[mikewarot]

Let's say that my silly BitGrid idea works, and the price of a Petaflop goes down by a factor of 1000. Then the other limiting factors that actually prevent AGI from taking place become exposed. Training data only gets you so far. A trained neural network is an optimization of the human knowledge its been fed, a premature optimization, one like the crisis in computer security (caused by settling on the wrong security model), will likely take decades to reverse.

---

You're worried about the wrong end of economics, though... it's the physical economy that's about to bite, hard. (I got that phrase from Lyndon LaRouche, he had a lot of interesting ideas, but that one stuck)

The global consumption of fossil fuels is about to hit its downslope, and a great simplification of society is going to be the result. We're extracting resources from the plant at millions of times the rate they are created, and because of the American model of corporate governance, only focusing on the next Quarter, planning to deal with this reality isn't happening.

---

It is possible to train an AGI today, with corporate level resources and a generational commitment. you need to put a really good anthropomorphic robot body through all the stages of childhood, being especially careful to emulate all the weaknesses, and the drive for survival, and the need for others to support you, as babies are completely helpless.