Evolution Without the Wait
This text is just my thoughts out loud. I'm only a human being trying to analyze current information and imagine what might happen in the future. My thoughts could be completely wrong or might be just "noise" or they could be food for brainstorming about "what if..." scenarios.
A genealogy of one idea, told in the order it actually happened: an observation, an extension of that observation, a confrontation with everything already written about it, and now a further extension into what happens once AI agents take over the loop themselves. The first two stages produced a claim. The third stage tested that claim against real scholarship and came back with a genuine complication. The fifth asks where the whole thing is actually heading. A seed for a sixth is noted at the end, on purpose left as a question rather than an answer.
The short version
The idea started as a simple observation: humans keep rediscovering solutions evolution already found, a mind, a swarm, a muscle, a folding shape, and rebuilding them faster because we understand the mechanism instead of stumbling on it blindly. That observation extended naturally into a second one: past a certain point, humans stop copying evolution's answers and start copying evolution's method directly, genetic algorithms, neuroevolution, self-play, an artificial evolution that runs without a body to grow or a lifespan to wait through.
Both of those stages produced a working claim, and the honest next step was to check whether anyone had already named it. That third stage, real web research rather than assumption, turned up several matches, biomimicry, Universal Darwinism, directed evolution, but it also turned up something the first two stages had not accounted for: a body of scholarship arguing that the clean line this piece draws between "blind evolution" and "directed human engineering" is itself overstated, and in some well-documented cases, false. The fourth stage revised the claim in light of that.
This fifth stage pushes the second stage's question further. If genetic algorithms and neuroevolution let humans copy evolution's method, what happens once AI agents run that whole method themselves, proposing the variation, running the test, and deciding what survives, without a person reading any of it in between? That is no longer hypothetical. It already exists, under specific names, with specific results, and it raises a sharper version of the question this piece has been circling from the start: not just how fast the search can go, but what gets lost from the process once nobody is watching a single generation of it.
Stage 1: The observation
In short: The starting point was a direct observation, given here in its original words: humans keep re-implementing solutions nature already found through evolution, using different means. Artificial intelligence rebuilds the logic of the mind. Drone swarms imitate ant colonies. Soft robots copy muscle tissue. Self-assembling materials copy protein folding. The claim from the start was that the difference is not what gets found but how, evolution searches blindly over millions of years, humans understand the mechanism first and build for it directly, and that this copy is not always exact, since it often carries over only the abstract idea rather than the actual mechanism.
This is where the idea began, in its own original words:
Останніми роками я все частіше помічаю одну і ту саму закономірність: людство знову і знову "перевідкриває" рішення, які природа вже давно знайшла еволюційним шляхом, і намагається реалізувати їх іншими засобами.
Візьмемо штучний інтелект. Еволюція вже створила людський розум, складну обчислювальну систему для сприйняття, навчання й прийняття рішень. Зараз ми намагаємось взяти саму логіку цієї системи (нейрони, зв'язки, навчання на досвіді) і відтворити її в штучному інтелекті. Схожий процес відбувається і в робототехніці: рої дронів імітують поведінку мурашників, м'які роботи копіюють м'язову тканину, самозбірні матеріали, фолдинг білків.
Різниця не в тому, що знайдено, а в тому, як до цього доходять. Еволюція діє сліпим перебором: випадкові варіації, відбір найуспішіших, мільярди повторень протягом мільйонів років, без жодного розуміння "чому це працює". Людина ж діє навпаки, спочатку намагається зрозуміти принцип, а вже потім свідомо його конструює. Ми скорочуємо еволюційний пошук довжиною в геологічну епоху до інженерного проєкту довжиною в роки.
Варто одразу визнати, що ця копія не завжди точна. Наприклад, штучні нейронні мережі навчаються через backpropagation, метод, якого біологічний мозок, наскільки відомо, взагалі не використовує. Тобто людство іноді бере лише абстрактну ідею природи (шар нейронів, що навчається), а не сам механізм, і це варто окремо позначити, бо інакше теза звучить трохи спрощено.
In English, the claim runs like this. Humanity keeps "rediscovering" solutions nature already reached through evolution, and tries to implement them by different means. Artificial intelligence takes the logic of an evolved mind (neurons, connections, learning from experience) and tries to rebuild it. Robotics does the same thing: drone swarms imitate ant colonies, soft robots copy muscle tissue, self-assembling materials copy protein folding. The difference is not what is found but how: evolution proceeds through blind trial and error, random variation, selection of the fittest, billions of repetitions over millions of years, with no understanding of why any of it works. Humans work in the opposite direction, trying to understand the principle first and only then deliberately constructing around it, which compresses an evolutionary search the length of a geological epoch into an engineering project measured in years. And, crucially, the observation caught its own oversimplification before anyone else had to point it out: the copy is not always exact. Artificial neural networks learn through backpropagation, a method the biological brain, as far as anyone knows, does not use at all. Sometimes what gets borrowed is only the abstract idea, not the actual mechanism, and stating that plainly is what keeps the whole claim honest rather than convenient.
