---
title: "Whole Earth Codec"
description: "Connor Cook, Christina Lu, and Dalena Tran propose the Whole Earth Codec as a foundation model trained on ecological data from planetary sensor networks."
type: "reading-notes"
authors:
  - "Connor Cook"
  - "Christina Lu"
  - "Dalena Tran"
published: "2023"
original_url: "https://codec.earth/"
canonical_url: "https://antikythera.wiki/work/studio/whole-earth-codec"
md_url: "https://antikythera.wiki/md/work/studio/whole-earth-codec"
last_updated: "2026-09-06"
site: "Antikythera Wiki"
---

# Whole Earth Codec

> 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:** Connor Cook, Christina Lu, Dalena Tran
- **Published:** 2023
- **Kind:** Studio projects · Studio paper
- **Venue:** Planetary Computation Studio
- **Original:** https://codec.earth/
- **Length:** 8.0k words
- **This page:** https://antikythera.wiki/work/studio/whole-earth-codec
- **Notes generated:** 2026-09-06

## Summary

[Connor Cook](https://antikythera.wiki/people/connor-cook), [Christina Lu](https://antikythera.wiki/people/christina-lu), and [Dalena Tran](https://antikythera.wiki/people/dalena-tran) propose the Whole Earth Codec as a foundation model trained on ecological data from planetary sensor networks. Its purpose is to discover relations among phenomena that human observation and language-centered models cannot adequately register, then make that knowledge operative in ecological management. The proposal joins a technical architecture to an account of distributed governance and a speculative theory of human–nonhuman interdependence. Sensors, shared representations, privacy mechanisms, funding arrangements, and restrictions built into design tools would together enable a biosphere that participates in its own modification. The authors also describe failures, contested decisions, and harmful uses. This is a design proposal with imagined future scenarios, not a report of an implemented system.

## The argument

The starting point is the Event Horizon Telescope, whose distributed instruments produced a black-hole image through computation. The authors distinguish an observation extending beyond any one location from knowledge extending beyond immediate perception. Terrestrial sensors already make Earth the object of an analogous observatory. What remains missing is an architecture that relates their heterogeneous outputs. Training foundation models principally on internet language excludes most of the information generated by the biosphere.

The Codec would encode ecological measurements and place them in a common latent space, a learned arrangement of numerical representations. Spatial and temporal correlations would bring data from different modalities into relation. Bioacoustic signals and atmospheric pollutants could thereby become computationally comparable without being translated into ordinary language. The authors treat transformer self-attention, which assigns contextual importance within a sequence, as an analogy for the planet observing its own relations. They call this recursive observation “folding the gaze.” Whether such correlations would produce the proposed higher concepts remains an open technical claim.

Knowledge and governance are coupled. A system capable of revealing ecological dependencies could guide interventions, produce harmful innovations, and change how responsibility is assigned. The authors therefore locate governance within infrastructure and information flow, alongside institutional agreements. They propose distributed learning and computation that require cooperation without giving any participant complete access. Funding and maintenance must support planetary purposes while accounting for the materials, energy, and labor that the model consumes.

The final step gives nonhumans a role in deciding what the infrastructure does. Here nonhumans include organisms and abiotic systems such as weather. The Codec would aggregate their needs, enable new exchanges, and negotiate conflicting objectives. Rather than merely tell people about ecological damage, it could constrain architecture, fishing, and extraction as those activities are planned. The argument does not end with restoring an earlier ecological condition: it proposes continuing evolution through new combinations of biological and technical systems.

The scenarios qualify this ambition. An apparent gain in whale comprehension drives harmful tourism; a supposedly equitable extraction decision still overrides local objections. The authors' projected improvements do not eliminate the political question of whose needs become measurable and whose objections can change an outcome. Better sensing, in their account, creates capacities for negotiation without guaranteeing agreement.

