The Universal Hierarchy of Life - Prof. Chris Kempes [SFI]

Summary

• Kempes’s central methodological call is that simulation power is not the same as scientific understanding. Biology, intelligence, and economies need physics’s “magic loop”: observations compressed into theory, theory producing “dangerous predictions,” and experiments testing what could not have been fitted from existing data. For AI investors, the distinction is material: detailed simulation may be useful without yielding a compact, transferable theory.
• A universal theory of life may emerge above chemistry, where radically different materials encounter common physical and evolutionary constraints. Kempes’s hierarchy moves from diverse substrates, through gravity, diffusion, temperature, and pressure, toward abstract optimization and learning dynamics. The wager is that this ascent “collaps[es] a huge amount of material diversity” into principles applicable to cells, culture, software, and extraterrestrial life.
• Substrate independence does not eliminate embodiment; every implementation still meets real constraints. Kempes accepts that “we need a material,” because embodiment connects software or concepts to energy and physics, yet argues that DNA, proteins, and lipid membranes may be only one hardware implementation. Language, human culture, and in-silico systems could count as “life living on a very strange substrate.”
• Evolution may be understood as learning under hard information limits. The error threshold constrains how quickly an evolving process can mutate while preserving enough information to adapt, given factors including information content and population size. The host calls this the Miller principle, which Kempes says is equivalent to the error threshold. Because that logic invokes neither biochemistry nor even particular physics, Kempes says it should apply across “language and culture and cells and genomes.”
• Complex functions spread through both transfer and convergence, with human culture unusually good at combining the two. Eyes evolved multiple times because focusing light creates a physical target; genetic material can also move laterally between bacterial lineages. Society’s “extra sauce” is rapid idea transfer plus flexible filtering: a bad idea may be remembered for “zero seconds.”
• Questions about whether simulated systems, viruses, or AI are alive remain underdetermined because life lacks a compact quantitative threshold. Kempes sees no principled reason sufficient computation, constraints, and environmental complexity could not produce life, but he keeps embodiment objections open. He also complicates simple parasite-based exclusions: humans outsource energy capture and contain microbiomes, making each person “a walking ecology.”
• Evolutionary discontinuities and assembly theory offer the most measurable routes toward a theory of complexity. Physical scaling limits create an “evolutionary wall” that may force new architectures, while assembly theory estimates complexity from the shortest recursively reusable construction path. Experiments “look like” they show a sharp abiotic-to-biotic cutoff—potentially useful for life detection, but not yet presented as a settled universal test.

Deep dive

1. Better theories compress reality and risk being wrong

• Kempes contrasts biology’s difficulty with physics’s relatively tractable early questions. Physics succeeded through a “magic loop”: observations revealed regularities, compact equations encoded them, mathematical exploration generated surprising predictions, and experiments returned the results to theory. A science of the biosphere needs the same circulation between abundant data and genuinely predictive principles.

• His three scientific cultures each contribute something different. Variance culture preserves diversity in observations; coarse-grained culture seeks simple principles; exactitude culture can “model everything” through vast simulations, from Earth systems and economies to artificial intelligence. Because each has trade-offs, researchers must “walk amongst the corners of that triangle,” with observation testing both compressed theory and detailed simulation.

• The host preserves Chomsky’s objection that deep learning can resemble “anything goes,” failing to demarcate what a theory excludes. Kempes’s answer is “compactness and compression”: equations are costly to learn how to decode, but afterward they transmit understanding efficiently. His strongest test is the “dangerous prediction” that cannot have been tuned to old data: “If X is true and Y is true, then certainly Z must be out there. Let’s go find it.”

• A universal theory of life would have to explain origins, increasing complexity, intelligence, major evolutionary transitions, and new kinds of life. Current ingredients include agency, semantic information, individuality, scaling laws, metabolism, computation, self-generation, and replication. Kempes’s honest hedge is decisive: “We don’t think we know what the theory is yet”; the eventual synthesis might be equations, concepts, or an unexpected weighting of those ingredients.

2. Universal life may sit above its particular materials

• Kempes and David Krakauer treat language and human culture as life evolving on human minds, and in-silico artificial life as another candidate living on computers built by humans. This is functionalism without rejecting matter: Kempes agrees that “we need a material” because embodiment connects abstract processes to energy, physical limits, and the external world.

• Earth biology has understandably centered DNA, RNA, ribosomes, folded proteins, and lipid membranes, but Kempes warns against mistaking the only implementation we know for the only possible one. A process may admit multiple algorithms, each algorithm multiple software and hardware implementations, much as different methods can all produce a sorted list. The hard question is mapping functions across those layers while preserving their constraints.

