What this covers

Rodney Brooks, MIT's emeritus robotics professor, discusses with Russ Roberts why both technologists and skeptics get the timeline of AI and robotics wrong. Brooks argues that observers systematically misjudge technological change through two linked errors: they overestimate what will happen in five years while underestimating what happens in fifty, and they conflate narrow machine performance with genuine understanding. The conversation ranges across driverless cars, the history of deep learning, GPS infrastructure, and how physical systems—unlike software—take decades to deploy because they must run reliably without failure.

Brooks illustrates his points through concrete examples: an AI that captions "kids playing frisbee" has no concept of games or childhood, yet observers credit it with human-level comprehension. He traces deep learning back through Geoffrey Hinton and Yann LeCun to show that the recent breakthroughs came not from new ideas but from more compute and better activation functions applied to old algorithms. He explains why the B-52 bomber will fly into the 2040s and why factories still run PLCs that emulate 1950s relays—physical capital persists because humans cannot afford to discard expensive, working assets. On superintelligence, Brooks contends we may be building ladders while claiming to approach the moon; the gap between current AI and genuine general machine intelligence could span centuries. He also challenges overconfident skeptics, noting that researchers were right to predict AlphaGo would win not through brute force but other techniques, so dismissing doubters entirely misses the nuance. Throughout, Brooks applies Amara's Law and invokes Clarke's Third Law to show why confidence in unlimited AI capability amounts to unfalsifiable faith rather than reasoned forecast.

Sharpest takeaway

Brooks argues that AI and other technologies are systematically misjudged: their short-run impact is overestimated while their long-run transformation is underestimated, largely because observers confuse narrow machine performance with general competence and treat advanced technology as limitless magic.

  • Amara's Law: overestimate short-run, underestimate long-run effects
  • Pundits confuse performance with competence in AI systems
  • Physical deployment of technology takes decades, giving humans time to adapt

The argument · threads7 threads · 33 claims
0.76

Exponential technological progress eventually hits limits and stops accelerating exponentially.

5 pointscentrality 5/5
  • We tend to overestimate the effect of a technology in the short run and underestimate its effect in the long run, because we immediately imagine fantastic near-term uses but cannot make the conceptual leap to how it ultimately reshapes the world.

    “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”

  • There are limits on humans and on everything we have ever developed; no technology has ever turned out to have no limits, so the assumption that future AI will be all-powerful and limitless has no precedent.

    “is any technology we've ever developed more powerful--have no limits? There are limits on humans. There are limits on everything we have developed.”

  • Deep learning was not a sudden breakthrough: backpropagation was buzzed about in the 1980s, ran into limits, and almost everyone abandoned it except Hinton and LeCun; only three small innovations—more compute, a better activation function allowing easy derivatives, and clamping/pre-structuring deep networks—made it suddenly work after 2008, becoming dominant within a decade after 20 years of prior work.

    “the main technical ideas of these algorithms, deep learning, were around in the 1980s: backpropagation is the technical term for how the learning applies.”

  • Most exponential trends are not exponential forever: Facebook grew exponentially until it used up everyone in the world, and Moore's Law itself ran out once feature sizes reached the point of counting atoms and could no longer be halved every two years, so it is a mistake to assume AI progress will remain exponential.

    “most exponentials are not exponentials forever. Facebook was exponentially growing for a while, but it's sort of used up everyone in the world now.”

  • The end of Moore's Law has been a great service to computer architecture: for 50 years no one could afford to deviate from the narrow path because a competitor riding Moore's Law would beat them, but now that it has run out there is a flourishing of computer architecture, exemplified by GPUs applied to deep learning.

    “Now that you no longer have Moore's Law we're seeing a flourishing of computer architecture for the first time in 50 years. And GPUs applied to deep learning are an example of that.”

0.72

AI capabilities are narrow, brittle, and heavily engineered rather than general intelligence.

4 pointscentrality 5/5
  • AI pundits make the mistake of seeing narrow machine performance and assuming the human-level general competence we would infer from a person doing the same task; an AI captioning 'kids playing frisbee' has no idea what a frisbee or a game is, cannot answer whether a six-month-old could play, and would not react to 'frisbees are tasty,' yet observers credit it with human-level understanding.

    “people, I think, saw that level of performance, and assumed that the system had the same level of competence as a human who could do that work. But, the system didn't even know what a game was, what a frisbee was.”

  • Words like 'learn' are suitcase words (Marvin Minsky's term) packed with many different meanings—learning to walk, ride a bike, speak Latin, do calculus all use very different mechanisms—so when an AI paper or press release says a system 'learned' something, people wrongly assume it can learn anything, when in fact it is a brittle, narrow capability; even 'deep' in 'deep learning' merely refers to number of network layers, not depth of thought.