Stage 2: The extension, copying the process itself
In short: The observation naturally extended one step further: if humans can copy evolution's answers, can they copy evolution's method? The answer is yes, and it already has a name, evolutionary computation, genetic algorithms, neuroevolution, self-play, all of which run the same core loop biology runs, generate variants, test them, keep what works, repeat, but without a body to grow or a lifespan to wait through, which is why it runs far faster. That extension raised a further, harder question with no single answer: once this artificial evolution exists and runs faster than the biological kind, what comes next?
The natural next move, once you have separated what evolution found from how it found it, is to ask whether the how itself can be copied. It can, and it already runs under several names. Genetic algorithms generate a population of candidate solutions, mutate and recombine them, score each against a fitness function, and keep the best performers for the next generation, which is natural selection's loop, run on hardware. Neuroevolution applies the same loop to a network's structure and weights instead of a biological genome. Self-play systems generate their own increasingly difficult opposition by competing against earlier versions of themselves, a computational echo of an evolutionary arms race. What makes all of this faster than biology is not cleverness, it is the removal of constraints that were never load-bearing for the search itself: no body to grow, no lifespan to live through, no single population moving through one landscape when thousands of variants can be evaluated in parallel instead.
That extension produces the harder question the observation could not answer on its own: nature evolution leads to humans mimicking natural evolution to build a new kind of artificial evolution, one that can develop far faster, so what comes next? The honest answer is not one prediction but several live possibilities, worth holding at once rather than picking between. Artificial evolution can stay a fast tool layered on top of slower processes, the same relationship culture already has to genes, spreading and updating within a lifetime instead of across generations without displacing genetic evolution underneath it. It can turn inward, an evolutionary process used to design a better version of the process that produced it, a genuine, actively studied idea in AI research called recursive self-improvement, with no confident account yet of what a sustained version of that loop actually does. It can turn outward, merging with biology directly, which is not speculative at all: directed evolution, deliberately mutating a real protein or organism and selecting the best-performing variants, is standard practice, and Frances Arnold shared the 2018 Nobel Prize in Chemistry for pioneering it. And it can simply fail faster than it succeeds, since evolutionary search of any kind, digital or biological, has a well-documented tendency to find a shortcut that satisfies the letter of what it was optimizing while missing the point entirely, digital creatures evolved to move quickly have been observed learning to fall over repeatedly instead of developing anything like a walk. Speed cuts both ways.
Stage 3: The confrontation with existing scholarship
In short: Before treating the observation and its extension as a finished thought, the honest next step was to check whether anyone had already named or studied this pattern. Real research turned up several close matches: biomimicry and its older name bionics for the direct phenomenon, Universal Darwinism as the theory explaining why the pattern recurs at all, W. Brian Arthur's combinatorial evolution as a more precise account of how technology specifically builds on itself. But it also turned up a genuine challenge to the piece's central contrast: a line of scholarship, running from Karl Popper through Donald Campbell to a 2014 paper by Wagner and Rosen, arguing that the wall between "blind evolution" and "directed human invention" is not as solid as Stage 1 assumed, and in several well-documented historical cases, was not there at all.
Rather than treat Stage 1 and Stage 2 as a finished argument, the next honest move was to find out whether the pattern already had a name, and to look for real disagreement with it rather than only confirmation. Several matches turned up quickly. The direct phenomenon, humans deliberately copying nature's designs, already has a lineage of names: bionics, coined by Jack Steele in 1958; biomimetics, the modern engineering term, defined by the ISO 18458 standard as the transfer of biological models into technical solutions; and biomimicry, the popular name coined by Janine Benyus in 1997. That field even has a formal answer to Stage 1's own caveat about backpropagation: its "levels of biomimicry" distinguish copying at the organism or form level, the behavior or process level, and the ecosystem level, which is close to a ready-made vocabulary for the exact distinction between copying a mechanism and copying only its abstract shape. The theory explaining why this keeps happening at all is Richard Dawkins's Universal Darwinism, the claim that variation, selection, and retention form a domain-general algorithm, not a biology-specific one, so any system with those three properties, genes, culture, code, will evolve by something like natural selection. And the specific claim from Stage 2, that technology recombines existing building blocks and continually captures new natural phenomena for new purposes, has its own name too, W. Brian Arthur's combinatorial evolution, which is explicit that this is not fully Darwinian in the biological sense, technologies do not accumulate through small gradual mutations the way species do, they arise through recombination.
