The Open Source Initiative (OSI) has published its official definition of truly open source AI based on the freedom to use, modify and share, open source AI must give access :
- details of the data used to train the AI
- the complete code used to build and run the AI
- training parameters and weights
Note that Meta's Llama does not meet all these criteria.
The #OSAID is deeply flawed exactly because of two #Meta employees (a Llama developer and a lawyer) that were invited to vote about the requirements and excluded training data.
Without training data, you cannot really modify any AI, just fine-tune them. Which is like saying that Windows is #OpenSource because you can tweak the registry.
And without training data you have no hope of being able to effectively study its bias or to identify backdoors planted in the system: https://arxiv.org/abs/2204.06974
#OSI is serving its sponsors, #Google, #Meta, #Microsoft... opening a loophole in the #AIAct.
