That’s a great question. I think a lot of the frontier model labs, of course, they take from open source, but they don’t contribute back, mostly because they were afraid that those frontier models can be used in a primarily open source manner. If they now, well, maybe Taiwan is ready with our digital signatures and everything, but I don’t think many other democracies are ready. I think that is the main thing that are keeping them gating their release. On the other hand, there are many other parts of deep learning that doesn’t have this offensive nature. I have in my mind, like the reward model, the constitutional alignment, these kind of things. these things, it doesn’t increase the base capability of the foundation model. And so open sourcing, it only gives more communities more ways to align the frontier model to their needs. But it doesn’t fundamentally increase capacity, for example. And so I think when looking at research, we need a way to just distill it to the essence of whether this is defense oriented or whether it’s offense oriented. And then we need to double down on investment on the defense oriented things and make it so that it’s open source.

Keyboard shortcuts

j previous speech k next speech