• His hierarchy therefore begins with potentially radical material diversity across planets. Lift one level, however, and living processes still operate under constraints such as gravity, diffusion, scale, temperature, pressure, chemistry, and acidity. A small cell or sphere of living material in fluid must interact with diffusion, whether or not it uses DNA, allowing theory to compress many substrates into a smaller “space of constraints.”

• The highest layer contains optimization-like principles. Evolution can be written as a learning dynamic, while the error threshold limits mutation rates relative to the amount of information being transmitted, population size, and the need to remain adaptive. Kempes’s claim is broad but conditional: such a material-agnostic constraint “should apply to language and culture and cells and genomes and all the rest.”

3. Physics drives convergence while culture accelerates transfer

• Similar functions can cross lineages in two distinct ways. Lateral gene transfer moves genetic material between bacterial genomes; convergence independently rediscovers solutions when physics provides a strong target. The eye is Kempes’s best specimen: despite its complexity, it evolved many times, with different construction details but the same basic task of focusing light and obtaining a picture at a distance.

• Human society is unusually effective at both mechanisms. Agriculture arose multiple times, resembling convergence, while a useful discovery made in one culture can propagate rapidly across languages and societies. Our abstract “cultural genomes” are also selective: an idea can “mutate my mind,” but the recipient evaluates whether to integrate it and what action, if any, should follow.

• The host asks whether an information phylogeny could ultimately transcend material substrates, like patterns moving between distinct basins of attraction. Kempes keeps the uncertainty intact: software and open-source projects may exhibit recognizable evolutionary dynamics, but beyond coarse-grained similarities, “we don’t know what those projections look like.”

• That gap defines what theory should accomplish. A successful projection either reveals that apparently different phenomena are “just one thing” viewed at the correct angle, showing where “sameness is hiding in plain sight,” or proves that certain differences cannot be removed. Kempes does not claim life, culture, and software have already been unified; he identifies the standard a unification must meet.

4. Life and intelligence can involve spectra, thresholds, and phase changes

• Kempes sees “no reason” sufficient compute, constraints, and environmental complexity could not produce life in an artificial world; computer viruses are already lifelike in their transmission, function, persistence, and possible evolution. Yet unlike checking whether a simulated gas obeys the second law of thermodynamics, researchers lack a compact calculation establishing when a simulation crosses into life. Strong materialist objections therefore remain open rather than resolved.

• The host’s virus challenge is that parasitism makes the organism inseparable from its host. Kempes complicates that framing: viruses occupy unusually complicated environments, but humans also do not photosynthesize—we consume organisms that captured and transformed energy—and our microbiomes make us “a walking ecology.” Under one projection, humans and viruses are comparable parasites; under another, human agency, prediction, technology, and lunar exploration clearly separate them.

• Instead of binary labels, Kempes wants quantitative spectra for intelligence, agency, and self-replication—perhaps something has “10 to the minus 20 intelligence” rather than none. The host’s pushback is compatible rather than contradictory: categories can contain spectra, while genuine phase changes still mark discontinuities, just as water and ice remain meaningfully distinct despite sharing an order parameter.

• Kempes locates those jumps at physical scaling limits. When a class of organisms reaches an asymptote—something tending toward zero or infinity—it meets an “evolutionary wall” requiring a new architecture. Eukaryotic organization involved one prokaryote inside another; true multicellularity required differentiated cells, regulation, communication, organs, and developmental programs. His group thinks Snowball Earth supplied conditions that induced multicellularity, after which that solution spread and became larger, though he cautions that such transitions may not always occur.

5. Assembly paths may expose complexity without assuming Earth biology

• David Deutsch’s separation of matter and logic explains why physical description alone is insufficient. A computer and a river can both be networks carrying currents, but the computer’s architecture repeatedly performs logical operations, maps inputs to outputs, and can be reprogrammed. Understanding why that architecture exists requires software-level description, not merely tracking electrons.

• Assembly theory, initially aimed at searching for life without presuming terrestrial biochemistry, asks for the shortest path that builds an object through recursive reuse of parts. Each creation or reuse of a part counts as a step. The resulting assembly measure serves as a lower bound: real synthesis may be more elaborate, but shortest paths permit comparison without knowing whether an object was made by life or shares its chemistry.

• Carbon-60 illustrates why molecular weight alone misleads: it is large but not especially complex under the assembly criterion. Because each additional step enters a rapidly expanding combinatorial space, one would expect a sharp cutoff somewhere. Kempes says the experimental data “look like” they show an abiotic-to-biotic threshold—an encouraging result, while still framed as experimental evidence rather than a completed theory of life.