---
title: "57 Ideas & Questions about Cognitive Infrastructures"
description: "This article-film records Benjamin Bratton's lecture midway through Antikythera's 2024 Cognitive Infrastructures studio."
type: "reading-notes"
authors:
  - "Benjamin Bratton"
published: "2025-05-10"
original_url: "https://coginfra-film.antikythera.org/"
doi: "10.1162/ANTI.5CZH"
canonical_url: "https://antikythera.wiki/work/journal/coginfra-film"
md_url: "https://antikythera.wiki/md/work/journal/coginfra-film"
last_updated: "2026-09-03"
site: "Antikythera Wiki"
---

# 57 Ideas & Questions about Cognitive Infrastructures

> 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:** Benjamin Bratton
- **Published:** 2025-05-10
- **Kind:** Journal · Journal article
- **DOI:** https://doi.org/10.1162/ANTI.5CZH
- **Original:** https://coginfra-film.antikythera.org/
- **Length:** 13k words
- **This page:** https://antikythera.wiki/work/journal/coginfra-film
- **Notes generated:** 2026-09-03

## Summary

This article-film records [Benjamin Bratton](https://antikythera.wiki/people/benjamin-bratton)'s lecture midway through Antikythera's 2024 Cognitive Infrastructures studio. Its fifty-seven ideas are proposals and questions rather than a unified theory, but they share a claim: AI is becoming part of the material systems through which cities, institutions, and ecologies sense and act, so people will increasingly live in AI rather than simply use it. Bratton places this shift within an evolutionary sequence from life and artificialization through intelligence and symbolic language to machine intelligence. He then applies that sequence to alignment, interface design, generative culture, and language. The stakes are planetary. Artificializing intelligence can amplify the short-term adaptiveness that produced the Anthropocene, or help biological and mineral forms of intelligence become mutually sustaining over longer timescales.

## The argument

Bratton begins with the Antikythera mechanism, an astronomical computer whose calculation oriented intelligence to its planetary setting. He asks the audience to imagine the Apollo Blue Marble image as a film of Earth's 4.8-billion-year history. In its final moments, the planet grows an exoskeleton of cities, sensors, satellites, and fiber-optic cables. The lesson is that evolution does not merely happen on planets. Planets evolve, and intelligence is one of the things a planet does. The Great Oxidation Event changed the atmosphere and enabled more complex life; much later, a biosphere produced a technosphere that can now alter the biosphere in return.

This planetary frame changes the meaning of computation. Computation is not only mathematical calculation or a collection of tools. It is a cognitive infrastructure that lets a planet perceive itself through systems such as climate science and intervene in its own processes. Bratton argues that computation was discovered as much as invented because artificial computers approximate processes already present in nature. Following Stanislaw Lem, he also calls computation an existential technology. A bulldozer extends an existing capacity, while a telescope or microscope changes what people understand themselves and the universe to be. AI matters in both senses, but the lecture concentrates on the second.

The lecture's main sequence begins before biology. Chemical evolution gives rise to life through autopoiesis, the capture of external matter, energy, and information to remake oneself. Living systems become better at persistence by making things outside themselves, a process Bratton calls allopoiesis or artificialization. Artificialization selects for intelligence because organisms that can model counterfactual possibilities can build more effective external structures. Intelligence selects for symbolic language because complex models must be communicated and coordinated. Machine intelligence is trained on those languages and then alters the production of language, intelligence, and artificialization. AI is therefore the artificialization of artificialization, the remaking of the capacity that has remade the world.

This sequence is recursive, not a ladder ending in AI. Each new capacity changes the conditions of the earlier ones, and AI will become material for further forms that cannot yet be predicted. Bratton calls AI "mineral intelligence" because humans have folded rock and metal into arrangements that let the lithosphere perform cognitive operations once associated with animal nervous systems. The substrate of complex intelligence now crosses the biosphere and lithosphere. The practical question is how these forms can support one another rather than enter a zero-sum conflict.

[Cognitive Infrastructures](https://antikythera.wiki/terms/cognitive-infrastructures) names AI's present evolutionary setting. Bratton rejects the image of intelligence as a private mind housed in a box. Natural intelligence develops in open environments through interactions among agents at many nested scales. Machine intelligence should also be understood as "AI in the wild," embodied through robots, phones, cities, industrial systems, and landscapes. The body of a distributed AI may be an entire artificial environment rather than an animal shape. As AI becomes infrastructural, infrastructures become cognitive: they respond to people, environmental conditions, and other machine systems. Some of their most consequential exchanges may occur among AIs without direct human participation.

