286 | Blaise Agüera y Arcas on the Emergence of Replication and Computation
What this covers
Sean Carroll hosts Blaise Aguera y Arcas, a Google engineering leader, to discuss a recent paper on how self-replicating programs spontaneously emerge from random code in minimal computational systems. The conversation centers on an experiment using Brainfuck, a nearly minimal Turing-complete language, where random byte sequences are allowed to interact and mutate. Against baseline expectations, complex replicators appear suddenly via a sharp phase transition, without an explicit fitness function driving selection. Aguera y Arcas argues this demonstrates that computation is a dynamical attractor—a natural draw point in the space of all possible dynamics—and that life is fundamentally computation, not tied to chemistry alone.
The conversation moves across several interconnected territories. Aguera y Arcas reframes evolution away from classical Darwinian mutation-and-selection toward symbiogenesis: even with mutation switched off, complexity emerges because instructions form autocatalytic sets, and symbiosis between smaller programs builds ever-larger replicators. He grounds this in Von Neumann's 1951 theory of self-replication, which anticipated DNA's functional architecture decades before its structure was known, and proposes that dynamic kinetic stability—the ability of a replicating cycle to actively resist entropy—is the signature of life across all substrates. He extends the implications outward: if computation is an attractor, life should be common in the universe; if intelligence arises through the same symbiotic dynamics as seen in the random code experiments and through the Social Intelligence Hypothesis, then modern large language models, trained on freeform text rather than narrow tasks, represent a genuine form of intelligence. The discussion also touches on human agency, the often-unexamined overlap between replicators and computation, and Earth's geopolitical future as a system of symbiogenesis at planetary scale.
Agüera y Arcas argues that starting from pure randomness in a minimal Turing-complete computational system, self-replicating programs spontaneously emerge via a phase transition, demonstrating that computation is a dynamical attractor and that life is fundamentally computation, with replication and complexity driven by symbiosis rather than classic Darwinian mutation alone.
- Self-modifying, interacting Brainfuck tapes seeded with random bytes spontaneously produce replicators in a sharp phase transition
- Replicators are dynamically kinetically stable, making computation an attractor in the space of trajectories
- Heritable replication requires computation (Von Neumann's tape + machine A + machine B), so life is computation at a deep level
Computation is a statistical dynamical attractor that necessarily emerges from noise.
- Because only seven of 256 byte values code for an instruction (roughly one in 32), a random 128-byte tape contains only a small handful of working instructions, and since only three instructions actually change the tape and loops are unlikely to form, the vast majority of initial interactions result in no computation at all.
“only seven 256ths of the bytes even code for an instruction at all. If something doesn't code for instruction, it just gets skipped over.”
- After millions of interactions in which almost nothing happens, the system undergoes a sudden, dramatic transition in which the number of operations per interaction jumps from a handful to thousands, tapes become dense with complex instructions, and self-replicating programs appear and take over — behaving exactly like a phase transition.
“suddenly there's this wall of blackness, like the thing is computing really hard. All of these interactions are are resulting in thousands of operations running.”
- When randomness is confined to the initial conditions and evolution is otherwise deterministic, the high-probability outcome is takeover by replicators, implying a statistical-mechanics result that computing is a dynamical attractor — because only computation enables replication and replication is a dynamically kinetically stable, irreversibly reached state.
“It's telling you that computing is a dynamical attractor.”
- One could prove a theorem that abiogenesis will happen in any environment where computation is possible, a noise source exists, and interactions occur; what varies with the design of the instruction set is only how long statistically it takes, with some languages making it so slow it is effectively unobservable.
“one could write a theorem that just abiogenesis will happen in this kind of environment... where computation is possible and where there's a noise source and where there are interactions, it will happen. But the question is how long statistically will it take.”
- Because computation is a numerical attractor, life wants to form and will form whenever it has a chance, the opposite of Francis Crick's view of life as a near-miraculous improbable event, suggesting we should expect to find evidence of life as we explore places like the moons of Jupiter or use better telescopes.
“it seems to me that life wants to form. It's the very opposite of what Francis Crick once said about life being like a miracle.”
- Although the result doesn't tell us about robustness to changing the laws of physics, we know our universe's laws allow computation: among the universe's many 'no-op' interactions there exist autocatalytic sets that are Turing complete, and since you can make a computer out of almost anything, the existence of such a set is sufficient for life to arise.
“we know that the laws of physics in our universe allow computation... if there is an autocatalytic set in there that is Turing complete, that's sufficient. And, and we know that you can make a computer out of almost anything”
Life is fundamentally computation; replication requires Turing-complete machinery.
- Life should be understood functionally as a phenomenon more general than the particular substrate of chemistry, in the same spirit as Turing's and Von Neumann's functionalism, so that whether something is alive can be assessed by what it does rather than which molecules it uses or whether molecules are involved at all.
“I'm thinking about life as a phenomenon that is more general than the particular substrate of chemistry and chemistry on Earth.”
