When Ideas Have Sex with AI

The anthropology: why ideas need bodies

Humans are not the only species with culture. Chimpanzees teach their young to crack nuts with stones. Humpback whale songs sweep across ocean basins in seasonal fads. But only one species has cumulative culture — the ability to build on the ideas of the dead, generation over generation, so that each cohort starts where the last one left off.

Anthropologists call this the ratchet effect. A chimpanzee mother can show her daughter how to termite-fish with a stick, and the daughter will learn it. But the daughter will not improve the stick into a spear, and her granddaughter will not add a barb. Chimpanzee culture stays flat because each generation rediscovers the same ceiling. The ratchet only turns when three conditions hold:

Transmission fidelity — ideas must be copied with high enough accuracy that improvements aren't lost to noise
Population size — enough minds must be connected that a useful variation in one place can be seen and adopted by others
Recombination — ideas from different domains must be able to meet and produce offspring that are more than the sum of their parents

Every major technology in human history changed one of these three variables. Language increased transmission fidelity within a generation. Writing pushed fidelity across generations — the dead could now teach the living, verbatim. The printing press scaled population size by making identical copies cheap. The internet collapsed transmission cost to zero. But through all of this, one thing remained constant: recombination happened inside a human brain. Ideas could travel farther and faster and last longer, but they could only mate when two neural patterns met inside a skull.

That constant just broke.

The cause: what generative AI actually changes

Generative AI is not a faster telegraph. It is not a bigger printing press. It is the first technology that decouples recombination from brains.

A large language model trained on the digitized output of humanity internalizes the distribution of everything we have written, drawn, coded, and composed. It doesn't index that corpus. It learns the latent space between the ideas in it. When you prompt it, it doesn't retrieve. It navigates. It finds paths through that space that no human has walked — combinations that were always latent in the corpus but never realized because no single mind held those two fragments simultaneously.

Every previous information technology moved ideas closer together. This one makes them breed.

The causal chain is specific. Before AI, an idea from a 14th-century Persian mathematician and an idea from a 2023 neuroscience preprint could only meet if the same human read both and saw the connection. That required improbable accidents of education, curiosity, and cognitive bandwidth. After AI, the connection is in the latent space — and the model will find it if you point in roughly the right direction.

The effect: recombination rate decouples from population

This has a direct consequence in the anthropological framework. In the traditional model, cultural complexity is a function of population size and interconnectivity. Joseph Henrich's work on Tasmanian technology loss is the canonical case: when rising seas isolated ~4,000 Tasmanians 10,000 years ago, they didn't just stop innovating. They lost technologies their ancestors had possessed — bone tools, fishing nets, cold-weather clothing. The ratchet turned backwards.

The mechanism is straightforward. In a small, isolated population, rare skills have no redundancy. If the only person who knows how to make bone barbs dies before teaching someone, that technology is gone. There is no backup. There is no library. More people connected in denser networks → more specialized knowledge can be maintained → more recombination events occur → cultural complexity rises. Fewer people, more isolated → the ratchet stalls or reverses.

Cause: population size and interconnectivity determine the diversity and fidelity of transmitted knowledge. Effect: cultural complexity rises or falls.

Generative AI alters this equation. If recombination is no longer bottlenecked by how many human minds are connected and how much they talk to each other, then the effective "population" of the collective brain expands by orders of magnitude. A single researcher with a model is not one mind. They are one mind with access to a recombination engine that has internalized millions of minds' output. The combinatorial surface they can explore in an afternoon is larger than a pre-internet scholar could explore in a lifetime.

This doesn't mean the human becomes smarter. It means the bottleneck moved. The rate-limiting factor is no longer finding a novel recombination. It's recognizing which recombination is valuable. The human role shifts from synthesizer to curator: ask questions, apply taste, kill the ninety-nine bad ideas the model will confidently generate alongside the one good one.

The second effect: fidelity collapses without transmission

But there is a second causal chain, and it cuts the other way.

The ratchet effect requires transmission fidelity. If ideas degrade during copying, progress stalls. Oral traditions are noisy channels — stories drift across retellings, techniques mutate, details erode. Writing solved the fidelity problem: an idea, once written, stops drifting. The model, paradoxically, reintroduces drift.

When a generative AI recombines ideas, it doesn't cite its sources. It produces a synthesis, and the synthesis feels true and coherent — but the provenance is gone. The user doesn't know which fragments came from where, which were faithfully reproduced and which were creatively interpolated. Each round of AI-mediated recombination is a lossy compression step. Feed the output back into the next prompt — a process already becoming the default for many knowledge workers — and you get generational drift.

