After Mutation Became Cheap

Evolutionary computation was designed when producing a candidate was work.

An algorithm encoded a population, perturbed its members, evaluated them, selected survivors, and repeated. Mutation had to be engineered carefully because variation was a scarce resource generated inside the optimization loop.

Generative models invert that economy.

Programs, hypotheses, designs, models, explanations, experiments, and strategies can now be proposed faster than we can understand them. Humans generate candidates. Agents generate candidates. Optimizers generate candidates. Running systems generate accidental candidates whenever conditions expose a new behavior.

Candidate production is becoming abundant. The surrounding machinery for evaluating, relating, and retaining those candidates has not kept pace.

The old loop

A classical evolutionary algorithm often assumes a clean object:

population
  -> evaluate
  -> select
  -> mutate / recombine
  -> next population

The individuals share a representation. The environment is fixed. The genotype-to-phenotype map is given. Fitness has a known shape. Generations are synchronized. The mutation operator stands outside the population and remains unchanged while its candidates evolve.

These assumptions make the optimization tractable. They also remove many of the things that matter in a long-running adaptive system.

Real candidates are heterogeneous. They encounter different environments at different times. Their evaluations have costs and uncertainty. Their representations change. Some are compositions of other candidates. Some remain useful as detectors despite being poor actions. Some should sleep for months and return when the context changes.

Most importantly, the machinery proposing and judging them can learn too.

A better covariance update inside CMA-ES touches only a small part of this change. The larger question is what evolution means when candidates, environments, evaluators, and mutation operators are all persistent objects.

The population already exists

Every active project contains a population, although it is rarely treated as one.

There are competing implementations, abandoned branches, rival explanations, parameterizations, notebooks, prompts, tests, models, diagrams, and partially successful experiments. Some are explicit alternatives. Others differ in ways nobody has recorded.

An AI-assisted project produces this population especially quickly. One conversation may generate dozens of plausible branches. Most become orphaned as soon as the next answer is selected.

The result is abundant proposals with weak heredity, inconsistent environments, and little provenance around selection.

Each candidate needs to retain:

• what produced it;
• what it inherited;
• where it behaved differently;
• which environments evaluated it;
• what evidence supports or contradicts it;
• what resources and authority it received;
• what descendants learned from its outcome.

This is why lineage is a memory you can walk. Lineage separates accumulated variation from an unordered collection of outputs.

Fitness as conditional evidence

Fitness is often presented as a number attached to an individual.

In an adaptive system, the more honest object is a distribution of evidence conditional on an environment:

candidate x environment x time x mode x constraints -> evidence

The evidence can include objectives, behavior, uncertainty, resource use, failures, and consequences. A utility function may reduce it for a particular decision, but that reduction should not overwrite the evidence from which it was made.

A candidate can therefore be dominated under one lens and indispensable under another. It may lose as an actuator but survive as a sensor. It may be economically useless at one fee tier and efficient at another. It may fail globally while occupying the only known niche that responds coherently to a rare condition.

Selection allocates resources under a stated context.

Survival may mean continued deployment, more evaluation, becoming a parent, retaining a niche, being compiled into a shared component, or remaining dormant with a precise reason for possible return.

In this vocabulary, death means loss of active resources while the memory remains available.

Evolution across scales

The boundary of an individual need not remain fixed.

A signal transform can be a candidate. A policy consuming several signals can be a candidate. A controller allocating among policies can be a candidate. A graph containing all three can be evaluated as a larger candidate. At a slower clock, the agent modifying that graph and the strategy it uses to choose experiments can also vary.

What appears to be an environment from below may appear to be a component from above.

This resembles a recurring pattern in living systems: competent units combine into larger units that operate on different spatial and temporal scales. The larger unit cannot process every lower-level event. It perceives coarse patterns, supplies constraints or signals downward, and relies on local competence to fill in the details.

Multi-clock intelligence gives this hierarchy a computational form. Fast artifacts react continuously. Slower evaluators accumulate evidence. Still slower agents alter representations, populations, and research programs.

Evolution can occur within a clock and across clocks. A successful slow insight may be compiled into a fast signal. A repeated fast failure may escalate into a change of language or evaluation at the slower layer.

When evolution evolves

Once a system can inspect its own search process, the mutation operator can enter the candidate set.

An agent may notice that numeric perturbations explore the wrong surface and introduce a structural mutation. It may discover that an important distinction is missing from the behavior descriptors. It may add a new evaluation environment, alter the language in which candidates are expressed, or replace the operator choosing parents.

The evolution skill becomes subject to evolution.

This recursion is powerful and dangerous. If candidates are allowed to reshape their own judge without preserving the previous evidence contract, adaptation collapses into result hacking. A system can become excellent at selecting the environments in which it already looks good.

Immutable provenance around each observation provides the anchor while fitness definitions change:

what was evaluated
under which contract
against which source
with which available information
and what actually happened

Evaluators, descriptors, and lenses may evolve, but they cannot rewrite old evidence into a stronger grade. A new lens can reinterpret an old result. It cannot claim that the result came from an experiment that never occurred.

Selection becomes the scarce resource

When mutation is cheap, candidate count ceases to be a useful measure of progress.

The scarce resources move elsewhere:

• environments capable of distinguishing candidates;
• exact evaluations with verified semantic paths;
• counterexamples that divide one apparent behavior into two;
• compute and real-world opportunities;
• attention able to understand why a branch failed;
• memory that prevents rediscovery from replacing accumulation.

The evolutionary operator should choose which uncertainty is most valuable to reduce before creating another child. A deliberately chosen experiment may contribute more than a hundred inexpensive mutations evaluated under weak pressure.

This changes the role of diversity too. Useful diversity is measured by consequential disagreement among candidates that retain some competence. Their disagreement increases what the system can perceive, as described in a map made of reactions.

The population becomes both a repertoire of possible action and an evolving instrument for constructing state.

An ecology of artifacts

The optimization loop becomes a persistent ecology of typed artifacts.

Candidates are born asynchronously. They inherit mechanisms through composition as well as mutation. They encounter overlapping but non-identical environments. Resources flow toward some and away from others. Useful fragments can move horizontally between lineages. Acquired information can be deliberately compiled into descendants.

Biology supplies structural clues rather than a specification. Intelligence can accumulate through populations whose interactions, boundaries, and environments change over time.

The central operation maintains a field in which evidence can change what receives attention, what reproduces, what acts, and what becomes part of the substrate itself.

Generative intelligence supplies variation. Lineage supplies heredity. Evaluation supplies pressure. Niches supply conditional persistence. Replay supplies memory.

One question remains: who or what decides where the system’s scarce resources and real authority should flow?

That is the right to act.