## Section by section

### I. Proposal: Inverting the Observatory and Folding the Gaze

The opening sections move from the telescope's distributed aperture to ecological sensors looking inward at Earth. The authors argue that language is only one part of planetary information. Energy, particles, waves, and fields can become inputs when sensors transduce them into records. Their objective is to connect these records through a foundation model that can support many subsequent applications.

Encoding converts patterned inputs into tokens; embedding assigns those tokens numerical representations through which relations can be learned. The authors distinguish this from simply collecting more observations. Their analogy to self-attention proposes a model whose internal relations disclose otherwise imperceptible ecological concepts. The rejection of a panoptic observer describes the intended relationship between model and planet; it does not by itself establish protection against surveillance.

### I. Proposal: Assembling the Codec

The architecture diagram specifies a distributed mesh of edge devices and regional data centers. Sensors include cameras, audio recorders, chemical instruments, lidar, and pressure or moisture detectors. Every sample receives a UTC timestamp and GPS anchoring. Separate encoders map modalities into a shared representation, while contrastive learning brings correlated observations nearer together. Decoders produce sequence predictions, and smaller labeled datasets support specialized models built on the common foundation.

Federated learning would keep data on devices while sending learned weights to regional centers. The authors distinguish their abstract representation from a digital twin intended to mimic its subject. They also stipulate a single pass over available data because of its scale. These are proposed design commitments; the essay supplies no training results or performance measurements.

### I. Proposal: Post-Codec Futures

Possible applications include generating weather patterns or genomes, detecting environmental damage, and mediating relations among species. The same capacities might facilitate bioweapons or extraction. More detailed attribution of ecological effects could alter insurance, litigation, and the assignment of blame. The authors also anticipate biospheric hallucinations: plausible outputs that falsely assert a genome's existence or a relation between rainfall and pollutants. These errors would have consequences beyond the model's reliability as an information service.

### II. Governance: Principles and the Governance Space

The principles require collective intelligence, accountability, maintenance, resistance to private capture, and goals spanning human and nonhuman interests. Infrastructure would initially combine existing sensors and facilities through memberships, partnerships, standards, and international agreements. The comparison with the internet and the Event Horizon Telescope supplies precedents for coordination, rather than a ready-made planetary constitution.

Information flow replaces a simple choice between open and closed access. Multiparty computation would allow separated parties to compute jointly; differential privacy would protect information about individuals through statistical treatment. The proposed architecture distributes training, weights, and inference so that no data center contains the whole system. The political aim is to make interdependence a technical condition of participation.

Resourcing includes energy, materials, capital, and labor. Training would favor cleaner grids and stronger economies rather than distribute burdens equally. Suggested funding includes a one-percent levy on downstream applications, carbon taxes, institutional investment, and insurance mechanisms. A purpose-directed trust would allocate resources through collective decision procedures and model input. Maintenance could involve specialist developers, grants, guilds, and bug bounties. These are alternatives and examples, not a settled financing plan.

Evolution would itself become governable. The Codec could propose protocols, identify sensor blind spots, and simulate the effects of changes before adopting them. Plural voting would admit nonhuman stakeholders, although the text does not specify how their credits or preferences become operational. The proposal couples machine self-modification to collective oversight without fully resolving conflicts between them.

### III. Interdependence: Nonhumans and the Codec

The authors argue that technological change already affects nonhumans while granting them little agency over its operation. They cite self-owning land and technologically augmented forests as preliminary attempts to alter that condition. The Codec would aggregate needs, connect populations, and support ecological evolution. Localized models could communicate within the wider system, making comprehensive representation compatible with differentiated participation.

Their objectives can conflict: increasing biodiversity, lowering carbon impacts, and avoiding genetic stagnation do not automatically select the same policy. Maintenance is only an initial goal. The larger ambition is a changing ecology that can formulate and pursue new arrangements rather than remain fixed at its historical baseline.

