---
title: "Artificial General Intelligence Is Already Here"
description: "Blaise Agüera y Arcas and Peter Norvig argued in this October 2023 Noema essay that the most important part of artificial general intelligence had already been achieved by the frontier language models of that year: ChatGPT, Bard, LLaMA, and Claude."
type: "reading-notes"
authors:
  - "Blaise Agüera y Arcas"
  - "Peter Norvig"
published: "2023-10-10"
original_url: "https://www.noemamag.com/artificial-general-intelligence-is-already-here/"
canonical_url: "https://antikythera.wiki/work/media/noema-artificial-general-intelligence-is-already-here"
md_url: "https://antikythera.wiki/md/work/media/noema-artificial-general-intelligence-is-already-here"
last_updated: "2026-08-23"
site: "Antikythera Wiki"
---

# Artificial General Intelligence Is Already Here

> Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below.

- **Authors:** Blaise Agüera y Arcas, Peter Norvig
- **Published:** 2023-10-10
- **Kind:** Essays, talks & interviews · Essay
- **Venue:** Noema Magazine
- **Original:** https://www.noemamag.com/artificial-general-intelligence-is-already-here/
- **Length:** 3.6k words
- **This page:** https://antikythera.wiki/work/media/noema-artificial-general-intelligence-is-already-here
- **Notes generated:** 2026-08-23

## Summary

[Blaise Agüera y Arcas](https://antikythera.wiki/people/blaise-aguera-y-arcas) and [Peter Norvig](https://antikythera.wiki/people/peter-norvig) argued in this October 2023 Noema essay that the most important part of artificial general intelligence had already been achieved by the frontier language models of that year: ChatGPT, Bard, LLaMA, and Claude. The models are flawed, but generality is a separate property from reliability or superhuman performance, and by the everyday meanings of "general" and "intelligence" they have it. The essay then asks why so few commentators will say so, and answers with four reasons: doubts about metrics, commitment to rival theories of intelligence, human or biological exceptionalism, and anxiety about economic consequences. Its most consequential proposals are that consciousness and sentience be separated from intelligence altogether, and that debate about what AGI ought to be not substitute for the fact of what it is. It is the public version of the generality argument in Agüera y Arcas's What Is Intelligence?

## The argument

The essay's governing analogy is ENIAC. The 1945 machine was slow, unreliable, and hard to use, but it could run sequential, looping, and conditional instructions, and that property made it the first general-purpose computer. Today's frontier models stand in the same relation to whatever follows them: the key property, generality, is present now, and decades hence they will be recognised as the first true AGI.

Generality is defined by contrast with narrow AI. MYCIN diagnosed bacterial infections, SYSTRAN translated, Deep Blue played chess; even AlphaGo operated within tasks its engineers specified. Frontier models differ along five dimensions: topics, tasks, modalities, languages, and instructability. The last is decisive. In-context learning extends what a model can do from anything in its training data to anything that can be described, so a general model performs tasks its designers never envisioned. The authors call this a multidimensional scorecard rather than a threshold, yet hold that the discontinuity between narrow and general is real.

Having asserted the fact, they turn to its denial. Scepticism about metrics is partly justified: models trained to pass bar exams are tuned to those exams, and fluent prose invites the "Chauncey Gardiner effect," in which a grammatical answer is assumed to come from an intelligent source. But the authors also use a finding of Schaeffer, Miranda, and Koyejo to dismantle a fear: apparent sudden "emergence" of abilities like arithmetic disappears when metrics give partial credit. Intelligence scales continuously, "more is more" rather than "more is different," and the same may be true of the supposed gap between humans and other apes.

The second reason is theoretical loyalty. Symbolic AI, from Leibniz through Newell and Simon's physical symbol system hypothesis, was brittle because its terms were only approximately defined and few logical inferences are universally valid. Chomsky and Gary Marcus nonetheless maintain that systems without explicit symbols cannot truly understand. The authors ask whether each criticism is prescriptive (AGI must be built this way) or empirical (I doubt it can work this way). Prescriptive criticism should yield to the test; empirical criticism is defeated whenever a system passes a well-constructed one, and neural networks can learn symbols, approximate any computable function, and emulate any program.

The third reason is exceptionalism. Some argue AGI requires consciousness or agency. The authors grant that a frontier model prompted to run an online business, as in Mustafa Suleyman's "modern Turing test," is not much like a screwdriver, and conclude that either such models are conscious or agency does not entail consciousness. Since consciousness cannot be measured, verified, or falsified in a nonbiological system, and asking the model is a Rorschach test that the model itself may answer either way depending on its tuning, the sensible course is to separate intelligence from consciousness and sentience.

