“Logic is an organized way of going wrong with confidence” Robert Heinlein
This is my opinion of what might be, not What THE FUTURE!!! Will Be!
My POV is hardware driven, I do electronic design. I don’t present myself as “an authority” on Artificial Intelligence, much less “an authority” on sentient artificial intelligence, until they are Real Things, there is no such thing as an authority in that field. That said, if the hardware doesn’t exist to support sentient AI, doesn’t matter how wonderful the software is.
The following is why I’ve been saying in a number of places that I expect hardware to be able to run a synthetic consciousness in ~20 yrs, @2045singularity on Twitter asked me to clarify what I meant.
1. I assume that if the physical operations of a human brain can be simulated in real-time, programs that simulate human consciousness in real time can be part of that simulation. I think that assumption is implicit in the thinking of just about everyone who believes sentient AI possible.
2. Simulation speed is driven by the available semiconductor technology which determines how fast CPUs, DRAM, internal (within chip and motherboard interconnects) can be run.
3. A recent research paper says they can simulate the operation of a human brain at ~1/1500 speed using the resources of a massively parallel data center class computing system. (this is an oversimplification – but close enough for brief discussion, recommend reading the actual paper). Another paper from Google discussing their emulation of a portion of the visual cortex said that they expect to be able to simulate a full human visual cortex in 5 years. As I see it, that paper roughly confirms IBM’s number.
4. Therefore, to get a data-center size CPU network capable of simulating a human brain with consciousness, we need computing systems running ~1500x as fast as current systems do. One can’t do this by simply throwing 1500x as many computers at the problem, electrical power problems aside (ask the NSA about that), the software is going to be hard enough to write without throwing in the kind of massive latency problems implicit in terms of implementing interconnections at the millions of processors level scattered over a significant geographic area.
5. Moore’s Law says transistor count per chip doubles every two years – this is a rough proxy for processor, etc. speeds. 10 Moore’s Law doublings, i.e. 2 to the 10th power = 1024 … I’m guessing advances in networking tech might make 1500x processing speed increase possible. Otherwise, 11 Moore’s Law doublings = 22 years.
6. 20-22 years can therefore be seen as a rough minimum to get to a hardware platform capable of running synthetic consciousness IF Moore’s Law continues to hold true – it might not.
7. Moore’s Law is driven by making profit off multibillon dollar CPU fabs with lots of equipment replaced on a 2 year cycle driven largely by growing Windows OS code, not due to physical limits so far. But physical limits may end Moore’s Law computing power growth in the next 20 years. Though spintronics and/or graphene transistors and FRAM and other new types of memory will probably drive a few more Moore’s Law iterations. IOW, we might not get to the “magic” 1500x number. Can we get to sentient AI if we don’t? Good question.
8. It might be possible to write a workable AI or software which will self-evolve into human-equivalent synthetic consciousness using a fraction of the computing resources I estimate. However, I think if this is true, it will be a *large* fraction of those resources. As in not 1/100 or even 1/10, more like 1/2 or 1/4. So instead of 20-22 years, 1/4 means 16-18 years.
9. Just because we have the hardware necessary to run human-equivalent synthetic consciousness does not mean any person or group will be able to write a workable AI or software which will self-evolve into human-equivalent synthetic consciousness by then. Maybe that will happen concurrently with the evolution of the hardware, maybe it’ll take another 100 years.