It's already no longer exponential and plateauing like moore's law with CPUs and other hardware, that's literally what your graph shows of them all bunchin up together at the top of the curve there, 4.6, 4.7 fable/mythos chatgpt 5.5 the new China one are all still on the same tier the difference between their capabilities and the previous ones are a fraction compared to the previous ones and another tier down
They've already shifted strategy from one giant parameter model to a bunch of smaller MOE mixture of expert models in a trench coat because they literally did hit a ceiling with the previous model training method
|