Moravec's paradox

• 1. Introduction

  ›

  Moravec's paradox is one of the most counterintuitive findings to emerge from decades of artificial intelligence research. A computer can beat grandmasters at chess, solve advanced algebra, and pass intelligence tests at an adult level. Yet give that same computer the task a one-year-old performs without thinking, like recognizing a familiar face or toddling across a room, and it struggles enormously. Hans Moravec put it plainly in 1988: 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." The paradox raises a question that cuts to the heart of what intelligence actually is. Why should the things we find hardest be easiest for machines, while the things we find simplest leave machines baffled? The answer, it turns out, is written in billions of years of evolutionary history.

• 2. A Billion Years of Practice

  ›

  Moravec's own explanation begins not with computer science but with deep time. Natural selection has been refining the human body and brain for roughly a billion years, according to his account, and that long process has quietly packed an enormous amount of engineering wisdom into the parts of the brain that handle sensing and movement. The deliberate act of reasoning, he argued, is "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." Abstract thought, by contrast, is a newcomer. Moravec placed its age at perhaps less than 100,000 years, a blink in evolutionary terms. Because it is so new, it has had little time to be optimized, which is precisely why it feels hard when we attempt it. The logic flows neatly from there: we should expect the difficulty of reverse-engineering any human skill to be roughly proportional to the time that skill has spent evolving in animals. Skills that evolved over millions of years, such as catching a ball, reading a face, or navigating a room, are deeply buried in unconscious processing. Skills acquired in historical time, like mathematics, logic, and formal reasoning, sit near the surface and are far easier to replicate in code.

  Marvin Minsky pointed to the same hidden depth from a different angle. "In general, we're least aware of what our minds do best", he wrote, adding that we are "more aware of simple processes that don't work well than of complex ones that work flawlessly." The very invisibility of perception and motor control is a sign of how thoroughly evolution has solved those problems in us. Steven Pinker extended the argument in his 1994 book The Language Instinct, calling it "the main lesson of thirty-five years of AI research" that "the hard problems are easy and the easy problems are hard."

• 3. Early AI's Wrong Turn

  ›

  Researchers who built the first AI systems in the mid-twentieth century were not being foolish; they were being misled by their own success. Early programs proved capable of using logic, solving algebra and geometry problems, and playing checkers and chess. Since those tasks are genuinely difficult for people and are widely treated as markers of intelligence, it seemed natural to assume the harder work was largely done. Rodney Brooks later described the problem bluntly: intelligence in that era was "best characterized as the things that highly educated male scientists found challenging," including chess, symbolic integration, and the proof of mathematical theorems. What four-year-old children could do effortlessly, like distinguishing a coffee cup from a chair or walking on two legs, was not counted as intelligence at all. Nor, Brooks noted, were aesthetic judgments included in the standard repertoire of intelligent behavior. The "easy" problems of vision and commonsense reasoning were expected to fall into place once the "hard" problems of logic had been cracked. They did not, and the field passed through what researchers would later call the AI winter.

  Allen Newell pushed back in a 1983 chapter on AI history, calling the entire framing a "myth." He wrote that "a myth grew up that it was relatively easy to automate man's higher reasoning functions but very difficult to automate those functions man shared with the rest of the animal kingdom and performed well automatically, for example, recognition." Newell's objection was not that the observation was wrong, but that treating it as a deep law risked misrepresenting what the field had actually learned. Arvind Narayanan later framed the paradox differently still, describing it as a statement about "what the AI community finds it worthwhile to work on" rather than a reliable prediction of which problems will be hard or easy for AI going forward.

• 4. Brooks and Nouvelle AI

  ›

  Rodney Brooks drew a practical lesson from the paradox and acted on it in the 1980s by abandoning the logic-and-symbol approach to intelligence entirely. He decided to build machines with "No cognition. Just sensing and action," deliberately leaving out what had traditionally been thought of as intelligence in artificial intelligence research. He called this new direction "Nouvelle AI." The bet was that perception and motor control, precisely because they had been so hard to replicate, might be a more productive foundation for building robots that could actually function in the physical world. Where earlier researchers had climbed the ladder of abstract reasoning and expected embodied skill to follow, Brooks started at the ground and worked up.

• 5. Moore's Law Catches Up

  ›

  Moravec had predicted as early as 1976 that raw computing power would eventually become sufficient to handle perception and sensory tasks. By the 2020s, computers were hundreds of millions of times faster than those available in the 1970s, and that accumulated speed had finally begun to make good on his forecast. Machine-learning techniques, fed by that vastly greater computation, started to close the gap on vision, speech, and movement. In 2017, machine-learning researcher Andrew Ng offered what he called a "highly imperfect rule of thumb": "almost anything a typical human can do with less than one second of mental thought, we can probably now or in the near future automate using AI." The one-second threshold maps loosely onto Moravec's original intuition, since fast intuitive tasks tend to be the oldest and most deeply evolved skills in the human repertoire. Whether the paradox has been resolved, or merely deferred, remains an open question. No consensus has yet formed among researchers about which tasks AI tends to excel at, leaving Moravec's original puzzle still alive at the edges of the field.