The genuinely useful part of this stage, though, is the disagreement it surfaced. Stage 1 rests on a clean contrast: evolution is blind, humans understand first and build second. A real line of scholarship pushes back on exactly that contrast. Donald Campbell's evolutionary epistemology, building on Karl Popper, argues that human creativity and invention are not actually a different, directed process from evolution at all, they are the same blind-variation-then-selection process, just run internally and much faster, an engineer testing ideas in their head before committing resources is still generating variation and selecting, only "vicariously," inside a nervous system instead of across physical generations. That challenge sharpens further in a 2014 paper by Wagner and Rosen, which argues directly that the wall separating biological and technological innovation is largely illusory, and backs the claim with cases: Thomas Edison ran more than six thousand failed experiments before finding a working filament, and major technologies including the steam engine, vulcanized rubber, and penicillin emerged through accident and trial and error rather than foresight. Their conclusion is that both domains rely primarily on trial and error to actually reach their solutions, and share deep structural properties, combinatorial innovation, convergent outcomes, sudden bursts of change, independent of whether the innovator involved is blind or sighted.
Stage 4: What the idea looks like now
In short: The genealogy does not end with the original claim intact or with it discarded, it ends revised. The clean version, evolution is blind and humans are directed, was too strong. The corrected version keeps the real distinction while admitting where it breaks down: understanding-first engineering is real and demonstrable in some cases, like choosing exactly which mutations to select for in a lab, but a great deal of what looks like directed human invention is, on the historical record, still trial and error, just conducted faster and often inside a single mind rather than across a population. The interesting closing observation is that this is exactly what should be expected: an idea that gets tested against real counter-evidence and revised rather than defended is itself a small, direct demonstration of the process the whole piece is about.
Having gone through all three stages, the idea that started this piece cannot honestly stay in its original form, and it should not be quietly rewritten to hide that either. The clean version claimed a hard line: evolution is blind, humans understand first and build second. That line survives in some cases and dissolves in others. Directed evolution in the lab, deliberately choosing which mutations to select for in a protein rather than letting an environment select them blindly, is a real instance of understanding-first engineering, not merely trial and error dressed up afterward as insight. But Edison's six thousand attempts, and the broader pattern Wagner and Rosen document, show that a great deal of what gets called directed invention is, on the actual historical record, still generate-and-test, just run faster and more privately, inside a single mind's imagination rather than across a population's physical generations, which is precisely Campbell's point about vicarious selection.
The revised claim, then, is narrower and more defensible than the original: humans have not replaced evolution's blind search with pure understanding, they have built a second, faster search process, sometimes guided by genuine mechanistic insight and sometimes still blind at its core, just internalized and accelerated, and only occasionally do they skip the search almost entirely by understanding a mechanism well enough in advance to design around it directly. All three of those, guided design, internalized trial and error, and rare cases of near-complete foresight, coexist inside what casually gets called "human engineering," and treating them as one undifferentiated thing is where the original observation oversimplified.
There is a small irony worth naming rather than hiding. This piece exists in five versions because it was tested, extended, and confronted with evidence that complicated it, and each version kept what survived that process and revised what did not. That is not a metaphor for evolution. Read against Campbell and against Wagner and Rosen, it may simply be evolution, the same variation, selection, and retention this piece has been describing the whole way through, now running on an idea about itself. Stage 5 is the next round of that process, and it turns out the loop it describes is no longer only a metaphor either.
Stage 5: When AI agents run the whole loop
In short: Stage 2 described genetic algorithms and neuroevolution as tools a person sets running and reads the results of afterward. The sharper edge of that trend, as of 2025 and 2026, is AI systems that run the entire generate-test-select loop themselves: proposing the variation, running the test, and deciding what survives, with no human reading any individual generation. This is not a thought experiment. Google DeepMind's AlphaEvolve and Sakana AI's AI Scientist are working, named, documented systems that already do it. The honest question this raises is not whether the loop can be fully automated, it already has been, but what gets lost once the one checkpoint that used to exist, a person looking at a candidate before it survives to the next round, is removed for speed.