The later parts ask how humans should meet this distributed intelligence. Conventional alignment assumes that better AI will reflect human intentions and values more faithfully. Bratton calls the underlying assumption reflectionism. One camp condemns AI because it mirrors the biases of society, while another condemns it for failing to resemble people closely enough. Both take human likeness as the measure. Productive disalignment instead asks whether departure from human cognition can disclose realities that human culture cannot yet absorb. Bidirectional alignment still requires systems to answer human needs, but it also accepts that societies will change in response to machine intelligence. An ethics adequate to an existential technology would protect this capacity for disclosure rather than train it away.

Interfaces are where this mutual adjustment becomes concrete. Human-AI interaction design cannot stop at chatbots because an interface is any arrangement that governs exchange and change between complex systems. Each side models the other and models how it is being modeled. These relations vary in both symmetry and number: one human can engage one AI, thousands of AI characters can engage one player, or one AI can speak with thousands of people at once. Such ratios produce different forms of social intelligence. Digital twins also cease to be one-way copies once their actions alter the people they represent and they begin interacting with other twins.

Generative AI changes culture because models do not simply automate separate media. Multimodal systems translate among text, image, sound, and executable instruction. Bratton argues that they model a cultural manifold rather than a collection of individually owned signs. This produces a new politics of representation. Twentieth-century movements often sought power through greater visibility. Data privacy and data poisoning pursue agency through deliberate withdrawal from models. If models govern the physical operations of a city, political action can also target their representations by staging a false reality for them to ingest. Bratton calls this the Potemkin politics of the model.

Language is the final test of the argument. Claims that AI research is running out of data usually mean that it is running out of freely available English text, a narrow sample of planetary information. Bratton provisionally defines language as "that which can be tokenized," then asks whether tokenization expands language into data and calculation or only reveals an old relation between them. Sumerian writing began as accounting, so calculation and writing were never cleanly separate. Human languages may occupy the full possible shape of language, a small region within a larger informational space, or an intermediate region that also includes nonhuman communication. AI will change human language as nonhuman speakers use it at scales that may soon outnumber human speakers.

The lecture closes with two versions of the same evolutionary problem. The bleak version is that artificializing intelligence may be highly adaptive in the short term and self-extinguishing in the long term. The hopeful version, drawn from James Lovelock's *Novacene*, treats abiotic intelligence as another phase in life's selection, complexification, and aggregation. Neither version makes humans irrelevant. Both deny that human intelligence is the final or self-contained form of intelligence.

## Section by section

### Part one: Introduction

The article-film opens by locating the lecture within the Cognitive Infrastructures studio at [Central Saint Martins](https://antikythera.wiki/people/central-saint-martins). The Antikythera mechanism supplies the program's founding image of computation as planetary orientation. The Blue Marble thought experiment then gives the talk its scale: the atmosphere, biosphere, and technological exoskeleton are successive transformations of one evolving planet. Bratton's formulation that intelligence is something a planet does establishes the premise for everything that follows.

### Part two: Planetary computation

Bratton defines planetary computation as both a way of knowing Earth and a medium for changing it. The emergence of AI may make the last fifty years of networked computation look preliminary. Lem's distinction between instrumental and existential technologies explains why capacity alone is not the subject. Computation changes concepts of life, mind, and humanity. Bratton also states the studio's method. Unlike the Russian Cosmists, whose ideas ran ahead of technical capacity, present designers face technologies that exceed the concepts available to describe them. Philosophy must therefore arise through direct technical engagement. The fifty-seven ideas function as reusable pieces rather than conclusions.

### Part three: The evolution of the artificialization of intelligence

This part develops the lecture's evolutionary chain. Bratton defines artificiality through "anomalous regularity," illustrated by the visible difference between planted and unplanned trees, then shifts to artificialization as allopoiesis. Autopoietic systems remake themselves; allopoietic systems build external things that help them capture resources and persist. Intelligence and symbolic language amplify that capacity. AI, built from symbolic languages, feeds back into every earlier stage. The sequence supports a distributed account of machine embodiment. Against the Turing test's single humanlike mind, Bratton cites the planetary ocean in Lem's *Solaris* and proposes a city or landscape as an AI's body. His ecological theory of automation treats factory automation as a simplified version of repeatable causal cascades already found in living systems.

### Part four: Cognitive infrastructures

The brief fourth part condenses the thesis. Natural intelligence emerges among nested agents at environmental scale, and artificialized intelligence does the same. AI becomes part of infrastructure while infrastructure acquires capacities to sense, model, decide, and respond. Bratton presents artificialization as material and life-like, not the opposite of nature. The resulting systems form an ecology of planetary intelligence whose decisive interactions may occur between machines.

