What If Intelligence Didn't Evolve? It "Was There" From the Start! - Blaise Agüera y Arcas
Summary
David Krakauer defines life by function: it is embodied, autopoietic computation, not a special kind of material. A kidney’s identity depends on what it does, whether built from tissue, tungsten or carbon nanotubes: “If you break a rock on a non-living planet, you now have two rocks. You don’t have a broken rock.” Function creates the distinction between working and broken—and between life and inert matter.
His BFF experiment shows random code undergoing an abrupt transition into a computational ecology. The system repeatedly pairs random 64-byte tapes written in a seven-instruction, embodied version of Brainfuck; in one run, average execution rises from two operations per interaction to 1,374, with the jump arriving near six million interactions. Simultaneously, the initially incompressible soup becomes highly compressible: “I think you have to call that phase of matter life.”
The episode’s central revision to standard evolutionary intuition is that novelty can arise through fusion even when mutation is exactly zero. Primitive one-byte replicators occasionally cooperate and begin reproducing as a larger unit, creating information about “how the two fit together.” Krakauer identifies this symbiogenesis with a gelation transition and argues that evolution is “symbiogenesis all the way down,” not merely mutation plus selection inside a predefined design space.
A causal intervention suggests rare compositional events, not gradual background activity, trigger the regime change. Capping replicator ancestry at a tree depth around 24 requires blocking only one interaction in roughly 1,000, yet prevents gelation; complex programs appear to require depths of about 20 or more. The investor-relevant pattern is discontinuity: a tiny fraction of interactions can determine whether an ecosystem stays stable or crosses into runaway complexity.
The proposed mathematics adds a merger operator to ordinary population dynamics. The R term captures reproduction, competition and Darwinian optimization, while K captures components combining into new units—“evolution” on the left and “revolution” on the right. Future fusion partners are already unusually cooperative and low-rank, while changes in the Jacobian’s leading eigenvalues can signal that the stable ecology is approaching a phase transition.
Krakauer argues that biology contains the same nested architecture at every scale. Only 1.5% of the human genome codes for proteins, while much of the remainder includes transposons and endogenous retroviral elements; his sharpest example is Arc, an endogenized viral element whose knockout in mice prevents new-memory formation. Genomes become “replicators made of replicators made of replicators,” with major transitions such as eukaryotes merely the most visible cases.
The AI-relevant call is that intelligence may be an ecology-level consequence of living computers composing and modeling one another. Fusion creates massively parallel computation that must represent both self and environment, especially other agents: “Life is never single player.” Energy still constrains complexity, but cooperation can improve energetic scaling and greater intelligence can unlock new energy sources—shifting attention from isolated capability toward the compositional dynamics of systems and ecosystems.
Deep dive
1. Function separates life from inert matter
Speaking in 2025, Krakauer revisits artificial life’s 14 open problems from 2000: how life arises, what open-ended evolution makes inevitable, and how life connects to mind and machines. Darwin could explain evolution after life existed but treated its beginning as almost equivalent to asking about “the origin of matter”; Krakauer suggests those origins might be the same problem.
Nineteenth-century chemistry displaced vitalism by showing that living matter contains no privileged substance. Yet strict materialism leaves the boundary between living and non-living unexplained. His proposed missing variable is function—information about what an arrangement does that cannot simply be “read off from the atoms.”
The artificial-kidney parable carries the distinction: its function survives whether the device uses cloned tissue, tungsten filaments or unknown technology, although it still requires a physical implementation. A rock broken in two remains two rocks; a broken kidney has lost what made it a kidney. Function is “like a spirit,” but one inseparable from matter.
2. Self-construction turns physics into computation
Von Neumann’s thought experiment asks how a robot surrounded by loose components could build another robot. It needs a tape describing itself, a universal constructor that follows the tape, and a copier that passes the tape to its offspring—including instructions for constructing both constructor and copier. Krakauer stresses that von Neumann derived this before DNA’s structure, ribosomes and DNA polymerase were understood.
