The OSI Has Had Enough Of Mark Zuckerberg’s BS

🌈 Abstract

The article discusses the controversy surrounding Meta's release of the Llama 3.1 language model, which was claimed to be "open source" by Mark Zuckerberg, but did not actually meet the existing Open Source Initiative (OSI) guidelines for open source software. The article also examines the OSI's newly proposed definition for "open source AI", which the author believes is too broad and should be more specifically targeted towards "Data Driven Generative Systems" (DDGS) like large language models and diffusion models, rather than AI in general.

🙋 Q&A

[01] Llama 3.1 and Open Source Claims

1. What were the key issues with Meta's claims that Llama 3.1 was "open source"?

• Meta's CEO Mark Zuckerberg claimed that Llama 3.1 was "open source", but it did not actually meet the existing OSI guidelines for open source software.
• The author argues that Zuckerberg's blog post defending the "open source" claim was "misguided".

2. What are the main components of an LLM that the OSI's proposed definition covers?

• The OSI's definition covers three main components of an LLM:
  • Data information: Detailed information about the training data, made available under open licenses.
  • Code: The source code for training and running the system, made available under OSI-approved licenses.
  • Weights: The model weights and parameters, made available under OSI-approved terms.

3. What are the author's thoughts on the OSI's requirements for disclosing training data?

• The author argues that disclosing the full training data is problematic, as it could open companies up to legal liability from those whose work was used in the training data.
• The author suggests that instead, the focus should be on allowing users to bring their own training data to use with the model.

[02] The OSI's Definition of "Open Source AI"

1. What are the author's concerns with the OSI's definition of "open source AI"?

• The author believes the definition is too broad, as it covers "AI" in general rather than being more specifically targeted towards "Data Driven Generative Systems" (DDGS) like LLMs and diffusion models.
• The author suggests the definition should be renamed to something like "open model" instead of "open source AI".

2. What are the author's thoughts on the OSI's requirement for open source training code?

• The author sees this as a positive, as it will allow companies to see how other companies make their LLMs perform so well.
• However, the author still believes that LLMs are too costly and complex for "normal people" to meaningfully contribute to, even with the training code being open.

3. How does the author's overall view of the OSI's definition change from the beginning to the end of the article?

• Initially, the author thought the OSI's definition was "really bad" for watering down the meaning of "open source".
• By the end, the author says they don't have much of a problem with the definition, aside from the wording and the use of "open source" instead of something like "open model".