need like ... noncanonical software? Not so much forks, but like ... Maybe software as like a cluster? an ecosystem? On-demand app store where features / forks are shared/upvoted/evolved by the community where the maintainers don't have to get burnt out, and when it inevitably becomes a ball of mud oh well it does the job? I really don't know!
I hope we can think about some answers and not get tribal though because this is really a huge problem and also a huge opportunity and so a minor reminder that there is a baby in that bathwater?
ctoth
I feel bad for people like him who get the brunt of dilettantes who can "code" polluting his time and focus. Reminds me of that mitch hedberg joke: "When someone hands you a flyer, it's like they're saying here you throw this away." but for PRs
fantasizr
I ran a bunch of different compilers on it, including some open source ones.
Some of them failed some tests, and it was natural to have my LLM (Claude Fable 5) root-cause the issues, and to double-check my test bench wasn't to blame.
But now I stood with all these patches that I couldn't just throw at the upstream maintainers all at once. I ended up just filing a few issues and moved on to other things.
It felt weird to just file issues when my LLM had already spent a lot of time root-causing and fixing the issues. But then, maybe they could just have their LLMs do the same.
Still not sure if it was the right call?
thomasahle
I maintain the hope that those technically minded who are really interested in coding and care about doing things properly using their own reasoning on all levels of detail will find each other and maybe become less diluted as a community by the coding-just-for-money crowd than in the past decade or two.
layer8
> My initial task when a new unexpected PR arrives is to determine if there is a person behind it or not, and luckily this is easy to figure out in just a few seconds.
OK. How? That would have been an interesting explanation to me.
stantaylor
But another issue is - AI disclosure (agent, model etc). I'm sure others tried similar approaches, but in case this is not common knowledge - I tried to see what happens if you ask agents to disclose themselves in the PR description / comments in a rule file.
It seems to work pretty well as most AI "assisted" PRs will be opened by agents using the gh cli or MCP on behalf of the user. (Of course this can be bypassed, but for someone who doesn't mind disclosing or doesn't care, this is a good step forward)
Example: https://github.com/arnica/depsguard/blob/main/AGENTS.md#ai-d...
(so far worked on PRs from both Claude Code and Codex - both got a footer disclosure of the agent name and model)
eranation
Review is indeed the main bottleneck now for open source, and we need to solve it. Introducing more friction is hardly helping.
kvark
This reminds me way back when you would see "Website written in Notepad" on websites.
devrundown