### Part five: Productive disalignment

Bratton reprises [Reflectionism](https://antikythera.wiki/terms/reflectionism) from his earlier alignment work. The Turing test's sufficient demonstration of intelligence became a cultural demand that all intelligence appear human. He connects the demand to AI grief and Freud's Copernican traumas, scientific decenterings that damage the species' preferred self-image. [Productive Disalignment](https://antikythera.wiki/terms/productive-disalignment) asks whether some distance from existing human values is beneficial, while bidirectional alignment describes mutual adaptation between societies and machines. The [Inverse Uncanny Valley](https://antikythera.wiki/terms/inverse-uncanny-valley) is the discomfort of seeing oneself through a machine's model. For Bratton, preserving such estrangement is part of AI ethics because it lets an existential technology revise human self-understanding.

### Part six: Human-AI interaction design

[Human-AI Interaction Design](https://antikythera.wiki/terms/human-ai-interaction-design), or HAIID, expands interface design beyond graphical controls and simulated conversation. Mutual mind-modeling produces social intelligence, even when each participant's model is a useful fiction. Bratton uses *Her* to show how quantity changes an encounter: Theodore imagines a one-to-one relationship while Samantha is speaking with 8,316 people. *2001: A Space Odyssey* supplies a one-to-one comparison, and games reverse the ratio with many AI characters for one player. The studio's Vivarium project adds simulated toy worlds and the difficult transfer from simulation to physical action. Embedding visualizations provide another interface, a representation of a model's representation of language, which itself represents reality.

### Part seven: Generative and synthetic AI

The discussion moves from DeepDream's dog-faced images to Hollywood labor conflict and Kuaishou's expected flood of generated video. Bratton's point is not that AI makes media faster. Multimodal models reorganize the relations among media and action. He contrasts a politics of increased representation with self-depresentation, the attempt to escape a model's sample. A model that governs city logistics can be contested by altering its internal representation, including through Potemkin performances designed only for machine perception. His story [The Myth of Blurope](https://antikythera.wiki/work/journal/hemstacks/blurope) imagines a Europe that protects sovereignty by withdrawing so completely from datasets that its existence becomes doubtful. The section ends with an unresolved cultural fork between convergence on shared generic outputs and divergence into private, continuously generated worlds.

### Part eight: Language ontology

Bratton questions whether "language model" names the right object. Training on free English text mistakes a small human archive for the world's information and may give AI the wrong initial conditions. Tokenization makes the boundary among language, data, and calculation porous. Research on structural similarities among human languages leaves open whether those similarities reflect reality's constraints or the narrow reach of human language. Whale communication suggests a larger biological field, while a robot language that became unintelligible to its designers points toward machine forms. Bratton distinguishes metalinguistics, what models disclose about language, from xenolinguistics, nonhuman uses of human language, and allolinguistics, the changes those uses induce in human speech and thought.

### Part nine: Conclusion: post-grief

The sad conclusion asks whether the same artificializing intelligence that succeeds over short timescales must fail over planetary ones. Long-term survival would require identifying and deliberately building the conditions under which advanced artificialization becomes symbiotic with its planet. The happier conclusion follows Lovelock's acceptance of abiotic intelligence as continuous with life. Human intelligence belongs to distributed structures of communication and to nested scales larger and smaller than individual brains. Machine intelligence changes the materials and organization of that process, not its evolutionary character.

## Key concepts

- **[Cognitive Infrastructures](https://antikythera.wiki/terms/cognitive-infrastructures)** — Material and institutional systems in which cognition is distributed across people, machines, and environments. AI becomes infrastructure while infrastructure gains capacities to sense and act.
- **[Artificialization](https://antikythera.wiki/terms/artificialization)** — The production of external structures through intelligence. AI recursively applies this process to intelligence itself.
- **[Existential Technologies](https://antikythera.wiki/terms/existential-technologies)** — Technologies that change what a civilization understands life, mind, and its place in the universe to be, not only what tasks it can perform.
- **[Mineral Intelligence](https://antikythera.wiki/terms/mineral-intelligence)** — Cognitive activity realized through arranged rock, metal, and electricity, extending complex intelligence from biological into lithospheric materials.
- **[Reflectionism](https://antikythera.wiki/terms/reflectionism)** — The assumption that machine intelligence is or should be a mirror of human thought, culture, and values.
- **[Productive Disalignment](https://antikythera.wiki/terms/productive-disalignment)** — The proposal that some departure from human cognitive forms can yield knowledge or capacities that strict alignment would suppress.
- **[Inverse Uncanny Valley](https://antikythera.wiki/terms/inverse-uncanny-valley)** — The discomfort of encountering oneself as modeled by a nonhuman intelligence.
- **[Human-AI Interaction Design](https://antikythera.wiki/terms/human-ai-interaction-design)** — Design of the interfaces through which human and machine intelligences model, affect, and adapt to one another.
- **[Self-Depresentation](https://antikythera.wiki/terms/self-depresentation)** — Deliberate withdrawal from computational sensing and modeling as a bid for privacy or sovereignty.
- **[That Which Can Be Tokenized](https://antikythera.wiki/terms/that-which-can-be-tokenized)** — Bratton's provisional functional definition of language under machine learning, which presses language toward data and calculation.