    “A suitcase word is something that Marvin Minsky, one of the founders of AI, came up with where, it's a word that has so many different meanings packed into it.”

  • When AlphaGo played the world Go champion, it was supported by roughly 200 engineers worrying about helping it, whereas the human champion's only support was a cup of coffee—illustrating that AI achievements are narrow, heavily engineered specialized tasks, not general intelligence.

    “you had 200 engineers there worrying about helping it, and supporting it... and the world Go champion had a cup of coffee. That was his support.”

  • AI progress has been driven by exponential hardware growth rather than new ideas: Alan Turing devised the essential algorithm for computer chess in the 1940s, MIT's 1965 Mac Hack embodied it but was beaten by people, and Deep Blue's defeat of Kasparov in the 1990s used essentially the same algorithm—so improvement came mostly from faster processors, not innovation.

    “It was exactly the same algorithm that Turing had come up with back in the 1940s.”

0.58

Physical technologies take decades to deploy because infrastructure must run reliably without failure.

4 pointscentrality 4/5
  • Deployment of technology involving physical capital takes far longer than software because, unlike Google or Facebook whose marginal cost for a new feature is zero (the code downloads into your browser), physical systems must keep running reliably and cannot afford failure; hence the B-52, built mostly in 1962, may fly until 2040 or even 100 years old, and factories still run PLCs that emulate 1950s electromagnetic relays, upgraded only about three times in 20 years.

    “We aim to do three upgrades every 20 years.”

  • Changeover to new technology takes a long time partly because of existing physical assets: houses last hundreds of years, and even the driverful cars people buy today will be around in 20 years because owners won't want to abandon an expensive asset, so the transition to driverless cars will be slow.

    “Even our driverful cars that we buy today are going to be around in another 20 years. People aren't going to want to give up that asset that they just spent a lot of money on today.”

  • Just as cars transformed cities and countryside, driverless cars will eventually restructure cities in ways we can't yet imagine, but the path there will be through special dedicated areas with different laws, where driverless cars gradually take over rather than directly replacing driver cars.

    “the way we'll get there is we'll start having special areas dedicated to driverless cars, where the laws are somewhat different from where we have cars with drivers”

  • Driverless cars are being overestimated in the short run; supposedly deployed systems like a Tokyo 'driverless taxi' actually run a fixed route four times a day with a human in the driver's seat, which is hardly a replacement for a human driver, and full deployment faces incredible practical problems.

    “it's a driverless taxi that makes four round trips per day on exactly the same route. Oh, and there's a person in the driver's seat, just in case.”

0.54

AI alarmists lack actual AI expertise and make unfalsifiable, faith-based claims about superintelligence.

3 pointscentrality 4/5
  • If even 1% of what AI alarmists believe AI is capable of today were actually true, Brooks—who has worked in AI for 40 years leading large teams—would be vastly wealthier than he is, which is his evidence that their claims make no sense.

    “If 1% of what these people believe AI is capable of today--if 1% is true of what they believe is true today... I would be so incredibly wealthier than I am.”

  • Per Arthur C. Clarke's Third Law, any sufficiently advanced technology is indistinguishable from magic, meaning an observer cannot tell what its limits are; even Isaac Newton, shown an iPhone, could not hypothesize how it works or know its limits, so AI pundits who assume the technology can do anything are making a faith-based, unfalsifiable argument.

    “Any sufficiently advanced technology is indistinguishable from magic. And, the argument there is that if it's a sufficiently advanced technology, you can't tell what its limits are.”

  • None of the prominent figures who warn that AI is an existential threat—Bostrom, Harris, Tegmark, Martin Rees—have ever done any work in AI itself; they are outsiders, and Bostrom in particular finds doom in everything from extraterrestrial searches to nuclear research, so AI worry is just one of his many catastrophe scenarios rather than special expertise.

    “none of these people who worry about this have ever done any work in AI itself. They've been outside.”

0.50

Reaching superintelligence requires conceptual breakthroughs we may not be capable of, not just time.

2 pointscentrality 4/5
  • Reaching the superintelligence pundits describe may not be just a matter of time; we have taken only baby steps toward it, like building ladders and claiming we're getting closer to the moon, and it could take hundreds of years—just as we've had chemistry for 2000 years and still don't understand a great deal of it.

    “maybe we're building ladders, and people are saying, 'Oh, yeah; we're getting closer to the moon. They'll get to the moon really soon.'”

