From Replication to Reasoning
Replication is necessary but not sufficient for intelligence. The gap between self-copying programs and abstraction-capable systems is the central mystery. Three approaches might bridge it: environmental coupling, meta-learning, and predictive coding.
BFF shows that self-replicators emerge spontaneously from random programs. ARC shows that intelligence requires abstraction from few examples. What's between them?
Can evolutionary dynamics produce systems that learn within their lifetime, not just across generations? Can replicators develop internal models that transfer to novel situations?
Abstraction emerges when environments demand it. If survival requires predicting patterns, organisms that develop internal models outcompete those that don't.
In a static environment, hardcoded responses win. But in environments with:
...organisms that model their environment gain advantage over those that merely react.
Evidence for: Coevolution in Avida dramatically increases complexity. Host-parasite arms races force organisms to develop more sophisticated behaviors.
Evidence against: Increased complexity ≠ abstraction. Avida organisms develop better heuristics, not transferable reasoning.
Learning is replication at the algorithm level. Just as genes replicate through organisms, learning algorithms replicate through problem-solving episodes.
Consider what a "learning algorithm" does:
This is autocatalysis at the algorithmic level. The algorithm catalyzes its own improvement. If evolution can produce molecular autocatalysis, perhaps it can produce algorithmic autocatalysis.
Evidence for: SOAR (Self-improving Language Models for Evolutionary Program Synthesis) achieves 52% on ARC by combining evolutionary search with self-improvement.
Evidence against: SOAR requires explicit meta-learning objectives. The learning loop is engineered, not emergent.
Friston's Free Energy Principle proposes that all adaptive systems minimize variational free energy—a measure of surprise. Both perception and action serve this goal.
Agents minimize F by either:
If replicators are implicitly minimizing free energy (staying alive = minimizing surprise), perhaps explicit learning is just efficient free energy minimization.
Evidence for: In vitro neural networks experimentally validated the free energy principle in 2023.
Evidence against: Free energy minimization is a framework, not a mechanism. It describes what systems do, not how to build them.
A comprehensive literature review on evolutionary learning theory is available: evolutionary-learning.md (15+ sources with full citations).
The question "Can evolution produce learning?" has a rich theoretical foundation. Key findings:
Livnat & Papadimitriou (2014) proved that sexual evolution under weak selection is mathematically equivalent to the Multiplicative Weights Update Algorithm—a powerful online learning algorithm. Evolution doesn't merely resemble learning; the equations are identical.
Hinton & Nowlan (1987) demonstrated computationally that learning "smooths" the fitness landscape. Organisms that learn can find solutions that pure evolution cannot reach, then genetic assimilation can make those learned behaviors innate. Learning guides evolution.
West-Eberhard (2003) argues that "genes are followers, not leaders." Phenotypic plasticity—organisms adapting to their environment—generates the variation that evolution acts upon. What organisms learn to do, their descendants may be born knowing.
AutoML-Zero (2020) evolved complete machine learning algorithms from basic operations, discovering backpropagation and dropout without human guidance. Clune's AI-GAs framework (2019) proposes this as the path to general AI: evolve the architectures, the learning algorithms, and the learning environments.
The literature suggests three key insights:
The remaining question is whether this transition can be observed spontaneously in minimal systems, without explicitly engineering meta-learning objectives.
Embed simple ARC-like tasks into BFF simulations. Reward programs that solve transformation puzzles. Measure whether evolved programs generalize to unseen variants.
Model abstraction as lossy compression of environmental regularities. Define metrics connecting Kolmogorov complexity, logical depth, and ARC performance.
Study how human children develop abstraction. What's the minimal environmental structure? What core knowledge priors are truly innate vs learned?
Combine genetic algorithms with program synthesis. Evolve populations of programs that solve ARC tasks. Use novelty search to escape local optima.
This research question has been asked for 30+ years. Tierra (1991), Avida (1998), NEAT (2002), POET (2019)—all explored whether evolution produces learning. None demonstrated spontaneous emergence of within-lifetime abstraction.
The "soup" framing is aesthetic, not functional. Population → variation → selection → repeat. Whether you call it soup, ecosystem, or GA, the dynamics are identical. No primordial soup has produced qualitatively different behavior than standard evolutionary algorithms.
Backprop, TD-learning, evolutionary strategies—all need fine-grained credit assignment. Evolution provides only coarse-grained credit (offspring success). No theoretical framework predicts spontaneous fine-grained credit assignment emergence.
A Turing-complete system can compute learning algorithms, but won't spontaneously produce them. The Kolmogorov complexity of gradient descent is hundreds of bits. Probability of random emergence: negligible.
For this research to contribute novelty, it must demonstrate one of:
Phase 1: Accelerate BFF complexity through environmental structure and coevolution.
Phase 2: Introduce external resources; measure within-lifetime adaptation.
Phase 3: Embed ARC-like tasks; test for transfer learning.
Create resource gradients in the BFF soup. Different regions provide different "nutrients" (tape values that aid replication). Measure: do organisms specialize? Do niches form?
Introduce parasites that exploit replicators. Measure: does an arms race develop? Does complexity increase sustainably?
Spatially isolated populations with periodic migration. Maintains diversity, enables parallel exploration.
Map tape regions to environmental "food." Programs that write to certain addresses gain replication advantage. Environment varies in learnable patterns.
Does the same organism improve performance during execution? Track: instructions executed before successful replication. Does this decrease over organism's "lifetime"?
Programs must solve simple transformation puzzles to access resources. Puzzles vary but share underlying rule. Measure: do programs generalize to unseen variants?
Use novelty search instead of fitness. Reward behavioral diversity. Escape local optima.