I don’t know what you mean by that analogy or your response to that quote.
I mean that, as far as I remember, it was a very good application, and remaining good for some time, however, it is not anymore.
Kolaru’s summary left out what for me is the most critical point: Keno’s wish, expressed in the State of Julia talk, that contributions to Julia would be either written by LLMs or checked by LLMs.
I think that asking people to use proprietary tools is antithetical to the open-source spirit. And it raises the question of how much the community here cares about open-source. I had always assumed that it was as passionate about it as I am. I mean, if I were ok with using proprietary software I’d have stuck with MATLAB, it was a lot of work to transition to Julia. I also want to emphasise that this is not a matter of money: I have research grants, I could use them to pay for whatever frontier models I wanted to.
Am I alone here?
Github is a proprietary platform for hosting open source software.
IMO, that is not a good representation of his commentary here. Here’s the closest he came to saying that:
“We’ve reached the point where, when I see a PR that somebody hasn’t at least run through an LLM and done an initial review, that might indicate there’s still some bugs to be found there. These tools have become unambiguously helpful.”
But perhaps even more importantly, he also included a lot of very explicit caveating in multiple ways: the fact that these are his personal musings, still in flux, the models still generate very subtle bugs, and this all still requires careful thinking around the community impacts.
Here’s a direct link to the timestamp on this segment:
And by far the worst part of developing Julia, and the most frequent subject in gripes. Coincidence? I think not.
The most popular tools have the most gripes, so no it’s not a coincidence.
Nothing about popularity here, github is simply a necessary part of developing Julia whether you want it or not.
And you seriously think there’s nothing to complain about github’s quality? It achieved a historic zero nines of uptime.
And you seriously think there’s nothing to complain about github’s quality?
Having an issue with GitHub’s uptime and having an issue with GitHub being closed source are two entirely separate issues.
You can use whatever tools you want, in particular for your own open source. And I understand there’s a very vocal “hard core” FOSS community (that’s FOSS in hyper-capitals). In my experience, that ideological hard core tends to get a major eye roll from the vast majority of the tech community, even the majority of the “open source” community (that’s “open source” in non-capitals), not to speak of the general public.
Generally, people use whatever tools they have access to, commercial or not. A pretty big percentage of developers runs macOS (or Windows, of course), which are commercial closed-source. And you can’t just gloss over GitHub: Yes, if you’re contributing to my projects, then I’m asking you to use GitHub, which isn’t any less closed source and commercial than any LLM.
If I know you have access to an LLM, then I can absolutely ask you to run it on your PRs without violating the “spirit of open source”.
That doesn’t mean that LLMs don’t have considerable negative externalities that are worth worrying about, but “not open source” seems pretty low on the list to me.
Wow. Ok, then indeed I am alone, I stumbled into the wrong community by mistake.
I mean, I don’t know that you’re alone, or that this mismatch means you’re in the “wrong community”. I might take the prevalence of the MIT license in Julia projects over the GPL as an indication of the ideological leanings of the community, but it’s not like I’ve run a representative poll.
In any case, I think there’s room for a variety of view points in the community, including more or less ideological FOSS viewpoints. But that doesn’t mean those views are necessarily going to be the basis of a consensus, or that I’m not going to tell you that in my opinion, it’s an unreasonable stance 
My own packages are MIT-licensed. This doesn’t mean I’m ok with proprietary software. I chose MIT simply because GPL makes the code harder to incorporate in other open-source software, and I want my software to be as easy to use as possible. I’m aware that this allows someone to incorporate it into proprietary software. I don’t think that’s likely given the arcane nature of my packages, but if it does happen then so be it, at least I’m not giving unnecessary headaches to legitimate users.
And yes, I think github and LLMs are in completely different categories of problem. Github is just a (shitty) git hosting service. It doesn’t have anything to do with the code. It’s certainly a blemish in the face of Julia, but could be switched for self-hosted gitlab tomorrow if anybody were willing to pay for it.
Now asking people to use a proprietary, expensive, cloud-hosted LLM to create the code itself? So people need the blessing of a megacorp to contribute? I’m at a loss for words. If you can’t see the problem - or the difference - then I’m really in the wrong community, and I have nothing to say to you.
@araujoms I’m not sure you are interpreting what @Keno said as he meant it, but hopefully he can clarify. I more took his comments as a suggestion that having an AI reviewer, for example as an automatic hook via Github CI or such that is on run when updating a PR, would be valuable in his opinion. I don’t think he was saying that individuals must themselves use AI.