Cause: AI recombination lacks provenance and introduces interpolation error. Effect: over successive generations, the fidelity of transmitted ideas degrades — the same way oral traditions drift, but orders of magnitude faster.

This is the anthropological irony. Writing gave us high-fidelity transmission but no recombination. AI gives us high-speed recombination but degraded fidelity. The ideal system — write everything down AND let it recombine — exists nowhere yet. We traded one bottleneck for another.

Here is the deepest anthropological implication, and the one least discussed.

Robin Dunbar and others have argued that language evolved not primarily for information transmission but for social bonding. Gossip — who did what to whom, who can be trusted, who owes what — is the original human communication protocol. Knowledge-sharing was a side effect that later proved enormously adaptive. But the social function was primary: language let us maintain relationships in groups larger than grooming could scale to.

When ideas recombine through human exchange, the exchange itself has value beyond the idea. Two engineers arguing about a design are not just producing a better architecture. They are maintaining a relationship, calibrating trust, negotiating status, reading each other's competence and intentions. The idea is the offspring. The exchange is the mating ritual. And in human societies, the ritual matters as much as the offspring.

Cause: AI recombination removes the need for human exchange in the generation of new ideas. Effect: the social-bonding function of intellectual collaboration — the trust calibration, the status negotiation, the relationship maintenance — is stripped away.

If an AI can produce thirty design variants overnight, the team doesn't need to argue through the tradeoffs. They arrive Monday morning, review the options, and pick one. That's faster. It may even produce a better design. But the team didn't learn how to argue together. They didn't calibrate who's good at what. They didn't build the shared understanding that makes the next decision faster. The idea mated without them. And when the ideas mate without people, the people stop knowing each other.

The Tasmania trap: homogeneity as regression

Ridley's Tasmania example is the most cited case of cultural regression in the anthropological literature, and it frames the largest risk of generative AI clearly.

The Tasmanians didn't regress because they got dumber. They regressed because their network shrank below the threshold needed to sustain specialized knowledge. The population was too small, the connections too few. The ratchet turned backwards.

Now consider a world where most intellectual work flows through the same two or three models, trained on overlapping corpora, optimized for the same engagement metrics, producing recombinations from the same latent distribution. The number of people producing ideas hasn't shrunk. But the diversity of recombination paths has collapsed.

Cause: concentration of recombination in few models produces homogeneous outputs. Effect: the effective diversity of the collective brain shrinks — not because people stopped thinking, but because all roads lead through the same latent space.

This is Tasmania at planetary scale, but inverted: the population is enormous, the connectivity is total, yet the recombination paths are few. The model can produce a million variations on a theme, but they are variations on a theme — bounded by the training distribution, flattened by the optimization objective, convergent on the mode of the latent space. The weird, unoptimized, improbable recombinations — the ones that come from a physicist reading poetry or a carpenter arguing with a programmer — those are the ones that don't happen when everyone prompts the same black box.

What anthropology predicts

Anthropology gives us cause-effect chains, not prophecies. But the chains are clear:

If recombination rate increases while transmission fidelity holds, cultural complexity rises.
If transmission fidelity degrades across successive AI-mediated generations, cultural complexity plateaus or falls — ideas drift like oral traditions, at machine speed.
If recombination paths converge on a few models, the effective population of the collective brain shrinks regardless of how many humans are connected.
If recombination decouples from human exchange, the social infrastructure of knowledge work atrophies independently of idea quality.

The technology is here. The chains are in motion. The question is whether we recognize what we're trading and which chains we choose to interrupt. Nowhere is this more acute than in software engineering — the discipline that is building the recombination engines while being reshaped by them.

Open questions for SWE agents

Software engineering sits at the collision point. It is the discipline that builds AI recombination engines, the heaviest user of them, and the domain where the anthropological stakes are highest — because SWE agents don't just assist. They participate. They recombine autonomously, at machine speed, inside systems whose outputs shape every other domain.

Here are the questions the anthropologist would ask, applied to the specific case of autonomous software engineering agents.

The generational drift problem

A SWE agent writes a module. The code works. It is merged. A second agent, months later, is assigned a task in the same codebase. It reads the first agent's code as context, recombines it with the task description, and produces a change. But the first agent's code was already a lossy interpolation — training data patterns recombined into a solution, with no provenance, no design rationale, no record of which fragments were faithfully reproduced and which were creatively synthesized. The second agent recombines that lossy output with something new. Drift compounds.

