"What is AI?" is a long-form feature by Will Douglas Heaven (Senior AI Editor, MIT Technology Review), published July 10, 2024 at technologyreview.com. Running roughly 12,000 words across eight named sections, it surveys the disagreement over how to define "AI" and the camps that hold competing definitions.
Author: Will Douglas Heaven (Senior AI Editor, MIT Technology Review) Published: July 10, 2024 URL: technologyreview.com Length: ~12,000 words; structured as a long-form feature in eight named sections.
Summary of argument
Heaven's premise is that there is no agreed definition of "AI," and that the disagreement is the subject rather than a side debate, because what people think AI is determines how it gets built, regulated, and adopted. He argues that AI is "an idea, a vision, a kind of wish fulfillment" — an ideal "shaped by worldviews and sci-fi tropes as much as by math and computer science." He characterizes the discourse as a polarized fandom war between AI acolytes (for whom AGI is in sight and superintelligence follows) and anti-hype critics (for whom it is smoke, mirrors, math, and harm).
The feature is organized around six anchor disputes, plus additional threads on interpretability, neural networks, and framing.
Key claims
Sparks of AGI vs. Stochastic Parrots
Heaven treats this as the central frame. Sébastien Bubeck and Microsoft Research published "Sparks of Artificial General Intelligence" (March 2023), reporting that an early GPT-4 spontaneously produced a verse-form proof of the infinity of primes and Latex code drawing a unicorn — examples Bubeck argues are "smoking guns of reasoning."
The opposing camp — Emily Bender (UW computational linguist), Alex Hanna (DAIR sociologist), and Gary Marcus — read the same evidence as marketing fluff. Bender calls the Sparks paper "a fan fiction novella." The methodology critique cites a non-public model checkpoint, no reproducibility, and reliance on hand-picked examples. The deeper critique holds that behavioral evidence cannot resolve whether the underlying process is reasoning or memorization.
The octopus thought experiment (Bender & Koller, 2020)
Two English speakers stranded on neighboring islands send messages over an underwater cable. An octopus that knows no English but is good at statistical pattern-matching wraps suckers around the cable, learns to predict next words, and eventually starts replying. Bender and Koller's claim is that a model trained only on text learns the form of language but not its meaning, because meaning consists of words plus the reasons they were uttered. The octopus cannot help the islander build a coconut catapult or respond to a bear attack, having no contact with the world the words refer to. Heaven presents this as the central thought experiment behind the "stochastic parrots" critique that LLMs are "mindless statistical tricks."
Stochastic Parrots (Bender, Gebru, McMillan-Major, Mitchell, 2021)
The other Bender paper, "On the Dangers of Stochastic Parrots," enumerates harms that Heaven says LLM developers ignore: compute cost and environmental impact, entrenched racism/sexism in outputs, and the danger of "a system that could fool people by haphazardly stitching together sequences of linguistic forms ... without any reference to meaning: a stochastic parrot." Internal Google conflict over the paper led to Timnit Gebru and Margaret Mitchell being forced out.
The TESCREAL bundle (Gebru and Torres)
Timnit Gebru and Émile Torres argue that understanding why frontier labs race to AGI and why doomers warn of catastrophe requires viewing the field through the TESCREAL framework: Transhumanism, Extropianism, Singularitarianism, Cosmism, Rationalism, Effective Altruism, and Longtermism. Common tenets they identify: an all-powerful technology (AGI/superintelligence) is within reach and inevitable; it could level up humanity; "if we don't build it, someone else will"; and, per Gebru/Torres, these worldviews share intellectual lineage with 20th-century eugenics movements. They argue the political consequence is that resources flow to fantastical futures (life extension, planetary colonization, AGI) while present-day costs (labor exploitation, racial bias, environmental damage) are dismissed.
The AGI naming history — McCarthy's "suitcase word"
The term "artificial intelligence" was coined by John McCarthy for a 1955 Dartmouth funding application. McCarthy's coauthors disliked it: Arthur Samuel thought "artificial" sounded "phony"; Claude Shannon preferred "automata studies"; Herbert Simon and Allen Newell called their work "complex information processing" for years. Marvin Minsky called AI a "suitcase word" capable of holding many divergent meanings. Cambridge historian Jonnie Penn lists alternatives that were in play: engineering psychology, applied epistemology, neural cybernetics, neuraldynamics, advanced automatic programming, and hypothetical automata. The Dartmouth proposal also encoded the symbolic vs. neural-network split that has divided the field since — what Cambridge/DeepMind researcher Harry Law calls "the core tension in AI."
