Open Source AI Definition
The OSI's 2024 attempt to say what open source means for an AI model, a compromise that left nobody entirely happy.
Version 1.0 was published on 28 October 2024 after two years of consultation. It requires a model to ship with its weights, its training and inference code, and "data information": a description of the training data detailed enough for a skilled person to build a substantially equivalent system. The data itself doesn't have to be released. As with ordinary software, use can't be restricted, which rules out OpenRAIL and the Llama licence.
Criticism came from both sides. Purists argued that a model without its training data is a binary without source. Companies argued that nobody can release data they often have no right to redistribute. The OSI's answer amounted to agreeing with both, and noting that a definition no model could meet would be pointless.