Fact checking Moravec's paradox
I have launched a YouTube channel in which I analyze AI developments from a normal technology perspective. This essay is based on my most recent video in which I did a deep dive into Moravec’s paradox, the endlessly repeated aphorism that tasks that are hard for humans are easy for AI and vice versa. Here’s what I found: Moravec’s paradox never been empirically tested. (It’s often repeated as a fact by many AI researchers, including pioneers I know and respect, but that doesn’t mean I’ll take their claims at face value!) It is really a statement about what the AI community finds it worthwhile to work on. It doesn’t have any predictive power about which problems are going to be easy or hard for AI. It comes with an evolutionary explanation that I find highly dubious. (AI researchers have a history of making stuff up about human brains without any relevant background in neuroscience or evolutionary biology.) Moravec’s-paradox-style thinking has led to both alarmism (about imminent superintelligent reasoning) and false comfort (in areas like robotics). To adapt to AI advances, we don’t need to predict capability breakthroughs. Since diffusion of new capabilities takes a long time, that gives us plenty of time to react — time that we often squander, and then panic! Watch the full argument here or read it below. Every week brings new claims about AI advances. How do we know what’s coming next? Could AI predict crime? Write award-winning novels? Hack into critical infrastructure? Will we finally have robots in our home that will fold our clothes and load our dishwashers? What will AI advances mean for your job? What will it mean for the social fabric? It’s hard to deal with all this uncertainty. If only we had a way to predict which new AI capabilities will be developed soon and which ones will remain hard for the foreseeable future. Historically, AI researchers’ predictions about progress in AI abilities have been pretty bad. We don’t really have principles that describe which kinds of tasks are easy for AI and which ones are hard. Well, we have one — Moravec’s paradox. It refers to the observation that it’s easy to train computers to do things that people find hard, like math and logic, and hard to train them to do things that we find easy, like seeing the world or walking. It comes from the 1988 book Mind Children by Hans Moravec, who was — and is — a robotics researcher. He wrote: It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility. In the early days of artificial intelligence, researchers focused on chess and other reasoning tasks, since these were thought to be some of the hardest and what made us uniquely human. But funnily enough, if you want to build a robot that can defeat human grandmasters, figuring out which moves to make is the easy part. Physically making the moves on the chessboard is the hard part. This is pretty well known today, so Moravec’s paradox seems to make a lot of intuitive sense. If Moravec’s paradox is true, the implications would be amazing. If we want to know which AI capabilities might be built next, we just have to see how hard they are for humans. So scientific research will get automated before folding clothes, and so on. But here’s the thing — Moravec’s paradox has never been fact checked. And that’s despite videos with hundreds of thousands of views, and TED talks all repeating it as a fact. When I dug into the evidence behind the so-called paradox, I found something surprising. In this essay I’ll discuss why the theory and evidence behind the paradox are flaky. Then I’ll explain why simplistic predictions about what is easy or hard for AI have misled AI researchers and tech leaders. It has led to alarmism on the one hand and false comfort on the other hand. (Now there’s a paradox.) And finally I’ll answer the question, if we can’t rely on Moravec’s paradox, then how should we prepare for AI advances and their impacts? How would we test Moravec’s paradox? We could take some sample of the tasks that are out there, determine how hard they are for people, how hard they are for AI, and make a graph. If we saw something like this, the paradox would be confirmed. But here’s the problem: Which set of tasks should we consider for our analysis? When AI researchers say Moravec’s paradox checks out, they are implicitly limiting their focus to problems that are considered interesting in the AI research community. There are an endless number of tasks that are easy for both humans and AI, but they are not interesting. How bright is an image? How to play tic tac toe? Thousands of these tasks get solved by programmers on a daily basis and coded into AI systems, and when people do so, they don’t make a fuss about them. There are also an endless number of tasks that are hard for humans and, as far as anyone knows, are also hard for AI. Identifying hit songs; predicting stock prices; cracking as-yet-undeciphered ancient scripts; even building a Dyson sphere. These are so hard that there is essentially no progress on these problems, although some of them attract junk research that tends to quickly get debunked. So these problems also don’t tend to get talked about as much. In fact, there are thousands of problems that computer scientists have proved to be “NP-complete”, which means we have strong mathematical reasons to think they will forever be hard for AI, so serious AI researchers don’t tend to work on them. They work on easier, approximate versions of the problems instead. The other two quadrants of the chart are different. On the top left, tasks like playing soccer, that are easy for humans but currently hard for AI, are extremely interesting to AI researchers. That’s because we know it’s possible to teach AI these skills, but we haven’t yet managed to do so, which makes them great tests for AI progress. On the bottom right, problems that are hard for humans but easy for AI, such as searching the web, are also interesting. These capabilities tend to greatly augment human productivity. Even though they are in some sense “easy”, the industry invests a lot in making web search and other tools work as effectively, quickly, and cheaply as possible. So research on these tasks is a big driver of AI progress. In short, when you’re thinking about the space of all possible tasks, if you basically ignore two quadrants of your 2x2 matrix because they are not interesting, then of course it will seem like what you’re left with shows a strong negative correlation between the two axes. To be clear, the reason AI researchers are drawn to Moravec’s paradox isn’t because it is empirically backed. It’s because it comes with an intuitively appealing story. In his book, Moravec provided this explanation: Encoded in the large, highly evolved sensory and motor portions of the human brain is a billion years of experience about the nature of the world and how to survive in it. The deliberate process we call reasoning is, I believe, the thinnest veneer of human thought, effective only because it is supported by this much older and much more powerful, though usually unconscious, sensorimotor knowledge. We are all prodigious olympians in perceptual and motor areas, so good that we make the difficult look easy. Abstract thought, though, is a new trick, perhaps less than 100 thousand years old. We have not yet mastered it. It is not all that intrinsically difficult; it just seems so when we do it. And then Moravec praises reasoning systems like STRIPS from the 1970s, and this is the kind of thing he has in mind when he says that reasoning is easy for AI. These are purely symbolic systems that solve problems like how to put blocks on top of each other in a certain way. But AI researchers have learned a lot in the half century since the heyday of STRIPS and other such reasoning programs. Th…
Send this story to anyone — or drop the embed into a blog post, Substack, Notion page. Every play sends rev-share back to AI Snake Oil.