---
title: "AI Is Evolving — And Changing Our Understanding Of Intelligence"
description: "Blaise Agüera y Arcas and James Manyika, writing in Noema in April 2025, set out five \"paradigm shifts\" that they say the success of modern AI forces on anyone thinking about intelligence:…"
type: "reading-notes"
authors:
  - "Blaise Agüera y Arcas"
  - "James Manyika"
published: "2025-04-08"
original_url: "https://www.noemamag.com/ai-is-evolving-and-changing-our-understanding-of-intelligence/"
canonical_url: "https://antikythera.wiki/work/media/noema-ai-is-evolving-and-changing-our-understanding-of-intelligence"
md_url: "https://antikythera.wiki/md/work/media/noema-ai-is-evolving-and-changing-our-understanding-of-intelligence"
last_updated: "2026-08-23"
site: "Antikythera Wiki"
---

# AI Is Evolving — And Changing Our Understanding Of Intelligence

> 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, James Manyika
- **Published:** 2025-04-08
- **Kind:** Essays, talks & interviews · Essay
- **Venue:** Noema Magazine
- **Original:** https://www.noemamag.com/ai-is-evolving-and-changing-our-understanding-of-intelligence/
- **Length:** 7.9k words
- **This page:** https://antikythera.wiki/work/media/noema-ai-is-evolving-and-changing-our-understanding-of-intelligence
- **Notes generated:** 2026-08-23

## Summary

[Blaise Agüera y Arcas](https://antikythera.wiki/people/blaise-aguera-y-arcas) and [James Manyika](https://antikythera.wiki/people/james-manyika), writing in Noema in April 2025, set out five "paradigm shifts" that they say the success of modern AI forces on anyone thinking about intelligence: computation is a natural phenomenon older than computers; brains are computers of an unfamiliar design that chips will soon copy; intelligence is prediction; general intelligence has already arrived and the goalposts keep moving; and intelligence at every scale is social. The essay is a general-audience précis of the argument Agüera y Arcas makes at length in What Is Intelligence?, published by Antikythera and [MIT Press](https://antikythera.wiki/people/mit-press) later that year, and it reports early results from his Paradigms of Intelligence team at Google. Its framing claim is that the hardest adjustment is not technical but Copernican: accepting that general, nonhuman intelligence may be commonplace.

## The argument

The authors compare the present to the collapse of geocentrism. New ideas are easy to adopt when they fit an existing worldview and painful when they do not, and the idea that intelligence is neither rare nor specifically human does not fit. Borrowing Bratton's phrase, they call the reaction an existential trauma and name its object "intelligence geocentrism." The five shifts dismantle it from the bottom up, starting with computation, which they take intelligence to be made of.

The first shift relocates computation into nature. Turing defined it abstractly in 1936; von Neumann showed in 1951 that any self-reproducing organism needs an instruction tape plus a machine to read it, which is the specification of a universal Turing machine, two years before DNA's structure was found. Life is computational by necessity, because growth, healing, and reproduction are what make a living thing more "dynamically stable" than inert matter. The authors' experiment, in which random 64-byte strings interpreted as Brainfuck programs begin to self-replicate after a few million random interactions, is offered as evidence that the transition from non-life to life is a phase change computation undergoes on its own.

The second shift is architectural. Unreliable vacuum tubes pushed early design toward binary logic, a single CPU, and sequential instruction-fetching, and Moore's Law made that design fast without making it brain-like. The brain is slow but runs 86 billion neurons in parallel on locally stored data, which is why it is efficient. The authors predict chips holding parameters rather than programs, able to be flashed but also to learn on the fly.

The third shift reads the large language model as a finding about minds. Most researchers expected general intelligence to require a special algorithm; next-token prediction at scale turned out to be enough. The predictive brain hypothesis explains why: nervous systems evolved to model the future of the environment, the body, and the self. The training-versus-inference split and fixed parameter counts are described as transitional; future models should develop, learn cumulatively, and unify planning with prediction.

The fourth shift concerns thresholds. Skeptics say LLMs fake intelligence; boosters say AGI is imminent; the authors reply that there is no threshold, or that it is already behind us. They argue functionally, as with wings on birds and planes, and treat objections about implementation as irrelevant to capability even where they matter for ethics. Their sharper observation is that the benchmark has quietly shifted from any human to every human: no individual has an LLM's breadth, so individuals are now less general than models.

The fifth shift grounds the rest in sociality. Theory of mind enabled cooperation at scale, which tipped humans into a superorganism; training on collective human output therefore already yields a superintelligence of breadth if not depth. The same pattern recurs inside the brain, where cortical columns are general learners that specialize and model one another, and inside models, where Mixture of Experts and emergent modularity show scale achieved by division of labor. Intelligence is a "social fractal." The conclusion the authors draw is that AI should be built as growing networks of continually learning, specializing agents rather than frozen monolithic models.

## Section by section

### Opening and the five shifts

The introduction lists the five shifts with a practical consequence for each, and says the greatest trauma will be how commonplace nonhuman intelligence turns out to be.

### Natural Computation

The section traces computation from ENIAC back to Turing's 1936 machine and outward into nature through Stepney's work on computation without a user, Wheeler's "it from bit," quantum computing, and above all von Neumann's universal constructor. The Brainfuck soup experiment is described as minimal artificial life emerging from randomness, and the authors insist nothing supernatural is involved. Multilevel selection and the Major Evolutionary Transitions close the section.

### Neural Computing

The authors explain why the classical computer diverged from the brain, recount GOFAI's failure and the reassuring narrative that brains are not computers, and credit the connectionist holdouts, Hinton, Hopfield, Rosenblatt, McClelland, and Fukushima. An aside on the word "model" argues that since the brain computes a model of the world, a full-size model of a brain would be the real thing, able to model us back. GPUs and TPUs are a partial correction; truly neural silicon is predicted within years.

