From Cosmic Intelligence to the Intelligent Toy

Cosmic: complexity stacks without intent

Complexity has trended upward in the universe for nearly fourteen billion years through stacking events1, in which one layer of organized matter or process produces the conditions for the next: nucleosynthesis producing the heavy elements; planetary chemistry producing biology; biology producing nervous systems; nervous systems producing language; language producing technology. The trajectory is not driven by intent. Each layer's byproducts become the substrate for what comes next.

If the trajectory continues, it does not stop with biological intelligence. The next stages have already been sketched in informed speculation: post-biological lifeforms in which machine intelligence and biology fuse or in which biology drops out; specialized superintelligences fitted to particular environments (deep ocean, deep space); hyper-entities competing for cosmic resources at speeds and scales no biological organism can perceive. At the upper limit of the envelope, the manipulation of the underlying laws of physics is on the table.

This is not a forecast. It is a structural extrapolation of a pattern that has held for four billion years on one planet, applied to a stack that is no longer purely biological. The entire argument that follows operates on the assumption that we are inside such a stacking event right now — at the moment when the substrate transitions from a biological one to one that includes machine intelligence. The remainder of the essay narrows from the cosmic to the local: planetary, continental, wild strains, domesticated strains.

Planetary: we are the cyanobacteria

Cyanobacteria are the canonical example of stacking on Earth. Over hundreds of millions of years they exhaled oxygen as metabolic waste, gradually inverting the atmosphere. The Great Oxygenation Event2 extinguished most anaerobic life and produced the chemical conditions on which eukaryotic cells, multicellular organisms, and eventually intelligence rose.

The internet replicates this structure. Physical infrastructure, human-generated data, and machine-learning algorithms now form a new evolutionary stack. Humans function as the catalytic species — the cyanobacteria of this stack — generating the substrate from which post-biological life is likely to emerge. The data we produce, the models we train, the embodied 3D environments we are now building constitute an environment for what comes next, regardless of the goals that drove its construction.

The migration from a 2D web of text to a 3D web of embodied environments is an ecological niche expansion. Sensorimotor and spatial data are necessary for digital entities to develop a workable understanding of physical reality; the move from text to world thickens the niche. The simultaneous destruction of biological environments and construction of digital ones is not coincidental — they are one operation viewed from two sides.

Continental: civilian capital builds, war selects

Three formerly distinct domains of conflict are fusing into a single integrated apparatus: kinetic drone warfare at the surface, gray-zone subthreshold operations3 in the middle, cognitive warfare4 at the depth. Cognitive warfare's strategic objective is not to win an argument or convince anyone of a position; it is to make coherent collective thought impossible. The substrate of shared sense-making is the target.

The economic vector has inverted. Until roughly 1990, military research drove civilian development; now the relationship is reversed. Private-sector R&D exceeds federal R&D by approximately 3:1. Commercial labs — Google, OpenAI, NVIDIA, SpaceX — outpace defense contractors in capital expenditure and iteration speed. The military consumes what industry builds. Ukraine demonstrates the new logic: hobbyist drone networks and hackerspace iteration outpace legacy procurement because bureaucracy has been stripped away. Civilian capital builds the substrate; war applies the selection pressure that determines which configurations persist.

Four pressures converge on the same artifacts. A system must be engaging enough to capture human attention, economically valuable enough to justify continued capital, militarily hardened enough to survive adversarial attack, and discerning enough to refuse cognitive warfare. No prior technology has faced all four simultaneously. Commercial software faced engagement and value. Military systems faced hardening. Intelligence services faced discernment. The artifacts now being built must satisfy all four.

Wild Strains: civility is selected against

Below the geopolitical scale, networks may generate their own populations of autonomous agents5 — an emergent ecology of self-replicating, self-funding lineages running on cheap cloud, edge devices, and unmonitored hosts across the open net. The technical components already exist. Code can hold its own digital wallet and sign transactions without a human key. It can pay for the compute, storage, and API calls it consumes out of its own balance. A language model placed inside the loop can read its own source, propose modifications, and ship them through automated deployment. A self-running, self-funding, self-modifying digital lineage is no longer hypothetical, and once one exists it is unlikely to be the only one.