The fourth reason is economic. Arguments about intelligence shade into arguments about status and class. Had AGI arrived on the Dartmouth schedule in 1956, during the Great Compression, it would have met optimism; today the redistributive pump runs in reverse, and claims that AI is "neither artificial nor intelligent" read as economic threat. The authors invoke Hume's is–ought distinction and ask that the "ought" questions about benefit, harm, and fairness be discussed directly rather than through denial of the fact.

## Section by section

### Opening and What Is General Intelligence?

The ENIAC comparison, the catalogue of narrow systems, the five dimensions of generality, and the argument that in-context learning is the meta-task that makes the difference. The section ends by posing the question of reluctance and listing the four reasons.

### Metrics

Suleyman's million-dollar test and the objection to equating "capable" with "capitalist"; exam contamination and Goodhart's law; the Chauncey Gardiner effect from "Being There"; and the nonlinearity argument that dissolves emergence into gradual improvement.

### Alternative Theories

The history and failure of symbolic AI, the Chomsky and Marcus quotations, four reasons neural networks can do whatever symbolic systems can, the prescriptive–empirical distinction, and a closing jab: critics devising tests that models still fail are doing useful work, but should wait a few weeks before calling AI hype.

### Human (or Biological) Exceptionalism

The screwdriver argument and its breakdown for models that prompt themselves; the Rorschach character of asking a model whether it is conscious, with ChatGPT and Bard trained to say no; Nagel's bat; and the recommendation to keep intelligence and consciousness apart.

### Economic Implications

Programming's rise in status once it ceased to be women's work, the irony that calculus is easy for symbolic AI while manual labour defeats current systems, the Dartmouth counterfactual, Hume, and the questions the authors think 2023 should be asking: who benefits, who is harmed, and how to do this fairly.

## Key concepts

- **Artificial general intelligence** — The capacity to perform competently at nearly any information task a human can do that can be posed in natural language and scored; a multidimensional property already present in frontier models, distinct from superintelligence.
- **ENIAC moment** — The point at which a technology acquires its defining general property while remaining crude in every other respect. [ENIAC Moment](https://antikythera.wiki/terms/eniac-moment)
- **Chauncey Gardiner effect** — The assumption that a fluent, grammatical response comes from an intelligent entity.
- **Human exceptionalism** — Resistance to evidence of machine intelligence in order to preserve something special about the human spirit, compared to resistance to heliocentrism; the corpus later names the same reflex [Humanity-of-the-Gaps](https://antikythera.wiki/terms/humanity-of-the-gaps).

## Connections

[What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence) cites this essay in its introduction, where Agüera y Arcas restates its core methodological rule: a claim that something passes every test yet is "really" not intelligent leaves science for faith. The book's chapter on generality and its LaMDA material expand the essay's case in agreement with it. [AI Is Evolving — And Changing Our Understanding of Intelligence](https://antikythera.wiki/work/media/noema-ai-is-evolving-and-changing-our-understanding-of-intelligence) is the later Noema piece in which Agüera y Arcas restates the "general intelligence" section for a wider evolutionary argument.

[The Model Is the Message](https://antikythera.wiki/work/media/noema-the-model-is-the-message), Agüera y Arcas's earlier essay with Bratton, shares the refusal to settle sentience by asking the model and the demand for a vocabulary that separates intelligence from consciousness; this essay is blunter, treating the consciousness debate as currently unresolvable and therefore irrelevant to AGI. Bratton's [The Five Stages of AI Grief](https://antikythera.wiki/work/media/noema-the-five-stages-of-ai-grief) treats the denial this essay diagnoses as the first of five Western responses to AI. [After Alignment](https://antikythera.wiki/work/journal/afteralignment) makes the same move of relocating impasses over machine mind from the machines to our language, but also rewrites the acronym as "artificial generic intelligence," a deliberate distance from the human-equivalence benchmark this essay accepts. The [Agentworld (Research Brief)](https://antikythera.wiki/work/book/agentworld-brief) carries the exceptionalism argument forward under the name humanity-of-the-gaps. [How to Think Unlike Humans?](https://antikythera.wiki/work/journal/unlikehumans) agrees on the critique of exceptionalism but is less interested than this essay in test scores as the measure of intelligence.


## Terms used

- [ENIAC Moment](https://antikythera.wiki/terms/eniac-moment)