Evolution Is an Optimizer, and Its Misses Matter for AI

Natural selection built eyesight, immune systems, and wings. It did not bother much with keeping humans mentally sharp at 70.

That sounds like a failure until you state the objective correctly. Evolution never promised long life, maximal intelligence, or graceful old age. It optimized reproductive success under brutal environmental noise, short time horizons, and a search process with severe limits. Once you look at it that way, some of biology’s strangest omissions start to look familiar to anyone training modern AI.

Jacob Kimmel of NewLimit has a useful way to frame the puzzle. The question is not “why didn’t evolution solve aging?” The better question is “why would we expect this optimizer to place much weight on aging in the first place?” That shift matters, because it turns biology into a giant case study in objective design, credit assignment, and constrained search.

If you build models, this should feel uncomfortably recognizable.

Evolution optimizes under the wrong reward for long life

People often talk about evolution as if it were an engineer refining a product. That language helps until it misleads. Natural selection is closer to a noisy optimizer that gets one scalar reward signal: did variants leave more descendants in a particular environment than nearby variants did?

That objective is narrower than most human intuitions about success. A body that survives childhood, reaches reproductive age, produces offspring, and helps enough kin may score well even if it falls apart later. From the genome’s perspective, a dramatic improvement in tissue repair at age 75 can matter less than a tiny reduction in infection risk at age 12.

This is the first lesson for AI. The solution space can contain beautiful outcomes that never appear during training if the loss barely notices them. People sometimes speak as if capabilities emerge simply because they are useful in some broad sense. Usefulness is not enough. The optimizer must receive enough signal, in the right place, often enough, to climb toward that outcome.

Biology gives us a vivid example. Long healthy life is obviously valuable to us. It was not necessarily valuable, in a strong selection sense, across most of human evolutionary history. High mortality from infection, injury, predation, childbirth, and plain bad luck meant relatively few individuals reached advanced ages. Traits whose benefits appear mostly after those ages sit behind a foggy reward signal.

That is not philosophy. It is optimization.

Long horizons erase the gradient

In reinforcement learning, long-horizon tasks are notoriously hard. If an agent gets a meaningful reward only after a long sequence of actions, credit assignment becomes weak and noisy. Which early action mattered? Which change to the policy improved the final outcome? The signal diffuses across time until training starts to look like guessing with expensive hardware.

Evolution has the same problem, except its training runs are measured in generations.

Suppose a genetic variant slightly improves brain maintenance after age 60. For that variant to spread, enough carriers must survive to 60, and their improved cognition must still affect reproductive success or inclusive fitness. In ancestral populations, that pathway existed, but it was thin. The gradient was faint.

Kimmel’s comparison to RL lands because the timeline is absurdly long. Ask a current model to optimize for an outcome that appears after twenty years, then provide reward as a single scalar tied to descendants, and suddenly the mystery flips. It becomes surprising that any useful long-term biological optimization happened at all.

This helps explain a pattern people find unsettling once they notice it. Fluid intelligence tends to peak relatively early. Repair systems that keep tissues pristine do not scale indefinitely. Many degenerative processes wait until later life to reveal themselves. If selection mostly “saw” the first half of adult life, then the second half was never going to receive the same engineering budget.

There is a nuance worth keeping. Humans are unusual among mammals because many individuals do survive well past reproduction, and kin can matter. The grandmother hypothesis exists for a reason. Older adults can improve descendant survival through care, knowledge, and social stability. That means late-life function was not invisible. It just did not carry the same weight as surviving infections, making it through childbirth, or staying fertile long enough to raise children.

AI has its own version of this weighting problem. Models can, in principle, learn extended planning, sustained reliability, and deep internal verification. In practice, if the training signal mostly rewards short-turn performance, those traits emerge weakly or inconsistently. We then act surprised when the system is eloquent for thirty seconds and flaky over a long workflow. Biology did that tradeoff first.

Aging can look like a length penalty

There is another twist, and this one is more counterintuitive. Even if long life had some positive value, indefinite maintenance may still be disfavored because bodies are expensive.

Every organism lives inside resource constraints. Calories spent on repair are calories not spent on growth, reproduction, immune defense, or offspring. Ecology imposes a budget. Evolution allocates that budget where the expected return is highest. The result is not a body optimized for permanence. It is a body optimized for enough maintenance to get through the relevant reproductive window, with some spillover beyond it.

Biologists have several theories that cash this out in different ways. Disposable soma says organisms invest in repair only up to the level that maximizes fitness under expected conditions. Antagonistic pleiotropy says variants that help early life can spread even if they damage late life. Mutation accumulation says harmful late-acting variants are weakly selected against because their costs arrive after many carriers have already reproduced. These are different models, but they rhyme.

The AI analogy is a length regularizer on reasoning. Give a model unlimited chain-of-thought and some tasks improve, but latency rises, cost rises, and failure modes can expand. So people impose penalties. Think longer when the gain is worth it. Stop when extra computation has diminishing returns.

Evolution appears to impose something similar on organismal persistence. Keeping every individual around, fully maintained, for a very long time slows turnover and ties up resources in bodies with declining marginal reproductive value. That does not mean aging was “chosen” in a simple conscious sense. It means the optimization process can prefer a world with faster replacement over one with maximal individual persistence.

