But your point hits on one of the first cracks to show in this story: We already have companies consuming much of the web and training models on all of our books, but the reports they produce are of mixed quality.
The article tries to get around this by imagining models and training runs a couple orders of magnitude larger will simply appear in the near future and the output of those models will yield breakthroughs that accelerate the next rounds even faster.
Yet here we are struggling to build as much infrastructure as possible to squeeze incremental improvements out of the next generation of models.
This entire story relies on AI advancement accelerating faster in a self-reinforcing way in the coming couple of years.
This is only true as long as you are not able to weigh the quality of a source. Just like getting spam in your inbox may waste your time, but it doesn't make you dumber.
The story is actually quite poorly written, with weird stuff about “oh yeah btw we fixed hallucinations” showing up off-handedly halfway through. And another example of that is the bit where they throw in that one generation is producing scads of synthetic training data for the next gen system.
Okay, but once you know everything there is to know based on written material, how do you learn new things about the world? How do you learn how to build insect drones, mass-casualty biological weapons, etc? Is the super AI supposed to have completely understood physics to the extent that it can infer all reality without having to do experimentation? Where does even the electricity to do this come from? Much less the physical materials.
The idea that even a supergenius intelligence could drive that much physical change in the world within three years is just silly.
If this happens, then we indeed enter a non-linear regime.