### III. Interdependence: Post-Anthropocene and the scenarios

The authors criticize ecological regulation for weak monitoring, inflexible metrics, and limited enforceability. The Codec would embed contextual limits inside the tools through which activities occur. In the Architecture Software scenario of 2028, a Turkish airport cannot select concrete because the model predicts watershed damage. A Vietnamese railway gains permission subject to land-repair conditions. These examples make governance a restriction on available design choices.

The Fishing Management scenario of 2050 complicates enforcement. Some nations ignore dynamic fishing maps, and the Codec cannot compel them. It redirects compliant fleets and eventually creates an economic advantage for participation. The narrative also includes resentment among compliant fishers and attempts to disable or spoof sensors.

Whale Communication, set in 2031, turns apparent translation into entertainment and tourism that disturb whale behavior. A public proposal leads to funding for restrictions on non-research generative use. This episode qualifies the earlier ambition to move beyond translation into human speech: intelligibility can stimulate exploitation as well as care.

Collective Iteration, set in 2048, supplies the sharpest institutional conflict. A European sensor expansion prompts mineral extraction in Australia. Community objections produce revised protocols, partial relocation to Chile, and benefit sharing, but Australian extraction proceeds. The scenario associates sites selected for extraction with weak sensor coverage, suggesting that what the model cannot adequately observe can be undervalued. Grassroots sensing becomes one proposed means of changing those decisions.

### III. Interdependence: New symbionts of human, planet, machine

Drawing on John Maynard Smith and Eörs Szathmáry, the conclusion links major evolutionary transitions to new forms of information transmission and individuality. The Codec might enable specialized entities to coordinate into larger biological–technical compositions. The authors speculate about a new inheritance medium and even planetary individuality; they do not demonstrate either. [Holly Jean Buck](https://antikythera.wiki/people/holly-jean-buck)'s criticism of the opposition between technological intervention and restorative climate justice supports their final rejection of a pristine nature outside technical mediation.

## Key concepts

- **[Folding the Gaze](https://antikythera.wiki/terms/folding-the-gaze)** — Planetary observation turning toward the relations within its own model, using self-attention as the analogy for recursive ecological comprehension.
- **Cross-modal latent space** — A shared numerical representation that brings spatially and temporally correlated observations from different kinds of sensors into relation.
- **Information flow** — Governance of who can know and compute what, using distributed computation and privacy mechanisms rather than a single open-or-closed access rule.
- **Biospheric hallucination** — An output with plausible structure but false ecological content, potentially misleading intervention and attribution of responsibility.
- **[Cognitive Infrastructures](https://antikythera.wiki/terms/cognitive-infrastructures)** — The coordinated sensors, computation, and institutions through which ecological intelligence would become an operative planetary capacity.
- **Interdependence** — A proposed redistribution of agency in which limiting unilateral human decisions could deepen relations among humans, nonhumans, and technical systems.

## Connections

- [After Alignment](https://antikythera.wiki/work/journal/afteralignment) directly presents the Codec. Bratton's discussion of data quality, de-individuation, and inward observation agrees with its central argument and identifies self-attention explicitly as an allegory. The project essay adds governance mechanisms and scenarios absent from that shorter presentation.
- [Planetary Vision](https://antikythera.wiki/work/journal/planetaryvision), published later, gives the technical and institutional account of the Event Horizon Telescope invoked here. Galison explains why correlations between instruments, rather than isolated observations, produce the image. That supports the emphasis on relational knowledge while distinguishing demonstrated astronomical integration from the Codec's proposed ecological generalization.
- [Antikythera](https://antikythera.wiki/work/journal/research) develops the shared claim that computation makes planetary conditions perceptible and actionable. Its discussion of climate models as instruments of knowledge agrees with the Codec's starting point. The Codec specifies one possible transition from that knowledge to constraints built into ecological and industrial practice.


## Related works

- [After Alignment](https://antikythera.wiki/work/journal/afteralignment)
- [Planetary Vision](https://antikythera.wiki/work/journal/planetaryvision)
- [Antikythera](https://antikythera.wiki/work/journal/research)