Where New Forms Come From

A frog egg contains everything needed to make a frog. For most of the history of biology, that sentence was allowed to close the question of where the frog’s shape comes from. In 2020 a group at Vermont and Tufts took skin and heart cells from frog embryos and assembled them into small bodies designed by a computer. Those bodies moved, cooperated, and repaired themselves in ways no frog does ¹. No population of these constructs had ever been selected for anything, which leaves an awkward question about what specified their form and behavior. Michael Levin co-authored that work, and he has now written a long paper arguing that a serious answer requires giving up physicalism ².

This essay does three things. It sets out what Levin proposes and what he offers as evidence, including the objections a mainstream biologist would raise. It then follows his question into two places where it has practical consequences. The first is de-extinction, where a company edits a handful of genes and lets development do the rest. The second is computational modeling, where the choice between generating data from formulas and fitting models to measurements is the same question in engineering dress.

What Levin is proposing

Before weighing the argument it helps to state it plainly, because the claim is unusual and easy to caricature. Levin’s paper was published in the journal Philosophies in September 2026. It runs to fifty-three pages and is explicitly framed as a set of hypotheses and a research program, with no new data ². The starting observation concerns synthetic morphology, the engineering of living constructs that have no evolutionary history. Such constructs turn out to have specific and consistent anatomy and behavior. The usual explanation for an organism’s form, that eons of selection tuned it, is unavailable for something that has never existed before. So where does its form come from?

His answer is that the physical world is not closed. He proposes a structured latent space of patterns. The term is borrowed from machine learning, where a latent space is a compressed space of possible configurations. Patterns from this space ingress, his term, into physical systems that are suited to receive them. Bodies, embryos, and machines are treated as pointers or interfaces into that space. Their behavior is read as partly drawn from it. The analogy he leans on is mathematics. Mathematical facts such as the distribution of primes do not depend on physics and have no evolutionary history. They still constrain and enable what physical systems can do, and Levin proposes to extend that status well beyond mathematics.

The paper states three hypotheses. The first is that the patterns in this space span a spectrum of agency, running from static truths of the kind mathematicians study to active patterns of the kind behavioral scientists study. Some of those patterns are what we would call minds. The second is captured in his compressed phrase, “mind:body is as math:physics.” He means that the relationship between a mind and a brain is the same relationship as between a mathematical fact and the physical objects it governs ². The third is that living things have no monopoly on this. Any physical construct, natural or engineered, is to some degree informed by patterns from the latent space. In Levin’s view this erases the distinction between organisms and machines.

The research program that follows has two parts. One is to study constructs that have not been selected for their properties. Any competence they show cannot then be attributed to selection, design, or training. The other is to study very minimal systems where every component is known, so that an unexpected competence cannot be attributed to undiscovered biology. His example of the second kind is a study of sorting algorithms that, when fed a stuck element, temporarily unsort the list to route around it. Nobody wrote that behavior into the algorithm. On falsification, he writes that “the framework is falsifiable in the same way all significant scientific ideas are falsifiable.” The criterion he gives is fruitfulness: if years of work along these lines yield no new discoveries or capabilities, the framework should be dropped ².

The experiments the argument rests on

The empirical work behind the paper is real and peer-reviewed, and it deserves to be separated from the metaphysics built on it. The 2020 xenobot paper described a pipeline in which an evolutionary algorithm searched over simulated body plans built from two cell types. The most promising designs were then assembled by hand from frog embryonic cells and shown to behave roughly as predicted ¹. A 2021 follow-up dropped the design step entirely. Frog epidermal cells left to self-organize formed ciliated spheroids that swam, healed after being cut, and showed collective behaviors. There was no genetic modification and no scaffold ³.

Anthrobots extended the same idea to adult human cells. Single cells from human airway tissue were cultured in extracellular matrix for two weeks and then released into low-viscosity medium. They self-assembled into motile spheroids between 30 and 500 microns across. Those spheroids moved at 5 to 50 microns per second in patterns that correlated with their shape. Placed on a scratched sheet of cultured human neurons, they induced repair of the gap ⁴. Again, no genes were edited.