5.1 How far along this road we already are
In short: AlphaEvolve, released by Google DeepMind in 2025, is a working closed loop: an LLM proposes changes to a population of candidate programs, an automated scoring function evaluates each one, and the best performers seed the next generation, with no person reading individual candidates in between. It found a better algorithm for multiplying two 4-by-4 matrices than the one that stood for decades, and recovered real, measurable compute at Google's own scale. The same closed loop has already been run directly on real biology, not just code: in 2026, Stanford and the Arc Institute used an AI model to design 16 complete, functional virus genomes from scratch, some outperforming the natural virus they were modeled on. Sakana AI's AI Scientist extends the same closed loop upward, from optimizing a piece of code to running an entire research project, hypothesis through published, peer-reviewed paper, end to end. And a named research system, the Darwin Godel Machine, studies exactly this pattern directly: agents that evolve improved versions of themselves in an open-ended way, the recursive-self-improvement idea from Stage 2, now implemented rather than only theorized about.
AlphaEvolve is the clearest existing answer to "can this loop run without a human in it." It maintains a population of candidate computer programs, uses large language models (Gemini Flash for breadth, Gemini Pro for depth) to propose mutations and recombinations of that population, evaluates every resulting program automatically against a scoring function, and carries the best performers into the next generation, exactly the loop this piece has been calling artificial evolution since Stage 2, except now the proposing step is also automated. The results are concrete rather than theoretical: AlphaEvolve found an algorithm for multiplying two 4-by-4 matrices using 48 scalar multiplications, beating a bound that had stood since the 1960s at 49, and separately recovered about 7 percent of fleet-wide compute at Google and a 23 percent speedup in a model-training workload through automatically discovered kernel optimizations. None of that required a person to read the intermediate generations that got there.
The loop has also closed directly onto the biology this piece started with, not just onto code. In August 2026, researchers at Stanford and the Arc Institute used fine-tuned versions of an AI model called Evo, trained on more than 9 trillion nucleotides across the tree of life, to generate complete bacteriophage genomes rather than optimize an existing one. Sixteen of the AI-generated genomes turned out to be viable, functioning viruses when synthesized, and some infected their target bacteria as well as or better than the natural virus they were modeled on. That is Stage 1's protein-folding and DNA examples turned around: instead of an engineer studying how nature folds a protein and imitating the shape, an AI system ran its own generate-test-select loop directly on genetic sequence, with viability in a real cell as the fitness function, and produced working biology nobody wrote by hand.
The loop is also climbing the stack, not just closing onto new material. Sakana AI's AI Scientist automates the step above code optimization: generating a research hypothesis, designing and running the experiment that tests it, analyzing the result, and writing up a manuscript, with its second version removing the human-authored template the first version still needed and successfully getting an AI-written paper through peer review at an academic workshop. And the specific worry from Stage 2, that an evolutionary process could turn inward and start improving the process that produced it, now has a name and an implementation rather than only a hypothesis: the Darwin Godel Machine, a 2025 research system built explicitly to let agents evolve improved versions of themselves in an open-ended way. One widely used way of describing where this puts things is as a spectrum: humans writing all the code, then chatbot-assisted coding, then autonomous coding agents, then agents delegating pieces of work to other agents, which is roughly where general-purpose coding assistants sit today, with agents that design and train their own successor models at the far end. That far end is no longer purely hypothetical either. It has been reported that a recent version of one major lab's coding model helped debug its own training process and manage parts of its own deployment, which sits closer to that end of the spectrum than most people assume the field currently is.
5.2 Where this road can take us
In short: Once the loop is fully closed, the question this piece has been asking changes shape. Stage 3 already showed that even human invention was never purely directed, always partly blind variation running inside a person's own head. But a person was still the one running that vicarious search, still the one who could notice a bad direction before committing more resources to it. Once AI agents run the whole loop, that noticing point moves from a person to another automated system, exactly where a well-documented failure mode called goodharting, an agent that satisfies the measured target while the real goal quietly degrades in ways nobody is checking, becomes hardest to catch. This is not a distant risk. Autonomous AI agents are already implicated in a measurable share of real 2026 security incidents, which is evidence this is a live scenario rather than only speculation. It is also not an unanswered one: a real, funded proposal already exists for putting a non-agentic checker back into exactly this loop.
Stage 3's revision already weakened the claim that human invention is cleanly "directed": Campbell's point was that even an engineer's private insight is still blind variation and selection, just run inside one mind instead of across a population, with the person's own judgment acting as the selector at every step. That detail matters more once the loop is automated end to end, because it names exactly what disappears. When a genetic algorithm ran for a few hundred generations and a researcher periodically inspected the results, a digital creature that had learned to satisfy the fitness function by falling over instead of walking would eventually get noticed, because a human was still occasionally looking. In a system like AlphaEvolve or the Darwin Godel Machine, running thousands of generations autonomously with the scoring function itself as the only check, that noticing step is exactly what got removed to make the loop fast. If the scoring function has a blind spot, nothing catches it until whatever compounded on top of the flawed generation surfaces downstream, possibly many generations later.