## Connections

- [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) is the studio collection for which this article-film provides a mid-studio conceptual statement. Its papers develop the proposals here rather than merely illustrating them. [Mutual Prediction in Human-AI Coevolution](https://antikythera.wiki/work/journal/coginfra/mutual-prediction-in-human-ai-coevolution) agrees with the account of reflexive modeling, while [Generative Topolinguistics](https://antikythera.wiki/work/journal/coginfra/generative-topolinguistics) turns the lecture's triple mirror and language questions into an interface research program.
- [After Alignment](https://antikythera.wiki/work/journal/afteralignment) supplies reflectionism, alignment overfitting, and bidirectional alignment. This article-film repeats that critique but places it inside a broader evolutionary account of infrastructures. [Existential Technologies](https://antikythera.wiki/work/journal/existentialtech) provides the Lem distinction that explains why AI can revise human concepts rather than merely extend human capacities.
- [AUTO–](https://antikythera.wiki/work/journal/auto) agrees that automation is an ecological and planetary process rather than the removal of human labor from an isolated task. The article-film extends this view to AI distributed across entire causal chains.
- [Speculative Philosophy of Planetary Computation](https://antikythera.wiki/work/journal/spoc) develops the claims that computation is found in nature, changes what intelligence can know, and exposes humanity to Copernican trauma. [Substrates Unbound](https://antikythera.wiki/work/journal/substrates) agrees that intelligence can cross biological and nonbiological materials, but examines that claim through biohybrid systems rather than planetary infrastructure.
- [Recursive Worlds](https://antikythera.wiki/work/journal/recursiveworlds) supplies the account of selection beginning in chemistry and reality being constructed through informational histories. The lecture draws this deep-time account forward to ask what machine intelligence will scaffold next.
- [What Is Intelligence?](https://antikythera.wiki/work/journal/wii-film) likewise treats intelligence as an evolutionary and collective process rather than a human possession. Its account centers prediction across nested agents, while Bratton centers the environments and infrastructures through which those agents interact.


## Terms used

- [Cognitive Exoskeleton](https://antikythera.wiki/terms/cognitive-exoskeleton)
- [Cognitive Infrastructures](https://antikythera.wiki/terms/cognitive-infrastructures)
- [Copernican Technology](https://antikythera.wiki/terms/copernican-technology)
- [Copernican Trauma](https://antikythera.wiki/terms/copernican-trauma)
- [Existential Technologies](https://antikythera.wiki/terms/existential-technologies)
- [Five States of AI Grief](https://antikythera.wiki/terms/five-states-of-ai-grief)
- [Hemispherical Stacks](https://antikythera.wiki/terms/hemispherical-stacks)
- [Human-AI Interaction Design](https://antikythera.wiki/terms/human-ai-interaction-design)
- [Inverse Uncanny Valley](https://antikythera.wiki/terms/inverse-uncanny-valley)
- [Mineral Intelligence](https://antikythera.wiki/terms/mineral-intelligence)
- [Productive Disalignment](https://antikythera.wiki/terms/productive-disalignment)
- [Reflectionism](https://antikythera.wiki/terms/reflectionism)
- [Scaffolding](https://antikythera.wiki/terms/scaffolding)
- [Sim-to-Real Gap](https://antikythera.wiki/terms/sim-to-real-gap)
- [That Which Can Be Tokenized](https://antikythera.wiki/terms/that-which-can-be-tokenized)

## Related works

- [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra)
- [Substrates Unbound](https://antikythera.wiki/work/journal/substrates)
- [AUTO–](https://antikythera.wiki/work/journal/auto)
- [The Long L](https://antikythera.wiki/work/journal/longl)
- [Recursive Worlds](https://antikythera.wiki/work/journal/recursiveworlds)