Complexity without copies is just randomness: Lee Cronin on assembly theory

Lee Cronin's third appearance on the Lex Fridman Podcast opens on a Nature paper with an unhumble title: *Assembly Theory Explains and Quantifies Selection and Evolution*, published with Sara Walker on 4 October 2023. A month later it had drawn about 2.3 million engagements on Twitter and a few hundred thousand downloads. Evolutionary biologists were angry, some physicists were grumpy, prebiotic chemists wanted more chemistry, and when creationists complained it was not creationist enough, Cronin figured he had done his job. Fridman tries a two-sentence summary: look at any object, count the steps it took to create it, and decide whether something like evolution built it by counting how many copies exist. Cronin says that is spot on, twice.

An object, as the paper defines it, is finite, distinguishable, persists over time, and can be broken into subunits so that the constraints needed to rebuild it are quantifiable. Cronin still wakes up some days unsure what an object is. The history, he says, is in the object. The assembly index is the minimum number of joining steps, with reuse of remembered motifs costing only one step, needed to reconstruct it from those elementary blocks. He treats the minimum as sacrosanct. Mathematicians later pushed him into a physical reason, but he already trusted it because, before this paper, his lab had shown you can *measure* the index of a molecule.

The chemist’s version is literal. Take thirteen copies of a molecule, cut every bond, lay the atoms out, and walk the shortest path back up, harvesting symmetry as you go. A Lego castle works the same way: the bricks are atoms, an offset pair can be reused, and the shortest walk on that graph is the index. Computing it gets combinatorially hard as molecules grow. The shortcut is physical. Infrared absorption in the fingerprint region, NMR counting distinct magnetic environments, and mass-spectrometric fragment counts each independently tend toward the same assembly index the math would give. That, Cronin argues, is why the number is not a metaphor. An instrument can read it.

Cronin layers four nested spaces. The assembly universe is combinatorial anything-goes: no physics, objects flying through mountains, faster than light. Assembly possible adds the laws, in chemistry the bonds, but still lets you summon every motif instantaneously, so in principle you could make anything. The interesting cut is assembly contingent: you do not get the highly assembled object in the future until the work has been done on the causal chain in the past. Assembly observed is simply what we see, from which we infer that a causal process built it. Compression in computer science, he says, assumes instant access to all of memory. Assembly theory says you do not get the next memory until you have done the work.

Fridman asks what in the universe actually does that work. Cronin reaches for a Tesla factory, a ribosome, a cell. He is not smuggling in morphic resonance or a woo-woo cosmic memory. Somewhere in physical reality there is a directed, constrained process making that object rather than every object. Then comes the harder question: what made the factory? Cronin ranks it as the first of four questions he wants to answer. His conjecture is that the factory emerges in the interplay between the environment and the objects being built.

He offers a chemistry thought experiment. Atoms are flung together in a box; molecules have half-lives; only a molecule that templates its own construction and survives the box will reign. If molecule A needs ten steps and molecule B needs five, and both are destroyed at the same rate, B copies itself faster. "The shortest path reigns supreme." That path, he says, looks like a conservation law for the minimum information required to propagate a motif in time and space. It is not absolute. Two objects share a joint assembly space, so the efficient compromise can make A slightly longer. Parallel construction can drive assembly depth below the assembly index, which is what his team has been watching in cells. In the limit, though, shortest paths are how you tell a factory from a random process.

The paper’s central equation defines assembly as the total amount of selection needed to produce an observed ensemble. It sums, over unique objects, terms that include the assembly index and the copy number, then subtracts one so that a single instance can still be a fluke. Two of a thing, Cronin says, is super important once the index is high. Fridman notes that Cronin had not previously emphasized copies. The intuition is that random processes will occasionally spit out something ornate; without copies you have no evidence of a factory.

That is also why the title caused a fight. The first two sentences of the abstract say scientists have tried to reconcile biological evolution with the immutable laws of physics, and that those laws underpin life, culture, and technology yet do not predict their emergence. Cronin would do it all again. He thinks physicists cannot yet run the Standard Model and get life out, and that evolutionary biologists should not claim ownership of the words *selection* and *evolution*. Prefacing them with "nonbiological" might have cooled the room. He did not think biology had the right to dismiss the origin-of-life gap as obvious.

