Evolve the learner, not the behaviour
An idea that runs through Rubedo's history is genome → interpreter → phenotype. The inherited "genome" wouldn't encode behaviour directly, and wouldn't just be a bigger model. It would encode compressed, useful structure, including the machinery for learning. Development turns that into a plastic agent, and lifetime experience does the rest.
This separates two things that are easy to confuse. Being well adapted now is about current performance. Evolvability is the capacity of an inherited organization to produce useful, viable variation in the future. Wagner and Altenberg's work on the evolution of evolvability is a good starting point for the difference: how variation is organized may matter as much as what's currently encoded.
The Baldwin effect
Literature
The Baldwin effect links evolution and lifetime learning without requiring learned answers to be inherited directly. Selection favours inherited conditions that make learning faster or more reliable. Over generations, what had to be learned can become easier to learn, or partly innate. Hinton and Nowlan's 1987 simulation is the classic computational illustration. Fernando and colleagues (2018) demonstrated a modern version, with evolution shaping initial parameters and learning hyperparameters so lifetime learning was fast. That's an operational connection, though it uses designed inner learning algorithms and task distributions.
The Baldwin effect is different from Lamarckian inheritance, copying a parent's learned parameters into its offspring. Both are options for an artificial system. They have very different consequences for diversity and for how fast competence accumulates.
Why naive evolution destroys competence
Classical genetic algorithms mutate the whole parameter vector. For a network with hundreds of millions of weights that has spent a lifetime learning, independently scrambling weights is roughly like giving the child brain damage. Accumulated competence is destroyed faster than selection can preserve it.
This connects to an old question in biology. Eigen's error threshold describes how replication fidelity limits the amount of information a lineage can maintain. Mutate too much and selection can't hold the population near a good solution. Rubedo was interested in this early on, as a constraint on how much inherited information a developmental system can carry.
The fix we find most convincing is to separate timescales:
Evolution acts on how the learner learns, not on the raw result of learning.
Viable variation also builds on existing organization. Evolution doesn't generate each generation from independent random bits. Recombination, development and modularity can preserve structure while varying it, though they can also disrupt it. Greater evolvability is possible under some conditions. It isn't automatic or monotonic.
Changing the amount of randomness vs. the structure of variation
A useful distinction from our recent discussions: turning up the randomness is different from changing what kind of variation happens. Suppose an agent has a decent mining strategy. Bigger random perturbations might occasionally find something better, but they can also break navigation, survival and tool use all at once. A better kind of variation would preserve those competences while changing how the agent investigates unfamiliar things, allocates practice, or combines existing behaviours.
The goal is to raise the probability that a change produces a coherent alternative worth testing, without knowing in advance which alternative will succeed.
| Control | Pushed too little | Pushed too far |
|---|---|---|
| Magnitude of changes | Search stays near existing strategies | Useful competence is repeatedly disrupted |
| Scope of changes | Important reorganizations become unreachable | Too much changes at once to judge any one change |
| Selection pressure | Weak alternatives consume resources indefinitely | Unproductive-for-now stepping stones disappear |
| Evaluation duration | Delayed benefits never become visible | Each candidate becomes prohibitively expensive |
| Preservation of alternatives | The population collapses onto one approach | Memory and compute fill with redundant variation |
| Cost of added complexity | Unnecessary machinery accumulates | Useful investment is rejected before it pays off |
No universal settings guarantee rising capability or escape from local optima. These controls interact with the environment, the representation, the budget and the stage of learning. And more complexity isn't the goal in itself. A simpler mechanism that learns faster is progress. What we want to grow is the range and efficiency of useful capabilities.
Letting the system set its own controls
Hypothesis
The stronger question is whether a system could adjust these controls itself, without us programming the schedule. We don't want to write "if progress stalls for 100 steps, double the mutation rate."
An evolutionary approach makes parts of the search process heritable and variable. Lineages carry different tendencies for exploration, plasticity, retention or variation, and selection acts on the consequences. A fixed inherited mutation rate is only the first step. The more interesting possibility is a conditional mechanism that responds differently to stagnation, uncertainty, changing conditions or accumulated competence, competing with other mechanisms and being retained when it works.
Literature There's real evidence this territory can be explored experimentally:
- Population-Based Training (Jaderberg et al., 2017) discovers changing hyperparameter schedules through population selection and perturbation, though its own exploration and replacement rules are engineered.
- Discovering Evolution Strategies (Lange et al., 2023) meta-learns evolution-strategy update rules that transfer to unseen optimization problems, and even uses a learned strategy to drive its own outer optimization.
- DiscoRL (Nature, 2025) learned a reinforcement-learning update rule from populations of agents in diverse environments. The rule performed strongly on benchmarks, including environments held out during discovery.
- Self-referential meta-learning (Kirsch and Schmidhuber, 2022) demonstrates networks that modify themselves on simple tasks, within a designed resource-allocation procedure.
None of these shows unbounded self-improvement. Together they show that useful learning machinery doesn't all have to be specified by hand. We can engineer the initial capacity to adapt, leave much of how adaptation proceeds learnable, and test whether those learned mechanisms transfer.
Some machinery always determines what can change, how it runs, and what counts as success. Biological evolution has such boundaries too: chemistry, physics, reproduction. The meaningful question isn't whether a system has literally no predefined mechanism. It's how much adaptive freedom exists inside the boundary.
A ladder of search
Put together, these results suggest a hierarchy we keep coming back to:
A conventional optimizer looks for the best parameters and, conceptually, stops. An open-ended system keeps finding new behaviours, representations, tools, niches, subgoals and ways of learning. That changes the question from "how do we optimize a model?" to "how do we build a process that keeps creating better search processes?" The top rung of this ladder is where the idea of recursive self-improvement lives.
Heredity and gestation
Earlier direction
Evolving a learner raises its own hard problems: what a genome should contain, how two genomes can be recombined without producing broken offspring, how much mutation inherited information can survive, and whether all of this can be afforded at all. These now have their own page, Heredity.
One idea from that line of work belongs here: artificial gestation. Biological organisms do an enormous amount of developmental computation before they interact meaningfully with the world. An artificial organism might similarly go through an internal "dreaming" phase before deployment, in which it builds its architecture, does predictive learning, consolidates inherited structure, simulates experience and calibrates its memory. If the genome carries learning machinery rather than finished knowledge, gestation becomes part of the developmental program that evolution can shape.
From evolving solutions to improving the search process
Open question A pretrained model can propose programs while an evolutionary outer loop selects and preserves alternatives. AlphaEvolve reports useful discoveries with this combination; the Darwin Gödel Machine reports improved coding-agent implementations. These support investigation of bounded process improvement, with different objects of search.
Rubedo's proposed first target is a small controller: mutation strength, archive selection, curriculum order or a tool workflow. Compare fixed, hand-tuned, random-search and adaptive controllers under equal full budgets. Keep evaluation outside the controller's write access and test successive rounds on new task families. A higher score on the tuning tasks is insufficient; an improved process should make later learning or discovery more effective after its own cost is counted.
Teacher knowledge, stronger base models, extra search and easier curricula can each explain apparent self-improvement. Verification costs and hardware constraints can also halt the loop. The frontier agenda keeps sustained recursive improvement as a hypothesis, and the meta-optimization pilot states a narrower test.
What we don't know
- How to vary representations, learning rules or developmental processes without losing accumulated knowledge.
- How to select for evolvability when its benefits take a long time to appear.
- Whether learned search mechanisms beat fixed ones at comparable total compute, including the cost of discovering them. Otherwise the improvement may just come from spending more.