In June 2024, Google's Paradigms of Intelligence team, collaborating with University of Chicago sociologist James Evans, published an unusual experiment on arXiv. The team created 131,072 short programs, each 64 characters long and composed of randomly selected letters. The programming language used was deliberately primitive, featuring only ten basic commands such as "add one, subtract one, copy from here to there." In each round, programs were randomly paired, concatenated, executed, and then split again. No tasks were assigned, and no scores measured success.1

In a representative run, a self-replicating program first appeared in the 2,354th cycle and quickly filled the entire computational environment with its copies. This transformation occurred in 40% of a thousand runs within 16,000 cycles. The frequency remained largely unchanged even when random letter modifications were entirely disabled, indicating that the replicators arose from programs altering each other's code.1

In this article:
  • How a Google executive defines life in a single sentence.
  • How a mathematician described the cell's replication mechanism five years before the structure of DNA was discovered.
  • The origin of the protein that forms the human placenta and a gene involved in memory.
  • What the same experiment yielded on a 1976 processor, and in which language it produced no results at all.

What does Blaise Agüera y Arcas consider central to his definition of life?

The paper's lead author is Blaise Agüera y Arcas, Google's Technology Director for Technology and Society and founder of the foundational AI research team. His 624-page book, What Is Intelligence?, was published by MIT Press on September 23, 2025.2 In a comprehensive interview with Big Think, he offers his own definition of life, acknowledging biologists' lack of consensus: life is computation that builds itself and complexifies through partnerships.3

In this definition, the material composition is secondary. What makes something alive is not its being made of carbon or protein, but rather its function as a system that grows, repairs itself, or reproduces, carrying its own blueprint within itself. Agüera y Arcas measures intelligence with the same lens, arguing that intelligence is tested by behavior: whether a system can solve a new problem in a way that would be impossible without an underlying world model.3

How did John von Neumann envision self-replication before DNA's structure was known?

The interview's backbone is John von Neumann, the Hungarian-born mathematician who, on September 20, 1948, at the Hixon Symposium in Pasadena, elucidated the logic of a self-replicating machine.4 The problem was a vicious circle: a machine needed to carry its own blueprint to build itself, but it also needed to copy that blueprint, leading to an infinite regress if the blueprint had to describe the blueprint of the blueprint.

Von Neumann broke this loop through a division of labor. His machine contained an instruction tape. Part A would read the tape and construct what was described, while Part B would copy the tape character by character without interpreting it. The blueprints for both A and B were also contained on the same tape. James Watson and Francis Crick published the double-helix structure of DNA in April 1953. In the cell, the tape corresponds to DNA, Part A to the ribosome that builds proteins, and Part B to the DNA polymerase enzyme that copies DNA.

Sydney Brenner, one of the founders of molecular biology, devotes considerable space to this lecture in his memoirs. He notes that if the history of ideas were written solely from published documents, it would seem that Watson and Crick drew upon von Neumann's work; however, the two parties were unaware of each other's research.5

Mavi, kırmızı ve yeşil hücrelerden oluşan, sol alttan sağ üste doğru basamak gibi dizilmiş üç benzer otomat ve onlardan sağa uzanan uzun talimat şeritleri
A self-replicating automaton by Renato Nobili and Umberto Pesavento, which expanded von Neumann's cellular rules and ran on a computer. The copy produced by the bottom-left machine has begun to build its own copy. Image: Golly screenshot, Wikimedia Commons, public domain.
Get science and technology articles in your inbox
Once a week, we send articles like this in a single email. No ads, unsubscribe with one click anytime.

What is the role of ancient viruses in the human genome?

The second half of the definition is partnership. According to Agüera y Arcas, complexity largely increases through the combination of separate entities forming a new whole, and traces of this process are evident in the human genome itself. According to the initial draft of the human genome published in 2001, only 1.1% to 1.4% of our DNA codes for proteins. In contrast, sequences resembling retroviruses constitute 8.3% of the genome.6

Some of these remnants are functional. In February 2000, Sha Mi and her team introduced a protein called syncytin in Nature. This protein is produced from the envelope gene of an ancient retrovirus named HERV-W and forms giant cells in the outermost layer of the placenta by fusing cells. Adding antibodies that inhibit this protein stops the fusion.7 The ancestor of this virus entered the primate lineage 25 to 40 million years ago.8

