It occurs to me that the thing that will probably limit advancements in AGI will be the availability of data to feed these systems.
If you believe that Moore's law is only on life support and not totally dead, there will be more processing power to harness in the future. The number of researchers and investment are clearly growing very quickly. The models used can endlessly be improved.
But on the other hand there are so many things that even today just aren't captured as digital data. I work as a mechanical engineer and there are many nuances to mechanical design that appear nowhere in print (or youtube video, or blog post for that matter). Learning these things takes a complex combination of of sight touch and intuitive leap. Even unsupervised learning requires some input to feed the net. I just don't see where it will come from.
I think you're mistaken about this. Once the philosophical and technical breakthroughs are made that allow us to build an AGI then it will get all the data it needs from its environment. It would be 'unsupervised' in the sense that human children are i.e. no pre-processing of data required but it would still need parenting.
It almost doesn't matter how smart the AGI is on it's own, if it can't participate in our social conversation and get it's information from the same sources humans do it's going to be stone dumb in practice. An individual human is far less intelligent than we generally give ourselves credit for: if it can't stand on the shoulders of our giants then it doesn't matter if it's ten times as tall as us individually.
I think part of it is we're feeding in really small subsets of existing data. Someday we'll be able to feed in 4K 60FPS video, and it'll be able to learn patterns from that. Right now we can only access a subset of the existing data.
If you believe that Moore's law is only on life support and not totally dead, there will be more processing power to harness in the future. The number of researchers and investment are clearly growing very quickly. The models used can endlessly be improved.
But on the other hand there are so many things that even today just aren't captured as digital data. I work as a mechanical engineer and there are many nuances to mechanical design that appear nowhere in print (or youtube video, or blog post for that matter). Learning these things takes a complex combination of of sight touch and intuitive leap. Even unsupervised learning requires some input to feed the net. I just don't see where it will come from.
Anyone think I'm way out of line here?