His broader sense of selection is stubbornly chemical. An object that forms and persists in an environment that would destroy weaker ones has been selected, but a rock on a mountainside for four billion years is not interesting. You need turnover: discovery time, when an object first appears, then production time, when it acts on its own chain of events and there is more of it. A self-replicating molecule in a warm pond is a proto-factory. What biology later adds is robustness, error correction, and the ability to survive many environments. The assembly equation, he says, is a way to ask how much selection is going on in a space, from a hydrothermal vent up through technology. In a petri dish of sugar and one *E. coli* cell, the amount of assembly starts low, then the rate of change goes sigmoidal as the cell eats, copies, and fills the box with factories until the food is gone.

Take a mass spectrometer to Mars, Titan, or Enceladus, Cronin says, with enough resolution to select a molecule by mass and be sure you are seeing identical copies. Fragment them, count fragments, weigh them. On Earth, a molecule heavier than about 350 with more than fifteen fragments is, in his view, an artifact only life produces. Raise the cutoff to 30 or 100 if you distrust the terrestrial number. At some point you find a molecule with a hundred different parts and a million identical copies. That, he says, could never happen in an infinite set of universes.

Copies are the cut between luck and a factory. Give someone one complicated molecule and they can always say it is just random. "Ultimate randomness and ultimate complexity are indistinguishable until you can see a structure in the randomness. So you can see copies." What it means for two objects to be equal depends on the microscope. Chemists are comfortable here: a mole of a molecule is Avogadro’s number of indistinguishable copies, overlaid bonding and configuration. Vibrational and rotational states do not count. Chirality does not count until you have a chiral detector; both hands are the same object until the instrument can tell them apart. Zoomed in, Fridman and Cronin are different; zoomed out they are the same species, then Earth life. If a box of objects is indistinguishable under your technique but the assembly index is low, you are at the wrong resolution.

Paul Davies, a colleague of Walker’s at Arizona State, keeps calling the instrument a life meter. Harsh surfaces burn complexity away, so you sniff for a gradient: warmer, warmer, life, then, in Cronin’s joke, Elon Musk smoking a joint at the bottom of a Martian cave. The device is a vacuum chamber with a chromatographic "nose," an electrostatic trap, and an infrared monochromator. Scoop, add solvent or heat, pull volatiles up the packed inlet, separate them, and on each band count mass-spec fragments and infrared absorptions. The scale runs from dead, through interesting, over a threshold, Earth life, up to "batshit crazy" alien intelligence. Mass spec only shows a peak once you have on the order of ten thousand identical molecules, so copy number is already baked in.

For decades, he says, the field obsessed over small terrestrial biomarkers, single amino acids, Earth-chemistry prejudice. Assembly theory’s pitch is simpler: complexity in abundance is evidence of selection. Chemists told him complicated molecules would of course form randomly off Earth, which he heard as the claim that Darwin’s book was typed by monkeys. Copy number is what made the argument land. His lab ran inorganic minerals, Scotch whisky, *E. coli* lysate, beer, and Taxol, plus five NASA blinds: a million-year-old fossil, a roughly ten-thousand-year-old Antarctic seabed sample, a Murchison meteorite, and others. Every sample was treated the same. Biological material produced molecules with assembly index greater than 15; every abiotic sample was under 15. NASA got the living/dead calls back correctly.

Whisky was not a stunt. On Islay, peat’s complex molecules are said to find their way into the barrel; that molecular complexity is the dark color and the flavor. Cronin ran peaty Ardbeg against Glengoyne, a milder lowland malt near his house. A complexity theorist mocked the beer result as proving beer is more complicated than a human. What they had actually done was break yeast cells, extract molecules, and look for anything over the threshold. Beer itself was not the object. The point of the whole battery was that you do not need to know Earth biochemistry to read the meter.

Cronin says, off the record and then on it, that the lab has mapped the tree of life with assembly theory. Classically you draw morphologies, as Darwin and Mendel did; then you sequence, as Craig Venter does from a yacht; then you do a little molecular archaeology. The fourth route is to take random biological samples, put them into a mass spec, and look not only at individual molecules but at coexisting sets and their joint assembly space. The signal is not that the most sophisticated cells make the most sophisticated molecules. Fungi, stuck in place, synthesize extremely complicated molecules on site; some animals are lazy and eat the fungi. Fingerprint the top high-molecular-weight species, fragment them, compare shared assembly paths, and the number of differences estimates how far back the common origin sits. When they realized they could recapitulate the tree with no gene sequencing and no drawing, the reaction was "huh."

That works on living samples. He wants old fossils next, because DNA and RNA are unstable on the scale of tens of thousands of years while some complex molecules may remain. After death, polymers racemize: the drift from a pure enantiomer toward a mixture is a half-life, a death date. Kinetic isotope effects offer another clock. Life enriches carbon-13 as metabolism runs, so in principle you could estimate how long an organism had been metabolizing, or when it stopped. He admits the timescale he sketched may be wrong, and that specialists already do sophisticated isotope work. The bet is that assembly theory adds a dimension: how long ago this molecule was produced by life, and how much it has since fallen apart.

