What this covers

Melanie Mitchell, a computer scientist and complexity researcher, joins Sean Carroll to examine why artificial intelligence excels at narrow, abstract games while stumbling at tasks humans perform unconsciously. The conversation traces the history of neural networks from perceptrons in the 1950s through the deep learning revolution, explaining how hidden layers and backpropagation enabled machines to build internal representations of data. Mitchell's core claim is that this success in well-defined domains masks a fundamental gap: modern AI lacks Common Sense — intuitive physics, object permanence, causal reasoning, and the ability to recognize sameness across different situations — capabilities that humans acquire innately or very early in life.

The discussion moves through concrete failures that reveal this gap. A system trained to play Breakout fails completely when the paddle moves up two pixels because it never abstracted the concept of a "paddle" as a persistent object. Deep neural networks cannot reliably recognize whether two things are the same without explicit training, yet humans do this effortlessly. Mitchell traces how big data training exposes machines to statistical quirks rather than robust concepts—fire trucks upside down fool recognition systems, while biased training data produces biased outputs in face recognition and self-driving cars. She argues that Moravec's Paradox (Easy Things Are Hard)) captures the inversion: what humans find trivial proves computationally intractable, while what we struggle with—chess, Go—turns out amenable to brute force. The pair discuss whether Embodied Cognition or built-in primitives might be necessary, a position contested among deep learning researchers, and examine DARPA's effort to build systems that replicate infant development.

Sharpest takeaway

Mitchell argues that modern AI's dramatic success at narrow, well-defined tasks like chess and Go masks its fundamental inability to acquire 'common sense' — the intuitive physics, causality, object permanence, and analogical reasoning humans take for granted — so claims of imminent human-level AI are overhyped.

  • Deep learning excels at narrow tasks but is brittle, easily fooled, and fails when context shifts because it lacks abstract concepts and causal models
  • Common sense, intuitive physics, and analogy may require built-in primitives or embodiment that pure data-driven learning cannot provide
  • 'Easy things are hard' — tasks humans do unconsciously are the ones machines struggle with most

The argument · threads4 threads · 39 claims
0.71

AI's game-playing victories are less remarkable than they appear because domains like chess and Go suit brute-force computation, whereas human success at them despite limited mental resources is genuinely impressive.

2 pointscentrality 4/5
  • Chess and Go are the most obvious places where computers should beat humans because they suit brute-force number crunching; the genuinely impressive thing is that humans with limited mental computation are good at them, so computer victories at games are less remarkable than they appear.

    “I've never been impressed with the fact that chess and go are regimes in which computers can beat us. I would think that those are the most obvious places where computers should be able to beat us. The impressive thing to me is that human beings, with our very limited number crunching capabilities in our head, are pretty good at chess and go.”

  • AI's most impressive successes are in well-defined finite domains like chess and Go where the rules are clear and there is no equivalent of an upside-down fire truck, but it is genuinely unclear whether these techniques will carry over to the messy real world.

    “this is one of the reasons why some of the most impressive successes of AI programs have been in very well-defined, finite situations like games... It's clear what the rules are. There's no equivalent of an upside down fire truck in a chess game.”

0.70

Current AI pattern-completion lacks the common-sense ability to privilege plausible patterns as humans and the world do, making it unclear whether techniques will transfer from finite games to messy real-world problems.

1 pointcentrality 5/5
  • Although mathematically any data sequence can be fit by infinitely many rules (the Duhem-Quine thesis), pattern-completion tasks like guessing the next number in a sequence are really science questions, not math questions — because not all patterns are equally plausible, the world or our minds privilege certain patterns as the right ones, which is exactly the common-sense ability computers lack.

    “Any set of data can be fit by an infinite number of theories... my response to him was that they weren't actually math questions, they were science questions. They were looking for patterns in the world, and not all patterns are created equal.”

0.56

Exponential-curve fitting tells us very little about actual AI trajectory since exponential-looking curves in nature almost always flatten or decay rather than continue climbing.

1 pointcentrality 4/5
  • Ray Kurzweil argues AI progress is on an exponential curve that will yield human-level intelligence within 10 years and superintelligence beyond, but exponential-looking curves in nature almost always flatten or decay, so curve-fitting tells us very little about the actual future trajectory of AI.

    “I think there are almost no exponential curves in nature, really exponential. The expansion of the universe is one of them, but most curves look exponential at the beginning and then flatten off, or decay away. So I don't think that the curve fitting tells us very much.”

0.44

As AI decision-making grows more confident and pervasive, predicting loan defaults and denying credit preemptively raises pre-crime concerns similar to those in Minority Report.

2 pointscentrality 2/5
  • We may overestimate how complicated people are; many individuals have a relatively limited set of things they would say or do in given circumstances, evidenced by how often Gmail's predictive suggestions correctly guess what someone is about to write, though the residual unpredictable fraction may matter a great deal.

    “Is it conceivable that many people actually have a relatively limited set of things they would say or do in any certain set of circumstances and that might be mimic-able in the foreseeable future?”

  • Using AI to predict who will default on a loan and denying it before any default has occurred is logically continuous with predicting who will commit a crime, raising a Minority Report-style pre-crime concern as AI decision-making grows more confident and pervasive.

    “what if we become really good at deciding who will default on a loan and therefore, we don't give them a loan even if they haven't defaulted yet... what is the difference between that and saying, 'Well you'll probably commit a burglary.'”