What Most People Get Wrong About Evolution | Akarsh Kumar

Summary

• Tim Scarfe’s closing correction is that evolution is “anything but random”: selection preserves partial wins, allowing some exponential combinatorial searches to become linear ones. In Core War, Akarsh Kumar’s LLM-guided evolutionary loop beat 96% of roughly 300 human-written programs even though the model was poor at zero-shot Redcode. The engineering signal is the combination of imperfect model priors, objective verification, and cumulative selection—not the LLM alone.
• Artificial life expands the research target from life as observed on Earth to “life as it could be,” searching across possible physics, chemistries, organisms, and intelligences. Kumar argues that one cannot derive general principles from a single specimen any more than one can inspect one car and explain all cars. For investors, the important optionality is a computational laboratory for testing which rule systems generate persistence, self-organization, and open-ended complexity.
• Kumar distinguishes today’s “statistical intelligence” from a path-dependent, regularity-based intelligence whose structure records how it was acquired. Humans and evolution complexify sequentially, while models can be trained on calculus without first learning arithmetic; Kumar allows that this alien approach “could just be better,” but suspects the path supplies unusually strong regularization, robustness, and generalization.
• Foundation models can serve as scalable proxies for human judgment when properties such as emergence, complexity, or interestingness resist clean mathematical objectives. The workflow is deliberately irreducible: “Just simulate it. See what happens,” let a vision-language model judge the result, then search again. This broad semantic target also makes optimization easier—“any cat” is more reachable than one exact pixel-level cat—and reduces the risk of finding what was specified rather than what was wanted.
• Open-ended behavior appears concentrated in narrow regions of rule space, not spread uniformly across it. After running and embedding all 260,000 rules in one experiment, Kumar’s team found a large island of non-open-ended systems and a much smaller island containing the most compelling simulations. That “Goldilocks zone” makes automated discovery valuable: the scarce object is not compute alone, but the right artificial physics.
• The simulations repeatedly show global organization emerging from minimal local interactions, including cell-like structures that exploit their artificial physics to persist. Game of Life needs only binary cells and a few neighborhood rules; Boids turns local steering into flocking; Particle Life uses a 6×6 interaction matrix among six particle types and can produce moving objects with membrane-like coatings and differentiated interiors.
• Artificial life remains a long-duration AGI wager rather than a near-term claim that particle-level simulation will directly yield intelligence. Kumar’s agenda is to extract an efficient abstraction of evolution that preserves its useful properties, potentially combining statistical models as critics with organically factored systems as the thing being created. The exchange also cautions against putting all eggs in one basket.

Deep dive

1. General theories require studying life and intelligence beyond their single known forms

• Kumar defines artificial life as the study of “life as it could be rather than life as it is.” Biology describes Earth’s current instance; artificial worlds let researchers seek mechanistic principles spanning “the space of all possible intelligences” and life forms. Looking only at the human brain, he argues, is like inspecting one car and declaring that all cars work identically.

• The unusual contrast is with contemporary model training: animals follow a natural sequence—“I’m going to learn this and then this and then this”—whereas a model can learn calculus without first learning arithmetic. Kumar concedes that this alien flexibility “could just be better,” but sees reasons to investigate whether the universe fundamentally rewards sequential construction.

• Ordinary curriculum learning is insufficient because a fixed syllabus already assumes the destination. Drawing on Why Greatness Cannot Be Planned, Kumar wants a “serendipitous curriculum” that can explore any path leading somewhere interesting while retaining useful discoveries accumulated along the way.

• Kumar’s honest uncertainty is central: he cannot yet explain why complexification should produce better representations. Natural evolution, human learning, and theories such as assembly theory all appear sequential, but proving that sequence is necessary—or proving it is not—“would be a huge discovery.”

2. Path-dependent regularities offer an alternative to purely statistical intelligence

• Tim Scarfe connects the argument to computational irreducibility: acquiring a complex representation may require synthesizing its building blocks because “there are no shortcuts.” Kenneth Stanley’s related framing is that benchmark performance records where a system arrived, while “it matters not just where you get but how you got there.”

• Novelty search illustrates why evolution is not equivalent to blind randomness. Random search forgets its history; novelty search knows what has already happened and avoids repeating it. That remembered path constrains future exploration, potentially steering it strongly toward structured representations rather than merely reducing the number of options.

• Scarfe frames this path dependence as unusually powerful regularization. He describes data as a form of regularization, weight decay as favoring low-norm solutions, and curriculum learning that respects learned symmetries as another, potentially much stronger, form. The discussion proposes studying “regularity-based intelligence” alongside statistical intelligence, rather than abandoning the latter.

• Scarfe presents the FE approach as automatic symmetry learning. CNNs encode translation invariance and transformers encode permutation invariance, but a hand-designed equation for daylight-versus-nighttime invariance is harder to imagine. In Picbreeder’s evolved solutions, the inductive bias need not occupy the whole search space: “The regularities are baked into the circuits.”