Asked how far up the hierarchy the theory can climb, from molecules to humans made of billions of other organisms, Cronin stays cautious. Next is cell surfaces, glycosylation, tissue types, disease cell types. Cellular-automaton *rules* already have assembly indices and copy numbers, because different people invented them at different times and copied one another with small edits. A human brain’s index is not comparable to a molecule’s; they sit in hierarchies of emergence. You might count shared morphological features across human, whale, dolphin, chimpanzee, and bird brains, with copy number equal to how many individuals of each there are, but you still need anatomy, or machine learning, to decide which features count. Fridman wants a NASA-style blind test on text, images, and culture. Cronin agrees that would be the first step, and that the framework will probably have to change and expand. The paper, he says, is a first stab at quantifying something biology is visibly doing. It is certainly wrong. He hopes it is less wrong than some other ideas.

Cronin used to go to meetings, he says, and find a way to offend everyone simultaneously, unifying factions that otherwise hated each other. This paper felt like that person at the party. Someone wrote to tell him it was an example of really bad writing. He would like to write that badly again: if you publish, you want people to read it. Fridman reads the pile-on as the scientific immune system, which can also overfire and shut down new conversation. The most exciting criticism, Cronin says, came from evolutionary biologists who told him the origin of life was a solved problem. That was how he knew he had hit a nerve. There is a gap: evolution had to occur before biology. Physicists were polite and unsure about initial conditions. His harder line is that the emergence of life was not, and cannot in principle be, encoded in the initial conditions of the universe.

Some of the noise was spite about *Nature* itself. He told critics this was a new idea that might change how we look at biology and the universe, and they heard hype, as if he had found an alien in a closet. What he wanted was a bold claim that was precise, testable, and correctable, not another woolly "information in biology" argument. Asked whether he had just put sexy names on something obvious, he is flat: nobody had measured a molecule’s assembly index before, and nobody had quantified selection, complexity, and copy number together in this primitive way. People told him it was rubbish and also that they had had the idea twenty years earlier. He asked which part, the rubbish or the revolutionary. "If it’s obvious, why have you not done it?"

No one, he says, has attacked the physical content. You can measure the assembly index; you can measure selection. If that is unhelpful the paper will sink. If it is helpful, people will scaffold on it toward a new paradigm. He wrote to the editor after the backlash to apologize for the controversy. The editor told him not to: it was a great paper, of course it would get backlash, he had said controversial things, and that was awesome. Papers, Cronin likes to say, are not gold medals. He wanted *Nature* because the claim is that something has to come before biological evolution if you are not a creationist, and this was a concrete quantification. Mechanism, autocatalytic sets, self-replicating molecules, comes next.

The standard comparison is Kolmogorov complexity: the length of the shortest program that outputs the object. That measure needs a Turing machine. Cronin issues a trigger warning for complexity theorists and then says algorithmic information theory will turn out to be an important *subset* of assembly theory. Selection produces factories, factories eventually produce computers, and AIT grows from there. Looking at life through information theory, he argues, skips causation. In computer memory every datum costs the same to access; you just compress. A high-index object means that somewhere in the past a real causal chain did the work. Fridman objects that the object does not come with a Wikipedia page; the theory can only infer the shortest history. Cronin accepts "infer." The shortest path is the worst case if you have no machine to make the object. It tells you how deep that object is in time.

Do different reaction probabilities along the chain matter? His first answer is no: the assembly index is a lower bound, a walk on a random graph with no reaction constraints. Then he goes further. He is writing a paper called "The Origin of Chemical Reactions." Reactivity exists, governed by quantum mechanics. Named reactions, the Wittig, the Suzuki, are compressed lists of constraints: Earth, 1 g, 298 kelvin, one bar, Earth’s composition, a glass flask, a pure reagent, temperature and pressure under control. He does not think the Wittig reaction is accessible on Venus. Assembly theory, on this telling, says nothing about chemistry because chemistry is itself a constraint structure. There was no chemist at the origin of life unless you believe in the chemist in the sky. Weak bonds can make you miscount an index when a molecule falls apart in the mass spec; infrared can catch that. The grammar of Earth chemistry is a reliable shorthand. It is still a shorthand.

