---
title: "What is Intelligence? (Lecture)"
description: "An article-film is the Antikythera Journal's format for a lecture: a filmed talk released as a journal article in its own right, with a DOI, the journal's design treatment, and a full transcript."
type: "reading-notes"
authors:
  - "Blaise Agüera y Arcas"
published: "2025-05-10"
original_url: "https://wii-film.antikythera.org/"
doi: "10.1162/ANTI.5CZC"
canonical_url: "https://antikythera.wiki/work/journal/wii-film"
md_url: "https://antikythera.wiki/md/work/journal/wii-film"
last_updated: "2026-08-23"
site: "Antikythera Wiki"
---

# What is Intelligence? (Lecture)

> 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
- **Published:** 2025-05-10
- **Kind:** Journal · Journal article
- **DOI:** https://doi.org/10.1162/ANTI.5CZC
- **Original:** https://wii-film.antikythera.org/
- **Length:** 15k words
- **This page:** https://antikythera.wiki/work/journal/wii-film
- **Notes generated:** 2026-08-23

## Summary

An article-film is the Antikythera Journal's format for a lecture: a filmed talk released as a journal article in its own right, with a DOI, the journal's design treatment, and a full transcript. This one records [Blaise Agüera y Arcas](https://antikythera.wiki/people/blaise-aguera-y-arcas)'s lecture previewing [What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence), delivered before the book was out and presented as the first collaborative project of Antikythera and the [MIT Press](https://antikythera.wiki/people/mit-press). The lecture's claim is that intelligence is the computation of one function, the probability of the future given the past, and that this same function is what life is. Brains compute it, bacteria compute it, and large language models compute it. From that identity Agüera y Arcas derives a chain of consequences: evolution is a learning algorithm, computation is an attractor because reproduction requires it, complexity grows by symbiosis, brains and free will and consciousness are products of mutual prediction, and present-day AI is a genuine major evolutionary transition rather than a counterfeit of intelligence. These notes cover what the film says; the book develops the same material at greater length and with more evidence.

## The argument

Agüera y Arcas begins from a practical surprise. His team at Google built Gboard, the phone keyboard that predicts the next word from the statistics of previous words. Everyone expected a ceiling, because next-word prediction is unboundedly hard: completing "after Ballmer's retirement, the company elevated ___" requires world knowledge, a sentence about Kilimanjaro in stacked pennies requires arithmetic, and a sentence about a grieving friend requires a model of her mind. Such completions were understood to be "AI-complete." The shock was that the same kind of model, made very large and trained on very large corpora, got them right. Scale was the only added ingredient.

This should not have been a shock, he argues, because computational neuroscience had long proposed that the brain does exactly this. Karl Friston's predictive brain, with antecedents back to Helmholtz, treats the cortex as a predictor of what comes next. The question is whether a system that predicts well is counterfeit intelligence or the real thing. Agüera y Arcas answers that it is the real thing, on Turing's functionalist grounds: intelligence is the function computed, input to output, not the material that computes it. He adds an asymmetry argument of his own. A student who knows mathematics can feign ignorance on a test, but no student who lacks it can feign competence.

The lecture's distinctive move is to extend this definition from intelligence to life. Anything that exists now and does not bring more of itself into existence is one coin-flip from extinction, so life must compute the probability of its own future. Chemotaxis in E. coli shows the function at its simplest, and a toy simulation shows evolution learning it from nothing. Life is therefore a viable model, one that predicts itself into continued existence. He concedes the claim is nearly tautological: only the things that successfully predicted their own future are still around.

The experimental centrepiece explains how such functions arise. A soup of random byte strings in the language Brainfuck, with no fitness function, undergoes a phase transition after a few million pairwise interactions and fills with self-replicating programs. Von Neumann's theory of self-reproducing automata explains why: reproduction requires a tape, a copier, and a constructor, and that arrangement is a computer. Computation is therefore an attractor wherever randomness and the capacity to compute exist, and complexity follows from symbiosis, since every replicator is an ecology for further replicators.

The second half makes prediction social. Brains evolve as sensory extensions that let cells predict one another, predators and prey escalate their mutual modelling, and human brain size tracks group size. Free will is critical instability combined with high-order theory of mind, consciousness is a self-model that selects among futures, and the unified self is narrated after the fact by an interpreter. Transformers are predictors of the same kind with deficits in state, memory, and online learning. He closes by arguing that AGI has arrived by the term's original meaning, that this is a creative transition of the same kind as eukaryotes or eusociality, and that the existential risks worth attention are nuclear weapons and climate.

