The article suggests that while there is a need for a public AI option and genuine open source, there are also partially open models that require definition. It acknowledges the importance of privacy-preserving, federated methods of machine learning model training. The OSI defends its stance by stating that it aims to facilitate open source AI in fields where data cannot be legally shared, such as medical AI. The article proposes the term "open weights" instead of "open source" to better represent this concept.
Key takeaways:
- The Open Source Initiative's definition of open source AI is criticized for allowing secret training data and mechanisms, and secret development.
- There is confusion as many AI models are open source in name only, and the OSI is accused of being influenced by industry players wanting corporate secrecy and the open source label.
- The author argues for a public AI option and real open source as a necessary component of it, while acknowledging the need for some sort of definition for partially open models.
- The OSI defends the exclusion of some training data in open source AI, citing legal restrictions on data sharing in fields like medical AI and the need to protect sensitive personal information and Indigenous knowledge.