Planetary VC's Ramez Naam argues true open-source AI requires training datasets and code, not just model weights

When people talk about "open source" AI models, what they usually mean is "open weight'. The model weights are open.

True "open source" models would be those where all the source code that trains, deploys, and serves the model is released, and arguably the training data (including RL data) and any software used to construct it.

Open weight models are great. I'm glad they exist. True open source goes beyond that, increases transparency, and allows others to build upon or improve on the model (and its training process, its performance when served, etc..) in much clearer ways than simply open weight.