Cognitive Infrastructures

Summary#

Cognitive Infrastructures is the compendium of Antikythera's 2024 studio, published in the journal in May 2025. Its framing essay argues that AI is a material force embedded in the world's decision-making systems rather than a disembodied brain, and that as AI becomes infrastructural, infrastructures become cognitive. Ten papers by the studio's researchers are grouped under five themes; each paper has its own notes.

The argument#

The essay begins from natural intelligence, which emerges at environmental scale, among many agents, in active landscapes as much as in brains. Artificial intelligence is said to be the same: spread through planetary computation, where it runs industrial processes, coordinates cities, and senses, models, and acts "in the wild." AI is infrastructuralized and infrastructures are made cognitive. Because interacting with the world now retrains model weights, to perform an action is also to represent it in a model; to play with the model is to remake it.

The rest of the framing is open questions. Can alignment mean societies evolving in relation to AI, not only AI bending to society? The studio calls these problems embryonic and sorts its research into five sections.

Productive Disalignments holds that complex intelligence arises among diverse minds, so AI's value lies in thinking orthogonally to human frameworks; the target is reflectionism, the assumption that AI must mirror human cognition, and the alternative is AI as an "existential technology" in Stanisław Lem's sense. Traversing the Uncanny Ridge names a "Canny Valley" of over-familiar generated images and proposes novelty as a loss function. Synthetic Counteradaptation models human and machine minds as reciprocally adapting, with AlphaGo's Move 37 and Lee Sedol's Move 78 as the case.

Post-Anthropocene Psycho-Physiologies reads human-AI relations through symbiosis and artificialization, the loosening of evolutionary paths into contingent intervention. Mutual Prediction in Human-AI Coevolution asks what follows when humans become the modelled rather than the modellers, with wheat and Tinder as examples. Xenophylum argues for xenomimicry, engineering robot bodies from functional parameters rather than biological templates.

Organs Without Bodies starts from the point that sensory organs drove the growth of brains, so thinking is bound to sensing. Organoid Array Computing imagines networks of lab-grown brain organoids as a grown rather than programmed intelligence. Cognition With and Beyond the Brain adapts Stiegler's endosomatization to epidermal and tissue-embedded media.

Planetary Time Computation observes that planetary computation standardizes time, through UNIX time for instance, while diversifying the timescales humans can perceive. Chronoseed treats a language model as a time capsule of human cognition for unknown future readers. The Chronoceptual Governor maps how technologies speed or slow perceived time and what that does to economic feedback loops.

Mimesis of Mimesis asks whether AI has concepts and whether replicating human mimetic processes is a doubling or something new. Minimum Viable Interiority uses a pandemonium architecture to ask what closure an NPC would need to stop being a philosophical zombie. Generative Topolinguistics turns embedding visualizations from maps into interfaces for acting on semantic space.

Section by section#

The framing essay has two layers. An untitled introduction of six paragraphs states the premise that AI is a material, infrastructural force rather than a disembodied brain, draws the parallel with natural intelligence as an environmental phenomenon, and then turns to open questions about interfaces, alignment, and what counts as intelligence. It ends by listing five thematic sections, which then structure everything that follows.

Each of the five sections has the same shape: a short thematic preface that sets out a problem and its key term, then two paper abstracts. Productive Disalignments introduces reflectionism and the Lem reference before abstracts on the Uncanny Ridge and synthetic counteradaptation. Post-Anthropocene Psycho-Physiologies introduces symbiosis and artificialization before mutual prediction and Xenophylum. Organs Without Bodies links sensing to thinking before organoid arrays and cognition beyond the brain. Planetary Time Computation pairs standardized and diversified time before Chronoseed and the chronoceptual governor. Mimesis of Mimesis asks whether AI has concepts before minimum viable interiority and generative topolinguistics. The abstracts are summaries, not the papers, and a numbered list of the ten papers and their authors closes the compendium.

Key concepts#

• Cognitive Infrastructures — the systems through which cognition is distributed beyond the individual mind, made reflexive by AI's retraining on use.
• Productive Disalignment — the principle that AI's value depends on its diverging from, not mirroring, human cognition.
• Reflectionism — the assumption that AI must reflect human cognition, blamed here for repeated impasses.

Connections#

The alignment questions restate After Alignment and draw on Modes of Cognition. The organoid and embodiment sections share ground with Substrates Unbound, and the infrastructure theme with AUTO–. The studio's own framing is presented in 57 Ideas & Questions about Cognitive Infrastructures. The coevolution papers feed the niche framing of the Agentworld (Research Brief).