Blog Post 15: A Letter to Stallman
Date: November 3rd, 2026 @ 1100 UTC-5:00
Author: lule34567
This page serves to document and archive the following conversation via email I had with Richard Stallman on his opinions of LLM usage in programming in free and open source projects. Any updates that are received back within this conversation will be placed here, alongsdie a subtitle referencing it is an update to this original article, alongside a date received.
About a week ago, I emailed Richard Stallman, "Chief GNUisance of the GNU Project and on the Board of Directors of the Free Software Foundation", on his opinions on two pieces of news that came relatively recently.
First, was the recent vibecoded clones of the Adobe suite of products, named Artcraft (link). Although it was open sourced under the MIT and Apache licenses, and was supposedly cloned under the clean room technique, the ethics of using LLMs heavily to create such a project is majorly debated.
Second, is the ongoing debate within the Linux kernel maintainer community, where Linus Torvalds has been increasingly becoming in favour of LLM-created contributions to the kernel. To quote Torvalds from an article off of FOSS Linux, "I realize that some people really dislike AI, but this is an area where I’m willing to absolutely put my foot down as the top-level maintainer. Linux is not one of those anti-AI projects, and if somebody has issues with that, they can do the open-source thing and fork it" and "We’re not forcing anybody to use it, but I will very loudly ignore people who try to argue against other people from using it."
At first, I discussed this with other fellow Linux kernel users on what Stallman's opinions would be on these two situations. One response I received states: "he's [as in Stallman] either going to be totally for it because LLMs do kinda open up a lot of software more than ever before, or he's going to be completely against it because he can't tell what the LLMs are doing in the background, and most are designed in such a way that even the open source models are not easily editable/viewed by humans," I then proceeded to email Richard Stallman about a day after, with the following email.
[The following email source was originally sent with OpenPGP encryption and headers, but has been stripped and simplified for reader convenience]
Date: Wed, 28 Oct 2026 13:33:32 -0400
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Content-Language: en-US
To: rms@gnu.org
From: Malingen <[REDACTED]>
Good afternoon Mr. Stallman,
I wanted to ask your on your opinions on two pieces of news that occurred recently involving LLMs:
First, a recent "vibecoded" project that reverse engineered the Adobe suite of products (which were proprietary) was released, in which remade the Adobe suite entirely with LLM-generated code, and was released under the Apache and MIT licenses. This marked a precedent in which proprietary projects can effectively be reverse engineered and output open sourced copies en masse, at the cost of being entirely LLM-generated or "vibe-coded". You can find the GitHub repository here: https://github.com/storytold/artcraft
Second, a debate within the Linux kernel contributor community has been gaining attention, where Linus Torvalds, the lead maintainer and a person in which you worked with before, has been using LLMs for the vast majority of contributions to the kernel. His opinion on LLM usage is mixed, where he doesn't mind LLM-assisted contributions and will treat it like every other PR, however, to quote from an article on the debate (https://www.fosslinux.com/159063/linus-torvalds-ai-linux-kernel-debate.htm): "I realize that some people really dislike AI, but this is an area where I’m willing to absolutely put my foot down as the top-level maintainer. Linux is not one of those anti-AI projects, and if somebody has issues with that, they can do the open-source thing and fork it" and "We’re not forcing anybody to use it, but I will very loudly ignore people who try to argue against other people from using it."
After discussing with a friend of mine on the subject and what your opinion would be, they said "he's [as in you, RMS] either going to be totally for it because LLMs do kinda open up a lot of software more than ever before, or he's going to be completely against it because he can't tell what the LLMs are doing in the background, and most are designed in such a way that even the open source models are not easily editable/viewed by humans"
In response, I decided to email you on your opinions on these two events, and how you feel about LLMs contributing to the free (as in libre) and open source repositories of software.
Sincerely,
Malingen
Aproximately three days later, I received a response from Stallman.
[The following email source was originally sent with OpenPGP encryption and headers, but has been stripped and simplified for reader convenience]
Content-Type: text/plain; charset=Utf-8
From: Richard Stallman
To: Malingen <[REDACTED]>
cc: rms@gnu.org
In-Reply-To: (message from Malingen on Wed, 28 Oct 2026 13:33:32 -0400)
Subject: Re: Regarding Recent "Vibe-coded" Ports of Proprietary Software, etc.
Reply-To: rms@gnu.org
Date: Sun, 01 Nov 2026 11:44:31 -0500
[[[ To any NSA and FBI agents reading my email: please consider ]]]
[[[ whether defending the US Constitution against all enemies, ]]]
[[[ foreign or domestic, requires you to follow Snowden's example. ]]]
> First, a recent "vibecoded" project that reverse engineered the Adobe
> suite of products (which were proprietary) was released, in which remade
> the Adobe suite entirely with LLM-generated code, and was released under
> the Apache and MIT licenses.
