"AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence", Clune 2019

It seems like to me that if AGI is eventually created via this paradigm, hardware overhang is almost guaranteed. That's because it requires meta-learning and the initialization of rich environments to train populations of agents, which is extremely computationally expensive compared to running the actual agents themselves. Therefore, if AGI is obtained this way, we could rather easily re-purpose the hardware used for training in order to run a large number of these agents, which could correspondingly perform vast amounts of economic labor relative to the human population.

Imagine you had an oracle which could assess the situation an agent is in and produce a description for an ML architecture that would correctly "solve" that situation.

I think for some strong versions of this oracle, we could create the ML component from the architecture description with modern methods. I think this combination could effectively act as AGI over a wide range of situations, again with just modern methods. It would likely be insufficient for linguistic tasks.

I think that's what this article is getting at. The author is an individual from Uber. Does anyone know if this line of thinking has other articles written about it?