Brief

To take AGI on directly is not to forecast it but to interrogate it, because the debate everyone is having — three years or thirty, scaling or wall — is conducted as prediction and is almost never empirical. The disagreements are definitional: people mean incompatible things by general intelligence and argue past each other about a target no two of them share, and getting the concept honest before arguing about the object is the whole discipline of being true to things. The term equivocates across at least three conceptions that come apart under any pressure. There is the economic — can do most of the economically valuable work a human can, the operational definition that dodges the question by pricing it. There is the cognitive-general — matches or exceeds humans across the full range of cognitive tasks, including genuinely novel ones (Chollet’s skill-acquisition efficiency; Legg & Hutter’s universality). And there is the metaphysical — a mind: something that understands, means, is answerable, has a stake, is someone home. Progress on the first tells you nothing about the third, and the discourse’s central sleight is to launder one into the other — to let a rising number on the first quietly stand in for arrival at the third. The thesis this publication can state and few others can: general intelligence in the strong sense is the view from nowhere rebuilt in silicon — intelligence unsituated, unstaked, applicable to everything from no standpoint in particular. But the epistemology of The View From Somewhere denies anyone that seat: no knower has ever occupied the view from nowhere; all real intelligence is from somewhere — embodied, historically placed, trained to notice some things and miss others, answerable to a world it is in contact with. So the strong target may be incoherent exactly the way Nagel’s view from nowhere is incoherent, and the way Haraway’s “god-trick” is a trick: not hard to reach, but confused as posed. What is coherent — and may well arrive — is intelligence answerable across many somewheres: broader than ours, faster, wider in coverage, and still no god’s-eye generality, because generality of coverage is not generality of standpoint, and a net cast over all of human speech is not a self that shows up anywhere. This dissolves the forecasting question rather than answering it. On the economic definition the thing is arriving, and the live questions are answerability and consequence — which is where The Answer Machine takes over. On the metaphysical definition, scaling coverage does not approach a mind; it approaches a superb answer machine — more competence, not one more ounce of standpoint. And it lets the essay refuse the oldest confusion of all, Spearman’s g: intelligence is not a scalar substance you have more or less of but a correspondence between a knower and a world (again The View From Somewhere), so “how general” was a mismeasure before we ever built the machines. Sharper than the rivals, and sharper by refusing their shared premise: the doomer and the booster both fight on the capability axis and both inherit the confused target (Good’s intelligence explosion, Bostrom’s superintelligence), it’s-just-autocomplete loses because it denies aptness the machines plainly have, and stochastic parrots is a claim about mechanism where this is a claim about the concept. The register is foundations, not society — the conceptual companion to The Answer Machine, not a replacement: this essay clears the ground by getting general intelligence honest; that one stands on the cleared ground and shows what the thing is and does. The discipline that keeps it from dating is the fallibilism of The View From Somewhere: hold every timeline and capability claim as dateable and cheap to be wrong about, and anchor the weight on the conceptual argument, which is timeline-independent — the essay must survive AGI “arriving” by one definition while its claim about another still stands.

Outline

• Doorway. Two serious people arguing whether AGI is three years off or thirty — and discovering, when pressed, that they mean different things by the word. The argument was never about evidence. (Candidate: a real exchange — a dinner, a thread, a boardroom — where the disagreement dissolves into definitions the moment anyone defines the term.)
• Three definitions. Economic, cognitive-general, metaphysical. Show they come apart; show the sleight that launders the first into the third.
• The general is the view from nowhere. Name the target the strong conception actually describes — intelligence from no standpoint — and tie it to Nagel and to The View From Somewhere.
• No knower is from nowhere. Situated, embodied, answerable (Haraway’s god-trick; Dreyfus on the frame problem; Brandom on answerability). So the strong target is confused, not merely distant.
• The secular god. Why the fantasy persists: the god’s-eye view is the old omniscience secularized — objectivity long imagined as the view sub specie aeternitatis, the knowing mind of God with the God removed. AGI is the attempt to build that knower back, a re-enchantment inside a disenchanted world (tie to the disenchantment thread). Names the essay’s title on the page: we are not scaling a tool toward generality, we are trying to rebuild a god — and the god’s-eye view was always a trick, a standpoint no one has ever stood in.
• What’s coherent instead. Intelligence answerable across many somewheres. Coverage-generality ≠ standpoint-generality. Reframe the milestone people are counting toward.
• Refuse g. Intelligence as correspondence, not scalar substance (Spearman as foil). “How general” was a mismeasure from the start.
• Dissolve the forecast. On the economic definition it’s arriving — hand to The Answer Machine. On the metaphysical definition, scaling approaches a better answer machine, not a mind.
• Close. The thing we fear or hope for was never on the coverage axis. What would have to be true for a machine to know from somewhere — and why that is a different project than the one being scaled.

Source ownership: this essay owns the definitional strand — Spearman, Turing, Good, Legg & Hutter, Chollet. Nagel and Haraway are owned by The View From Somewhere and read here as the same target rebuilt in silicon. Brandom is owned by The Answer Machine.

Reading List

• Spearman, “‘General Intelligence,’ Objectively Determined and Measured” (1904)
• Turing, “Computing Machinery and Intelligence” (1950)
• Good, “Speculations Concerning the First Ultraintelligent Machine” (1965)
• Legg & Hutter, “Universal Intelligence: A Definition of Machine Intelligence” (2007)
• Chollet, “On the Measure of Intelligence” (2019)
• Nagel, The View from Nowhere (1986)
• Haraway, “Situated Knowledges” (1988)
• Searle, “Minds, Brains, and Programs” (1980)
• Dreyfus, What Computers Still Can’t Do (1992)
• Brandom, Making It Explicit (1994)
• Chalmers, “The Singularity: A Philosophical Analysis” (2010)
• Bostrom, Superintelligence: Paths, Dangers, Strategies (2014)