I am not imagining things. The ban on non-US citizens happened. It didn’t work as the government intended, but it has illustrated how easy it is to have the rug pulled from under you if you rely on LLMs that are not under your control.
Of one particular model of one particular LLM provider (and for entirely stupid reasons). I don’t think this really affected anyone’s ability to contribute to open source in a serious way.
But yeah, when Claude is down or Github is down for a few hours, US citizens or not, a lot of us have gotten ourselves into a situation where we’re going to be twiddling our thumbs (or spending the time by engaging in this discussion). That might be something to be concerned about at a personal level. But if this became a permanent situation, or if both Anthropic and OpenAI really were to limit access to the US only (I’m not sure what the scenario for that would be, but I suspect it might be a situation where “how do I find food” becomes more pressing than “how can I access my LLM”), we would find other ways.
I suspect 5 years from now we’ll be having distilled special-purpose LLMs running on our laptops that will be more than capable to review your Julia PRs. This technology isn’t going to go away, and it’s going to be a major part of software development.
Claude/OpenAI down is really not a problem even right now, since there are chinese models doing the same for almost all tasks, at a fraction of the cost (and probably of the compute), which are even self-hostable. Also, I think the strategy currently adopted by those companies are much more aligned with FOSS than most/every american hyper-scalers. Speaking of particular providers at the moment doesn’t really makes sense (fortunately).
(Just to be clear, as I mentioned some model names in my previous post directed to Matt, I have absolutely nothing against Chinese models, I was referring specifically to the business model and the structure of the licence.)
My 2 cents: I think we are overreacting. People, we are living a world transformation second only to the dawn of the internet that we, millennials, saw. By that time, it was almost impossible to predict where the world would be in 1 year. We saw extremely wrong guesses from both sides. We will see extremely wrong guesses now either.
So, instead of predicting the future, whether we are living on an AI bubble, or there will be no more manual programming, let’s try to experience more of the present. Currently, we have access (somewhat cheaply) to amazing tools that can help us bring the Julia ecosystem to match what other languages have today but took several more decades. Furthermore, there is no enforcement, each developer is free to do what they think is correct.
When things change, we assess and adjust the course 
I mean, in one scenario (AI dominates), programming languages will be a thing of the past, the models will just write executables way better. Hence, Julia development, open source, etc. will be a pointless discussion. In other scenario (AI bubble ends), we will have used this time to make Julia compete to the top mainstream languages.
I just affirmed there’s plenty to complain about Github, and you’re ironically making these sort of assumptions about one of the few people here who doesn’t use it (or need to). To clarify, my point was that much of the open source community, a far broader group than Julia users, do use proprietary tools in the process; proprietary operating systems were already mentioned, and same goes for hardware and architecture.
Open source is not a monolithic philosophy, and evidently, most do not share your open-source spirit. Perhaps that was once true, but by now, open-source contributions are mostly funded by corporations, not fickle donations. They don’t do that out of generosity; they’re putting that (sometimes free) labor and technology right back into their products and pulling people into their ecosystems. To bring it back to AI, Github hosted the vast majority of repos for free, and in return they got a trove of training data for LLMs for free. AI companies are subsidizing and sometimes handing freebies to open-source developers for the same reason. Of course there are complaints about this state of the world, but this didn’t happen by chance, and “just stop” isn’t realistic.
I really doubt this. Inference has been heavily optimized in the last few years, but models have apparently outpaced it. If people are cooperating on cloud services anyway, there’s no reason for constraining compute like this. It’d also be easier to stick to one standard instead of just trusting that every laptop runs a proper model.
I still don’t think there’s any basis for big superficial predictions like this. Whether Julia is even on the right track is still up for debate, and many people have reasonable arguments for saying “no, these other languages are.” Throwing more agents at a problem doesn’t solve it if the problem isn’t the right one to begin with. Assuming that it does, much more compute is put into mainstream languages, so they’ll be solved long before niche languages are.
Not necessarily. The penchant for breaking and changing needs to be there to get anything really changed for most languages. For example, I don’t think Python can add strict typing at this point, even though it would clearly be better for an AI world to have it.