The first submission to *Nature* came back with professional skepticism: not that interesting, a hard concept, what are you smoking, politely. Sara Walker and Cronin took a year. The preprint was crowded with twenty-three equations; Cronin told first author Abhishek that they all had to go except the assembly equation. They went through something like 160 versions, then at version 40 started again from a blank page and asked, sentence by sentence, what they wanted to say. They expected desk rejection. It went to review. The comments were deep and critical but not dismissive: explain this, are you sure it is not Kolmogorov? Three rounds, then the editor said it was in. An earlier mass-spec paper on measuring complexity had gone through six rounds at *Nature* and still been rejected, with a chemist insisting they must be committing fraud.

Fridman asks what you do at ten in the evening on rejection number five. Cronin does not usually get emotional about papers; this one compressed five years of angst. Working remotely from Walker in Arizona and people at the Santa Fe Institute, he felt like the *Beautiful Mind* portrait minus the genius, just gibberish, expanding expressions with no mathematical ability until copy number fell out of the equation and everything collapsed. A part of him was devastated. A part was excited: they could not tell him why he was wrong. That feeling, he says, goes back to school. In a learning-difficulties class he kept asking the teacher what he could do today to prove he was smart.

Primary school had flagged poor handwriting, apparent inability to read, and weak mathematics, even as he took electronics apart and tried to build computers. He could read; he refused to read what he was told. He was caught teaching another pupil to read. Teachers would not enter him for higher GCSEs; his father offered to pay, and they admitted they would only charge if he failed. He passed, was barred from A-level maths by a C at GCSE, and reached university through chemistry and a good teacher. At eight he tried to build a carbon-dioxide laser, set a fire to harvest CO2, and burned half the garage. He likes not understanding things. He likes hanging around smart people and confusing them, provided they do not simply call him an idiot and they give different reasons. With assembly theory, nobody offered a consistent "why." What he wants is actionable criticism: the equation fails here, the method fails there, so he can triangulate. If someone is persistent enough, he hires them. Someone once gave him a chance.

A fight with Walker about whether time exists, and the work of physicist Nick Gisin, pushed Cronin to a triad. If you want free will, time has to be fundamental. If time is fundamental, you give up Platonic mathematics for intuitionism: Hilbert allowed infinite numbers, Brouwer insisted all numbers are finite. The combinatorial space that produces life and technology is, he claims, larger than a Sean Carroll or David Deutsch multiverse can hold. Max Tegmark’s mathematical block universe, initial conditions you can run to the end and rewind, is not correct. "The universe is not big enough to contain the future." The law of the excluded middle works on last week’s snow in New York. It does not work on next week’s. That failure, for him, is why time is real.

The argument is not ordinary chaos. A golf ball a few centimeters across can store coordinates down to atomic dislocations and no further. A finite object cannot contain an infinitely precise number, so in principle the initial conditions of a collision cannot be specified. Classical and quantum mechanics, on this telling, share an uncertainty principle: not ignorance, but the universe generating new things. Looking back, you can say true or false. Looking forward, you cannot. Boltzmann brains get probability zero because they skip the causal chain evolution would need. Life, Cronin says, is a novelty miner: it mines novelty from the future and actualizes it in the present. Taxol is unique to Earth, an accident of interplay, not a decision written into the Big Bang. He would invert Carl Sagan’s pale blue dot. Do not look at Voyager to see how big the universe is. Look at Voyager, which came from Earth, which came from LUCA, and see how big Earth is. Combinatorially, Earth is the largest place we know in the universe.

He still guesses that every star has planets and that life is probably emerging on them. The combinatorial spaces are so different that causal cones will not easily overlap, in time or in the ability to communicate. His biggest fear is that life is everywhere and we have become infinitely lonelier because of our scaffolding. That is why he wants a planet simulator: Earth 2.0 with a different genetic alphabet, a different protein alphabet, cells and evolution anyway, so that life and selection can be pried off Earth’s chemistry. Then you would know where to point James Webb, and how to read a signal that looks random but has copy number and evolutionary structure. Cellular automata do not refute this. You need initial conditions, rules, *and* time as a resource to iterate the novelty out; some rules you cannot predict. Without that coordinate, you cannot get the complexity out. If you believe in free will, he tells Fridman, the only conclusion is that time is fundamental.

Cronin calls current systems autonomous informatics, on purpose, to make people grumpy. He sees no evidence of AGI anytime soon and calls the doom scenario nonsensical in the extreme. You can compute the chance of an asteroid strike because it has happened and the mechanism is known. People who worry about AG, antigravity, emptying the atmosphere have no mechanism and do not bother; AGI, he says, is in the same epistemic box. Eliezer Yudkowsky’s roughly 95 percent, or Scott Aaronson’s 2 percent if the internet had it right, are not numbers he will treat as knowledge. His own figure is zero. What he does worry about, on the record, is fake data and fake users. He wants authentic people and authentic data, and he is willing to be regulated for risks that exist.