The most crucial question about this is a question your description does
not answer: is the output a derivative work of the Adobe programs?
If a human reverse engineers a program and then writes something similar,
that is a derivative work, meaning it can't be free. And if it can't
be free, it changes nothing.
This is why in real reverse engineering projects we need to use,
carefully, the "clean room" technique. That means, the reverse engineer
writes down the crucial specs, then shows them confidentially to
a programmer who gets no other non-public information, and then writes
the new program. If they did this and did it right, the new program
should be an independent work, so it can be made free.
If it is free, it is a great step forward for freedom.
> You can find the GitHub repository here:
> https://github.com/storytold/artcraft
Github is a bad place to host any software!
See https://www.gnu.org/philosophy/whats-wrong-with-github.html.
Please spread the word.
As for the use of an LLM, it is an injustice _to its users_ if it is
not free. And almost none of then are free. "Open weights"
is basically "open binaries" -- a joke.
I do not advocate "open source", nor categorize software that way.
(See https://gnu.org/philosophy/open-source-misses-the-point.html.)
I advocate free/libre software.
But nearly all open source programs are free/libre, so nearly all of
them are ethical. By contrast nearly all "open source" LLMs are not close to
being free/libre.
> how you feel about LLMs contributing to the free (as in
> libre) and open source repositories of software.
LLMs are not persons, they are tools. It makes no sense to speak of
their "contributing" to anything. The question is whether we ought to use
them in various ways.
--
Dr Richard Stallman (https://stallman.org)
Chief GNUisance of the GNU Project (https://gnu.org)
Founder, Free Software Foundation (https://fsf.org)
Internet Hall-of-Famer (https://internethalloffame.org)
I was initially misinterpreted his response as simply "well, if the LLM-generated code is contributing to a F(L)OSS [free (as in libre) and open source software] project, then I approve of it." However, I was corrected by another Linux kernel user I spoke to, where Stallman dissaproves of the project (among other vibecoded repositories) due to the equivalent, closed source nature of LLMs as we know today.
Essentially, LLMs can be split into two parts: The LLM itself, or the "weight", and the datatset used to train that LLM. We have already seen LLMs be open sourced as "open weights" by numerous parties, but you need to imagine that these LLMs are pre-trained; similarly to pre-compiled software, except much, much, much more difficult to "decompile" and, practically, "pick its brains". On the other hand, the datasets used to train these LLMs are often huge in size, due to how these LLMs also train off of the internet itself. If you want to make a dataset for a LLM free and open source, you would need to seek consent from essentially every party that created and uploaded something from the internet that an LLM has seen and interpreted: whether that'd be websites, art, music, published documents, video, etc.
I thereby sent the following response.
[The following email source was originally sent with OpenPGP encryption and headers, but has been stripped and simplified for reader convenience]
Date: Sun, 1 Nov 2026 15:49:19 -0500
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
To: rms@gnu.org
Content-Language: en-US
From: Malingen <[REDACTED]>
My colleague also mentioned that all LLMs should be released freely, but also mentioned the datasets used to train these LLMs should also be released freely if it weren't for how large they were. Given the majority of LLMs nowadays use the internet for training and scrape off of it, you could effectively say you would need to freely (as in libre) open source the entire internet that the LLM trained off of. However, as my colleague states, "if it were a thing, it would probably drastically the decrease AI scrapping problem".
Releasing "open weights" is effectively the same as releasing a pre-compiled executable, without the source code; the difference here being that is nearly impossible to decompile an LLM given how complex its behavior is with its millions and billions of equivalent neurons. Therefore, to quote my colleague, "I think it [as in having open weights be released] is the closest way to kinda right the wrongs of the massive amount of widespread copyright violations that are LLMs"
I strictly do not use LLMs in any capacity due to my opinions on how they are a negative both in real life (communities being affected by data centers) and online (LLM-produced content being extremely prevalent online to the point of creating mis/disinformation and low quality content over high quality content). However, I believe from a F(L)OSS perspective that LLM companies/providers should do all in their power to freely and openly release both the LLM itself (the "open weight") and the datasets used to train them, even if those datasets are incredibly, incredibly large and impractical to store in its entirety in consumer level hardware. In fact, I believe these datasets could be partitioned for more practical download sizes, whether that be based on topics (language, math, programming, etc.), license types (proprietary vs. MIT vs. GPL vs. etc.), or other organizational methods. I'm aware of that datasets exist that are free to use and contain zero copyrighted material, but never at the size in which popular LLMs would actually train on solely.
As of writing (03/11/2026), I am currently awaiting a response back, and will update accordingly.