One of the world’s two top conferences in AI and machine learning (ML) took place recently in Stockholm: ICML 2018. By some estimates, approximately 6000 people from all over the world descended on Stockholm for the event and there were a total of about 4500 papers submitted! It was a stimulating (and at times overwhelming!) event with tutorials, workshops, invited and contributed talks covering several of the frontiers of AI and ML – deep generative models, optimization, reinforcement learning.
It was also an occasion for me to reflect again on a contentious issue that has garnered some space in public debate, thanks to celebrities such as Stephen Hawking, Bill Gates and Elon Musk, namely Superintelligence – the idea that AI is rapidly moving towards a singularity when machine intelligence will exceed human capabilities and that this will constitute grave existential risks for the future of humanity. At the opening of Vetenskapsfestivalen this year, another celebrity, Max Tegmark (with his trademark rock star image, beamed live from MIT via a hologram) to talk about it and his new book Life 3.0 (of which I’ve written a skeptical review). Even though I was thus prepared for the hyperbole, I was shocked by his claim, that according to most AI researchers, superintelligence was just around the corner, at most a few decades away and that it would be upon us way sooner than climate change! This is a preposterously irresponsible claim and it is also totally wrong! Superintelligence proponents like to cite surveys they conducted to gather opinions from AI researchers, but the trouble is that they conducted them in all the wrong places! They should have conducted a survey at ICML 2018 and the results would leave no room for confusion.
But beyond surveys which are not always very meaningful, there are substantial reasons for why AI researchers don’t pay too much attention to the superintelligence talk. In this post, I want to point at some of these.
Some local proponents of superintelligence have made the claim that the reason AI researchers don’t take them seriously is nothing more than the inability of the AI researchers to solve the problem themselves! Since this comes from a mathematician, let me first start with an analogy. Suppose I took it upon myself to philosophize on the future of mathematics. Reading about the progress in the twin primes conjecture, I might marvel at the “exponential progress” of the “polymath” project in going from 70 million down to 246 in a few months and thus predict that the full conjecture would just need a few more months at worst. A mathematician working in the area wouldn’t take my prediction seriously at all and would actually find it ridiculous. Is that simply because she can’t solve it herself?! Er … not quite. The mathematician would note that I based my prediction on nothing more than a blind interpolation without any idea of the underlying technical details of the methods involved in the proofs, and that there are fundamental obstacles in the method that would prevent the proof from going all the way down to 2.
At least the problem is well defined in the case of the twin prime conjecture. With “intelligence” the problem is much worse since the very concept is ill defined. We may all say of “intelligence” that “I know it when I see it” as US Supreme Court Justice Potter Stewart famously said about about pornography, but beyond that it is extremely hard to pin down concretely. The most famous suggestion is the Turing test from the seminal paper by Alan Turing in 1950. In that article, Turing starts by noting that it is very hard to precisely define what it means to ask “can a machine think?” and suggests that we “replace the question by another, which is closely related to it and is expressed in relatively unambiguous words”, namely, the Turing test. He concludes by speculating about a time when machines will compete with humans on numerous intellectual tasks and suggests tasks that could be used to make that start such as playing chess.
Modern AI has indeed taken up the suggestion and we have seen impressive achievements along these lines. First, IBM’s Deep Blue defeated the world chess champion Gary Kasparov in 1997 and since then chess programs regularly beat human grandmasters even given a pawn advantage. More recently, in a dramatic breakthrough, DeepMind’s AlphaGo program decisively beat the world champion at Go, a game considered to be much more difficult than chess. The latest version of AlphaGo learnt to play Go by itself without any human input and discovered completely new strategies for Go.
Impressive as these achievements are, there is still a huge difference between the kind of narrow AI represented by programs that are designed to solve a precise narrowly defined task such as playing games like chess and Go, and general AI which can learn to solve general tasks as we humans do. To show why, I could do no better than to point to a series of long and thoughtful essays by noted robotics researcher Rodney Brooks on Superintelligence.
Here is an interview with Deep Learning Guru Yann LeCun where he says:
“We’re very far from having AI technology that would allow us to build machines like this. And it’s basically because machines today don’t have common sense.”
Two keynote talks at ICML also gave indications about how much more difficult the more general problem is. First Joyce Chai gave a talk about interacting with robots to train them to perform specific tasks such as making a smoothie. Her focus was on making the robot understand natural language via interaction and demonstration with humans. Making a smoothie is not as easy as you might think! In the end, the robot could be trained to recognize a peeled orange from an unpeeled one. Next was a very interesting talk by Josh Tennenbaum on Building Machines that Think like People. As he observes, “no machine system yet built has anything like the flexible, general-purpose commonsense grasp of the world that we can see in even a one-year-old human infant.” His talk gave examples from their work on teaching machines to learn about the physical world using an interesting combination of simulation and probabilistic programming.
Not only is the superintelligence debate more suitable for a conversation over beer than serious AI research, it may actually be positively harmful by distracting from more urgent problems. Consider another example that was touted as an example of general AI in the mathematician’s post linked to above, Google’s Duplex. This is undoubtedly yet another example of an impressive engineering feat by Google, but as any NLP researcher would tell you, it hardly counts as a real advance in genuine language understanding. Indeed this system is trained on a very narrow range of tasks and is thus yet another example of narrow AI as pointed out by cognitive scientists Marcus and Davis on Google Duplex.
On the other hand, Google’s Duplex is a very good example of a real and present danger of AI, namely automation! As the demo shows, such narrowly defined tasks are now within the range of automated systems, for example, we could envision call centers being automated using that technology. The threat posed by AI is here about job displacement about which we wrote a commentary titled AI Dangers: real and imagined. I would also recommend the books The Second Machine Age and Rise of the Robots. The threat of automation and job displacement calls for serious thinking, not only from from AI researchers, but also economists, social scientists and policy makers. Talking about superintelligence only distracts from this much more urgent calling.
Returning to AI safety, the AI community has started taking the risks of AI seriously, for instance the recent work from Google and DeepMind. These efforts are grounded in concrete AI research, for, as they write,
in our opinion, one need not invoke … extreme scenarios to productively discuss accidents, and in fact doing so can lead to unnecessarily speculative discussions that lack precision … We believe it is usually most productive to frame accident risk in terms of practical (though often quite general) issues with modern ML techniques.
We would do no better than to end with the wise words of Francois Jacob:
The beginning of modern science can be dated from the time when such general questions as “How was the Universe created” … “What is the essence of Life” were replaced by more modest questions like “How does a stone fall?’ How does water flow in a tube?” … While asking very general questions lead to very limited answers, asking limited questions turned out to provide more and more general answers.
Do you believe factors such as Moore’s Law or Conway’s Law will eventually and/or abruptly cease our progress in the path towards creating the perfect General AI? If yes (or no), why (not)?
LikeLike