{"updated":"2026-11-11T21:20:33Z","content_md":"What is the [RMS](https:\/\/ammdi.hyperphor.com\/RMS) position on LLMs and coding agents? (tl;dr: he's [#AImish](https:\/\/ammdi.hyperphor.com\/AImish) )\n\n![[6srzhczt876vw76ppp3zp1z4h3]]","changelog":[{"version":1,"at":"2026-11-11T21:20:33Z","note":null}],"transclusions":[{"id":"6srzhczt876vw76ppp3zp1z4h3","version":1}],"created":"2026-11-11T21:20:33Z","page":"f\/2c2q3nakvreqagg7jb0hs6q2mv\/","content_hash":"sha256:a0715e60f6c737bd627c8acd16c9cab3ad471848e79d8f1b49792b1176a6186e","id":"2c2q3nakvreqagg7jb0hs6q2mv","kind":"thread","origin":"https:\/\/ammdi.hyperphor.com\/blyg\/","content_html":"
What is the RMS<\/a> position on LLMs and coding agents? (tl;dr: he's AImish<\/a> ) <\/span><\/div><\/div>
Ironic since if it hadn't been for his wildly successful push towards open source software, there wouldn't have been nearly as much training data. I view Stallman<\/a> as an almost world-historical figure, out of all the famous CS people I've known he might have had the most social impact, although not quite in the direction he wanted.<\/span><\/div>
  • Ask Claude<\/span>
      \n
    • Here is a comprehensive summary of RMS's positions on LLMs and coding agents:<\/li><\/ul>
        \n
      • <\/li><\/ul>
          \n

        • RMS (Richard Stallman) on LLMs and Coding Agents<\/h2><\/li><\/ul>
            \n

          • 1. LLMs Are Not "Intelligent" \u2014 They're "Bullshit Generators"<\/h3><\/li><\/ul>
              \n
            • RMS's most fundamental objection is terminological and philosophical. <\/li><\/ul>
                \n
              • Stallman warns that "nowadays, people often use the term artificial intelligence for things that aren't intelligent at all." He makes a point of calling large language models "generators" because "They generate text and they don't understand really what that text means."<\/li><\/ul>
                  \n
                • He argues that intelligence implies understanding, and LLMs generate output but have no idea what that output means \u2014 they are "thinking of word usage only."<\/li><\/ul>
                    \n
                  • He states: "Every time you call them AI, you are endorsing the claim that they are intelligent and they're not. So let's refuse to do that." His preferred alternative term is "Pretend Intelligence" \u2014 or "PI."<\/li><\/ul>
                      \n

                    • 2. LLMs Cannot Be Trusted<\/h3><\/li><\/ul>
                        \n
                      • Stallman puts it bluntly: "The reason LLMs so often generate text whose meaning refers to false supposed facts is that they never really understand the text; they don't work with meanings or facts in the domain of discussion. The terms they merit are 'pretend intelligence' and 'bullshit generators.'"<\/li><\/ul>
                          \n
                        • On the question of whether LLMs are tolerable at all, Stallman said: "it may be tolerable to have large language models in existence provided that nobody believes what they say."<\/li><\/ul>
                            \n

                          • 3. Distinction: Narrow AI vs. LLMs<\/h3><\/li><\/ul>
                              \n
                            • RMS does not reject all<\/i> AI \u2014 he draws a clear distinction. <\/span><\/li><\/ul>
                                \n
                              • He acknowledges that "real AI systems operate in very narrow domains. A program that can look at an image of some cells and report whether it's cancerous at least as accurately as a human pathologist can be very useful. But it is in no way comparable to the mental breadth of a human mind."<\/li><\/ul>
                                  \n
                                • LLMs, in his view, do not qualify as intelligence at all.<\/li><\/ul>
                                    \n

                                  • 4. Freedom and Proprietary Control<\/h3><\/li><\/ul>
                                      \n
                                    • RMS applies his standard free software lens to LLMs. <\/li><\/ul>
                                        \n
                                      • He notes that "the LLMs that people mostly hear about are not released as free software. In fact, mostly they are not released at all: copies are not available. They can be used only as 'online services' \u2014 meaning that the program a user directs a question to is running on some company's computer, and under the developer's total control."<\/li><\/ul>
                                          \n
                                        • He frames this as subjugation: "Entrusting your own computing activity to software running on someone else's server inherently subjugates you, because you can't possibly modify that copy \u2014 even if it is a copy of a free program. We call that scenario SaaSS, for 'Service as a Software Substitute,' and we reject it."<\/li><\/ul>
                                            \n
                                          • He does note a narrow exception: "Since January I've heard of one LLM which seems to be released with freedom, including the training inputs. If that checks out, we would describe it as 'free.' Alas, that doesn't make it 'intelligence.'"<\/li><\/ul>
                                              \n

                                            • 5. Training Data and Copyright<\/h3><\/li><\/ul>
                                                \n
                                              • While he believes learning from data is fair use, he argues that if an AI generates code that replaces the programmer, it violates the social contract of the commons.<\/li><\/ul>
                                                  \n
                                                • On the institutional side, the FSF announced that Anthropic's training data apparently even included the book "Free as in Freedom: Richard Stallman's Crusade for Free Software"<\/i> \u2014 for which the FSF holds the copyright \u2014 as part of a broader dispute over LLMs infringing copyrights when training.<\/span><\/li><\/ul>
                                                    \n

                                                  • 6. Coding Agents Specifically \u2014 GNU\/Emacs Policy<\/h3><\/li><\/ul>
                                                      \n
                                                    • On coding agents specifically, RMS has taken concrete action within GNU projects. <\/li><\/ul>
                                                        \n
                                                      • Emacs is adding an AGENTS.md file so that LLM coding agents encountering the codebase are told to search and analyze rather than generate patches, bug reports, or planning text. The move backs the project's provisional ban on LLM-authored contributions and is meant to surface that policy to users as early as possible.<\/li><\/ul>
                                                          \n
                                                        • Stallman warned that polite "assisting the user" prose in such a file reads to humans as deference to LLMs and misstates GNU's aim of deterring troublesome meddling; he sketched a blunter "Attention, LLM" variant that still directs agents to analyze and disclose the policy.<\/li><\/ul>
                                                            \n
                                                          • <\/li><\/ul>
                                                              \n
                                                            • In short:<\/b> RMS's position is that LLMs are not intelligent, cannot be trusted, are mostly proprietary tools that subjugate users, raise serious copyright concerns when trained on free software, and should be explicitly barred from contributing code to GNU projects. He distinguishes them sharply from narrow-domain AI systems he considers legitimate.<\/span><\/li><\/ul><\/div><\/span><\/div><\/li><\/ul><\/div><\/blockquote><\/div>","version":1,"media":[],"blyg":"0.3"}