## Section by section

### The talk and its author

The book's title pays homage to Schrödinger's *What Is Life?*. Agüera y Arcas recounts learning to program on machines simple enough for a child to understand, and remarks, citing Benjamín Labatut, that his generation watched the passage from a world we could understand to one we cannot. He did computational neuroscience with Bill Bialek at Woods Hole, sold a startup to Microsoft, and moved to Google in the early 2010s when neural networks began to displace hand-written computer vision. His team produced DeepDream and the on-device models that include Gboard.

### Next-word prediction and its unbounded hardness

Language statistics are built up as tables: letter frequencies, bigram tables that generate fake words like "felogy," word-pair tables with thirty thousand rows, then conditioning on many previous words. The audience completes "Humpty" and "Helter" without effort, and he uses that to show where histograms stop. The Ballmer, Kilimanjaro, wet-keyboard, shipping-container, and grieving-friend examples each need a different capacity: specific knowledge, arithmetic, common sense, optimisation, theory of mind. Making the same models bigger removed the expected roof, and he names this the central empirical shock of his career.

### Functionalism and the Turing Test

Friston's theory of cortical responses is the neuroscientific version of the same function. Turing's 1950 paper is read as the canonical statement of functionalism: if questioning a machine convinces you it is intelligent, there is no further fact about "fairy dust" inside. Agüera y Arcas acknowledges the position is controversial and offers the mathematics-test asymmetry as his reason for holding it, then proposes to go further than Turing by treating life as a function of the same kind.

### Why life computes: E. coli and the toy bacterium

To reproduce, E. coli must eat, and chemotaxis gets it to sugar with a one-bit control: flagella bundled and swimming forward, or flying apart and tumbling to a random orientation. Howard Berg's rule, keep swimming while concentration rises and tumble when it falls, produces statistical ascent toward food. Jeff Stock's phrase "hair brain" names the protein network that computes this. The platform is irrelevant; what evolution selects is behaviour, the probability of action given stimulus. In the simulation, stationary toy bacteria die out, while bacteria with heritable action-given-stimulus tables learn chemotaxis in diverse styles, from cautious followers to reckless explorers. None maximises a score; the only prize in an infinite game is continuing to play. Internal signals such as hunger count as stimuli alongside external ones, which is where he locates the origins of dopamine and serotonin, and actions feed back as future stimuli, making this active inference in Friston's sense.

### What computation is

Turing machines and the Church-Turing thesis establish platform independence. The section then insists that computation is not rationality: MANIAC's bomb calculations used Monte Carlo methods, quicksort pivots randomly, and Turing had a hardware random-number instruction built into the Ferranti Mark I. The image of the computer as HAL or Data is a holdover from Leibniz's "calculemus" and from Boole, who meant his algebra as a theory of mind. Ada Lovelace's 1844 ambition to derive laws for the molecules of the brain was realised in a form by McCulloch and Pitts in 1943, who read neurons as logic gates. Rosenblatt's perceptron, with motorised volume knobs as weights, is the rival road. Computer science took the logical fork and produced everything from spreadsheets to phones; cybernetics took the other and, after decades of obscurity, produced modern AI.

### Computational Life: self-replicators from noise

Brainfuck has eight instructions, so random 64-byte strings contain on average two valid ones. Pairs are drawn from a soup of 8,192 strings, concatenated, run as self-modifying code, split, and returned, with occasional random byte flips. After several million interactions, well-formed replicators appear, thousands of copies of the dominant one, and operations per interaction climb from two into the thousands. Kolmogorov complexity, approximated by zip-ratio, drops sharply at the same moment, because structure compresses and noise does not. He calls the new state a machine phase, or computronium, or life, and says the transition occurs at a random time but always occurs. Because everything is platform independent, life in this sense is independent of the physics that implements it.

### Why computation is an attractor: von Neumann

Von Neumann's self-reproducing automaton needs a tape, a machine that copies the tape, and a machine that builds from it, with both machines described on the tape. DNA, polymerase, and ribosomes fill these roles, and von Neumann described the scheme before DNA's function was known. Building from a tape requires loops and conditionals, so reproduction requires a computer. He conjectures that any universe with randomness and the capacity to compute will evolve life, and that life is common, as matter seeking its most stable state.