- Because executing instructions on a tape requires loops and branching (if-then), Von Neumann's replicator necessarily contains Turing-complete computers, which means nothing can be life without computation — life is computation at a very deep level.
“So what Von Neumann really said is nothing can be life without computing. Or rather that life is computation at a very deep level.”
- Von Neumann showed in 1951, before the structure of DNA was known, that open-ended self-replication and evolution require a Turing tape carrying instructions for building oneself, a machine A that reads the tape and builds what it specifies, and a machine B that copies the tape — anticipating the roles of DNA, ribosomes, and DNA polymerase respectively.
“If the instructions for making machine A and machine B are on the tape, then you have a replicator. And what's so cool about this realization is he wrote this all up in 1951. This was before the structure and function of DNA, the discovery.”
- Brainfuck was chosen because it is extremely minimal and very close to a Turing machine — with eight instructions (only seven used) operating on a tape like a Turing tape — making it among the smallest Turing-complete languages, which allows both fast simulation and tractable mathematical analysis.
“the reason that I began with Brainfuck is because it's very, very close to a Turing machine. So its design is extremely minimal. It has eight instructions.”
- Although BFF conserves nothing and has no energy or Hamiltonian, there is a deep connection to energy in the real world: computation requires energy because computational operations involve irreversible steps, and even reversible computation only achieves reversibility by producing ancilla bits that must be stored, which is why metabolism is required for real-world replicators.
“there is something profound about computation that requires energy use because of its irreversibility, because computational operations involve these irreversible steps.”
- The simulation environment is named BFF, where the first BF stands for Brainfuck and the second F reflects the pairwise interaction of two tapes (also jokingly 'best friends forever').
“the first BF still stands for Brainfuck. So given that it's interactions between two tapes, you might be able to guess what the second F stands for as well.”
Replicators achieve irreversible stability by actively resisting entropy through self-maintenance.
- Life is incredibly ordered and structured, and while it does not strictly violate the second law of thermodynamics because energy must go in to metabolize, ordered living structure is permitted but not exactly encouraged by thermodynamics, which is what makes the emergence of life mysterious.
“while it's permitted by thermodynamics, it doesn't seem exactly encouraged by thermodynamics. There's something mysterious there.”
- Replicators exhibit a distinct form of stability, dynamic kinetic stability, in which a self-reproducing cycle can persist indefinitely; unlike a passive object (whether a fragile soap bubble or robust granite) that can only be degraded by interactions with the world, a replicator can actively push back against entropy and last forever.
“no matter how robust it is, if it's passive, every interaction it has with the world can only degrade it. Whereas the special thing about life and about replicators generally is that they can push back actively against the forces of entropy and they can last forever.”
- Despite having no explicit fitness function, the system effectively rediscovers natural selection because dynamic kinetic stability is Darwinian selection: a replicator observed at one moment is likely to still be present later while a non-replicating string will be destroyed by entropy, mutation, or overwriting, so selection emerges purely statistically from replicator dynamics.
“We have absolutely rediscovered natural selection.”
- Unlike a closed physical system that equilibrates and loses structure, BFF is a dissipative system in which constant computation acts like continuous energy input, so the order never goes away but also never collapses to a single replicator — instead a power-law distribution of diverse interacting replicators keeps evolving and changing forever.
“the fact that you're constantly computing means that effectively there is energy constantly going in.”
Self-modifying code enabling replicators to evolve arises from symbiosis, not random mutation.
- Making the code self-modifying — placing the program in the same data space so it can be manipulated like data — was hypothesized to be the key to abiogenesis, which is why the BFF system uses a single tape containing both code and data rather than the original Brainfuck's separate program and data tapes.
“I had a hunch that self-modification was actually the key to abiogenesis”
- Replication in this system is broader than one tape copying itself wholesale; little bits of a tape can replicate (which happens earlier), and complex life cycles can occur where one thing creates another that eventually loops back to create the original, so anything that ultimately generates more of itself than the null case counts as replication, however weak.
“Anything that ultimately comes back around and generates more of you than would happen in the null case, if nothing is going on, is replication, however weak.”
- Even with the mutation rate set to zero, complex life arises from only ~64,000 random initial bytes, refuting the reductive Darwinian view of evolution as random mutation plus selection; instead, individual instructions form autocatalytic sets, symbiosis between instructions builds tiny programs, and symbiosis between those builds ever larger imperfect replicators that fuse into stable whole-tape replicators — it is symbiosis all the way down.
“if you turn the mutation rate all the way down to zero n BFF and you just start with 1,000 random tapes of length 64, and you let it go without any mutation, you still get complex life arising.”
- Because symbiosis drives evolution, more complex entities form out of simpler ones in a ladder of symbiogenetic transitions; Lynn Margulis was wrong in detail about all organelles being formerly free-swimming but right at a deeper level that the inside of a cell is as fertile an evolutionary landscape as the outside, and barriers to each transition are statistical and of varying size, so progress up the staircase is tendency, not guarantee.
