IThe foundation: technology is evolution outside the body
Biological evolution is slow, so humans found a shortcut: we evolve around the body instead of within it[20]. A tool is not an object. It is a stabilized behavior pattern: the hammer is repeated force, writing is repeated memory, the drawing is repeated instruction, the computer is repeated calculation. Each technology freezes a piece of cognition so it can be reused, transmitted, and improved without waiting for biology.
This produces a recurring loop that has run for the whole of human history:
Each cycle externalizes another human capacity: matter (body tools), meaning (symbols), coordination (institutions), labor (machines), computation (software). The sequence is not random. It is cumulative culture compounding: every innovation builds on an archive of prior discoveries.
IIWhere AI sits in this sequence: externalized imagination
Generative AI is the cognitive phase of the same trajectory: repeated imagination. For the first time, the capacity being externalized is not force or memory but the generation of possibilities itself.
The four-year arc of generative models (2022–2026) revealed something diagnostic: image quality improved on a predictable compute curve and was therefore never the story. The story was control: the slow re-attachment of human authorship to machine generation through geometry locks, masks, instruction-following, editable representations. This matters philosophically because it confirms what tools have always been: not autonomous forces, but couplings[19]. The relationship is a mirror loop. Human intent shapes the model's continuation; the continuation bends the human's next thought. Neither side is passive. The mirror is not neutral.
IIIThe field's current direction: evolution's algorithm, made explicit
Examine what the major laboratories actually built in the last twelve months, and a pattern emerges that none of them individually claims but all of them jointly demonstrate. Every layer of the stack is converging on the same structure:
- Interface: from turn-based chat toward copresence, continuous and simultaneous human-machine coupling (Thinking Machines, "Interaction Models") [10]
- Substrate: from sealed pixels toward operable worlds: editable, persistent, physically valid spatial state (World Labs' "3D as code" and Marble; VAST's Project Eden) [5][6][7][13]
- Loop: generated environments in which agents act and are evaluated, generating new environments in turn (DeepMind's Genie 3 + SIMA, an "endless training dojo") [8][9]
- Engine: systems that conduct experiments on their own methods and improve them (recursive self-improvement; the Darwin Gödel Machine; the discovery-automation labs) [1][2][3][4][11][14]
Strip the branding and every one of these is the same machine: generate variation → evaluate → archive → recombine → repeat.
That is not a new invention. It is evolution's own algorithm (variation, selection, inheritance, memory) extracted from biology, run on silicon, and accelerated by orders of magnitude. The deep claim of the technology-as-evolution thesis is thereby confirmed from an unexpected direction: the endpoint of externalizing human capacities is the externalization of the search process itself. Technology began as evolution's product; it is becoming evolution's implementation. The frontier labs converged on this architecture not because they read evolutionary theory, but because there is apparently only one general algorithm for open-ended discovery[12]. Culture, science, and now AI are successive substrates running it.
The historical progression closes a circle: adaptation to nature → control of matter → energy → information → cognition → control of the search over possibilities.
IVThe self-referential turn, and its limit
The frontier question: what happens when a system generates its own environment? When it builds its own laboratory, simulates its own experiments, evaluates its own hypotheses?
Mechanically, the loop multiplies: simulation depth × adaptation speed, freed from human timescales. This is the moment technology stops being externalized capability and becomes externalized epistemology: science outside the mind.
But here the evolutionary frame exposes the structural limit. Evolution requires an environment it did not choose. Selection pressure must come from outside the organism; that exteriority is what makes fitness mean something. A system that generates its own selection environment is breeding against an imagined niche, improving its skill at inhabiting its own beliefs. Errors in its world-model become its own training signal. Self-play conquered Go because the rules were external and perfect; a self-generated world has no external rules.
A model that builds its own lab can self-improve its skill. Only contact with reality can self-improve its truth. The dream, left alone, audits itself. Nature, eventually, audits the dream.
VWhat remains outside: the body as instrument
Where does reality enter the loop? Through verification, and verification has a structure. Judgment stratifies into layers of decreasing automatability:
- Crystallized rules: constraints so thoroughly verified they became computable (codes, standards, formulas). Machines inherit these completely.
- Learnable patterns: regularities extractable from labeled experience. Machines approximate these increasingly well.
- Embodied experience: what a space feels like, what scale is right, what an atmosphere does to a person[16][18]. Here the measuring instrument is the human body itself: proprioceptive, scaled, mortal, moving through the world[15]. A simulation can reproduce the photons and the physics; it cannot stand in the room. Prediction of embodied response is possible, but it is parasitic on ground truth that only bodies generate[17].
And the key insight binding the layers: the rules are fossilized experience. Every stair ratio and ceiling minimum is centuries of bodily experiment, verified until it crystallized into code[21]. The first layer is the third layer, archived. This is the general mechanism of cumulative culture: felt experience → repeated verification → rule → machine inheritance. Machines accelerate everything downstream of crystallization; the upstream source remains embodied life: new felt experience, newly verified.
This is why "judgment" persists at every scale of the analysis: choosing what is worth trying, deciding what to trust, knowing when to stop. Execution externalizes; the calibration of value does not, because value, for humans, is ultimately calibrated in bodies.
VIThe synthesis
Technology is evolution outside the body, and it has now externalized so much that it is building evolution itself: variation engines, generated environments, archives of discoveries, self-improving search. The trajectory is directional even though every event in it is contingent: more perception, more memory, more simulation, more coordination, faster adaptation. Short timescales are noise; the pressures are persistent.
But the more complete the externalization becomes, the more clearly its dependence shows. A search process needs an environment it did not invent; an evaluator needs ground truth it cannot generate; the dream needs an audit. The human role does not vanish in this story. It concentrates: from doing, to directing, to verifying, back toward the original instrument, the body in the world, where meaning has been calibrated since before the first tool.
Technology is evolution running outside the body; it has now externalized imagination and is externalizing discovery itself. As the loop closes toward self-reference, the scarce and load-bearing element is the one thing that was never externalized: embodied contact with reality, and the judgment that grows from it.