Three Timescales of Adaptation
For most of the history of life on Earth, adaptation operated on a single timescale: biological evolution, measured in generations. Roughly fifty thousand years ago, a second inheritance system emerged — cultural evolution, measured in years. Machine learning now adds a third: computational adaptation, measured in hours and days. These are not merely different speeds of the same process. They are qualitatively different adaptation systems with different mechanisms, dynamics, and properties. The interaction between them produces phenomena that none of them produces alone — and the mismatch in their speeds is now a structural feature of the AI transition rather than a temporary condition.
The Substrate Dependencies — Contrarian ^ Opus
There is a parallel reading that begins from the material conditions required to sustain computational adaptation. Unlike biological evolution (which requires only sunlight and organic matter) and cultural evolution (which requires only human communities), computational adaptation demands massive industrial infrastructure: semiconductor fabs, power grids, cooling systems, rare earth mining operations, and the complex supply chains that connect them. This substrate is not self-sustaining. It requires constant human maintenance, geopolitical stability, and enormous capital flows. A single fab costs twenty billion dollars; a training run consumes the electricity of a small city. The timescale mismatch Agüera y Arcas identifies may be less significant than this dependency mismatch.
More fundamentally, computational adaptation operates only within the narrow band of tasks we choose to optimize for. While biological and cultural evolution explore the entire possibility space of survival and meaning, machine learning systems optimize for metrics we define: prediction accuracy, loss minimization, benchmark performance. The speed of adaptation matters less than the profound narrowness of what is being adapted to. The institutional lag may actually protect something vital — the slow, inefficient processes through which humans negotiate what matters, what counts as knowledge, what deserves preservation. The mismatch in timescales could be reframed as a mismatch in values: the computational system optimizes for efficiency and accuracy, while human institutions optimize for legitimacy, meaning, and social cohesion. The real risk is not that our institutions cannot keep up, but that in trying to keep up, they abandon their actual function: to be the slow variables that prevent the system from optimizing itself into a corner.
In the AI Story
Biological evolution provided the hardware — the brain, the vocal apparatus, the social instincts. Cultural evolution provided the software — language, institutions, accumulated knowledge. Computational adaptation now provides a third layer: systems that traverse, synthesize, and extend the knowledge landscape at a speed that makes cultural evolution look glacial. A large language model can be trained on the accumulated output of human culture in weeks. It can be retrained in hours. The binding constraint has shifted.
The institutional consequence is severe. The rate at which AI systems generate knowledge now exceeds the rate at which human institutions — universities, regulatory bodies, professional organizations, editorial boards — can evaluate that knowledge. The gap is not closing. It is widening with each capability improvement. The institutions were designed for cultural-evolution timescales; they now operate in a computational-evolution environment, and the mismatch produces specific failure modes: regulatory frameworks that address last year's capabilities, curricula that train for skills the machine now performs, certification systems that credential competencies the market no longer values.
The framework illuminates The Orange Pill's central frustration with institutional response. The EU AI Act, American executive orders, emerging frameworks — all address supply-side questions (what companies may build) on cultural-evolution timescales while the computational evolution of capabilities outpaces them. The demand-side questions (what citizens, workers, students, parents need to navigate the new environment) receive even less attention, because demand-side institutions — education systems, professional bodies — are even slower to adapt.
The solution is not to slow computation. That horse has left the barn. The solution is to accelerate the adaptive capacity of human institutions — but institutions are themselves products of cultural evolution and carry the inertia of their history. A university does not redesign its curriculum in a quarter. A certification body does not redefine its standards in a year. The structures that shape how humans relate to knowledge are deeply embedded and resistant to rapid change. Which means the AI transition's most consequential challenge may be not technical but institutional: whether existing structures can adapt faster than their history suggests is possible.
Origin
The three-timescale framework draws on Joseph Henrich's cultural evolution synthesis, extending it to computational adaptation. Agüera y Arcas's specific articulation emerges from his Santa Fe Institute work on complex adaptive systems.
Key Ideas
Three qualitatively different systems. Biological, cultural, and computational adaptation operate on different mechanisms, not merely different speeds.
The mismatch is structural. Institutions designed for cultural timescales cannot keep up with computational ones, and the gap widens rather than closes.
Evaluation, not generation, is the binding constraint. The species that defined itself by accumulating knowledge now faces the challenge of evaluating knowledge generated faster than any institution can process.
The demand side is neglected. Policy focuses on what companies may build; citizens must navigate the new environment largely without institutional support.
Appears in the Orange Pill Cycle
The Variable Weight of Speed — Arbitrator ^ Opus
The relative importance of these three timescales depends entirely on which question we're asking. For raw capability development, computational adaptation dominates completely (95%) — Agüera y Arcas is right that knowledge generation has fundamentally shifted to machine timescales. But for questions of implementation and social integration, the contrarian's substrate dependencies matter enormously (70%) — computational systems remain tethered to physical infrastructure that operates on human timescales. The fab takes years to build regardless of how quickly the models train.
The institutional lag question splits more evenly. For technical standards and safety protocols, the mismatch is as severe as the entry suggests (80% Edo's view) — regulations addressing last year's models are indeed obsolete on arrival. But for deeper institutional functions like determining professional ethics, educational values, or social priorities, the contrarian's point holds (65%) — the slowness serves a purpose, preserving space for human deliberation about what we actually want from these systems. Speed is not always a virtue when the destination remains contested.
Perhaps the synthetic frame is this: we're witnessing not a simple hierarchy of speeds but an ecology of timescales, each with its own role. Computational adaptation excels at exploring the possibility space; cultural evolution excels at selecting what to preserve; biological evolution provides the substrate of meaning and motivation that makes any of it matter. The challenge is not to accelerate everything to computational speed but to design interfaces between these timescales — mechanisms that allow rapid capability development while preserving slow deliberation about values. The mismatch becomes problematic only when we assume all adaptation should converge to the fastest timescale, rather than recognizing that different types of change require different temporal rhythms.
Further reading
- Henrich, Joseph. The WEIRDest People in the World (Farrar, Straus and Giroux, 2020)
- Agüera y Arcas, Blaise. "Cultural Evolution at Machine Speed." Noema, 2024
- Toffler, Alvin. Future Shock (Random House, 1970)
- Boyd, Robert and Richerson, Peter. The Origin and Evolution of Cultures (Oxford, 2005)