Aaronson, Scott

professor at UT Austin, director of the Quantum Information Center, former MIT professor; his second appearance on the podcast

Scott Aaronson: Computational Complexity and Consciousness

Lex Fridman Podcast · Oct 12, 2020

• Scott Aaronson
• consciousness
• Integrated Information Theory
• Roger Penrose
• computational complexity
• P vs NP
• GPT-3
• quantum computing
• Gödel's incompleteness theorem
• cancel culture

Core ideas

Aaronson opens by dismantling the "simulation" question as largely unanswerable in principle (a perfect simulation is by definition indistinguishable from base reality) but treats the related, better-posed question of whether the universe is computable (satisfies the Church-Turing thesis) as a live and so-far-well-supported empirical hypothesis. He's skeptical of "theory of everything" projects like Eric Weinstein's and Stephen Wolfram's not because unification is a bad goal (he notes even leading string theorists like Ed Witten pursue exactly that) but because premature claims of having solved it draw deserved skepticism. The episode's most developed argument is Aaronson's demolition of Giulio Tononi's Integrated Information Theory (IIT) of consciousness: he distinguishes the genuine "hard problem" (why is there subjective experience at all) from what he calls the "pretty hard problem" (given a physical system, determine whether and how much it is conscious), and argues IIT fails even at the latter, more modest goal — its central quantity "phi" is asserted via unmotivated axioms with "nothing even resembling a derivation," and worse, it implies that a giant, functionally trivial grid of XOR gates could have far higher phi (and thus be "more conscious") than a human brain. He notes Tononi's response — biting the bullet and insisting this grid genuinely is superconscious — as remarkable in its boldness but disqualifying: a theory that can't get "obvious" cases right (a blank wall isn't more conscious than a person) isn't tracking what we mean by the word.

On Roger Penrose's consciousness theory, Aaronson carefully unpacks how much more radical it is than "quantum effects matter for the brain": Penrose requires the brain to exploit unknown, uncomputable laws of quantum gravity (going beyond even quantum computation, which is still simulable classically given enough time), requires an "objective collapse" mechanism modifying quantum mechanics itself (a testable claim that experimentalists are actively probing), requires consciousness to bias that collapse non-randomly, and rests on a Gödel's-incompleteness argument that Aaronson says the "overwhelming majority" of logicians and computer scientists reject — since nothing stops an AI from provisionally adopting new axioms and reasoning from a stronger vantage point, exactly as humans do, with no guarantee of infallibility either way (a rebuttal Alan Turing already made in 1950). On GPT-3, Aaronson calls it a genuine, surprising leap over predecessors like Eugene Goostman (which he'd publicly criticized as no real advance over 1960s-era ELIZA) — describing it via Scott Alexander's line that OpenAI "ground up the entire internet into a slurry" — while noting its strange profile of childlike failures (basic arithmetic, spatial/logical reasoning, balancing parentheses) alongside fluent, sometimes better-than-human prose, which he takes as evidence that predictive next-word processing alone is insufficient and that the brain likely couples that kind of predictive mechanism to other specialized reasoning subroutines.

Turning to complexity theory (his home field), Aaronson calls universality — the discovery that a handful of simple primitives (NAND gates, or Turing machine instructions) suffice to express any computation whatsoever — the most beautiful idea in computer science, both a "flattening" of the field (there's no infinite hierarchy of ever-more-powerful languages waiting to be discovered) and a liberation (anyone with a universal system has the same expressive power as the greatest programmer). He walks through the P/NP distinction using factoring as the running example (believed hard to solve, easy to verify) and states flatly he'd bet against P=NP (giving it maybe 2-3% odds), while carefully explaining that even a proof of P=NP wouldn't guarantee a practically usable algorithm — invoking Leonid Levin's darkly funny "universal search" argument that a P=NP witness could in principle be found via an impractical but formally polynomial brute-force enumeration of all algorithms. He surveys further complexity classes (PSPACE, #P, BPP, BQP, and his own favorite, SZK — statistical zero-knowledge, illustrated via a graph-non-isomorphism protocol where a party can be convinced something is true without learning why) as the "elements" of theoretical computer science.

The conversation closes with Aaronson's reflections on COVID-19 (comparing reality unfavorably to the movie Contagion, whose writers failed to imagine that institutional incompetence and conspiracy theorists would actually be running the response, rather than a "hyper-competent" government), his fear of post-election political violence and institutional breakdown, and an extended, personal defense of open discourse against "cancellation" dynamics — describing being called racist and sexist on social media for defending unpopular ideas and people, and arguing that cancellation culture is "spectacularly ineffective at its own professed goal" of fighting racism and sexism, while admitting (in contrast to Steven Pinker's apparent equanimity under similar attacks) that the pile-ons genuinely wear on him psychologically. He ends on love for his family as the one area where, as a theoretical computer scientist, he modestly declines to offer any special technical insight.

Excerpts