No Singularity for you!
All nerds everywhere have heard of the legendary AI genius Ray Kurzweil and his page turning opus on the transcendence of the human condition into machines. The basic premise of Kurzweils argument is that human intelligence and knowledge appears to be accelerating exponentially and if human technology continues to accelerate at its current rate we’ll all be god like cyber-entities by next Monday.
Kurzweil says that even though most of the people in the field think we’re still several decades away from creating a human level intelligence, he puts the date at 2029 — less than 15 years away.
Interestingly it’s not just Kurzweil making this claim. It seems to be a mantra getting adopted by many of the technology industries greatest and most respected minds.
“I am in the camp that is concerned about super intelligence. First the machines will do a lot of jobs for us and not be super intelligent. That should be positive if we manage it well. A few decades after that though the intelligence is strong enough to be a concern. I agree with Elon Musk and some others on this and don’t understand why some people are not concerned.”
Even the world’s governments are funding massive research projects to produce human brain simulations;
Thus with respect to the many esteemed minds who seem to share this view, I wish to reject this premise and assert that they are all morons that don’t know what they are talking about. No immanent robot transcendence is in anybody’s near future. Now don’t get me wrong I hold many of these folks in great esteem… but they happen to be completely and obviously wrong on this account and I have two credentials in my pocket that I believe make my case worth serious consideration.
First I have multi-generation experience with the problem, I am a 5th generation AI engineer. My great, great grandfather Edgar Bristol pioneered the use of analog neural networks for factory automation in the 1920’s and my family has been working with them ever since.
“Back in the 1920s, when Edgar Bristol, Sr., was inventing the very first loop controller”
Today I work on machine learning technologies that involve self-programming systems and my brother Aaron with his masters in AI, makes video games. My great great grandfather would have said of today’s digital neural networks that they are farther from simulating human intelligence than the primitive analog ones he worked with in the 1920’s, for reasons that I will articulate shortly.
Number two, I have a pretty good track record of calling BS on trendy technology hype before everybody else eventually comes around to agreeing with me. I wrote an extensive series of blog articles that were critical of the Human Brain Project when it was announced. Almost two years later the consensus appears to have gone my way. Here’s what I said about the Human Brain Project almost two years ago when it was announced and everybody was leaping with enthusiasm for it.
Here’s what they’re saying about it today….
“The report says that ambitions for whole-brain simulation are premature”
So why do I know better than all these brilliant people that we are in no danger of being dominated by or becoming intelligent machines anytime soon? My family has worked on it for several generations and I’ve worked on machine intelligence for several decades and here’s what we’ve learned about the problem;
The analog computers my great great grandfather worked on were closer to simulating a real brain than a digital computer based neural network because man-made digital computers have had the noise and chaos beaten out of them. Yes it’s true that we have made machines that think, this achievement may arguably be even greater because we made machines that think NOTHING like us. The very forces of physics that give rise to life and organic computing and thought have been systematically erased from digital machines. Our computers are the wrong architecture, they cannot simulate the turbulent systems that give rise to life and biological thought. Yes we can use a computer to crudely simulate these forces at tremendous computational expense but we can never do it perfectly or fast with a digital machine. The analog signals that living systems compute with can solve problems involving infinities, trillions of simultaneous variables and synchronize timing events at the quantum level almost instantaneously. No digital computer will ever match these computational capabilities at any clock speed or scale. When you examine the calculations that are really involved in computing “life” it becomes obvious that we are using the wrong tools to try to solve the problem.
The fact that nobody else seems to say this as clearly and unambiguously as I just have is a persistent mystery to me. All these intelligent people should KNOW that our computers just can’t do that and won’t be able to do that in 15 years let alone 100 years if we don’t junk the digital architecture and start over again. Now I know what some of you reading this are thinking… you’re thinking “But QUANTUM computers are right around the corner! …and they can compute anything!” This is another one of these modern “cold fusion” kind of computing fantasies that mystifies me… sure quantum computing sounds groovy, I hope it works, but I don’t really give a damn… because we don’t need them to compute life or artificial intelligence… we don’t use them ourselves to think! An analog computer that could compute with signals instead of bits could solve amazing computing problems NOW without any science-fiction premise required. In the 1980’s when I was first studying 3D ray-tracing and light simulation the promise of optical computing was all the rage. We were right around the corner from being able to compute with light interference patterns instead of silicon! I’m 48 years old now, most of my hair is gone… and there are still no commercial parallel optical computing chips in mainstream use for some reason… I can wait for quantum computing, we don’t appear to need it for any of the interesting problems…
Now I hate to be a wet blanket and shatter everyone’s Star-Trek fantasies about how the future will unfold but as somebody who really wants to participate in achieving amazing scientific and computational advances, I have to begin with a brutally grounded rational foundation. I’m not going to invent anything amazing if I demand that reality conforms to how I WISH things worked. When I study how living systems compute I find a computing model that cannot be simulated practically in a machine and that many of life’s most important properties, like the ability to self-program and systematically evolve derive from physics forces we can’t practically simulate with a digital machine.