The Turing test, Blockheads, and behaviorism
Turing's 1950 imitation game sidestepped defining intelligence by looking for "its shadow" in behavior. Ned Block's 1981 "Blockhead" counterexample imagines a vast lookup table of every possible question-answer pair: by Turing's test it would appear intelligent, but it has "the intelligence of a toaster." Block's argument is that whether behavior is intelligent depends on how it is produced, not how it appears. Heaven notes that the unresolved version of this problem now sits inside every benchmark used to score frontier models.
Interpretability, neural networks, and framing
Heaven discusses Anthropic's Golden Gate Bridge result, Chris Olah's May 2024 interpretability paper identifying neurons that fire on Golden Gate Bridge inputs; turning the feature up makes Claude obsessed with the bridge. He uses this as the most concrete current evidence that LLMs encode something more structured than pure pattern-matching, while noting Olah's caution: "It's a relatively limited picture, and the analysis is pretty hard."
He also reports Ellie Pavlick's "lookup table" finding (Brown University): an LLM learned the same encoding for France→Paris as for Poland→Warsaw, having discovered the abstract relation. Heaven quotes the open question: "But what does this show? Is encoding its own lookup table instead of using a hard-coded one a sign of intelligence?"
On neural networks, Heaven sets Hinton ("Deep learning is going to be able to do everything") against Marcus ("Neural network people have this hammer, and now everything is a nail"). Marcus's case draws on cognitive science: brains are not blank slates but have innate structure. Marcus still believes AGI is on the horizon — he just thinks today's neural-network fixation is wrong. Margaret Boden's framing holds that AI needs not just powerful computers but "powerful ideas — new theories of how thinking happens, new algorithms that might reproduce it. Maybe we will, maybe we won't."
Two cultural artifacts round out the piece. The Shoggoth meme — a Lovecraftian "tentacled monster wearing a smiley-face mask" — is shorthand for the gap between ChatGPT's friendly UI and the unfathomable model behind it. Mustafa Suleyman's "new digital species" framing (TED, April 2024) casts AI as "so universal, so powerful, that calling it a tool no longer captures what it could do for us"; Heaven flags this as the kind of framing that "makes my spidey sense tingle" given Microsoft AI's commercial interest. As a counter-example to the Turing-test-as-design-target tradition, Heaven cites Dor Skuler, whose company Intuition Robotics builds elder-care companion robots (ElliQ) that explicitly do not try to pass as human ("Why is it in the best interest of humanity for us to develop technology whose goal is to dupe us?").
Heaven's own position
Heaven states his view directly:
"AI is many things. But I don't think it's humanlike. I don't think it's the solution to all (or even most) of our problems. ... It's an idea, a vision, a kind of wish fulfillment."
He cites his 2020 GPT-3 line approvingly: "the greatest trick AI ever pulled was convincing the world it exists." He is sympathetic to Pavlick's wait-and-see position, that behavior is the only thing measured reliably and that theoretical commitment is loaded.
Provenance
The feature names several frameworks that recur in policy and academic discourse: TESCREAL as an analytic frame for frontier-lab ideology and the AGI-race dynamic; Stochastic Parrots and the octopus as the standard counter-frame to capability-extrapolation arguments; "Sparks vs. Parrots" as two-camp shorthand carried into policy discussion; and the "suitcase word" and the symbolic (GOFAI) vs. neural-network divide as the structural split running through roughly 70 years of AI history. Heaven is a senior journalist whose work recurs in the news cycle.
Relationships
- supports: Stochastic Parrots (Bender, Gebru, McMillan-Major, Mitchell, 2021) and the Octopus Test (new) — Bender/Koller octopus + Bender/Gebru et al. Stochastic Parrots paper
- supports: TESCREAL Framework (Gebru and Torres) (new) — primary expository source for the framework
- supports: AGI Timelines — surveys both the Sparks (capability-extrapolation) and Parrots (capability-skeptical) sides
- supports: Sparks of Artificial General Intelligence: Early experiments with GPT-4 — provides the journalistic context and counter-arguments
- related: Mechanistic Interpretability — Anthropic Golden Gate Bridge feature
- related: General-Purpose AI (GPAI) — definitional ambiguity at the heart of "general-purpose AI" categories in EU AI Act and elsewhere
- related: Emily M. Bender (new), Will Douglas Heaven (new), Timnit Gebru, Sébastien Bubeck, Gary Marcus, Chris Olah, Geoffrey Hinton, Mustafa Suleyman
- contradicts: Sparks of Artificial General Intelligence: Early experiments with GPT-4 (in part) — gives the strongest journalistic articulation of the methodological objections to the Sparks paper