Computational Abiogenesis

The thesis that life’s origin is best understood as the spontaneous emergence of self-replication in a computational substrate — and the experimental program (centered on Google’s Paradigms of Intelligence team, 2023–present) that studies this emergence in soups of random programs. The strong form of the thesis holds that computation is a dynamical attractor: wherever a physical environment can support computation, replication will tend to arise, because replicating entities are more dynamically stable than non-replicating ones and replication requires computation. ¹

The intellectual foundation

Darwin declined to touch the origin of life (“it is mere rubbish, thinking, at present, of origin of life,” 1863). The modern framing descends from two results that predate molecular biology’s details: Turing’s 1936 universal machine, which defined computation itself, and john-von-neumann’s late-1940s theory of self-reproducing-automata, which showed that any self-replicating system must separate a description (copied passively) from a constructor (reads the description and builds the machine). DNA polymerase and ribosomes turned out to instantiate von Neumann’s copier and constructor respectively — before anyone knew DNA’s structure. ²

Von Neumann’s result carries the load-bearing implication: a universe whose physics does not support computation cannot contain life. The open question is the inverse — given computation, how readily does life arise?

The bff experiments

Agüera y Arcas and colleagues (Google Paradigms of Intelligence + University of Chicago) built “primordial soups” of random 64-byte tapes interpreted as programs, principally in bff, a dialect of Brainfuck in which code and data share one tape and I/O heads sit on the tape itself, so programs are self-modifying. Two tapes are repeatedly drawn at random, concatenated, executed as one 128-byte program, and split back apart. There is no fitness function and no seeded replicator; a low-rate background byte randomization is optional. ³

The reported sequence (arXiv 2406.19108, Aug 2024):

1. Quiescent phase — few valid instructions execute; changes are mostly random, but imperfect replicators (short code fragments with any nonzero probability of generating more code) appear early and rise steadily. Code begets code: only code can produce more code, so a selection process operates before any true replicator exists.
2. Transition — after millions of interactions, whole-tape replicators emerge and take over the soup; computation per interaction jumps from a few operations to hundreds (reproduction requires loops, hence branching instructions).
3. Complexification — replicators diversify, evolve robustness to mutation, and interact symbiotically; the soup’s compressed size collapses as one lineage dominates.

The result reproduced across substrates: Forth variants, Z80 and 8080 machine code (including a 200×200 grid visualization by Mordvintsev), with and without background mutation (≈40–50% of runs transition within the epoch budget in the main configurations). Counterexample: SUBLEQ, a one-instruction language, admits hand-written replicators (60 bytes; 25 in a variant) but showed no spontaneous emergence even after billions of executions — the authors conjecture replicator length relative to tape size matters without establishing it. ⁴

The team’s interpretation: computation is a dynamical attractor because replication is a kinetically stable state and only computation yields replication. Life, on this view, is the predictable outcome of a statistical process — “minimal conditions: a physical environment capable of supporting computation, a noise source, and enough time.” ⁵

The 2026 critique

Knierim, Versari, Obryk, Agüera y Arcas, and Saurous revisited the claim with direct self-replicator detectors (the original paper inferred emergence from soup compression, i.e., takeover). Findings (arXiv 2607.01483, July 2026): ⁶

• Emergence and takeover are distinct: replicators arise under random mutation alone; interactions are what spreads them through the soup.
• A distributionally tuned random-walk mutation search finds BFF self-replicators at least as easily as the paired-interaction soup — so interaction is not an unusually powerful search operator in this system.
• Capping ancestry-tree depth/width does not prevent emergence, only takeover.

The attractor thesis survives in weakened form: computational substrates do generate replicators readily, but the bff soup’s specific interaction mechanism is not doing special work — random search in program space already does it. The interesting variable shifts to detection and diffusion, not creation.

