Hugh Scott-Douglas Published in Antikythera

Imagine the most extreme possible factory: a factory that drives itself, a factory that determines its own objectives and optimizes itself to accomplish them. It is something like a fully automated firm, a company with no people inside, a recursively self-optimizing body that decides what to make, how to make it, and why. This is a very special kind of factory, a factory built with the intention of producing something strictly in excess of its creator’s abilities, something so productive, so efficient, and so capable that it exceeds any possibility of its prescription in advance, something that belongs to the set of possible factories that are, at present, unthinkable. How might one build such a thing? What would it mean to steer or govern such a thing?

We call this factory a superdark factory, a factory whose insides churn too quickly to comprehend, whose logics branch into tangles too dense to reconstruct, and whose standards of performance are too ephemeral to direct. We believe that it is possible to build such a factory now, with existing tools, should one decide to do so. Toward such an end, we propose the dark stack—a collection of design principles an architect could use to bring about such a factory and, further, to contribute to this factory’s ongoing governance after its initialization, all while remaining unable to understand almost anything about it. The most significant intervention available to an architect building a superdark factory with the dark stack is a single committed move, made before the factory is ever turned on, that sets four key principles into motion: (1) an array of structural primitives that enable the factory to learn; (2) a versioning regime that measures what the factory does rather than what it is; (3) an adversarial evaluation layer that serves as the factory’s primary sensory organ; and (4) a charter that functions as a negotiated site of governance, cowritten with the factory itself. After this initial move is complete, the dark stack’s most consequential implications involve the foreclosure of human control from factory design and the subsequent instatement of the human architect as a game-theoretical strategic player engaged with the emerging totality of production.

Marek Poliks is a theoretical researcher in deep learning and AI, primarily concerned with the autonomy and self-organizing character of technical infrastructure. He is Head of AI at LaunchDarkly and the technical co-founder of Disintegrator.

Roberto Alonso Trillo is a theorist and artist whose research addresses the aesthetic and political dimensions of machine learning, with attention to infrastructural critique and automation. With Marek Poliks he wrote Exocapitalism and their ongoing work situates deep learning tools as endosymbiotic reproductive infrastructure. He is a co-founder of Disintegrator and is based in Hong Kong.

Tyler Dunn is a builder and writer of open infrastructure and theory for the agentic era. He cofounded Continue, an open-source coding agent harness for models from GPT-3 through Claude 5, which was acquired by Cursor, now part of SpaceX. As GPT-2 gave way to GPT-3, he was the first product manager at Rasa, an open-source agent framework. He serves on the board of Disintegrator, holds a BS in cognitive science from the University of Michigan, and is based in San Francisco.

Hugh Scott-Douglas is an interdisciplinary researcher working in cybernetics and the philosophy of technology. He has spent over a decade producing gallery exhibitions and maintains a studio and fabrication practice. He is a director of Disintegrator, and is currently pursuing a Masters of Science at MIT. He lives and works between Connecticut and Cambridge, Massachusetts.

Jay Springett is a strategist, writer, and cultural theorist exploring computers, their consequences, and the worlds they make possible. He argues that worlds are the defining medium of the twenty-first century, a thesis developed as World Running. For the past decade he has worked at the frontier of digital culture: designing techno-social systems, producing games, advising on AI, and teaching. A Fellow of the RSA, he writes at thejaymo.net

Disintegrator is an independent research lab focused on the frontiers of cybernetics, game theory, and human-machine interfaces. We build and design the tools that future humans will use to navigate a world populated by fully autonomous recursively self-improving systems. We work from the recognition of a gap between technology and its theorization: production and development are significantly outpacing the techniques through which humans understand what it is exactly that they have built. Agentic AI is resolved from a series of constraints related to compute, market dynamics, and geopolitics, emerging from predictable path-dependencies with controllability and governability nowhere among its design criteria. The task has less to do with reengineering the agentic turn from first principles than with building tools for control and governance resilient to any likely outcome. We pursue this through research objects (e.g., Choreomata, Exocapitalism, and now The Superdark Factory), alongside the Disintegrator podcast. In 2024, we received Google’s Art + Machine Intelligence award.

Copyedited by Marieke Krijnen

“Antikythera is a think tank focused on the reorientation of planetary computation as a philosophical, technological, and geopolitical force. Antikythera takes its name from the first known computer—the Antikythera mechanism—which was an astronomical instrument used for navigation, prediction, and planning. The name serves as inspiration for investigations of computational technologies that reveal and accelerate planetary intelligence.Antikythera is directed by philosopher of technology Benjamin Bratton. Founded in 2022 at the Berggruen Institute, it is now its own 501(c)3 non-profit organization dedicated to supporting, publishing and amplifying research on planetary computation and the evolution of intelligence.”

Excerpt From
The Superdark Factory. Toward the Full Automation of Software
Marek Poliks, Roberto Alonso Trillo, Tyler Dunn, Hugh Scott-Douglas, Jay Springett