Open Weights Are Good. Open Source Is Better

The piece addresses the debate over releasing model weights versus fully open-sourcing AI systems, explaining that an AI model comprises training data, model weights/parameters, and the code used to prepare data and train. Releasing weights gives users meaningful new options: running models locally, avoiding provider lock-in, and fine-tuning with private data. Those benefits make open-weight models clearly preferable to fully closed systems. Yet weights alone do not provide the four core software freedoms - use, study, modify, and share - because withholding training data and training code prevents full inspection, modification, and verification of behavior, a limitation that matters for trust, safety, and cybersecurity as AI spreads into critical systems.

Open Source AI, by contrast, releases weights, training code, and either the training data or a detailed account of how it was constructed, enabling unrestricted use, study, modification, and redistribution. That completeness unlocks collaborative innovation, reproducible research, and deeper forensic and safety analysis; for example, a fully open large language model allowed researchers to inject data and observe memorization dynamics. The argument concludes that while open weights are an improvement, real progress and societal value - already tied to an $8.8 trillion open-source demand-side ecosystem - depend on embracing full Open Source AI and educating policymakers, producers, and users to support it.