  • Humans may simply not be smart enough to build genuinely intelligent systems, just as dolphins are not dexterous or smart enough to build a robot dolphin; we are not as smart as we like to think, and a sufficiently advanced alien intelligence might view us as we view dolphins.

    “Maybe we're not as smart as we like to think we are; and I think that's true of many people. We're not as smart as we like to think we are.”

0.44

Technology adoption depends on prior familiarity and metaphors that break down at complexity.

7 pointscentrality 2/5
  • Technology adoption builds incrementally on prior familiarity: the smartphone worked when it arrived around 2007 because people had already gotten used to touchscreens in ATMs and airports (the first appeared in research labs in 1988), so the only genuinely new thing to learn was the pinch-to-zoom gesture, while button-touching merely emulated physical buttons.

    “The reason smartphones worked when they came along--I think it was 2007--the reason they worked was people had gotten used to the idea of a touchscreen.”

  • AI researchers correctly predicted that brute-force search—which works in chess—would never beat a human at Go; AlphaGo and AlphaZero in fact won not by brute force but by using other techniques, so the claim that skeptics were simply wrong about Go is itself mistaken.

    “What was said was: Using brute force search will never beat a person in Go. Whereas brute force search works in chess. And in fact, Alphazero, Alphago, don't just use brute force search.”

  • Following Lakoff and Johnson's 'metaphors we live by,' much of human language and our technology is based on physical metaphors derived from how our bodies move through the world (e.g., 'tackle a problem'), but our brains and nervous systems are not built on those metaphors, so when phenomena get complex—like quantum mechanics being forced into 'wave or particle'—the physical metaphor breaks down.

    “Much of our language is based on physical metaphor for how our bodies move around, the flow of time, how we tackle a problem... Everything in our language is based really on metaphors, our physicalness in the world.”

  • The food supply system will be transformed over the next 50 years by indoor farming and synthetic meat, driven by climate change shifting where crops grow and by the unsustainable carbon footprint of meat production—livestock holds 1.6 times the volume of carbon as humans, who in turn hold 9 times that of all other mammals.

    “I think we are going to have a transformation in our whole food supply system over the next 50 years. Indoor farming and synthetic meat.”

  • Because there were roughly 100 promising approaches in the 1980s and only deep learning 'popped,' we cannot tell in advance which of the thousands of existing research ideas will become the next dominant AI technique; something else will be the hottest flavor in a few years, but no one knows which.

    “we couldn't tell from other 99 this one popped. Maybe one of those others is going to pop some day”

  • Hollywood movies about technology depict a world identical to today with just one thing changed (e.g., Bicentennial Man's family reading physical newspapers alongside an intelligent robot), but in reality many things change together, so the fantasy of suddenly waking up to a single evil superintelligence is nutty—we would first see disdainful, then grumpy, then nasty AI systems and would change the world before any such endpoint.

    “when you watch a Hollywood movie involving technology, usually the world is identical to what it's been like, but one thing changes.”

  • 1960s science fiction like the original Star Trek, set centuries in the future, depicted computers as mainframes with flashing lights that could be defeated by hard questions, grossly underestimating progress that would actually arrive within 20 years; only 2001: A Space Odyssey came close with small screens and speech interfaces.

    “their estimation of what computers would be like in that future were grossly less than they were even in the 1990s”

0.40

Driverless cars will face novel regulatory and cultural obstacles requiring slow, gradual adoption.

3 pointscentrality 3/5
  • Removing the human driver creates entirely new cultural and regulatory edge cases that don't exist today—such as whether a 12-year-old or a dementia patient riding alone is allowed to instruct the car and thus counts as 'driving' it—and resolving these will take a long time with publicized bad incidents along the way.

    “Is the kid allowed to tell the car what to do, or change what's going on? If the kid is doing that, is the kid now, in scare quotes, "driving the car?"”

  • GPS is so deeply embedded in critical infrastructure that its loss would cause months of serious disruption: the super-accurate clocks on GPS satellites synchronize the entire electrical grid, so without them the grid would have to be broken into isolated sub-pieces and could no longer ship electricity across the country.

    “our electrical network, infrastructure, uses those clocks on those GPS satellites in order to synchronize the whole grid. If the GPS satellites go away, our grid is going to break”

  • GPS does not simply run on its own; it is actively operated from Colorado Springs by a US Air Force team, and if they stopped work, within about a week car GPS would not know the right street and within 2-3 months it would have the town wrong—showing how much constant adjustment is needed.

    “If they stop work, in about a week, your car GPS wouldn't know exactly what street it was on. And in 2 or 3 months it would get Town wrong”