"The first AGI will be Federated" -- come hear my keynote at the @AcmSensys conference this Wednesday morning. I'll explain why when we finally achieve AGI, and flick the switch to bring it online, it will have been trained using some variety of federated AI. And not the training methods of today, DDP and the like, that are limited to being deployed in single data centers -- which are themselves fast reaching the limits of their own scalability as we struggle to co-locate and power north of 100k GPUs in one physical place. In support of my talk, here is a preview of a few results coming in hot from a new arxiv paper produced by @flwrlabs and @CaMLSys (@Cambridge_Uni) that will go live also Wednesday. We show our unique approach and system (aka Photon built on Flower) produces LLMs that are both better (by perplexity) and faster to train than DDP (when network speeds are equal, as measured up to 7B). Furthermore, we show this can scale effectively as we add more weakly connected GPUs -- while also being more efficient (eg better wall clock training times) than any existing known methods for training under decentralized/widely-distributed settings. Folks, it's time to train different. Larger and more complex LLM settings are underway. Link to arxiv Wednesday.