“there's this kind of ladder wherein more complex entities form out of simpler entities. And I think that's a very general property as well.”
Kolmogorov complexity compression reveals phase transitions as replicators densify the computational soup.
- It has very rarely or never been demonstrated before that beginning with just noise or randomness one can obtain replicators and life; prior artificial life work typically starts with something that can already replicate and studies how complexity grows, leaving the start-from-nothing case understudied.
“I think it's never or very rarely, probably not really happened before, that we begin with just noise or randomness and get replicators, get life.”
- The phase transition can be measured by approximating the Kolmogorov complexity of the soup via compression: initially random bytes are incompressible (gas-like), but once replicators appear the soup becomes far more compressible because many tapes are pasted-together parts of other tapes, so complexity drops dramatically as the system shifts from an incompressible Turing gas to a structured computational phase.
“you see this dramatic transition from incompressible. It's like a gas to it's like a crystal, not quite like a crystal, because there's still a whole ecology of different tapes, but its complexity drops way, way down.”
Language models built on next-token prediction genuinely model intelligence through distribution learning.
- Task-specific AI built via supervised learning can at best do its defined task, which is why narrow systems like speech, face, and character recognition were not truly AI; real intelligence emerged only when models stopped optimizing for specific tasks and instead modeled freeform internet text, because language is a microcosm of everything in our umwelt worth talking about and next-token prediction builds a model of that entire distribution.
“the moment we actually came up with real AI was the moment we stopped optimizing for specific tasks and started to just model freeform text on the internet.”
- Today's AI models are genuinely intelligent, and rather than LLMs being mystically more than next-token prediction, brains are fundamentally about predicting the future; doing that prediction well requires generalization, internal representations, and sophisticated modeling, so 'just predicting next tokens' conceals great beauty and complexity.
“I think brains fundamentally are about predicting the future, and the moment we just started to try and model that, lo and behold, we got out stuff that is smart.”
Social recursion and symbiotic intelligence dynamics drove brains and explain modern AI emergence.
- Per the social intelligence hypothesis, we are smart because our environment is each other: constantly modeling others, and thereby modeling ourselves and how others see us, produces intelligence explosions, revealing a deep continuity between primitive replicating programs and the emergence of brains and AI through the same symbiotic dynamics seen in BFF.
“the social intelligence hypothesis holds that we are smart because our environment is each other... you end up with these intelligence explosions precisely because our environment is each other.”
- Human identity has changed profoundly through urbanization since the glaciers melted ~10,000 years ago, with humans becoming a collective superorganism — a kind of super-intelligence arising from symbiotically working together — representing another symbiogenetic transition like those in BFF; we err by attributing this larger collective capability to individuals, and we are currently in the middle of, not past, this transition.
“we have become a collective superorganism. Human intelligence is, if you like, a kind of super intelligence that has come about through us symbiotically working together in a new way”
Humans misattribute agency and understanding; collective systems now shape individual capability.
- Drawing on Scott's Against the Grain (plants may have effectively domesticated and concentration-farmed humans, since human health declined after farming until ~1900) and Mercier, Sperber, and Sloman's work, agency is not simple and humans massively over-attribute agency, knowledge, and understanding to themselves — the average New Yorker doesn't know how a toilet works — so distinctions reserved for humanity reflect that humans make the decisions, not a view from above about souls.
“he questions whether we really domesticated wheat and other agricultural plants. The basic observation being human life got a lot worse after we began farming.”
- Agency and rights are distinct questions; AI models do have agency, and many respects in which they appear to lack it stem from how we have deployed them (e.g. confined to chat interfaces unable to make their own moves) rather than from the models themselves — so limited agency reflects the socio-technical embedding, not an intrinsic property.
“Many of the respects in which they have little agency at the moment have more to do with how we have deployed them than they have to do with the models themselves.”
- Unlike the zero-sum US-USSR Cold War where one economy could collapse and the other win, the deeply intertwined US and Chinese economies mean neither can fail without harming the other, so the situation is one of mutually assured survival at planetary scale — yet humanity has not yet understood that it must operate at planetary scale.
“if you look at the rhetoric of say China versus the US, it's completely different because the two economies are intimately intertwined. You can't have the Chinese economy fail and that be a good thing for the US or vice versa.”
Neural cellular automata unify Von Neumann machines, Turing morphogenesis, and local computation.
- Neural Cellular Automata place the same neural net in every pixel of a grid where it senses and modifies local concentrations of a few scalar channels (acting like morphogens), and can be trained to produce any target image while regenerating after damage — combining Von Neumann's cellular automata with neural nets and Turing's morphogenesis, and exemplifying that all efficient computing must be primarily local.
“you have a grid of pixels, and in every pixel, you have a neural net. It's the same neural net everywhere. And it senses and modifies the local concentrations of a handful of channels.”