Ah, consensus. Consensus has been discussed here before. I’d like to sum up my view and express my thanks for participating in this discussion. I support higher adoption. I believe broader community engagement is the priority. Specifically regarding LLMs trained for Julia, and ideally or in the consecutive next step development of LLMs written explicitly in Julia.
I just don’t think that’s a reasonable framing.
When you submit a PR, I’m going to run it through CI, and insist that you fix any failing tests or wrong formatting that it identifies. That runs on a closed source commercial platform, but is a complete standard practice in open source.
Now, I’m going to recommend that you have your LLM (if you have access to one) to help you prepare your PR (just like I recommend that you run the test suite locally). And I’m probably going to have an LLM do a review of your PR and ask you to fix problems that it may identify.
Does that mean you need the “blessing of a megacorp”? I don’t think so.
And it’s not like “free CI” doesn’t have negative externalities either: I have long advocated that running every PR for every package through “Julia nightly” and every version of every operating system is a waste of resources. We treat CI minutes as “free” because Microsoft subsidizes them, but the resource usage from unnecessary CI minutes across the entire Julia ecosystem is likely significant. Granted, it’s likely dwarfed by the resources from the current LLM craze, and that’s a problem, but a problem that needs to be solved with things like pricing carbon and strong environmental regulation.
I think you’re blowing this out of proportion.
You do not need to use LLMs to contribute to Julia, and I also don’t like using proprietary software for Julia development. However currently in the hands of experts these tools can find lots of bugs. I think in the hands of non experts or given the wrong context they will just create annoying noise. There are other open source bug finding tools like Supposition.jl, writing careful tests, and thinking. But in general it is a safe assumption that most PRs are not perfect and can be improved with careful review and testing.
@araujoms — I would hope that you see in Keno’s comments (and in particular his emphasis on community impact and how strange and fast this all is) his own humanity and good-faith endeavor at finding ways to best improve Julia using the most effective tools at hand without alienation. That’s what I see in them.
And I’d hope that you see in this topic a group of people who are similarly trying to figure out how to do so.
I certainly don’t want your perspective excluded here. I actually agree with you that the closed and proprietary (and geo-politically-locked!) nature of companies at the frontier is problematic… but in framing that commentary as “excluded” out-of-hand you are removing yourself from the discussion and ironically unilaterally re-framing those issues to seem so too!
These are completely different things. The test suite is open source and part of the package. You can easily run it locally, and indeed that’s standard practice. That it also runs on Microsoft’s servers when people make a PR is just a detail of how the system is set up right now.
Now if you don’t have Anthropic’s blessing you won’t be able to run Claude by yourself to prepare the PR. And that’s not an idle concern, the US indeed tried to ban non-US citizens from using it a couple of months ago. And if Julia doesn’t have Anthropic’s blessing it won’t be able to use Claude to review the PR either. Now if Microsoft decides to blacklist some person or organisation from github, the code is still there, it will run on gitlab tomorrow. If Anthropic blacklists you? I guess you can try OpenAI, but the list of alternatives is very short and none of them is under your control.
The freedom to create software by yourself is quite fundamental. No wonder GCC was one of the very first GNU projects.
As an addendum Matt, do you prefer the new Kimi K3 license from Moonshot AI, the one regarding Model as a Service with a 30% revenue share and a $20m total company / affiliate revenue limit? This is a rhetorical question. I don’t intend to dive into this or start a whole discussion about it. Could be a speculation. I just wanted to point out the potentially far reaching repercussions such a business model might have. Thanks again for the opportunity to participate.
I share your and @mbauman concern about the geo-politics, monopolization, and likely a whole range of other issues.
But come on: Nobody is making it mandatory for you to use a tool that you don’t have access to. We’re merely recommending that you make sensible use of the tools you do have access to, and as a maintainer, I am going to make reasonable use of tools that I have access to.
We’re far from a situation where your ability to contribute to open source is entirely dependent on your access to a particular LLM provider. I agree that would be a problem.
If we ever get to that situation (and I suspect that we won’t), then that likely means that LLM technology has proven itself so generally useful that it also has become widely democratized. And hopefully, we’ll have gotten a handle on mitigating negative externalities. I’m just saying that if LLMs become truly indispensable on a large scale, countries and individuals have huge incentives to obtain access to that technology, one way or another.
You’re imagining a cyberpunk dystopia that doesn’t currently exist and hopefully will never exist. Which isn’t saying that everything is smooth sailing and we don’t have anything to worry about, or that we don’t need to continue to fight politically against all kinds of bad things, from corporate greed to environmental devastation.