The second example in the interview is the Arc gene. Agüera y Arcas refers to it as a virus, though sources are a bit more cautious. According to a study published in Cell in 2018 by Elissa Pastuzyn and her team, Arc evolved from mobile genetic elements called Ty3/gypsy, which are ancestors of retroviruses; thus, it is a close relative rather than the virus itself. The Arc protein forms capsules resembling viral capsids and transfers RNA between nerve cells within these capsules, playing a role in neural communication.9 Mice with the Arc gene deleted in 2006 retained their short-term memory but could not form lasting memories.10

Where does artificial intelligence fit into this evolutionary narrative?

Agüera y Arcas places artificial intelligence within the same narrative, but his own criteria temper his enthusiasm. A phone can perform any calculation, yet it cannot reproduce itself. Current AI models also do not construct themselves; humans build them in factories and data centers. He acknowledges this, recalling a statement by astronomer Robert Jastrow in the early 1980s: we are the reproductive organs of computers.3

He extends the idea of partnership back to the steam engine. Fossil fuel energy increased the world population from one billion to eight billion; he estimates that seven out of eight people alive today exist thanks to those machines. The United Nations announced that the eight billion threshold was crossed on November 15, 2022.11 Agriculture, medicine, and clean water also played significant roles in population growth, so this ratio should be read as a rough estimate rather than a precise measurement.

More ambitious statements in the interview, such as the claim that general artificial intelligence has already arrived, represent the author's personal views and are highly contentious in the field. The portion demonstrated in the lab is narrower: self-replicating code can emerge from aimless interaction in a primitive language. The team also replicated the experiment in the Forth language and using the instruction set of the Zilog Z80 processor, which was released in 1976, and replicators again appeared. However, in the SUBLEQ language, which relies on a single subtraction command, no self-replicating programs spontaneously emerged in any run.1

Notes and Sources

  1. Blaise Agüera y Arcas, Jyrki Alakuijala, James Evans, Ben Laurie, Alexander Mordvintsev, Eyvind Niklasson, Ettore Randazzo, Luca Versari, "Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction", arXiv 2406.19108, 2024. arxiv.org ↩
  2. Blaise Agüera y Arcas, What Is Intelligence? Lessons from AI About Evolution, Computing, and Minds, MIT Press / Antikythera, September 23, 2025, 624 pp. mitpress.mit.edu ↩
  3. "Google CTO: A unifying theory of life, intelligence, and technology", Interview with Blaise Agüera y Arcas, Big Think, Full Interview series. bigthink.com ↩
  4. John von Neumann, "The General and Logical Theory of Automata", Cerebral Mechanisms in Behavior: The Hixon Symposium, ed. L. A. Jeffress, New York: Wiley, 1951, pp. 1-41; talk September 20, 1948. See also A. W. Burks (ed.), Theory of Self-Reproducing Automata, University of Illinois Press, 1966. ↩
  5. Sydney Brenner, My Life in Science, London: BioMed Central, 2001, pp. 33-36. historyofinformation.com ↩
  6. International Human Genome Sequencing Consortium (Lander et al.), "Initial sequencing and analysis of the human genome", Nature 409, 860-921, 2001. doi.org/10.1038/35057062 ↩
  7. Sha Mi et al., "Syncytin is a captive retroviral envelope protein involved in human placental morphogenesis", Nature 403, 785-789, February 17, 2000. doi.org/10.1038/35001608 ↩
  8. Valérie Cheynet et al., "Synthesis, Assembly, and Processing of the Env ERVWE1/Syncytin Human Endogenous Retroviral Envelope", Journal of Virology 79(9), 5585-5593, 2005. doi.org ↩
  9. Elissa D. Pastuzyn et al., "The Neuronal Gene Arc Encodes a Repurposed Retrotransposon Gag Protein that Mediates Intercellular RNA Transfer", Cell 172(1-2), 275-288, 2018. pubmed ↩
  10. Niels Plath et al., "Arc/Arg3.1 is essential for the consolidation of synaptic plasticity and memories", Neuron 52(3), 437-444, 2006. doi.org ↩
  11. United Nations, "Day of Eight Billion", November 15, 2022. un.org ↩