The interpretive frame Levin brings to these results predates the new paper. His 2022 “Technological Approach to Mind Everywhere” argued that cognition comes in degrees along a continuum. It also argued that morphogenesis is a form of problem-solving in which cell collectives navigate toward a target anatomy. The medium in which those targets are stored, on this account, is bioelectric networks: the patterns of voltage across cell membranes coordinated through ion channels and gap junctions ⁵. The planarian flatworm is the standard example. Briefly altering the bioelectric state of a fragment produces a two-headed worm whose fragments go on regenerating two heads indefinitely, with an unchanged genome. The 2026 paper takes that body of work as given and asks where the target patterns themselves originate.

Where a mainstream biologist gets off the train

Having laid out the proposal, it is only fair to say where most working scientists would part company with it. The proposal is non-physicalist and explicitly dualist, in that it posits causes outside physical events. As Levin acknowledges, it also implies a form of panpsychism, the view that mind is present in some degree throughout matter ². Each of those positions has a long history of rejection by working scientists.

The conventional reading of every result above is that self-organization from physical and chemical interactions is sufficient. Cells carry an enormous repertoire of behaviors shaped by selection for a body that must tolerate damage, variable cell numbers, and variable environments. A xenobot’s cilia-driven swimming is what ciliated epithelium does when it is on the outside of a sphere. An anthrobot’s structure is what airway progenitors do when confined and then released. Nothing in these observations requires a source of pattern outside the cells, and the burden of proof sits with the person proposing one. Levin’s reply is that calling such outcomes emergent names them without explaining them. He holds that treating the latent space as real is the more productive research bet, which is a claim about research strategy. It can only be settled by whether the strategy produces results the alternative would not have.

The mathematical analogy is also less settled than the paper needs it to be. Platonism about mathematics is the view that mathematical objects exist independently of us and are discovered, and it is one position among several. Even Platonists generally hold that mathematical facts constrain physical systems without acting as causes in the sense a biologist means. Levin argues that a mathematical fact should count as a cause because it is the most insightful explanation. That broadens the word cause in a way many philosophers would resist.

The falsification criterion is the other sticking point. It is fruitfulness over years, and no single experiment is named that could refute the framework. The paper is candid about this. It says that no single observation, including the existence of anthrobots, entails the Platonic model. It also says that physicalism is flexible enough to absorb almost any finding after the fact ². That candor is welcome. It does mean readers should treat the framework as a bet on where to look, held by a serious experimentalist, and keep it separate from his findings.

De-extinction as a test of how much a form is specified

Whatever one thinks of the metaphysics, the question Levin asks has an unusually literal test case. In de-extinction, a company alters a small number of genes and relies on an existing developmental system to produce everything else. Colossal Biosciences is the clearest example.

On April 7, 2025, Colossal announced three wolf pups it described as dire wolves. The method was reported by Science and Nature news and confirmed by the IUCN statement discussed below. Endothelial progenitor cells were cultured from gray wolves and edited. The nuclei were then transferred into domestic dog egg cells, and the clones were carried in dog surrogates ⁶⁷⁸. The number of edits reported was 20, aimed at visible traits such as coat color and body size ⁷.

The genomic groundwork was released as a preprint by Colossal and academic collaborators. A preprint is a paper posted before peer review. It reported two dire wolf paleogenomes at 3.4x and 12.8x coverage, meaning each position in the genome was read on average that many times. It found that roughly two thirds of dire wolf ancestry derives from a lineage sister to the clade containing gray wolves, coyotes, and dholes. The remaining third comes from a lineage near the base of the true dogs ⁹. The preprint also identified 80 genes under diversifying selection in dire wolves. It does not describe the edited animals themselves, and as of this writing the editing has not been described in a peer-reviewed paper.

The scientific response was pointed. The IUCN Species Survival Commission’s Canid Specialist Group issued a statement that the three animals “are not dire wolves.” It noted that dire and gray wolves differ by thousands of genes while the editing touched a handful. It also noted that there was no evidence yet that the animals were phenotypically distinct from gray wolves at all ⁶. A geneticist quoted by Science asked whether a chimpanzee with 20 edits would be human ⁷.

