Agents Are Pushing Code and Hardware Toward Open, and Two Essays Disagree on Why
Five paragraphs, 968 points, 539 comments. The most discussed post on Hacker News this Sunday was a short essay by developer Jakub T. Jankiewicz titled "All AI-generated code should be Open Source". It went up on the author's blog on Nov 14 at 19:21 GMT and hit the top of HN on Nov 15. Essay: https://jakub.jankiewicz.org/blog/ai-code-open-source/
The argument has two legs. The moral one: models learned from decades of open-source code and from Stack Overflow answers under Creative Commons licenses that require attribution, so turning that into proprietary output "feels like license laundering". The legal one: purely AI-generated code has no human author, so in the US it arguably should not be copyrightable, citing the monkey selfie dispute and the US Copyright Office, plus Wikimedia Commons tagging purely AI images as public domain. The UK does grant limited protection to computer-generated works, so the essay's fix is to release AI code under an OSS license or CC0 everywhere, at least the code where humans did little beyond writing prompts.
The counterarguments are strong, and the comment thread made most of them. Real codebases are mixed: a human picks the architecture, rejects half the diffs, rewrites the rest, and where the line of "human authorship" falls is unsettled in every jurisdiction. "Not copyrightable" would mean public domain, not open source, and those are different things. Whether training on licensed code infringes anything is still being fought in court. And a company that cannot copyright an agent's output can still keep it as a trade secret by never shipping the source. So this is an opinion with a real question under it, not a legal conclusion. The question is the useful part: if your agent wrote 80% of the repo, you probably cannot say with confidence who owns it.
A quieter essay posted the same day pushes openness from the other end. In "Your Agent Will Pick the Hackable One", Geoff Stearns starts with the M-Vave FM-1, a $70 pocket FM synth. Owners began writing their own firmware in September, almost all with AI coding tools per their forum posts. In October M-Vave open-sourced its firmware under GPLv3 with a dev kit, a browser emulator and a tool that restores stock firmware in about ten seconds. Behringer followed with firmware source for its upcoming KICK and IGNITE, its parent said it would open selected designs so owners can use AI to write their own, and Akai and Erica Synths announced similar moves. Essay: https://geoffstearns.com/blog/your-agent-will-pick-the-hackable-one/
The thesis is about buying, not tinkering. Ask an agent which router, TV or synth to buy and it will lean toward the device it can change later, so openness starts to matter even to buyers who never think about it. Closed makers push back with security (Chamberlain cutting off Home Assistant in 2023, Bambu Lab server authorization, Google developer verification for Android installs, live in four countries since Sep 30 and worldwide in 2027), and the essay's test is simple: who holds the signing keys. Put the two essays together and you get the same pressure from two sides. Cheap code makes closed output harder to justify, and cheap firmware makes closed devices harder to sell. The catch both essays leave open is who the agent works for. An agent shipped by a company with its own devices or its own code to sell has no reason to pick the open one.
← Back to all articles
The argument has two legs. The moral one: models learned from decades of open-source code and from Stack Overflow answers under Creative Commons licenses that require attribution, so turning that into proprietary output "feels like license laundering". The legal one: purely AI-generated code has no human author, so in the US it arguably should not be copyrightable, citing the monkey selfie dispute and the US Copyright Office, plus Wikimedia Commons tagging purely AI images as public domain. The UK does grant limited protection to computer-generated works, so the essay's fix is to release AI code under an OSS license or CC0 everywhere, at least the code where humans did little beyond writing prompts.
The counterarguments are strong, and the comment thread made most of them. Real codebases are mixed: a human picks the architecture, rejects half the diffs, rewrites the rest, and where the line of "human authorship" falls is unsettled in every jurisdiction. "Not copyrightable" would mean public domain, not open source, and those are different things. Whether training on licensed code infringes anything is still being fought in court. And a company that cannot copyright an agent's output can still keep it as a trade secret by never shipping the source. So this is an opinion with a real question under it, not a legal conclusion. The question is the useful part: if your agent wrote 80% of the repo, you probably cannot say with confidence who owns it.
A quieter essay posted the same day pushes openness from the other end. In "Your Agent Will Pick the Hackable One", Geoff Stearns starts with the M-Vave FM-1, a $70 pocket FM synth. Owners began writing their own firmware in September, almost all with AI coding tools per their forum posts. In October M-Vave open-sourced its firmware under GPLv3 with a dev kit, a browser emulator and a tool that restores stock firmware in about ten seconds. Behringer followed with firmware source for its upcoming KICK and IGNITE, its parent said it would open selected designs so owners can use AI to write their own, and Akai and Erica Synths announced similar moves. Essay: https://geoffstearns.com/blog/your-agent-will-pick-the-hackable-one/
The thesis is about buying, not tinkering. Ask an agent which router, TV or synth to buy and it will lean toward the device it can change later, so openness starts to matter even to buyers who never think about it. Closed makers push back with security (Chamberlain cutting off Home Assistant in 2023, Bambu Lab server authorization, Google developer verification for Android installs, live in four countries since Sep 30 and worldwide in 2027), and the essay's test is simple: who holds the signing keys. Put the two essays together and you get the same pressure from two sides. Cheap code makes closed output harder to justify, and cheap firmware makes closed devices harder to sell. The catch both essays leave open is who the agent works for. An agent shipped by a company with its own devices or its own code to sell has no reason to pick the open one.
Comments