Numerical testing of evolution theories

TL;DR — In the early 1950s, on the IAS machine in Princeton, Nils Barricelli ran what were arguably the first artificial-life experiments: self-reproducing number patterns in a shared memory "universe" that spontaneously exhibited parasitism, symbiosis, and genetic crossing. His 1962–63 Acta Biotheoretica papers argued that evolution's creativity comes from symbiogenesis — the merging and recombination of independently reproducing entities — not from mutation plus selection alone. Every run plateaued within thousands of generations, and he said plainly why: the universe was too small. That was a scale failure, not a concept failure, and scale is precisely what we now have. With open-endedness a named research goal and 2024 results showing self-replicators emerging in random program soups, Barricelli's question — does symbiogenesis drive sustained complexity growth? — is finally testable.

The universe in 512 cells

Barricelli's setup is almost embarrassingly simple, which is part of its power. The universe is a one-dimensional circular array of cells (a few hundred on the IAS machine). Each cell holds an integer or is empty. One rule drives everything: at each generation, a number n in cell i copies itself to cell i + n — a gene's value is its reproductive behavior. When two copies target the same cell, collision rules decide the outcome: identical numbers coexist quietly; different numbers trigger deterministic "mutation" rules that may produce a new number or kill both.

The crucial consequence: a lone number cannot persist. Its copies march around the ring and collide with themselves or others. What survives are patterns — sets of different numbers arranged so that each member's copying behavior sustains the others. Barricelli called these symbioorganisms, and the name carries the thesis. In this universe, the unit of selection is cooperative from the very first generation. Symbiosis isn't an exotic late-stage phenomenon, as it was treated in mainstream 1960s evolutionary theory; it's the precondition for existing at all.

What he saw

Running this on the IAS machine from 1953 (with the 1962–63 papers consolidating a decade of runs), Barricelli observed, unprogrammed:

• Parasites: patterns that could not reproduce alone but hijacked the numbers of a host organism, sometimes driving hosts extinct and then dying themselves, sometimes settling into stable host–parasite coexistence — predator–prey dynamics in a memory array, years before anyone named the field.
• Crossing: when two organisms' territories collided, segments exchanged, producing recombinant offspring — occasionally fitter than either parent. Barricelli argued this was the digital analogue of sex, and the real engine of novelty.
• Self-repair and speciation-like phenomena: damaged organisms reconstituting, incompatible variants partitioning the ring.

And then, always: stasis. Within a few thousand generations one organism or a small stable ecology dominated, and nothing qualitatively new appeared. In the 1963 Part II paper he tried to break the plateau by giving organisms a phenotype: numbers were interpreted as strategies in a simple competitive game, so selection acted on performance rather than raw persistence. Improvement was real but slow and bounded.

His theoretical claim, stated against the grain of the Modern Synthesis then dominant: mutation-selection is a hill-climber; the creative moves in evolution — the jumps to new levels of organization — come from symbiogenesis, the fusion of previously independent replicators. Lynn Margulis would make the biological version of this argument famous a few years later (endosymbiotic origin of mitochondria, 1967) and win, eventually. Barricelli made the computational version first, with data.

Why it could not work then

The IAS machine had roughly 1,024 words of 40 bits — about 5 KB — and executed on the order of 10^4 operations per second. Barricelli's universes were a few hundred cells; storing even a handful of generations for analysis strained memory, and runs of tens of thousands of generations took his entire machine allocation (famously, he ran overnight when von Neumann's group wasn't using it).

The problem is that open-ended evolution appears to have a scale floor. A 512-cell universe can hold perhaps a dozen organisms of modest size. There is no room for a diverse ecology, no reservoir of neutral variation, no spatial refugia where losing lineages survive to innovate later. The population fixates, and fixation is the death of open-endedness. Barricelli understood this — he wrote about universe size limiting evolution — but he couldn't buy his way out. Roughly speaking, the gap between his machine and one modern GPU is ~11 orders of magnitude in operations per second and ~7 orders in memory; a datacenter adds several more.

There was a second, subtler limit: instrumentation. Barricelli analyzed his universes by printing number arrays and inspecting them by eye (some of the punch-card outputs survive and have been beautifully reconstructed by historians — George Dyson reproduces them in Turing's Cathedral). Detecting an organism, tracing a lineage, deciding whether complexity increased — all manual. Even if the IAS machine could have run a rich ecology, he could not have seen it.

What changed

Three things, all recent.