3. Artificial universes turn emergence into a searchable scientific object

• Artificial life need not be grounded in our world because researchers can choose the grounding: alternate physics, artificial chemistry, or altered element interactions. Kumar’s target question is counterfactual—why does one physics generate open-ended complexity while a nearby one collapses into “just random noise”?

• Conway’s Game of Life is the clean specimen: a 2D grid contains only alive or dead cells, with a few neighborhood rules such as a dead cell becoming alive when exactly three of its eight neighbors live. From this minimal setup come oscillators, spaceships, long-lived interactions, and “insane amounts of complexity.”

• The community even measures moving patterns against the system’s causal speed: information travels at most one cell per time step, called (C), so a glider might move at (C/16). Its toy scale is an advantage—an artificial universe simple enough to study rather than a claim that the toy already captures reality.

• Scarfe asks whether recurring gliders are “shadows” of something fundamental or artifacts of grids and interpreters. Kumar keeps both possibilities open: different worlds produced radically different gliders, including one resembling a train leaving tracks, but convergence across every rule within a substrate might reveal a deeper substrate-level constraint.

4. Local interaction rules can grow repair, flocking, and persistence

• Lenia makes cellular automata continuous and can yield patterns that creep across the screen. Neural cellular automata instead replace the local 3×3-patch update with a small neural network, creating a searchable space of physics in which local rules can repeatedly self-organize a target form.

• The 2020 “Growing Neural Cellular Automata” gecko captures the difference between appearance and generative competence. Delete part of the continuously updating lizard and it reconstructs itself: the system has learned an ability to rebuild the form from local interactions, not merely “what the lizard looks like.” Kumar’s extension is: learn the process that grows intelligence, not intelligence’s snapshot.

• Boids derives collective motion from local decisions. Kumar recalls fish shifting around an approaching whale shark: perhaps one fish detects it and tugs nearby fish left, propagating coordinated motion. Parameterizing each agent’s decision rule lets researchers ask how altered local policies reshape the global flock.

• Particle Life assigns particles six color-coded types and a 6×6 matrix of attractions, repulsions, or neutral interactions. Some matrices produce bland noise; others yield moving, cell-like objects with red boundaries and differentiated interiors. Kumar reads them as structures “exploiting the physics” to persist—not universal life, since persistence appears only in some simulations.

5. Foundation models make vague emergence objectives operational at scale

• The paper Automating the Search for Artificial Life with Foundation Models replaces one hand-coded simulation with a parameterized family. Specifying 100 possible elements without prescribing their interactions is easier than inventing an entire periodic table; an automated process can search the missing physics.

• The method embraces computational irreducibility rather than predicting dynamics analytically: “Just simulate it. See what happens. Plug it into a foundation model.” The model supplies a scalable approximation of human judgments such as complexity or interestingness, where brittle formulas invite Goodhart’s law—“something that you asked for but not something that you wanted.”

• Semantic objectives widen the solution set. Asking for “a cat” avoids supervising one exact cat pixel by pixel; asking for any open-ended simulation avoids demanding a predetermined state sequence. The foundation model’s fractured internal representation is acceptable here because it acts only as a critic whose output bootstraps a more organically structured system.

• The resulting rule map was sharply uneven. The team ran all 260,000 candidate rules, embedded their outputs with CLIP, and found a large island of non-open-ended behavior beside a small island containing the “coolest simulations.” A more ambitious thought experiment—explicitly speculative—would search quadrillions of particles for physics progressing from replicators to predator-prey dynamics and alien animals.

6. LLM-guided evolution converts weak proposals into cumulative program search

• Scarfe presses on the gap between intriguing simulations and useful general intelligence. Kumar calls artificial life “a very long-term bet”: the goal is not to simulate every low-level particle, but to extract an ultra-efficient abstraction of evolution that preserves the properties needed for intelligence—something efficient enough to run 10 years from now.

• Core War supplies a concrete test bed. Programs called warriors execute from random memory locations in a Turing-complete assembly-like language called Redcode; each tries to crash rivals by placing an invalid instruction before their execution thread while defending itself. Scarfe’s summary: “It’s like Game of Thrones, Turing edition.”

• The evolutionary loop uses an LLM as mutation operator and MAP-Elites because the landscape is deceptive and non-local. It evolves one warrior to beat the initial program, another to beat both predecessors, then repeats, preserving an expanding adversarial lineage rather than greedily optimizing against only the latest opponent.

• The first round beat 96% of a dataset of roughly 300 human warriors, despite poor zero-shot and best-of-(N) Redcode generation. Later warriors increasingly defeated unseen humans while behavioral variance declined toward a generalist. Evolution needs mutations to work only “1% of the time”: selection retains each useful fragment, enabling the next search to build on it.