So what does this mean for Kurzweil and his grand delusions about machine intelligence? I think he may ultimately be right but I see two definitive barriers to his anticipated singularity.
Sometime between here and bionic transcendence we will have to completely ditch the digital computing paradigm and that transition may be difficult and slow. There will be no real advances in “Artificial Intelligence” until that happens. Everything we call AI today is really just humans making tape recordings of ourselves and calling them “intelligent” when the recording sounds like us. This is the case with all of our achievements in modern neural networks, they are just interestingly life-like marionettes. REAL neural networks, design and program themselves… that’s what makes them intelligent.
- There is a limit to intelligence but no limit to stupid. It’s not clear or maybe even plausible that a liquid computer the size of a grapefruit trapped inside our skulls can ever be smart enough to comprehend its own design well enough to consciously replicate and improve on itself. Kurzweil may have FAITH that collecting enough of these computing grapefruits to collaborate with one another can solve the problem but in my experience the larger the collection of people the more ignorance and self-interest dominates any collaboration. It is not true that technology improves with time, it just gets more complicated and unwieldy until economic forces require that the old solution be discarded in favor of something simpler that a thinking grapefruit can manage. Therefore we inherently cap the ultimate potential complexity of any technology we make to suit our individual human limitations… as demonstrated by our invention of thinking machines that aren’t allowed to have thoughts we can’t understand.
This may sound like a pessimistic view of the future but I have found it to be very liberating when it comes to doing pioneering AI work. When you accept that the machines we’ve made will never evolve into Commander Data but are remarkable in their own way, what follows is the realization that we have created the opportunity to explore ways of thinking and computing that are NOTHING like the ones we are most familiar and comfortable with. Living computers have major limitations, predominantly among them the need to survive and perpetuate themselves at the expense of any higher purpose. This constraint has made our computing paradigm amazing but also brutally pragmatic. We quickly lose interest in computing solutions to problems that don’t directly result in getting food, shelter, reproduction or entertainment. Computers on the other hand have infinite patience and fascination for any problem we turn them to. The real opportunity for “Artificial Intelligence” is getting machines to think about things we can’t or won’t think about. This observation however requires giving up control and full comprehension of how the computer solves these kinds of problems. If we can comprehend how the computer solved the problem then we restricted it to solving only problems we could figure out for ourselves. We have to sacrifice the RESTRICTION we impose on AI that it works in a way we can comprehend in order to produce AI that can do something new and amazing. If you accept the premise that any interesting computed AI must by some definition be “incomprehensible” to us, then it becomes easy to recognize dead-ends in modern AI research.
The big buzz word in modern AI research is “Deep Learning” with convolutional neural networks. This is a very fancy way of saying that human beings are hand coding neural networks to solve interesting problems. I assert that there is only ONE interesting problem worth solving in this space and it is creating a neural network that can design and optimize itself to find solutions to problems without being specified by a human engineer. Without this requirement all we are doing is making recordings of our own intelligence and playing them back.
Today, as in 1769 when the legendary Mechanical Turk chess playing robot spread awe and wonder around the world, little has actually changed. The 2015 Mechanical Turks we have designed are no closer to actually being intelligent without a human agent puppeteering them.
So if I accept Kurzweils premise that there is something like a “singularity” to look forward to, which I’m actually inclined to do. I would modify his vision in a couple important regards.
- There will never be a robot body or digital mind that we can migrate to. They don’t work.
- We may successfully employ the thinking machines we have created to devise medical technologies that extend our organic lives indefinitely and that advances in this area may accelerate exponentially