### Complexification and major transitions

Every replicator creates niches for sub-replicators, including within the random padding of its own tape, and the soups show nested replicators arising. Mitochondria within archaea, ribs within a snake, and cortical columns within a brain follow the same pattern. Maynard Smith and Szathmáry's major transitions are read as symbioses: moments when independent replicators replicate jointly even at the cost of independence.

### Brains as mutual prediction

Synchronising fireflies use eyes to predict their neighbours; jellyfish muscle cells lack eyes and grow tendrils to feel what distant cells are doing, which is how nerve nets arise as sensory organs for coordination. In bilaterians the front end meets the world first, so sensory neurons cluster there. The brain is not a homunculus commanding the body but the body's sensory endpoints gathered in one place. Cambrian predation drives an arms race of theory of mind at increasing orders, under the principle that cooperation within allows competition without. The Second World War is offered as a Cambrian explosion of warfare from which cybernetics emerged, and Wiener's 1943 paper already contains the predictive hypothesis. Humphrey's gorillas and Dunbar's neocortex-to-group-size relationship ground the claim that human intelligence grew by out-predicting conspecifics, and his own group finds theory of mind improving with model size.

### Free will, consciousness, and the interpreter

The scene in *Before Sunrise* in which two strangers agree to meet on a platform in six months shows psychology predicting where bodies will be better than physics can. Free will is critical instability, life holding itself where a whisper can redirect a body for months, combined with high-order theory of mind including a model of oneself. Consciousness is a self-model that selects among counterfactual futures. Its unity is then dismantled: split-brain patients never report a second person and confabulate reasons for actions of the mute hemisphere; blindsight patients navigate while denying they see; Johansson's choice-blindness subjects fluently justify choices they did not make, and only about a third notice the swap. "We're all stochastic parrots," he says. The self is a fractal of mutual predictors on the same team, and the interpreter narrates. He adds that the internal predictors' disagreement, their not being aligned, is what makes the whole exceed its parts.

### Transformers

Transformers are predictors of this kind but lack durable state, long-term memory, and online learning. A chatbot that reaches a right answer and then gives a wrong explanation is exhibiting the interpreter, not proving it lacks intelligence, contrary to critics like Gary Marcus. QuietSTaR, Gemini's drafts, and chain-of-thought prompting are steps toward an inner voice. Jumping spiders plan hour-long attacks serially in tiny brains, which reverses the usual ranking: slow serial thought is the cheap option and parallelism the advanced one. Libraries are society's externalised chain of thought.

### Is there anything it is like to be a chatbot?

Any social being that models other models, including itself, at higher orders will appear conscious, and large models pass standard theory-of-mind tests because prediction over human text demands it. Agüera y Arcas removes the "appear": distinguishing appearing in every way from being is leaving science behind, which he takes to be the lesson of the Turing Test.

### AGI and the transition we are in

"Artificial general intelligence" was coined to contrast general with narrow capability, and by that meaning it has arrived; the goalposts have since moved toward something mystical. With [Peter Norvig](https://antikythera.wiki/people/peter-norvig) he has argued this is a major evolutionary transition, and he adds later transitions to the canonical list: eusociality and the technosphere, electronics, the internet, and AI. Transitions are creative rather than eliminative; bacteria survived eukaryotes. The panic reflects a false self-image of humans as individual top predators, when individually we cannot explain a toilet and collectively we form a superorganism that includes wheat, cattle, and trucks. The risks he takes seriously are climate collapse and armed nuclear arsenals, alongside disruptions to economies, governance, energy, and identity that a planetary-scale intelligence is needed to manage.

## Key concepts

- **Probability of future given past** — The single function intelligence and life compute: predicting what comes next from what came before. See [Predictive Intelligence](https://antikythera.wiki/terms/predictive-intelligence).
- **[Functionalism](https://antikythera.wiki/terms/functionalism)** — Intelligence is the input-output function computed, not the substrate; platform independence and the Turing Test follow.
- **Viable model** — A predictor that predicts itself into continued existence; his definition of life. See [Dynamic Stability](https://antikythera.wiki/terms/dynamic-stability).
- **[Infinite Game](https://antikythera.wiki/terms/infinite-game)** — The evolved bacteria maximise no score; the only prize is continuing to play.
- **[Machine Phase](https://antikythera.wiki/terms/machine-phase)** — The state of a random soup after the transition to self-replicating computation, detectable by a sudden drop in Kolmogorov complexity.
- **[Major Evolutionary Transition](https://antikythera.wiki/terms/major-evolutionary-transition)** — Maynard Smith and Szathmáry's moments of joint replication, extended here to the technosphere, electronics, the internet, and AI.
- **[Cooperation Within, Competition Without](https://antikythera.wiki/terms/cooperation-within-competition-without)** — Cooperation among cells or individuals is what makes competition between organisms or groups possible.
- **[Theory of Mind](https://antikythera.wiki/terms/theory-of-mind)** — Modelling others' models at increasing orders; the engine of brain growth and the reason large models pass Sally-Anne tests.
- **[Critical Instability](https://antikythera.wiki/terms/critical-instability)** — Living systems hold themselves where tiny inputs can redirect them; one of the two ingredients of free will.
- **[The Interpreter](https://antikythera.wiki/terms/the-interpreter)** — The narrating part of the brain that confabulates reasons after the fact, in split-brain patients and, on his reading, in chatbots.
- **[Superorganism](https://antikythera.wiki/terms/superorganism)** — The collective human-plus-technosphere entity in which modern intelligence actually resides.