Von Neumann’s deeper identification was that a universal constructor and a universal Turing machine are “literally one and the same thing.” Anything living must construct itself through some combination of growth, maintenance, healing or reproduction. That autopoiesis makes life embodied computation: “No computation, no life.”
“Embodied” here means closure between computational medium and computer, not merely a robot possessing a body. A laptop manipulates abstracted symbols but cannot extrude another laptop; the relevant machine combines laptop and 3D printer, with atoms serving as memory. Computation itself requires a concise mapping between physical dynamics and logical states, plus free energy and waste-heat disposal as logical entropy is reduced.
Krakauer groups three category errors together: the Sapolsky error confuses reversible physics with irreversible computation, where causation becomes meaningful; the early Wittgenstein error forgets that “there are no birds in physics” without an observer’s model; and the early Leibniz or old-fashioned-AI error assumes airtight logical propositions can yield intelligence, despite real-world reasoning resting on imperfect patterns and regularities.
3. A seven-instruction soup bootstraps programs from noise
BFF asks whether abiogenesis can occur inside a minimal artificial system. Krakauer modifies Brainfuck from eight instructions to seven and merges its code and data tapes, allowing programs to read and overwrite their own instructions. That removes the original language’s barrier to making a copy of itself.
A typical experiment starts with 1,024 random tapes, each 64 bytes long. Only about one byte in 32 represents a valid instruction, leaving roughly two instructions per tape. Each interaction selects two tapes randomly, concatenates them into 128 bytes, executes the result, separates them and returns them to the soup.
After millions of repetitions, “you go from noise to programs”: dense, functional sequences appear that require genuine reverse engineering. In one 8,000-tape snapshot, the leading sequence occupies 5,000 tapes versus 297 for the runner-up. Self-copiers persist while overwriting non-copiers, making complex cycles dynamically more stable than simple fixed states—“the second law of thermodynamics, but doing something unexpected.”
4. Abiogenesis appears as a gelation phase transition
Early interactions execute only about two operations on average; by the end of one run, the average reaches 1,374. The soup has not merely selected one short replicator—it has become “intensely computational,” with far more executable code than its random starting density.
In Krakauer’s ten-million-point plot, the jump occurs near six million interactions. A compression-based entropy estimate changes at the same moment: random tapes begin essentially incompressible, then become highly compressible because programs are copying themselves and one another. He calls it a phase transition rather than a smooth optimization curve.
The pre-transition material resembles a gas because its parts are uncorrelated. The post-transition phase is neither liquid nor solid: it contains functionally differentiated structure at every scale. Krakauer calls it life, describing it as “self-dissimilar”—closer to a multifractal than a repeating fractal.
Across runs, transition times land roughly between one million and seven million interactions. Their distribution fits a 12-step process in which successive prerequisites have long-tailed difficulty. The delay implies hidden stepping stones before visible life, supporting his hedged claim that “pretty much any universe” with randomness and computation could eventually evolve life through dynamical stability.
5. Fusion supplies novelty even when mutation is zero
Krakauer originally included random mutation because evolution is conventionally taught as “chance and necessity”—generate variants, then retain what works. Turning mutation all the way to zero produces the same complexification, exposing what fixed-species equations miss: Lotka–Volterra dynamics can optimize rabbits and wolves forever, but cannot create a third species or enlarge the predefined design space.
Replication exists from the start in minimal form: every copy instruction transfers at least one byte. When two unreliable, one-byte replicators happen to meet, they may reproduce more successfully together than separately. Once the pair begins copying as a unit, a symbiogenetic event has created novelty without altering either component through mutation.
The governing analogy is Smoluchowski coagulation: monomers merge into dimers, trimers and larger clusters through balanced merger-gain and merger-loss terms. When the sticking kernel scales with an exponent above one, cluster size reaches a finite-time singularity and the system gels, as gelatin sets in a refrigerator. Krakauer identifies BFF’s life transition as this generalized gelation.