### Predictive Intelligence

The authors record their own surprise that next-token prediction sufficed and their retrospective view that it should not have surprised anyone who accepted that brains compute. LLMs are said not to be black boxes; "artificial neuroscience" can record and ablate any part of them. Prediction is traced through the motor loop of grasping a cup and thirst as a species-level prediction. They then list expected changes: self-constructing architectures, persistent test-time improvement, vision-language-action models in robotics, and chain-of-thought as a tree of simulated futures they tentatively link to free will.

### General Intelligence

Functionalism answers the skeptics: cochlear implants, anaerobic respiration, and trucks versus railroads show function served by different mechanisms, and there is no homunculus that makes a person irreplaceable. LLMs lack bodies, life stories, and kinship, which matters for legal status but not capability. Economic definitions of AGI, OpenAI's "most economically valuable work" and Suleyman's million-dollar test, are rejected as arbitrary. A researcher from 2002 confronted with today's models would, they claim, declare AGI achieved. Unsupervised pretraining is named as the source of generality.

### Collective Intelligence

The longest section moves from primate brain-size and group-size correlations to the claim that humanity is a superorganism. Humanity's Last Exam, written by nearly a thousand experts, shows median human performance near zero while models score 3 to 19 percent. Minsky's Society of Mind and the visual word form area support the "social cortex" picture; scaling laws and emergent modularity extend it to AI. The authors concede that open-ended learning alarms the safety community, note that in-context learning already makes today's models mesa-optimizers, and cite Infini-attention and long-term memory as steps toward persistence. Self-models and consciousness are derived from the need to model other selves, and LLMs are reported to pass theory-of-mind tests at human level.

### Beyond AI Development As Usual

A short coda calls for "knight moves" into adjacent fields and restates the aim: intelligence that helps humans understand themselves as ecologies of smaller intelligences and constituents of larger wholes.

## Key concepts

- **Natural computation** — the claim that computing exists in physics and biology before any engineered computer, with von Neumann's self-reproducing automaton as the clearest case. See [Computation](https://antikythera.wiki/terms/computation).
- **Dynamic stability** — the property by which anything that heals or reproduces persists where inert matter degrades; the authors' explanation for why life is computational. See [Dynamic Stability](https://antikythera.wiki/terms/dynamic-stability).
- **Major Evolutionary Transition** — a functional symbiosis in which independent entities become interdependent parts of a larger whole. See [Major Evolutionary Transition](https://antikythera.wiki/terms/major-evolutionary-transition).
- **Predictive intelligence** — intelligence as statistical modeling of the future, including one's own actions, on the evidence of the past. See [Predictive Intelligence](https://antikythera.wiki/terms/predictive-intelligence).
- **Functionalism** — judging intelligence by what a system does rather than how it is built. See [Functionalism](https://antikythera.wiki/terms/functionalism).
- **Anyone to everyone** — the shift in how AI is benchmarked, from any single human to all human experts combined.
- **Superorganism** — humanity as an interdependent whole whose collective intelligence is already superhuman. See [Superorganism](https://antikythera.wiki/terms/superorganism).
- **Social fractal** — intelligence organized as communities of agents at every scale, from neurons modeling neighbors to societies of specialists. See [Social Fractal](https://antikythera.wiki/terms/social-fractal).
- **Theory of mind** — modeling what others see, know, and intend; the proposed origin of selfhood and a capacity the authors report LLMs possess. See [Theory of Mind](https://antikythera.wiki/terms/theory-of-mind).

## Connections

- [What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence) is the full statement of every thread here; the essay compresses the book and adds Manyika's policy voice, in agreement throughout.
- Its first chapter, printed separately as What Is Life? Evolution as Computation, develops the Natural Computation section at length.
- [Artificial General Intelligence Is Already Here](https://antikythera.wiki/work/media/noema-artificial-general-intelligence-is-already-here), by Agüera y Arcas and [Peter Norvig](https://antikythera.wiki/people/peter-norvig), is the argument the General Intelligence section summarizes.
- [What Is Intelligence?](https://antikythera.wiki/work/journal/wii-film), the Antikythera article-film, covers the same ground visually.
- [The Five Stages Of AI Grief](https://antikythera.wiki/work/media/noema-the-five-stages-of-ai-grief) supplies the "Copernican trauma" framing borrowed in the first paragraph; the two essays agree that denial of machine intelligence is a reaction to decentering, and Bratton in turn cites Agüera y Arcas's "Marcus Loop."
- [Mutual Prediction in Human-AI Coevolution](https://antikythera.wiki/work/journal/coginfra/mutual-prediction-in-human-ai-coevolution) takes the same predictive-processing premise and asks what happens when humans and models predict each other; an extension rather than a dispute.
- [Modes of Cognition](https://antikythera.wiki/work/journal/modesofcognition) and [Substrates Unbound](https://antikythera.wiki/work/journal/substrates) share the substrate-indifferent view of intelligence but are more cautious about treating LLM success as a settled theory of mind.
- [After Alignment](https://antikythera.wiki/work/journal/afteralignment) is in tension with the essay's mild treatment of mesa-optimization and its reliance on human output as the route to superintelligence; Bratton argues that reflecting human values is the wrong orientation for AI, a question the authors largely set aside.


## Terms used

- [ENIAC Moment](https://antikythera.wiki/terms/eniac-moment)
- [Predictive Intelligence](https://antikythera.wiki/terms/predictive-intelligence)
- [Social Fractal](https://antikythera.wiki/terms/social-fractal)