Fridman offers the usual ladder: intelligence has let humans dominate other species, so a system ten times smarter, given the grid and the weapons, is hard to bound. Cronin fails the argument on two category errors. Agency and intention come from humans, from the prompter and the programmer; talking as if the model decides is a mistake, especially when we do not understand human free will. Superintelligence does not follow from capability. A chess computer is very good at chess. Eric Drexler’s atomically precise self-replicating nanobots, he says, made the same error: engineering principles shoved down to a scale with no resources and no precision. The only working molecular nanotechnology we know was produced by evolution. The human brain uses about 20 watts in a small volume; a ChatGPT cluster burns thousands of watts and still needs human correction. Paperclip worlds fail the resource math. A virus too deadly nukes a small radius and dies. When doomers are asked for a mechanism, they make something up. Yann LeCun’s line about regulating jet engines before they are invented is the one Cronin likes.

He reads AI doom and AI salvation as twin new religions, and suspects some of the volume is cover for damage social media has already done. Fridman grants a social role for people who shout that we are all going to die: 95 percent in the mouth of a prophet may average to 0.03 percent in the public, which might be the right background concern, the way nuclear weapons in a war involving a nuclear power are currently underweighted. Cronin grew up with *Threads* and actual existential dread. Cleaning sulfur from the air, he notes, had the unintended effect of accelerating warming; removing every nuke might unleash conventional war. Their half-serious thought experiment is to find the minimum number of weapons, even a virtual mutually assured destruction in a simulation with catastrophic economic penalties, that drives large conflict toward zero. He is joking and he is not. The through-line is that he wants knowledge generated, not frozen by science fiction.

Chem Machina, named for *Ex Machina*, is Cronin’s attempt to understand intelligence as evolution actually produced it: origin of life, multicellularity, locomotion, senses, then the ability to remember the past, interrogate the present, and imagine the future. He does not expect to build intelligence in gel. There are no neurons. He does want reprogrammable hybrid hardware that processes information far more cheaply and transfers knowledge across domains. Large language models, he says, mine the past. The human brain appears to mine the future. Train a model on every prime between zero and 1000 and it will not know what a prime is, and it will not produce anything outside the dataset. Reasoning requires intention, and intention in today’s systems comes from outside, from the prompt and the causal chain that started the session. If a system, human or chatbot, produced something truly novel that he could not predict even with a full audit of the past, he would be sold. Shallow novelty that is still in the training set is interpolation.

The constructive counterexample is GPT-for-electron-density. Molecules are not graphs and not continuously differentiable; they are electron density. The lab computed densities for millions of molecules, trained a generative model on what density looks like, then pours noise into a protein pocket and asks the model to turn the noise into density that maximizes electrostatic stickiness and minimizes steric clash. A second, ChatGPT-like module translates that density into a SMILES string for a robot chemist to make. It is leaky, but it is end-to-end. They trained on molecules with at most nine heavy atoms; the model generated plausible twelve-atom structures. "This hallucination I will take for free." The molecules are new to us and still not truly novel, because the host pocket constrained them. Genuine novelty, assembly theory’s version, is what happens when objects with different histories collide. The more different the histories, the more novelty they generate going forward. That is also his picture of the universe: determined looking back, nondeterministic looking forward, because the future will not fit in the present. After the fact, the new thing joins the dataset and the past looks as if initial conditions had planned it.

Is there room for God? God, he says, exists in our minds, created by selection, the way humans created the concept of superintelligence. There is a creative force in the universe, but he does not think it is sentient. He is an atheist, not an angry one, and he will not call religious people stupid. Selection is the force that creates novelty: objects that could decay instead persist through operations that maintain structure. Free will sits at the boundary of past and future. Imagination has causal consequences the initial conditions cannot predict, which already breaks physics as currently written. What he finds beautiful is that the universe seems undecided and open; every time he nears an answer, more questions appear, and the chase is enough. He does not hate that it will be over. Is it over for Newton? Newton still has causal power. Cronin wants to leave Easter eggs. A faithful Lee Cronin language model would be an interesting artifact and a boring companion, because the interesting part of knowing someone is not knowing what they will do next. Selection produces life; universes with life in them are physically different from universes without. In a thousand years, he suspects, that will look obvious. Assembly theory already looks obvious to its enemies. That, he thinks, is why they got so angry.