Open Source Initiative presents AI definition
The Open Source Initiative has defined what information AI providers must disclose in order to be truly open source.
The world in data.
(Image: Outflow_Designs / Shutterstock.com)
Meta's AI model family Llama is not considered open source according to the newly published definition of the Open Source Initiative (OSI). Meta has disclosed a great deal of information about how its AI is structured, such as the architecture and weights. However, the training data is missing. This is likely to lead to discussions between the OSI and Meta as well as other AI providers. The AI Act has also already moved away from a kind of free pass for open source models. The new definition (OSAID) could lead to further restrictions.
According to OSI, AI that is considered open source must provide access to the training data so that others can understand it. In addition, the entire code must be available. All settings and weights are part of the information that must be freely accessible. Ultimately, the question is whether a model can be replicated. For months, the OSI asked for submissions and opinions for the definition. The aim was to get tech companies, users and politicians on board.
Videos by heise
The definition also includes the ability to use and modify open source AI and to create something of one's own based on it. In future, the OSI wants to denounce AI models that do not meet the requirements but claim to be open source. However, the initiative cannot impose sanctions or penalties if someone misuses the title of open source in its eyes.
Llama is not open source, Gemma is just open
Meta does describe the Llama family as open source, and they have indeed provided significantly more information than other providers. Nevertheless, the training data is not known. Meta also requires that providers who make Llama available for platforms with more than 700 million monthly users obtain permission. The definition does not allow this either. There are numerous freely available AI models on Hugging Face; Aleph Alpha and Stability AI, for example, also describe their AI models as open source. They too will at least have to check this.
Google offers an open model with Gemma. Unlike Meta, they do not specifically say that it is open source, but only that it is open. OpenAI explains that it cannot make its own models openly available due to security concerns.
The AI Act, which came into force in August, provides for exceptions for AI models that are open source. The question now is whether the policy will adapt to the new definition of open source AI. However, there have already been exemptions before. AI models with too high a risk should also fall under certain obligations, even if they are open source. Anyone who charges a fee for the use of AI models cannot be an open source provider either.
(emw)