Reproducibility is the New Copyleft: Defining AGI-oriented Reproducible Builds
A new paper proposes a novel approach to software licensing for advanced AI systems, including AGI. It argues that traditional copyleft licenses, like the GNU GPL, are insufficient due to the complex and non-reproducible nature of AI model artifacts such as code, data, and weights. The paper suggests that a functional equivalent of copyleft for AGI should be based on reproducible builds, ensuring bit-exact reconstruction from declared inputs. It outlines seven requirements for such builds and proposes that a 'protocols, not platforms' governance model is more suitable than traditional licensing for AI-to-AI coupling mechanisms. AI
IMPACT Proposes a new framework for AI governance and licensing, potentially impacting open-source AI development and AGI control.