Early BFF replicators are mostly “inanimate,” where copying code and copied sequence are disjoint, or “viral,” where they overlap only partly. Fully cellular replicators contain their complete copying machinery inside what gets copied. They cannot appear until roughly halfway through a normalized run, then rise sharply near gelation—emerging from symbioses among incomplete replicators.
6. One blocked interaction in 1,000 can prevent the transition
The full model combines R, ordinary population dynamics, with K, the merger kernel. R handles self-reproduction, cooperation, competition and overwriting within shared byte niches; K creates new replicators from several precursors and can produce an output unlike their simple concatenation. Krakauer’s shorthand is “evolution” for R and “revolution” for K.
A sandboxed interaction can reveal whether a new replicator would form and trace which earlier copied bytes constitute its ancestry. Rejecting interactions whose ancestry exceeds a depth around 24 blocks only about one in 1,000 interactions, yet prevents gelation. Complex programs require tree depths of roughly 20 or greater, making deep symbiogenesis causally necessary rather than a retrospective description.
With fusion clamped, replicator populations follow noisy logistic curves around a steady state. Their correlated fluctuations allow reconstruction of R: the strong diagonal records self-replication; negative off-diagonal entries are largely symmetric because niche competition is mutual; positive entries are asymmetric because A helping B does not imply B directly helps A. The resulting ecology contains directed cycles of enablement.
Replicators about to fuse occupy unusually low-rank submatrices: they are already cooperating rather than behaving independently. Under low ancestry ceilings, the Jacobian’s leading eigenvalues remain negative and the system stable; allowing deeper compositions pushes more leading real parts positive, announcing the runaway transition before it occurs. “The more you evolve these things, the more they begin to cooperate.”
7. Symbiogenesis gives evolution an arrow
Krakauer sees conditional Kolmogorov complexity as a bridge between algorithmic information theory and assembly theory. Fusion preserves the replicators and adds information describing “how the two fit together.” That information comes from random encounters, not mutation: symbiogenesis selectively turns the soup’s thermal randomness into durable algorithmic structure and therefore biases evolution toward complexity.
Major-transition accounts associated with Eörs Szathmáry and John Maynard Smith emphasize eight or 12 dramatic events, including eukaryotes and multicellularity. Krakauer regards these as “the tip of a gigantic iceberg.” Most fusions are lopsided and visually modest, but he makes the categorical claim that symbiogenesis is the source of novelty throughout evolution.
His biological specimen is the human genome: only 1.5% codes for proteins, while much of the rest includes transposons and endogenous retroviral elements reproducing within genomic ecosystems. Arc, he says, was endogenized in the mammal lineage and is now important enough that mice lacking it cannot form new memories. The genome is “replicators made of replicators made of replicators.”
8. Intelligence emerges when living computers model one another
His closing definition is “an embodied autopoietic computation arising and complexifying through symbiogenesis.” Each fusion makes computation more parallel and forces the combined system to model itself, its partners and its environment. Once modeling others begins, intelligence and theory of mind become fundamental: “Life was intelligent from the start” and “life is never single player.”
Asked whether biology requires a privileged hierarchical scale, Krakauer says no: humans can be modeled as bacterial colonies or as reproductive units, just as lichens and insect colonies admit competing boundaries. Moving structures between R and K is a coarse-graining choice. Like temperature and pressure, higher-level entities are observer-dependent yet essential for seeing phase transitions and building ecologically useful models.
An audience objection—coupling can increase without higher-order function—draws a three-part response. Composing functions creates a more complex function, much as software imports and combines prior modules; energy limits how much complexity survives; and teamwork can improve energetic scaling. Under suitable environmental conditions, greater computational intelligence may also unlock new energy sources, expanding the budget for another major transition.
Another challenge notes that BFF’s strings receive their functionality through an external goal or CPU. Krakauer agrees that the physics–computation boundary is perspective-dependent: minimal embodiment only requires code to operate on and copy itself, enabled here by merging program and data tapes. He extends the argument downward through particles, atoms and molecules, with a decisive transition when available primitives become rich enough to form a Turing-complete instruction set.