OpenAI with the launch of o1 has inadvertently proved out the crypto x AI intuition I wrote in this essay exactly one year ago: A model like o1 now has a reasoning token budget of 25,000 for a single prompt - representing atleast a 20x increase in average per prompt token budgets. This is a parallelized computation task, where distributed and decentralized networks can scale infinitely. And - the intelligence as perceived by a user actually improves! Consider if a network like Hyperspace grows from the 20k+ nodes today to 20 million nodes one day, it would be amongst the top 5 AI clusters in the world where you would get the most intelligent answers from. And it would also be cheaper at the same time as it’s a function of local electricity costs plus real-time market economics. There are no middle-men. Even micropayments between agents would be managed through a protocol run by a fraction of these nodes. Big pre-trained models on small clusters are dinosaurs in that world. Too slow, not that intelligent, even as they please some people in the academic community. What eventually wins are the biggest networks. Since intelligence is now a function of parallelized scaling of inference, nearly all of AI is going to be distributed. This is a once-a-20-year-market. Our priorities are clear: we want to build the most impactful distributed AI venture and we will not compromise on that vision. That work thus spans from new models, AI products, distributed systems, networks, clients, and protocols. We scale using software. cc @ylecun