This is easier to see outside humans. Annual plants race through a season and die. Pacific salmon make a single massive reproductive push and collapse. Even in species with longer lifespans, the amount of maintenance varies wildly depending on ecological niche. Nature does not have one policy. It has many local bargains between repair and reproduction.

Humans complicate the picture because culture changes the payoff structure. An elder in a foraging society can contribute knowledge that no younger individual has. In modern societies, older adults often create far more value than their calories cost, which is an embarrassingly low bar if we are being honest. But cultural evolution moves much faster than genetic evolution. A genome tuned for Pleistocene hazards does not instantly update because pension funds and cloud computing appeared.

That lag matters. It is one reason the aging problem may contain tractable interventions. If selection underweighted late-life maintenance, there may be repair pathways that can be improved without violating deep physical laws. They may have been ignored, not impossible.

The search process limits the biology you can reach

There is a deeper lesson beneath the objective. Even if evolution “wanted” more longevity or more intelligence, it still had to find them through a weak and local search process.

Think about mutation rate first. In machine learning terms, it behaves a bit like step size. If variation is too low, adaptation crawls. If variation is too high, the system destabilizes. In multicellular organisms, pushing mutation too far is not an abstract risk. It shows up as developmental failure, loss of function, and cancer. Evolution cannot simply crank exploration and hope for the best.

Population size plays a similar role. Larger populations sample more variants in parallel and make rare useful combinations easier to discover. Bacteria enjoy this advantage. Humans do not, at least not in the same way. Long generation times make the search slower still. A species that reproduces quickly can explore a landscape that is effectively closed to one that reproduces slowly.

Then there is loss weighting. Evolutionary pressure is not distributed evenly across problems. Historically, infectious disease, injury, fertility, and early development commanded enormous weight because they killed or disabled individuals before reproduction. If your optimizer is spending most of its budget on surviving pathogens and getting babies through birth, there is less pressure left for pristine arterial maintenance at 85 or perfect episodic memory at 70.

This is where AI researchers should resist a common simplification. Universal approximation is not the same as practical learnability. A function class can represent a solution without training ever finding it. Biology gives us the same warning. The existence of long-lived species proves that the broad space of life can support extreme maintenance. Bowhead whales can live for centuries. Naked mole rats resist cancer unusually well. Some organisms regenerate astonishingly. Existence proofs are real. They do not imply your lineage can reach that basin of attraction easily.

That gap between representable and reachable may be the most important connection between evolution and AI. We often ask whether a model architecture can, in principle, do some task. The more important question is whether the objective, data, optimization dynamics, and compute budget make that outcome likely. Evolution says no, loudly, across millions of years.

Intelligence follows the same logic

People are tempted to treat human intelligence as evidence that evolution relentlessly drives cognition upward. The actual pattern is narrower. Evolution increased intelligence where it paid off under real ecological pressures: social coordination, tool use, language, planning, and adaptation to changing environments. It did not optimize for general cognitive excellence across the whole lifespan.

That distinction helps explain the asymmetry many people notice with age. Our species is cognitively extraordinary, yet individual cognition is fragile. Processing speed falls. Working memory narrows. Neurodegenerative disease remains common. If evolution had a strong persistent signal for maintaining top-tier cognition deep into late life, these declines should be rarer and later. Instead, the maintenance curve looks exactly like a trait whose returns were front-loaded.

It also suggests caution about AI benchmarks. A system can look impressive on the tasks the training regime rewards and still remain mediocre on adjacent capabilities that matter to users. Human evolution produced a brain that is astonishingly good at social inference, pattern completion, and language, while still being terrible at remembering where it put the keys. Optimization creates peaks with weird contours.

The broader point is that intelligence is not one knob that gets turned up. It is a bundle of subcapabilities, each weighted differently by the training environment. Biology and machine learning both produce jagged profiles, not smooth superiority.

The opportunity sits inside the blind spot

Once you stop treating evolution as a perfect engineer, a more interesting possibility appears. Some of its misses may be shallow misses.

That does not mean aging is easy to solve, or that late-life cognitive maintenance is one clever intervention away. Biological systems are entangled, and many apparent improvements collide with cancer risk, developmental tradeoffs, or metabolic cost. When you remove one limit, another often introduces itself like an uninvited dependency in production.

Still, the optimization framing changes the default. If a trait was weakly selected, then poor performance on that trait does not prove the problem is intrinsically hard. It may simply mean the search process never spent much compute there. In AI terms, the model was undertrained on a capability users later decided they cared about.

That is why the analogy is useful for both biotech and AI. In biotech, it suggests there may be genuine room for progress by directly optimizing repair, resilience, and cellular maintenance beyond what natural selection bothered to discover. In AI, it warns against reading current model behavior as the limit of what the architecture can become. Many failures are not statements about possibility. They are statements about gradients, objectives, and budgets.

Evolution remains the oldest optimizer we know, and it is less a master builder than a patient tinkerer with a noisy scoreboard. Its record is impressive because life persisted, adapted, and diversified under absurd constraints. Its blind spots are impressive too. They show how much can be left on the table when the reward arrives late, the regularization is harsh, and the search barely reaches the region of the map where the better answers live.