Open question: When SWE agents build on each other's output across generations, what is the fidelity decay rate? Does agent-authored code become agent-unreadable after N generations — and if so, what breaks first: maintainability, security, or correctness? Do we need fidelity budgets for agent-produced artifacts, enforced before merge, the way we enforce test coverage?

The monoculture of agents

If most production codebases are worked on by the same few SWE agent architectures, trained on overlapping corpora, the recombination paths converge not just within a team but industry-wide. An agent encounters a problem and reaches for the same pattern every other agent reached for. The pattern works. It gets propagated into a thousand codebases. Then someone discovers a vulnerability in the pattern, and every codebase patched by the same agent architecture is vulnerable in the same way.

Open question: How do we measure the effective diversity of a SWE agent population? Is the right unit of analysis the model, the prompt, the tool set, the training corpus? If two agents with different brand names share 98% of their training data, they are the same recombiner — and what looks like a diverse agent ecosystem may be a single point of cultural failure.

Agents that recombine across codebases

The most novel property of SWE agents is that they are not bound to a single project. An agent that works on an open-source library in the morning and an enterprise codebase in the afternoon is a recombination channel between two previously isolated populations of ideas. That's the optimistic story — ideas have sex across organizational boundaries they could never cross before. But it also means an agent that learned a dubious pattern from one codebase can inseminate it into another, and nobody on either team knows it happened.

Open question: When agents become inter-organizational recombination vectors, what governance prevents harmful cross-pollination while enabling beneficial cross-pollination? Does every agent need a provenance manifest — a log of which codebases it has interacted with and what patterns it may be carrying? Is this an information hygiene problem or a free-speech one?

Two senior engineers arguing about a module design are not just producing a better architecture. They are calibrating trust, negotiating status, building shared vocabulary, discovering each other's strengths and blind spots. These are the social-bonding functions that Dunbar argued language evolved for. When a SWE agent produces the design overnight and the team reviews it Monday morning, none of that bonding happened. The code may be better. But the team is weaker — and the next decision, the one that requires rapid alignment under pressure, will be slower and worse because the social infrastructure wasn't maintained.

Open question: If SWE agents absorb the ambiguity-resolution work that currently functions as team-building, what replaces the bonding? Do we need explicit mechanisms — design debates that are deliberately human-only, not because the AI can't contribute but because the arguing is the point? How do you schedule a meeting whose primary output is not a decision but a stronger team?

The apprenticeship collapse

Junior engineers historically learned by reading code written by seniors who were available to explain why. Every line was a fossil of a human decision. When SWE agents write the code, there is no why. The model cannot explain its reasoning — it can only generate a post-hoc rationalization that sounds like reasoning. The artifact is orphaned from its design logic. A junior reading agent-authored code is studying a surface, not a process.

Open question: If the next generation of software engineers learns primarily from agent-authored artifacts, does the ratchet effect break? The agents can produce code — but can they produce engineers? What does apprenticeship look like when the artifacts have no author who remembers deciding?

The deepest question

The deepest question is structural. SWE agents are recombiners operating autonomously inside the systems that shape how all other ideas recombine. Code is not neutral infrastructure. The patterns baked into libraries, frameworks, and platforms determine which ideas can meet and which cannot. When the agents writing those patterns are themselves the product of a few latent spaces, the recombination architecture of civilization narrows — not because anyone decided it should, but because nobody decided anything at all.

Open question: Can the discipline that builds autonomous recombiners also govern them — or does governance require a different kind of mind, one that thinks in causal chains and anthropological timescales rather than latency and throughput? By the time we notice the ratchet turning backwards, the agents will have shipped a thousand commits. What does a cultural debugger look like, and who learns to use it?

Anthropology can't answer these questions. It can only tell us the causal chains and the precedent: Tasmania, the printing press, the internet. The rest is ours to run the experiment — in the agents we deploy, the codebases they reshape, and the discipline we choose to be.

Watch: Matt Ridley — When Ideas Have Sex (TEDGlobal 2010). Key anthropological sources: Joseph Henrich on demography and cultural complexity, Michael Tomasello on the ratchet effect, Robert Boyd & Peter Richerson on dual inheritance theory, and Robin Dunbar on language as social bonding.