Evolution Is All You Need
Creativity, AI, and Open-Endedness
Creativity has long been regarded as one of humanity's defining traits, a mysterious, almost magical faculty. Yet, I propose that creativity—both human and artificial—reduces fundamentally to evolutionary processes: variation and selection.
Creativity as Evolution
At its core, evolution consists of two essential steps:
Variation: Generating novelty (random mutations, recombinations, or trial ideas).
Selection: Systematic retention of successful variations based on context-specific criteria.
Importantly, while variation is random (or unbiased regarding future utility), selection is decidedly non-random, systematically filtering ideas toward increasingly sophisticated outcomes.
This distinction is crucial: creativity isn't randomness. It's randomness harnessed by purposeful selection. Misunderstanding this distinction is a primary reason creativity feels mysterious.
Responding to Common Misconceptions
Critics often object to evolutionary explanations of creativity by equating evolution to mere randomness. In fact, only variation is random; selection introduces purposeful guidance. Consider these illustrative examples:
Biological Evolution: Genetic mutations occur randomly; natural selection retains those mutations that enhance fitness in a particular environment.
AlphaGo and AlphaZero: Moves explored through Monte Carlo sampling are random, but selection (choosing winning strategies) is structured, explicit, and goal-directed.
Generative Adversarial Networks (GANs): Random variations in output images become highly realistic through rigorous selection pressures imposed by a discriminating network.
These examples illustrate that creativity emerges through directed search: randomness provides novelty, selection provides purposeful direction.
The Open-Endedness Challenge
However, critics like David Deutsch and Brett Hall argue we don't fully understand creativity or evolution because they exhibit a property termed "open-endedness." This means:
Evolution continuously generates novel complexity with no fixed endpoint or predefined goal.
Solutions to existing problems create new environments and challenges dynamically.
Deutsch and Hall note that simulations of evolution typically fail because they use static fitness functions and environments, unable to replicate true long-term innovation. They interpret this as evidence of a conceptual gap.
Addressing Open-Endedness
Their critique rightly highlights a challenge—but it's not a conceptual gap. The evolutionary algorithm—variation and selection—is fully understood conceptually. Open-endedness is not mysterious; it naturally emerges when:
Variation and selection operate recursively at multiple hierarchical levels (e.g., genes, organisms, ecosystems, ideas, cultures).
Fitness criteria evolve dynamically, driven by earlier solutions altering the landscape of future selective pressures.
Real-world creativity shows this hierarchical recursion clearly:
The evolution of flight created entirely new adaptive niches.
Market innovation continuously creates and destroys industries, redefining the competitive landscape.
Scientific progress continuously changes criteria of successful theories.
Thus, open-endedness emerges naturally from applying variation and selection iteratively and dynamically across multiple scales and shifting environments.
Computational Limitations vs. Conceptual Understanding
The difficulty in replicating true open-ended evolution computationally arises from intrinsic computational complexity, not from conceptual mystery. Full computational replication of indefinitely recursive, multi-level evolution might be practically infeasible (like fully simulating the weather indefinitely), but this does not mean the underlying conceptual model is incomplete or misunderstood.
Notably, significant partial simulations exist:
Avida and Tierra: Simulations that spontaneously generated entirely novel reproductive strategies and interactions, demonstrating partial open-ended innovation.
OpenAI Five, AlphaZero: Systems exhibiting genuine innovation by dynamically redefining their selective landscapes through iterative learning.
These examples, though limited, reinforce the robustness of the conceptual model.
Clarifying the Misplaced Expectation
Deutsch and Hall’s critique implicitly expects simulation completeness for conceptual validity. But simulation infeasibility doesn't equal conceptual incompleteness. The evolutionary process—variation and selection across recursive, dynamically shifting landscapes—is conceptually sound, even if computationally demanding.
Conclusion: Why Evolution Is Indeed All You Need
Creativity is neither mysterious nor merely random. It is evolutionary:
Randomness generates diverse variations.
Systematic, purposeful selection guides these variations toward increasingly sophisticated outcomes.
Open-endedness, far from disproving evolution as a complete explanation, highlights evolution’s remarkable generative power arising naturally from repeated, recursive application across shifting criteria and scales.
Deutsch and Hall's critique strengthens rather than weakens the evolutionary view of creativity, clarifying the profound simplicity and conceptual completeness of evolution as an algorithm of creativity.
In short: Evolution is indeed all you need.
References
Dawkins, Richard. The Blind Watchmaker. W.W. Norton & Company, 1986.
Dennett, Daniel C. Darwin's Dangerous Idea: Evolution and the Meanings of Life. Simon & Schuster, 1995.
Deutsch, David. The Beginning of Infinity: Explanations That Transform the World. Viking, 2011.
Hall, Brett. Creativity and Consciousness.
Silver, David, et al. "Mastering the Game of Go with Deep Neural Networks and Tree Search." Nature, 2016.
Goodfellow, Ian, et al. "Generative Adversarial Nets." Advances in Neural Information Processing Systems (NeurIPS), 2014.