Cessation of public development of Kefir C compiler
An independent developer of the Kefir C compiler has halted public releases, arguing that large language model companies are the primary beneficiaries of unpaid open‑source work and that current copyright norms no longer match authors’ intentions. Commenters debate whether training LLMs on GPL and other FOSS code violates either the legal or social “contract” of open source, with some deciding to stop publishing code or put projects behind authentication, and others insisting that broad reuse — including by AI — is inherent to free software. The exchange reflects a wider shift from a high‑trust to a low‑trust digital environment, where creators in software and other fields increasingly withhold work to avoid uncredited or unwanted AI training.
Project Status & Community
- Kefir C compiler’s public development is ending; prior source remains available.
- It appears to have been a one-person project; without public updates it’s viewed as effectively “dead” as open source and now a private “toy,” though this is seen as the author’s right.
- Technically, it’s praised as a small but correct C compiler, passing GCC torture tests and having well-crafted source.
Motivations for Ending Public Development
- Central motivation: discomfort that unpaid work is primarily benefiting companies training LLMs, contrary to the author’s intent in using GPL.
- Some commenters express similar decisions: stopping publication of code, art, or writing due to scraping and AI training.
- Others see this as irrational or inconsistent with the long-standing reality that free software can be used commercially in ways the author may not like.
Impact of LLMs on Open Source & Licensing
- One view:
- LLM training on GPL/FOSS code breaks the “social contract” of attribution, copyleft, and reciprocity.
- Models and their outputs are argued to be derivative works that evade GPL obligations, effectively treating code as public domain.
- This reduces incentives to publish and may push knowledge into closed “guilds.”
- Opposing view:
- GPL governs redistribution, not use; training is just another form of use.
- Models store patterns, not copies; outputs are not automatically derivatives unless they reproduce substantial code verbatim.
- FOSS always carried the risk of others benefiting without giving back; LLMs don’t fundamentally change that.
- Legal status is described as unsettled; some insist infringement must be proven with concrete examples of regurgitated code.
Trust, Incentives, and the Creative Economy
- Several see this as part of a shift from a higher-trust to lower-trust digital world, accelerated by disrespectful scraping and robots.txt violations.
- Others argue the internet has effectively been low-trust for decades (spam, moderation, authentication).
- Concern that widespread AI use will demotivate creators, shrink the pool of publicly shared work, and possibly entrench existing models.
Technical & Philosophical Views on AI Coding
- Some developers report avoiding AI for personal projects, finding more joy and “soul” in handwritten code.
- Others say LLMs make this “the best time” to write software by automating mundane tasks and letting them focus on higher-level design.
- Debate over creativity:
- Critics call LLM output inherently derivative and worry about stagnation or self-regurgitation.
- Supporters note most developers already reused patterns and Stack Overflow; true novelty was always rare.
Mitigations and Policy Ideas
- Practical defenses: putting sites behind authentication, requiring email for access, throttling bots.
- Policy suggestion: a “sender/initiator pays” regime for unsolicited automated requests, modeled on anti-spam-fax law; critics doubt enforceability and suggest strong penalties would be required.
- Some frame LLMs as actually fulfilling the open-source ideal—if, and only if, models remain broadly accessible rather than enclosed.