I think Anthropic agree that such tooling for Python would help in an AI world given they bought the company developing this typechecker: GitHub - astral-sh/ty: An extremely fast Python type checker and language server, written in Rust. · GitHub
Though the issue with Python’s types aren’t necessarily the checking but the semantics. The semantics of its types are incapable of expressing what can be optimized for example. You’d need to break its type algebra to change that.
Is it difficult?
Strongly disagree. Coding Julia is joyful. Using agents is joyful. Using agents for coding Julia is double joy.
No, programming languages are not going to be a thing of the past even when AI completely dominates the industry. AI needs high level machine code representation as much as humans need, which programming languages essentially are. Imo, with spec driven AI development, we are converging to a workflow where stating a problem in natural language, planning implementation and implementing through a chain of intermediate code representations of different levels of abstraction (from high to low) are combined into a single highly automated compilation tool chain, with occasional human-in-the-loop interventions at different levels.
Happy to see that AI discussions in Julia community are going around how to use AI in more productive ways, and not around how to ban or restrict it like in some other dev communities.
I worry more about Github than Claude for some reason. I switched to self hosted Gitlab for my stuff for that reason.
Another part that we’ve been trying to do to avoid “endangering the community” here is to moderate these discussions to avoid flamewars and direct attacks on each other. I appreciate how folks have largely conducted themselves here thus far — thanks! But this is one of those conversations that could continue endlessly; let’s plan on a weeklong time-box for this topic (and the split one, too).
Very nice! This is by far the most specific, quantified, and quality-tested showcase of LLM use in a mature repo that I’ve seen.
Nonetheless, it seems wholly contingent on getting $315,000/yr in free compute from AI companies who are all hemorrhaging money.
Reducing that portion of the talk to just his words doesn’t convey the full extent of the communication - this is a core developer in THE annual talk about Julia making strong statements about AI to the community, some of them not with caveats, etc., and weak caveats don’t really mean much from people speaking from positions of authority.
I don’t think the personal stump section, as performed, was at all appropriate for the State of Julia talk, and I found it to be poorly reflective on the community as a whole in my opinion. It doesn’t seem like I’m alone on that.
With which parts of Keno presentation do you disagree? “Not enough caveats” is rather unspecific. And the Julia community appears to be robust against some "authority"s talking points.
Not completely on topic, but a good example, where I think that AI has been digested as a tool, but had no impact on the big picture of what Julia wants to be is RFC: experimental session save/restore - Pull Request #62809 - JuliaLang/julia - GitHub
From the PR (“what this is for”):
Interactive work. Ending the day mid-exploration and getting the same REPL back tomorrow: loaded data, defined functions and types, loaded packages,
usingscope, including methods you defined onBaseor package functions. Cheap checkpoints before risky steps in a long exploratory session. Keeping a few alternative working states side by side and switching between them.
Our engineers and all the other software developers I know have moved from the REPL to the prompt, where there’s no such thing as “end of the day.” It seems like people are feverishly deploying AI on autopilot to solve the problems that have plagued them for ages, never looking up to see that those problems have disappeared.
It would be cool to reactivate the Julia Developer Survey to find out how other Julia users/developers have changed their work approach. This survey would be a good basis for (de-)motivating features.
No a side note, I didn’t see much forward looking statements in the State of Julia talk, which, given the name, is understandable. Could someone give me pointers (or links) to the talks that cover the core developers goals?
I think that this would be a valid concern for critical infrastructure. It would be irresponsible to transition to a framework that needs heavy subsidies that may disappear at any point.
But my understanding is that this specific proposal is about an extra check for PRs using agents. If, for some reason, this becomes unavailable, then we lose the benefit of this service, but we will be no worse off.
I would consider this a practical proposal that is difficult to evaluate on an a priori basis, arguing from first principles. The real test would be empirical — then, after about 1000 PRs, we will be in a position to see how much it helps, how much noise it adds, what the lags are, etc.
At the same time, I see no problem if the Julia maintainers decide to wait a while with implementing this. This is a boat one can jump on anytime. There should be no fear of missing out.
Yes, I just noticed there was apparently no survey for 2026. This is unfortunate, given that AI tools have become overwhelmingly adopted in software development over the last 12 to 18 months. It would be interesting to know what the broad community actually does, and how do people actually think. Those kinds of politicized discussions sometimes have a tendency to be hijacked by a vocal minority.