Everything’s possible until you have to actually make it work, then surprisingly the actual possibilities end up quite limited.
That said, despite micro-nuclear propulsion being a fantasy, the bomb was real.
I also have to wonder how far throwing a shitload of data at a statistical machine will get us, without a much stronger understanding of how these systems work and what the goals should be to achieve "AGI", whatever that means.
But,
> AI is expensive as hell to run, and more advanced systems might require more and more computation. We can only build so many GPUs
While certainly not using techniques like LLMs, we do know you can get some decent intelligence out of a 23 watt analog computer, the human brain.
So it is plausible that we could find algos + hardware which use 100W of power, and are quite a bit smarter than a human. Multiply that times a multi-megawatt datacenter, and the availability of some level of super intelligence might appear to be relatively “limitless.”