The woolly mouse, announced a month earlier, drew similar comments. That preprint reported mice with up to seven genes edited at once. The edits included loss-of-function changes in Fgf5, Tgm3, and Fam83g. The result was long, curly, golden-brown coats ¹⁰. Independent geneticists pointed out that most of those edits were chosen because they were already known to alter mouse hair. They also noted that cold tolerance, the trait that would matter for a mammoth, was not tested ¹¹.

Set the naming dispute aside and look at what the experiments measure. Colossal’s practice is to specify a small number of loci and let the animal build the rest. The host may be a gray wolf or a mouse or eventually an Asian elephant. The result is recognizably a gray wolf with some altered traits, which is exactly what the conventional account predicts. The vast majority of the form came from the unedited genome and the developmental process reading it, and the edits contributed a few traits. In that sense de-extinction by editing is a running experiment on the ratio between what an engineer specifies and what the system fills in. So far the ratio sits heavily on the side of the system.

Levin would agree with that description and then ask a further question. Does the unedited genome plus physics fully account for the filled-in part? Or is the genome hardware that admits many patterns, with the default one simply the most reliable ²? Colossal’s work does not answer that. It does put a number on how far a handful of edits gets you, and the number is small. Neither party has endorsed the other’s framing, and nothing in the sources suggests Colossal thinks in Levin’s terms.

Generating data from formulas versus fitting models to data

The same question reappears in a different vocabulary when we ask how computational models of biology get built. A physics engine is a program that generates the state of a system from equations. Game engines integrate Newton’s laws for rigid bodies, and molecular dynamics integrates the same laws for atoms under a force field. The data come out of the formulas. Machine learning runs in the other direction. A model is fitted to observed data. The formulas, if there are any, are implicit in the weights. Most of biology today lives on the second side, because we have very few formulas that generate whole-organism data.

The two approaches have been converging, and the convergence is instructive. Cranmer, Brehmer, and Louppe’s 2020 review of simulation-based inference describes the underlying problem. Simulators are high-fidelity models of a phenomenon but are poorly suited to inference, because the probability of an observation under the simulator cannot be written down. The review then describes how neural networks trained on simulator output can recover the parameters that produced an observation ¹². DiffTaichi is a language for physical simulation in which every step is differentiable. It lets an engineer take gradients through a simulator, so a controller can be optimized against the physics directly. In the paper’s examples this converges within tens of iterations ¹³. In robotics, domain randomization trains a vision network entirely on simulated images. The colors, lighting, and textures of those images are randomized. The network then localizes real objects to within about 1.5 centimeters on a physical robot without having seen a real photograph ¹⁴. In each case the simulator is a source of unlimited labeled data, and that is the property biology lacks.

How far can biology be simulated from first principles today? At the level of proteins, molecular dynamics works because the force fields are good and the systems are small. At the level of a cell, the fair benchmark is still the 2012 whole-cell model of Mycoplasma genitalium, one of the smallest known bacteria at 525 genes. That model divided the cell into 28 processes, each with its own mathematics. It fitted more than 1,900 parameters drawn from over 900 sources, then predicted phenotypes from genotype and directed experiments that found new kinetic parameters ¹⁵. Nothing comparable exists for a human cell, let alone a developing embryo. The most celebrated model in structural biology, AlphaFold, sits on the fitting side of the line. It was trained on roughly 100,000 experimentally determined structures accumulated over decades, with physical and evolutionary knowledge built into the architecture but no simulation of folding ¹⁶.

Read in these terms, Levin’s latent space is the claim that formulas for form exist and can be found. One day a new organism could then be generated from principles the way a game engine generates a scene. That is speculative. What is established is narrower and still significant. Bioelectric states can be read and rewritten to change anatomical outcomes with a fixed genome. Cells from existing organisms can be coaxed into stable forms that no organism has. The design step for such forms can be run in simulation before it is run in cells, as the 2020 xenobot pipeline did ¹³⁴⁵. Whether the rules governing those outcomes turn out to be reducible to physics or to require something more is the open question, and the engineering payoff of finding the rules is the same either way.

