Comment Re:Seems reasonable (Score 1) 294
Letsee here. The Kurzweil model is suggesting that we can get to smart machines by way of brute force. Not necessarily the only way, but one that's hard to argue against as it's just extending today's neural net simulators to faster hardware. Using the open source NEST model, a supercomputer in Japan, the K computer, simulated a second's worth of "brain" activity in 40 minutes. That was a network 1% the size of human brain. So you'd need at least 240,000 times the CPU power to do this at 100% in realtime. Except maybe a few more zeros, since growing a neural network isn't linear, even if you're able to split off subsection for different work, as the brain seems to do. Sometimes.
So 2029 is 15 years away. If we take the erroneous but popular idea that Moore's Law is both a real law and directly about CPU performance (neither of which is true), that's a doubling of performance every 18 months. So by 2029, we only have computers 1024x faster than today's. But by 2045, computers will be a million times faster, at least based on these bad assumptions. So maybe we have a supercomputer than can run a human brain sized neural net in realtime. That get us Skynet by brute force, but not Commander Data. That's another 20 years off.
Of course, I started low... anyone ran run NEST. But it's by far not the most aggressive model. IBM built a more efficient model, modeling a whole artificial brain the complexity of the human brain on a Blue Gene/Sequoia Q supercomputer. It ran 1053x slower than realtime.... which suggests a realtime version might be possible around 2029. IBM actually say it might be as early as 2023, as they're building chips that implement their "neurosynaptic cores" in hardware. The model has over 2 billion neurosynaptic cores, and it's very intentionally designed to be a brain, though not a strict emulation of a human brain. There are dozens of projects around the world doing similar things. One team in Europe has a realtime honeybee scale brain running, and hopes to have a rat scale brain done this year. Another team has a non-realtime model similar to a cat's brain... can hatz cheezeburger?
So it sure looks possible to have Skynet by 2029. Self-contained thinking mobile machines, probably not for a decade or two beyond. And that's assuming no technological roadblocks in scaling our hardware. But also no huge leap away from the brute force approach. And no hardware design help from IBM's realtime brain of 2023. But of course, it won't even graduate college before 2030, assuming a few upgrades along the way. And a few years after that, we may not even understand the improved brain it's getting us to build for it...