Wolfram, Stephen

computer scientist, mathematician, theoretical physicist; founder of Wolfram Research, creator of Mathematica and Wolfram|Alpha; fourth appearance

Stephen Wolfram: ChatGPT and the Nature of Truth, Reality & Computation

Lex Fridman Podcast · May 9, 2023

• Stephen Wolfram
• ChatGPT and large language models
• Wolfram Language and Wolfram Alpha
• computational irreducibility
• the observer and computational boundedness
• semantic grammar and laws of thought
• second law of thermodynamics
• the ruliad and physics project
• cellular automata and Rule 30
• AI risk and apex intelligence

Core ideas

Wolfram opens by drawing a sharp distinction between what ChatGPT does and what his own computational language project does: ChatGPT is "wide and shallow," continuing a prompt based on statistical regularities learned from a trillion words of human-written text via a comparatively shallow neural-net computation, whereas Wolfram|Alpha and the Wolfram Language aim to be "deep," representing knowledge symbolically so arbitrarily long chains of formal computation can derive genuinely new answers rather than just reproducing patterns already present in training data. He traces this symbolic representation approach back to his 1979 design decisions building SMP (predecessor to Mathematica) — representing everything as symbolic expressions (a function applied to arguments, not necessarily evaluated) — which turned out, unexpectedly, to be both a good match for how humans conceptualize the world and, decades later, unusually easy for AI systems to parse and generate, since the language was built to be coherent and human-readable rather than purely machine-executed.

Central to his worldview is computational irreducibility: for many systems, the only way to know what a set of rules will produce after many steps is to actually run them — there's no shortcut. He connects this to his major physics-project discovery that science and cognition are only possible because there exist "pockets of reducibility" within a universe that is otherwise irreducible — regularities (like gas pressure, or Newton's laws) that let bounded observers like humans compress overwhelming microscopic detail into simple, usable narratives. He extends this into a broader theory of "the observer": a computationally bounded observer is one that (1) equates many different microscopic configurations into single coarse-grained aggregate descriptions (the way pressure ignores which specific molecules hit a piston) and (2) insists on believing itself persistent through time (a single continuous thread of experience) even though, in his physics model, both the universe and the observer's own substrate are constantly being "rewritten." He argues this same interplay between computational irreducibility and the computational boundedness of observers is what derives all three pillars of 20th-century physics — general relativity, quantum mechanics, and the second law of thermodynamics (statistical mechanics) — as necessary consequences of observers with our specific characteristics, rather than arbitrary features the universe happens to have. He recounts his 50-year personal history with the second law, from a childhood attempt (age ~12) to reproduce a textbook diagram of gas molecules dispersing, through his 1980s cellular-automata work (including Rule 30, discovered in 1981, which produces apparently random behavior from an utterly simple rule and remains formally unproven as truly random), to his recent reformulation of entropy increase as the story of a computationally bounded observer failing to track a computationally irreducible molecular system.

On why ChatGPT specifically works, Wolfram's central claim is that GPT has empirically discovered a "semantic grammar" of language — a layer of regularity beyond syntactic (part-of-speech) grammar that governs which sentences are meaningful, not just grammatical (contrasting "the happy electron" — syntactically fine, semantically empty). He frames this as a direct sequel to Aristotle's discovery of syllogistic logic (patterns abstracted from listening to orators, independent of the specific content) and George Boole's 1854 generalization of it into Boolean algebra ("The Laws of Thought"), arguing GPT is discovering further, still only partially charted "laws of thought" — and that neural networks turn out to generalize the way humans do (making the same category judgments, e.g., about handwritten digits) largely because their architecture, dating to 1943, was loosely modeled on the brain in the first place. He's candid about GPT's failure modes stemming directly from its word-by-word generation process: it can talk itself into a wrong answer via a bad "chain of thought" yet immediately recognize the error once the whole output is fed back to it, and he shares two personal anecdotes of being fooled by highly plausible-sounding but factually wrong outputs (a persuasive but fabricated essay claiming wolves are blue, and a "HAL 9000 death song" that turned out to be "Mary Had a Little Lamma"). He argues the reinforcement-learning-from-human-feedback step is what pushed language models past some threshold into being genuinely useful and human-aligned, a threshold he says nobody, including people inside the field, predicted would be crossed as early as 2022.

Wolfram frames the practical integration of ChatGPT with Wolfram|Alpha/Wolfram Language as solving a truth and precision problem: natural language is a fuzzy, socially-defined "transport layer" between minds, good for communicating loosely defined intent, but computational language is a precise intermediate representation you can actually test, verify, and compute consequences from — the LLM's job is translating fuzzy human intent into that precise form, at which point the deep, deterministic computation happens in Wolfram Language rather than being hallucinated by the LLM itself. He extends this into a vision of radically democratized access to deep computation (comparing it to how Mathematica removed the need to "delegate to a programmer" back in 1988) and predicts this reshapes education: rather than everyone learning the "trade" of programming syntax (which he expects to become obsolete like Assembly language), what becomes essential is a broader literacy he calls "computational X" or "CX" — the general skill of formalizing any domain (color as three numbers, sound as a spectrogram, preferences as rankings) into a form computation can act on — and he's currently trying to write a textbook defining that curriculum, guessing roughly a year of coursework would suffice.

On AI risk, Wolfram positions himself as an intentional optimist relative to people like Eliezer Yudkowsky, comparing doomsday "eventually a smarter AI will emerge and destroy us" arguments to ontological arguments for the existence of God — logically clean but resting on an oversimplified model — and argues there is no "apex intelligence" even in principle (just as there's no longest-running halting Turing machine), because the space of possible computation/rule-space ("the ruliad") is infinite. He does take seriously narrower, nearer-term risks: an AI-generated exploit for manipulating Congress members that he says he "did not think about... before" and generated almost instantly while discussing it live; the fundamental impossibility of a perfect sandbox once a system is a universal computer (since computational irreducibility guarantees exploitable unexpected corners); and the risk that increasingly persuasive, personalized AI "auto-suggestion" (drafting your replies, nudging your choices) gradually removes humans from the decision loop even without any dramatic takeover. He also discusses the philosophical question of whether AI or ordinary computers already have some form of experience/consciousness, framing consciousness not as the pinnacle of possible computation but as a specialization tied specifically to the "single thread of experience" and boundedness that human-like observers have — and reflects, prompted by recently getting his own full-body MRI scan, on the strange disjunction between seeing his brain as mere physical structure and still subjectively "feeling the way he feels."

Excerpts