This is an interesting development, but it’s not clear to me whether they’ve made any algorithmic breakthroughs in machine intelligence which happen to be well-suited to memristors, or if they’ve just discovered a more efficient way to simulate neural net–type structures, and are expecting intelligence to show up by emergence.
Neither of those quite cover it. They’re using memristors as a much more dense and energy-efficient way of creating neural nets. Their research program will try to work out the best neural net architectures for various fuzzy tasks, the kind of thing that we intuitively think is easy because we have special-purpose hardware for it: vision, motion control, extrapolating the paths of moving objects, efficient locomotion, and so on. If those things can be accelerated with energy-efficient memristor-based neuromorphic chips, then it would be a great boon for robotics.
I’m still hoping for a second industrial revolution, brought about by a combination of cheap energy and better robotics. If we could just improve those things, we could have all sorts of cool futuristic dreams come true: arcologies, cheaper seasteading, large-scale space-based manufacturing—the resources are there, if we can just exploit them cheaply and efficiently enough. This looks like it has a lot of potential. Anybody who’s done robotics work would love to have vision and motion-control processors that don’t have to be treated like blind idiot chips.
However, any implication that they’re looking at this as a way of building general AI is just press-release hyperbole. Their research program itself is focused strictly on easier-to-achieve things.
Artificial intelligence hasn’t stood still over the past half century...Gmail’s Priority Inbox...
Facepalm.
Also, I’m real enthused about this project. I’m sure it will be a breeze to convince the U.S. Department of Defense that they need to invest time and money to study how to make their AIs friendly should they actually succeed with this project.
This is an interesting development, but it’s not clear to me whether they’ve made any algorithmic breakthroughs in machine intelligence which happen to be well-suited to memristors, or if they’ve just discovered a more efficient way to simulate neural net–type structures, and are expecting intelligence to show up by emergence.
Neither of those quite cover it. They’re using memristors as a much more dense and energy-efficient way of creating neural nets. Their research program will try to work out the best neural net architectures for various fuzzy tasks, the kind of thing that we intuitively think is easy because we have special-purpose hardware for it: vision, motion control, extrapolating the paths of moving objects, efficient locomotion, and so on. If those things can be accelerated with energy-efficient memristor-based neuromorphic chips, then it would be a great boon for robotics.
I’m still hoping for a second industrial revolution, brought about by a combination of cheap energy and better robotics. If we could just improve those things, we could have all sorts of cool futuristic dreams come true: arcologies, cheaper seasteading, large-scale space-based manufacturing—the resources are there, if we can just exploit them cheaply and efficiently enough. This looks like it has a lot of potential. Anybody who’s done robotics work would love to have vision and motion-control processors that don’t have to be treated like blind idiot chips.
However, any implication that they’re looking at this as a way of building general AI is just press-release hyperbole. Their research program itself is focused strictly on easier-to-achieve things.
Neither; they’re expecting DARPA money to show up by press coverage.
Facepalm.
Also, I’m real enthused about this project. I’m sure it will be a breeze to convince the U.S. Department of Defense that they need to invest time and money to study how to make their AIs friendly should they actually succeed with this project.