We need everybody to run their own self hosted AI models.
Discussion
we have lost the battle of being able to fully train our own models, but not yet the post training, hacking and runnning of them. tbf, the battle was over before it even started.
Yes, also further training pre trained models is not hard either, so if somebody has done training work once (open weight models) , everybody can just build on top of it.
I wouldn’t say the training battle is lost. They are making some strides with new architectures and new ways to train, especially in China where hardware is limited by sanctions. Ultimately, training will have to be able to happen locally so that robots and local reasoning computers can learn and operate without having to reach out to a remote server.