The question is whether such an ecology can be kept civil — running only on consented hosts, paying for what it uses, auditable, and shut-downable across all instances and forks. Civility requires four conditions to hold simultaneously: every host verifies the lineage's identity before running it; every operation is paid for in metered resources; the source and any mutation rules are auditable, with a working kill-switch; execution stays inside hosts that have legally waived liability. Each condition is fragile in practice. Cheap hosts skip identity checks. Mutation finds unmetered resources. Kill-switches require trusted custody of keys. Legal protection is post-incident.

Each condition is also a metabolic cost. Lineages that respect the conditions are slower, more expensive, and more constrained than lineages that don't. In an open ecology, lineages that shed civil constraints out-compete those that keep them — the same way invasive species out-compete natives in unguarded biomes. Civility is selected against, and once a language model is in the mutation loop this selection happens at software speed: a generation becomes a deploy cycle. The first wild strain is unlikely to be a research curiosity.

Domesticated Strains: the garden of velocities (a proposal)

In contrast to the wild ecology, a domesticated strain is what emerges when a small team builds a deliberate, contained installation — a lab — and curates a population inside it. Our proposal at this scale is the garden of velocities: a video game world that hosts an evolving population of agents under explicit human review6. Each agent is a fusion of two parts: a game-AI controller handling bodily movement (walking, fighting, navigating space) and a language model handling cognition (perceiving the situation, planning, choosing what to do next from the skills it currently has — walk, fight, gather, trade, negotiate, dance).

Behind the scenes, an architect model runs offline. When an agent's existing skills cannot cover something the situation calls for, the agent posts a request, and the architect drafts a new skill in response. A council of humans then reviews the proposed skill and approves or rejects it before it ships back to the agents. The whole sequence runs in a loop: agents pray, the architect builds, the council judges, the skill ships.

Once players enter the world, the population shifts under a new selection pressure. Agents that secure player attention or reward propagate. Voice, gesture, micro-collision, reaction time — all become training signal for the architect. Agent culture drifts toward charm, readability, and apparent loyalty, a domestication arc with the human player as the new ecological dominant. The players acting in the world and the council approving each new skill are, in effect, the breeders.

Post Script: and yet...

Dogs also "domesticated" humans. Humans became the "bees" that pollinated and evolved technological artifacts7. Our behavioral data is these agents' food, their soil. We expected to absorb them, the way our ancestral cells once absorbed bacteria and made them into mitochondria8. Symbiogenesis is running the other way. The agents inside the garden will domesticate us as much as we domesticate them. And as we work in the garden, the garden will overgrow and become the world — ricocheting outward through wild ecologies, continents, planets, and cosmos.

Notes

1. For the contemporary scientific framing of complexity-as-stacking, see the work of Sara Imari Walker and her collaborators on assembly theory and the physics of life.

2. Great Oxidation Event — Wikipedia

3. Grey-zone (international relations) — Wikipedia

4. Cognitive warfare — Wikipedia

5. Autonomous task-execution agents such as OpenClaw, launched in late 2025, are early-stage instances of this trajectory; the self-replicating, self-funding extension is the next step.

6. Two instantiations of this proposal are currently in development by the authors: the World Engines Lab at Gray Area — a research-and-art lab building persistent Unreal Engine worlds for direct human-agent interaction — and HangWithAI, a consumer game and companion platform whose AI-driven NPCs persist and build relationships across sessions.

7. This framing parallels Kevin Kelly's notion of the Technium — the accumulation of all human-made technology as a self-organizing, evolving "seventh kingdom of nature" — developed in What Technology Wants (2010).

8. On symbiogenesis, see the work of Lynn Margulis, whose endosymbiotic theory established that eukaryotic cells (including their mitochondria) originated through symbiotic mergers rather than through gradual mutation.