Framing: the major transitions

The broader theoretical frame is Maynard Smith and Szathmáry’s major evolutionary transitions (1995): a small set of events that changed how information is stored and transmitted — replicating molecules → compartmented populations, unlinked replicators → chromosomes, RNA → DNA/protein, prokaryotes → eukaryotes, asex → sex, unicellularity → multicellularity, solitary life → colonies, primate societies → language. Their shared signature: entities that once replicated independently can afterward replicate only as part of a larger whole. ⁷

Agüera y Arcas extends this list with technology: the human–technology symbiosis is, on his account, the most recent major transition — “technology is not distinct from nature or biology, but merely its most recent evolutionary layer.” This is Margulis-style symbiogenesis generalized past the cell, and it rhymes with Teilhard’s noosphere (a new layer accreting on the biosphere) reached from the opposite, mechanistic direction. ⁸

The functionalist corollary

The article’s philosophical payload: if what matters is function, substrate is incidental. A neuron replaced by a faithful computational model would leave the brain — and “you” — unchanged; “we are made out of functions, and those functions are made out of functions, all the way down.” This is the substrate-independence premise of intelligence-as-tool (“meat that wields mind”) with an experimental origin story attached. Nathan’s marginal note on the DNA-ribosome passage flags the tension: the (nuclear) proteome’s emerging importance complicates the clean tape/machine picture without touching the general insight. ⁹

Limitations

• No chemistry, no energy. The soups model no catalysis, no metabolism, and no thermodynamic cost — the article gestures at Landauer’s principle (irreversible computation dissipates free energy) but the experiments abstract it away. Origin-of-life theories centered on autocatalytic networks or metabolism-first are unaddressed. ¹⁰
• Substrate dependence. SUBLEQ’s non-emergence shows the result is a property of particular instruction sets, not of computation as such; which substrates permit spontaneous replication is open.
• Critique’s deflation. Per the 2026 follow-up, the soup mechanism itself is not demonstrated to be special — random mutation suffices for the creation step.
• Simulation is not proof. As Sean Carroll’s podcast discussion with Agüera y Arcas put it, this is a computational model of the pre-life→life transition, not evidence that biological life began this way.

Connections

• self-reproducing-automata — the theoretical ancestor; bff soups are von Neumann’s question run as an experiment (emergence rather than hand-design).
• john-von-neumann — the copier/constructor architecture the experiments implicitly vindicate.
• intelligence-as-tool — substrate-independence of function, applied to minds.
• noosphere — the technology-as-evolutionary-layer thesis in its mystical register; same claim, different epistemic style.
• evolutionary-game-theory — imperfect replicators competing in a soup as the zero-institution baseline that game-theoretic dynamics later elaborate.
• technological-singularity — if replication-with-complexification is an attractor, intelligence explosions are one more transition, not a rupture.

Sources

• Blaise Agüera y Arcas 2024 — In the Beginning, There Was Computation — highlight-linked primary (Nautilus, Aug 2024; Readwise body)
• Blaise Agüera y Arcas, Jyrki Alakuijala, James Evans, Ben Laurie, Alexander Mordvintsev, Eyvind Niklasson, Ettore Randazzo, Luca Versari 2024 — Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction — primary research (arXiv 2406.19108v2)
• Charlotte Knierim, Luca Versari, Robert Obryk, Blaise Agüera y Arcas, Rif A. Saurous 2026 — BFF: Simple explanations for complex phenomena — primary research (arXiv 2607.01483; the critique)
• Derek Turner 2019 — Philosophy of Macroevolution (Stanford Encyclopedia of Philosophy) — SEP reference (major-transitions framework)

Footnotes

1. Blaise Agüera y Arcas 2024 — In the Beginning, There Was Computation ↩

2. Blaise Agüera y Arcas 2024 — In the Beginning, There Was Computation ↩

3. Blaise Agüera y Arcas, Jyrki Alakuijala, James Evans, Ben Laurie, Alexander Mordvintsev, Eyvind Niklasson, Ettore Randazzo, Luca Versari 2024 — Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction ↩

4. Blaise Agüera y Arcas, Jyrki Alakuijala, James Evans, Ben Laurie, Alexander Mordvintsev, Eyvind Niklasson, Ettore Randazzo, Luca Versari 2024 — Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction ↩

5. Blaise Agüera y Arcas 2024 — In the Beginning, There Was Computation ↩

6. Charlotte Knierim, Luca Versari, Robert Obryk, Blaise Agüera y Arcas, Rif A. Saurous 2026 — BFF: Simple explanations for complex phenomena ↩

7. Derek Turner 2019 — Philosophy of Macroevolution (Stanford Encyclopedia of Philosophy) ↩

8. Blaise Agüera y Arcas 2024 — In the Beginning, There Was Computation ↩

9. Blaise Agüera y Arcas 2024 — In the Beginning, There Was Computation ↩

10. Blaise Agüera y Arcas, Jyrki Alakuijala, James Evans, Ben Laurie, Alexander Mordvintsev, Eyvind Niklasson, Ettore Randazzo, Luca Versari 2024 — Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction ↩