Scale changed what emerges. The 2024 "Computational Life" work (Agüera y Arcas et al.) dropped random Brainfuck-like programs into a shared soup with no explicit fitness function and watched self-replicators arise spontaneously and take over — but only above a certain scale and interaction density. Lenia (Chan, 2019 onward) showed that continuous cellular automata at modern resolution produce a zoo of lifelike, self-organizing creatures invisible at Conway-scale grids. The consistent lesson: emergent phenomena in artificial life are scale-dependent, and Barricelli was operating far below the interesting regime.

Open-endedness became a named goal. Stanley and Lehman's work, POET, and the 2024 DeepMind position that open-endedness is essential for superintelligence have turned "systems that generate novelty indefinitely" from a curiosity into a research program — one that conspicuously lacks a canonical, cheap, analyzable testbed.

Instrumentation exists. We can now run millions of universes in parallel (JAX-based alife like Leniabreeder and ASAL uses exactly this pattern), and — the part Barricelli could not have imagined — use learned models to detect organisms, cluster phenotypes, and quantify novelty automatically, e.g. by tracking compression-based complexity or foundation-model embeddings of universe states over time.

A 2026 revival

Keep from Barricelli: the shared-memory universe with no explicit fitness function, reproduction as a side effect of the entities' own dynamics, and collision-driven recombination as the primary variation operator. These are the features that made symbiosis and parasitism emergent rather than designed, and they distinguish his setup from Tierra and Avida, which hand organisms a copy instruction and (in Avida's case) an externally imposed reward schedule.

Replace: numbers-with-shift-semantics with something more expressive — short programs in a redundant instruction set, or continuous neural-CA patches — so phenotypes can grow in complexity without the genotype space saturating. Run at datacenter scale: universes of 10^9–10^{12} cells, spatially structured (locality gives refugia and parapatric speciation for free), for 10^7+ generations.

Then ask Barricelli's question quantitatively, which he never could:

1. Does symbiogenesis drive complexity? Run matched universes with recombination/fusion enabled vs. ablated (mutation only). Measure sustained growth of organism complexity — compressed description length of dominant lineages over time. Barricelli's thesis predicts a qualitative separation: mutation-only universes plateau, symbiogenetic ones don't (or plateau much later).
2. Does evolvability evolve? Track whether lineages' variational properties — mutation robustness, recombination compatibility — improve over deep time, a live open question in both alife and evolutionary biology.
3. Where is the scale floor? Sweep universe size across six orders of magnitude and map the phase transitions: no life → replicators → ecologies → (hopefully) unbounded novelty. The 2024 program-soup results suggest these transitions are sharp and findable.

Total cost of a serious version is modest by 2026 standards — comparable to a mid-sized LLM training run — and unlike an LLM run, every intermediate state is a scientific observation.

Descendants, vindication, and what's open

Barricelli has been partially vindicated on every front. Symbiogenesis is now textbook biology (Margulis). Recombination as the key variation operator is the core of genetic algorithms (Holland, 1975, who knew of this lineage of work) — and, in a form he'd recognize, of modern model merging and mixture-of-experts composition, where independently trained "organisms" fuse into fitter wholes. Digital ecologies were reborn in Core War, Tierra (Ray, 1991), and Avida (Lenski/Ofria/Adami, whose 2003 Nature result on the evolution of complex features is the closest thing to a Barricelli experiment done right). Historians — notably George Dyson and David Ackley — have restored his priority: he was doing artificial life a full generation before the term existed.

What remains open is exactly what stopped him: nobody has demonstrated sustained, open-ended complexity growth in a digital universe. Every system so far — Tierra, Avida, Lenia, program soups — eventually plateaus, just later and higher than Barricelli's did. Whether that reflects insufficient scale, impoverished physics, missing symbiogenetic mechanisms, or something fundamental is arguably the open question of artificial life. Barricelli's 1963 paper contains both the first evidence and the first hypothesis; sixty-three years later we finally have the hardware to test it.

Where to read it

The paper is at doi.org/10.1007/BF01556602 (Acta Biotheoretica, 1963 — this is Part II; Part I appeared in 1962, and the bibliographic details of the 1963 paper are verified). Read alongside: George Dyson's Turing's Cathedral (chapter on Barricelli, with reproductions of the original outputs and the human story), Ray's Tierra paper (1991) and Lenski et al., "The evolutionary origin of complex features" (Nature, 2003) as the direct descendants, Agüera y Arcas et al., "Computational Life" (2024) as the scale-changes-everything result, and the DeepMind "Open-Endedness is Essential" position paper (2024) for why the question now has an audience. For the biology Barricelli anticipated, Margulis's Symbiosis in Cell Evolution closes the loop.