## Connections

- [What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence) is the work this film previews; the book develops every section at length, and the film's compressions on free will and consciousness are best checked against it.
- [Artificial General Intelligence Is Already Here](https://antikythera.wiki/work/media/noema-artificial-general-intelligence-is-already-here) is the Norvig co-authored essay the closing section summarises; [AI Is Evolving — And Changing Our Understanding Of Intelligence](https://antikythera.wiki/work/media/noema-ai-is-evolving-and-changing-our-understanding-of-intelligence) restates the predictive thesis for a general audience. Agreement throughout.
- [After Alignment](https://antikythera.wiki/work/journal/afteralignment) and [How to Think Unlike Humans?](https://antikythera.wiki/work/journal/unlikehumans): the aside that internal predictors' non-alignment makes the whole exceed its parts replies to Bratton's talk of the previous year, and the two agree that alignment as uniformity is a loss.
- [Mutual Prediction in Human-AI Coevolution](https://antikythera.wiki/work/journal/coginfra/mutual-prediction-in-human-ai-coevolution) takes the film's account of social intelligence as its premise and extends it to human-AI pairs; [Minimum Viable Interiority](https://antikythera.wiki/work/journal/coginfra/minimum-viable-interiority) tries to engineer the self-modelling the film says constitutes consciousness. Both are extensions in agreement.
- [Substrates Unbound](https://antikythera.wiki/work/journal/substrates) and [The Long L](https://antikythera.wiki/work/journal/longl) share the platform-independence premise; [The Noocene](https://antikythera.wiki/work/journal/noocene) shares the reading of AI as a major transition.
- [Speculative Philosophy of Planetary Computation](https://antikythera.wiki/work/journal/spoc) and [AUTO–](https://antikythera.wiki/work/journal/auto) agree that computation is discovered rather than invented, but a difference of emphasis remains: the film's closing call for planetary-scale intelligence is asserted rather than argued, and Bratton's work supplies the political and infrastructural argument it lacks.
- [The Five Stages Of AI Grief](https://antikythera.wiki/work/media/noema-the-five-stages-of-ai-grief) is the "five stages" Agüera y Arcas refers to when naming the challenges of purpose and identity at the close.


## Terms used

- [Computation](https://antikythera.wiki/terms/computation)
- [Cooperation Within, Competition Without](https://antikythera.wiki/terms/cooperation-within-competition-without)
- [Critical Instability](https://antikythera.wiki/terms/critical-instability)
- [Functionalism](https://antikythera.wiki/terms/functionalism)
- [Infinite Game](https://antikythera.wiki/terms/infinite-game)
- [Life](https://antikythera.wiki/terms/life)
- [Machine Phase](https://antikythera.wiki/terms/machine-phase)
- [Predictive Intelligence](https://antikythera.wiki/terms/predictive-intelligence)
- [Social Fractal](https://antikythera.wiki/terms/social-fractal)
- [Superorganism](https://antikythera.wiki/terms/superorganism)
- [Symbiogenesis](https://antikythera.wiki/terms/symbiogenesis)
- [The Interpreter](https://antikythera.wiki/terms/the-interpreter)

## Related works

- [AUTO–](https://antikythera.wiki/work/journal/auto)
- [Latent Spacecraft](https://antikythera.wiki/work/journal/latentspacecraft)
- [Substrates Unbound](https://antikythera.wiki/work/journal/substrates)
- [Speculative Philosophy of Planetary Computation](https://antikythera.wiki/work/journal/spoc)
- [Antikythera](https://antikythera.wiki/work/journal/research)