One pattern, many substrates
Evolution can genuinely be read as a search process. A population explores a space of organisms: genotype → phenotype → fitness. Mutation generates nearby possibilities, selection favours some trajectories, and death removes the rest from further reproduction. That really does resemble stochastic optimization. Daniel Dennett called natural selection an algorithm, a mindless procedure that reliably produces design.
Science has the same shape: hypothesis → experiment → selection → new hypothesis. So do markets: strategy → profit or loss → capital allocation → new strategies. So do immune systems, cultural traditions and engineering practice. Underneath all of them:
Richard Watson and Eörs Szathmáry pushed this further in "How Can Evolution Learn?". They argue that evolution doesn't just resemble search but can be formally related to learning, so that evolving systems can, under some conditions, generalize from past selective environments the way a learning system generalizes from data. That's one of the papers we find most relevant to the whole of Rubedo.
No free lunch
Literature
If everything is search, is there a best search? Wolpert and Macready's No Free Lunch theorems say that, averaged over all possible problems, no optimization algorithm beats any other. Any algorithm that does well on some problems pays for it on others.
That matters a great deal for the idea of a "generalized optimizer." It means general optimization can't come from a universal algorithm that's good at everything. It has to come from matching the structure of the world we actually live in: locality, regularity, compositionality, the fact that the past predicts the future. Which brings us back to inductive biases (see Compute). Evolution's real achievement may be accumulating biases that suit this universe, not escaping bias.
Physics that looks like optimization
Literature Speculation
Several serious thinkers have suggested that something optimization-like is built into physical processes. These are contested ideas. We list them because they sharpen the question, not because we endorse them.
- Lotka's maximum power principle (1922): natural selection tends to favour organisms that capture and use more of the available energy flow.
- Dissipative adaptation (England, 2013): driven physical systems may tend to settle into configurations that are good at absorbing and dissipating energy, which could be a thermodynamic tilt toward self-replication.
- Causal entropic forces (Wissner-Gross and Freer, 2013): a system that maximizes the diversity of its accessible futures shows behaviour resembling tool use and cooperation in simple simulations, a physical cousin of empowerment.
- The free-energy principle (Friston): self-organizing systems that persist can be described as minimizing a bound on surprise. Whether it's an empirical theory or a reframing remains hotly debated.
- The universe's computational capacity (Lloyd, 2002): a physical estimate of how many elementary operations the universe could have performed. It's a reminder of how much "search" physics has done, and the backdrop to our compute question.
The speculative leap
Speculation
Our discussions have also wandered further, and we want to show that openly because it's part of how we think. The question is whether the metaphysics of reality could itself be described as an optimization problem:
- Death as a pruning mechanism, ending unsuccessful branches of a search.
- Quantum superposition as a set of candidate choices, with something like a simulation exploring several possible continuations from a given point and continuing only the promising ones.
- Reality as a state engine, stepping forward under rules that happen to reward persistence and adaptation.
We want to draw a hard line here. There is no scientific evidence that death exists to prune simulation branches, or that quantum superposition is an optimizer exploring possible universes. Everett's many-worlds interpretation describes branching, but nothing in it involves evaluation or selection of branches. Bostrom's simulation argument is a philosophical argument about probabilities, not evidence of any particular simulator or its goals.
As a conceptual model, though, it's intelligible. An optimizer able to branch simulations would plausibly behave like this: expand a state into several successors, evaluate the trajectories, and keep the promising ones. That's Monte Carlo tree search, beam search, evolutionary search. The analogy is mathematically coherent even though the metaphysical reading is speculative. We find it useful as a lens and nothing more.
A more defensible version: optimization becoming physical
Hypothesis
There's a reframing of the same intuition that we think holds up better:
If intelligence is the ability to search possible futures and preferentially bring about better ones, then intelligent systems may gradually make reality behave more like an optimization algorithm. Not because reality began that way, but because they acquire the power to turn more and more of the causal world into a controllable search space.
A primitive organism controls almost nothing. A person controls their immediate surroundings. A company controls factories and supply chains. A civilization controls rivers, power grids, agriculture, chemistry, satellites and information networks. As capability grows, the gap between predicting possible futures and choosing which one becomes real shrinks. Optimization becomes more and more physically instantiated.
On this view the "universe as optimizer" isn't a claim about physics. It's a claim about what intelligence does to the part of the universe it can reach. That connects directly to the physical substrate of power and to the question in Civilization as an optimizer: if more and more of the world is being optimized, what is it being optimized for?
Why a research project cares about any of this
Three reasons. First, the generate–evaluate–select–amplify pattern is a reminder that the evaluator is half the system. Every search is only as good as what it selects for, which is the incentives question in another form. Second, the No Free Lunch theorems keep us honest. There's no universal optimizer to discover, only better matches to the structure of a world. Third, being explicit about where analogy ends keeps the rest of the work clear. We'd rather show our speculation, labelled, than hide it inside claims that sound more scientific than they are.