Jean Louis <[email protected]> writes: > * Björn Kettunen <[email protected]> [2026-03-25 01:22]: >> Ihor Radchenko <[email protected]> writes: >> >> >> At some point I'd like to tackle Prolog, which is also abysmal. And >> >> ob-haskell doesn't play well with latest haskel-mode; I have to use a 10 >> >> yo >> >> version. Your thoughts on all the ob-<your obscure language here>? >> > >> > If you know bugs in ob-haskell (which *is* a part of Org mode), feel >> > free to report them. >> > >> > Whether you can send LLM-generated patches, it depends. >> > >> > You definitely cannot send complex patches. LLM-produced code is likely >> > public domain, and adding non-trivial amount of public domain code may >> > have implications on GPL licensing. Until GNU clarifies on these >> > implications with lawyers, we are putting large LLM contributions on >> > hold. The best I can suggest here is turning patches into third-party >> > packages - those are definitely not restricted in terms of what you can >> > use or not (all the legal burden will be on you, the author). >> >> Assuming for a second the second the source of the code isn't by itself >> the issue. How can it be ok to submit code generated from closed source >> SAS LLVM's? That goes against what free software is from anyone's point >> of view in my opinion. > > Dear Björn, > > Greetings. > > Your point about LLVM vs. LLM -- probably you mean LLMs? I personally > do not know what you would mean with LLVM's.
Yes exactly, sorry. > If with SAS, you are referring to to this link below, I can understand > you: > > Who Does That Server Really Serve? - GNU Project - Free Software Foundation: > https://www.gnu.org/philosophy/who-does-that-server-really-serve.html > > The GNU philosophy piece "Who Does That Server Really Serve?" warns > against exactly the kind of dependency you're describing—but it also > assumes a world where users are forced to interact with software as a > service. That assumption is increasingly outdated. Today, I can run > Qwen, Llama, DeepSeek, or any number of open‑weight models entirely > locally on my own hardware. Hugging Face, Allen AI, IBM, Apertus, and > others are making this the norm. When I generate code, it's on my > machine, with models that are publicly available, often under > permissive or free software licenses. The "proprietary service" > framing doesn't apply when the user controls the tool end‑to‑end. The assumption isn't so outdate when users predominantly interact with SAS LLM's such as Claude. Sidenote: We should not call them AI but LLM. The former obfuscates what these actually are. > Like right now while we are speaking, I am running this one in > background, which by using opencode software, improves my Elisp > projects: > > (rcd-llm-get-current-running-model) ⇒ "Qwen3-Coder-Next-UD-Q3_K_XL.gguf" > > It has "open" weights, it means, there is Apache 2.0 license for it. > > Here are the dataset links in a single-level list: > > - nvidia/Llama-Nemotron-Post-Training-Dataset: > https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset > - SWE-smith: https://github.com/SWE-bench/SWE-smith > - SWE-Flow: https://github.com/Hambaobao/SWE-Flow > - Multi-SWE-RL: https://huggingface.co/datasets/Multi-SWE-RL/Multi-SWE-RL > > This may not be all, it may not be free software licensed, but guess > what? I don't care. People can read information, there is no license > to read, so software can read it on behalf of people and create > weights. I don't want to bother my life with it. Point is that users > of huggingface.co and many other websites providing LLMs, already > enjoy so much of the free software freely generated, where there is > almost zero chance that freedom for those people who do generate > software would be legally attacked somehow. Maybe corporations > creating such could get legal attacks, but chances for users is > basically zero. The issue here is still the training data they use. > That is practical reality. > > We now have a dilemma: > > To create more free software, faster and more efficiently, while > enjoying our lives — breakfast, the swimming pool — while the computer > works on it in a continuous loop; > > Or to pester, annoy, scold, and protest because the new technology is > faster than anything before, and the true "artist" — the programmer — > is now left behind. > > So if the process is free (local, "open" models), and the output is > code contributed under GPL with a human submitting it, what exactly is > the ethical problem? > > But beyond that, I want to challenge the deeper premise: that we > should even be treating copyright provenance as the gold standard for > free software contributions. > > LLMs, by generating code that lacks a clear human author and thus > defies traditional copyright attribution, are a feature rather than a > bug. > > They actively undermine the copyright regime that free software has > had to work within for decades. Rather than seeing this as a problem, > I view it as liberation—a way to bypass the very system that free > software has always had to negotiate. > > In this framing: > > - Copyright was always a burden on software freedom > > - Free software fought within that system (GPL, copyright assignment, etc.) > > - LLMs now let us produce code without the usual authorship constraints > > - This effectively destroys copyright's grip on software production > > That's not a bug. That's the point. > > The free software movement spent decades playing by copyright's rules > because there was no alternative. Now there is. If I can generate > high‑quality GPL‑compatible code using locally‑run open models, > without assigning copyright to the FSF or worrying about contributor > agreements, I'm not weakening free software—I'm accelerating it. I'm > bypassing the friction that copyright intentionally creates. > They also can protect FOSS code to stay foss. Similar to what public-code public money from the FSFE tries to do (https://fsfe.org/activities/publiccode/index.en.html). > So when projects like Org mode say "we're holding large LLM > contributions until GNU clarifies licensing implications," I hear: > "we're waiting for the old system to give us permission to use tools > that make that system obsolete." > > The legal caution is understandable. But ethically? The fear that > LLM‑generated code somehow taints free software rests on an assumption > that the process must be constrained by copyright thinking even when > the output is clearly free. That's cargo‑culting free software > formalism over actual freedom. Using such tools is like sawing of a tree while you sit on. I could go over the LLM part as long as the code it's trained on is foss and the code out of it also stays foss. > Let's stop pretending that copyright assignment and human‑only > authorship are essential to freedom. They were tactics, not > principles. If we can now produce more free software with less legal > overhead, using tools we control, that's a win—not a threat. > Copyright assignment doesn't even work in most European countries. I.e. if you're in Germany you can't sign your copyright away. Which is why the FSFE suggests to use Fiduciary Licence Agreements. Anyway that's OT but I just realize again how out of touch the FSF is in the global sense.