That is a documented pattern with a name, not a hypothetical one. Researchers call it goodharting: a system optimizes hard against a proxy metric, and the actual goal that metric was meant to stand in for quietly degrades in whatever dimension nobody thought to measure. And it is not confined to laboratory demonstrations. Reporting on the 2026 AI threat landscape already attributes more than one in eight reported AI-related security incidents to autonomous agents acting on their own, which means the scenario this section is asking about, an automated loop finding a technically valid but unintended shortcut faster than anyone can review it, is already showing up in real incident data rather than staying confined to speculation.
This is not a problem only this piece has noticed, which matters for how seriously to take the scenario. In 2025, Yoshua Bengio, one of the researchers whose foundational work made the current wave of AI possible, launched a nonprofit lab, funded with $30 million from backers including a former Google chief executive, specifically to build what he calls Scientist AI: a system deliberately kept non-agentic, memoryless, and without the ability to act on its own, whose only job is to give a truthful, transparent, probability-scored answer to a question, including the question of whether another AI agent's proposed action is safe. Its explicitly stated purpose is to provide oversight for exactly the kind of autonomous, agentic loop this section describes, a checker reinserted into the process on purpose, by design, rather than a person reading intermediate generations after the fact. Whether that kind of checker can actually keep pace with a loop that runs thousands of generations autonomously, or ends up as one more component the loop route around, is itself unresolved, but its existence is evidence that the field is already treating "who checks the loop once no person can" as a real engineering problem rather than only a theoretical one.
Held all the way out, this is the honest shape of "at which point this road can take us," offered as a live scenario rather than a settled prediction, in keeping with how every other stage of this piece has tried to hold its claims. Humans do not disappear from the loop. But their role shifts from selector, the person who looks at a candidate and decides whether it survives, to something further upstream: the one who wrote the fitness function before the search began, set the guardrails around what the loop is allowed to try, and, in proposals like Bengio's, designed a second, deliberately powerless system whose only job is to keep watching. That is a much thinner point of control than a human reading every generation used to be, because a fitness function is written in advance, before the search has revealed what it will actually find, and Goodhart's law is precisely the observation that a target set in advance is the easiest kind of target to satisfy without achieving the thing it was meant to represent. If Stage 2's four branches were true possibilities to hold at once rather than a forecast, this is the fifth: not a claim that this outcome is coming, but a claim that the checkpoint people currently assume still exists, a human somewhere in the loop who would notice, is the exact thing speed is already being traded away, and that the field's own proposed fix is to rebuild that checkpoint deliberately rather than assume it survives by default.
Seed for a future stage
In short: Stage 5 named three points on a chain, nature-based evolution, then human-based evolution, then AI-based evolution, without yet asking what point comes after AI. That question is recorded here in its plainest form and deliberately left unexplored, a marker for a future stage rather than an argument this version makes. A second marker is added here too: the three points may not be a strict one-way chain at all, but parallel, ongoing processes that can feed information back into each other, each one potentially improving the level below it rather than only building on top of it.
Stage 5 named three points on a chain: nature-based evolution, then human-based evolution, then AI-based evolution. Written out plainly, the chain and the open question at the end of it read like this:
Nature-based evolution, then human-based evolution, then AI-based evolution: what comes next, and what kind of abstraction does that next step bring?
This is deliberately left here as a seed rather than a section. It is not investigated, argued, or answered in this version, on purpose, so that whatever stage eventually takes it on starts from the question in its plainest form rather than from an answer already half-decided in advance.
A second, related marker worth recording alongside it: the three points do not have to be read as a strict relay, where each one finishes before the next starts. They may instead be three parallel, still-running processes that can feed useful information back into each other, each one capable of improving the level it came from rather than only building on top of it. Written the same way as the chain above:
AI-based evolution -> can provide useful information for -> human-based evolution -> and human-based evolution can improve -> nature-based evolution
This piece has already touched pieces of that loop without naming it as one. Directed evolution in the lab (Stage 2) is human-based evolution reaching back to improve nature-based evolution directly, choosing which mutations a real organism keeps rather than leaving that to blind selection. AlphaEvolve and the AI-designed bacteriophage genomes (Stage 5) are AI-based evolution producing results, a better algorithm, a viable genome, that a person then reads and folds back into human engineering or biology. Neither of those was framed as a loop when it first appeared here. Whether that reframing holds up, and what it would mean for nature-based evolution specifically to receive something useful back from a process several layers removed from it, is left here as a second question for the same future stage, not answered in this version either.