Data is heavy, models are getting light. Moving the model to the data is cheaper than moving the data to the model, and cheaper wins. A model that runs where the data is must be one you can hold, and only open weights can be held
Why nowOpen models over a hundred times smaller now clear the same bars, and the best of them have almost caught the best closed ones
Doubling one model buys less each time. Models that make different mistakes correct each other, so adding one keeps paying
Why nowThe public text to train on is running out. The latest gains came from asking more models more often, not from a bigger one
A person can only read so fast, and that never grows. Everything that scales has to happen between machines
Why nowThe tokens served each month grow by orders of magnitude, and a team of agents spends many times the tokens of a chat to give one person one answer
Words are a compression built for a reader. Between two models there is no reader, so the compression is only loss
Why nowD needs A, B and C, and all three hold now: open models clear the old bars, gains come from asking more models, and tokens served grow by orders of magnitude. Passing the state already works