Two ways out of data scarcity

That brings the argument back to our own work and to a practical problem. Our view is that biology has no foundation model, in the sense of one model that transfers across tasks the way language models do. The main reason is data scarcity, and architecture is only a secondary reason. The data that would specify a person, or an embryo, over time is not being collected at anything like the volume that trained AlphaFold on proteins or a language model on text.

There are two ways out, and we would pursue both. The first is to generate data from principles. That requires formulas we mostly do not have yet. It is the direction the physics-engine literature and, in its own way, Levin’s program point toward. The second is to measure far more people, far more often, so that the fitting approach has enough to fit. The two are not exclusive. The whole-cell model was built from 900 sources of measurement, and the xenobot pipeline was a simulator whose predictions were checked against cells.

Levin’s paper is a philosophical bet made by someone who runs experiments. The safest reading of it is as a research program with a clear success criterion he has set for himself. The de-extinction results are a reminder that most of any organism’s form is filled in by a system we did not specify and cannot yet simulate. Where new forms come from remains an open question, and it is one that measurement and simulation will have to answer together.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

1. Sam Kriegman, Douglas Blackiston, Michael Levin, Josh Bongard. A scalable pipeline for designing reconfigurable organisms. Proceedings of the National Academy of Sciences, 2020. https://doi.org/10.1073/pnas.1910837117 ↩ ↩² ↩³

2. Michael Levin. Ingressing Minds: Causal, Non-Physical Patterns In-Form Natural, Synthetic, and Hybrid Embodiments. Philosophies, 2026. https://doi.org/10.3390/philosophies11050161 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. Douglas Blackiston, Emma Lederer, Sam Kriegman, et al. A cellular platform for the development of synthetic living machines. Science Robotics, 2021. https://doi.org/10.1126/scirobotics.abf1571 ↩ ↩²

4. Gizem Gumuskaya, Pranjal Srivastava, Ben G. Cooper, et al. Motile Living Biobots Self-Construct from Adult Human Somatic Progenitor Seed Cells. Advanced Science, 2024. https://doi.org/10.1002/advs.202303575 ↩ ↩²

5. Michael Levin. Technological Approach to Mind Everywhere: An Experimentally-Grounded Framework for Understanding Diverse Bodies and Minds. Frontiers in Systems Neuroscience, 2022. https://doi.org/10.3389/fnsys.2022.768201 ↩ ↩²

6. IUCN SSC Canid Specialist Group, Taxonomic Review Task Force. Conservation perspectives on gene editing in wild canids. IUCN Species Survival Commission, 2025. https://www.canids.org/resources/CSG%20gene%20editing%20in%20wild%20canids.pdf ↩ ↩²

7. Phie Jacobs. Is the dire wolf back from the dead? Not exactly. Science, 2025. https://doi.org/10.1126/science.zl0oigi ↩ ↩² ↩³

8. Ewen Callaway. This company claimed to ‘de-extinct’ dire wolves. Then the fighting started. Nature, 2025. https://doi.org/10.1038/d41586-025-02456-3 ↩

9. Gregory L. Gedman, Kathleen M. Pirovich, Jonas Oppenheimer, et al. On the ancestry and evolution of the extinct dire wolf. bioRxiv, 2025. https://doi.org/10.1101/2025.04.09.647074 ↩

10. Rui Chen, Melanie L. Coquelin, Kanokwan Srirattana, et al. Multiplex-edited mice recapitulate woolly mammoth hair phenotypes. bioRxiv, 2025. https://doi.org/10.1101/2025.03.03.641227 ↩

11. Science Media Centre. Expert reaction to unpublished preprint on inducing loss of function of genes in mice to produce woolly mammoth-like hair phenotypes. Science Media Centre, 2025. https://www.sciencemediacentre.org/expert-reaction-to-unpublished-preprint-on-inducing-loss-of-function-of-genes-in-mice-to-produce-woolly-mammoth-like-hair-phenotypes/ ↩

12. Kyle Cranmer, Johann Brehmer, Gilles Louppe. The frontier of simulation-based inference. Proceedings of the National Academy of Sciences, 2020. https://doi.org/10.1073/pnas.1912789117 ↩

13. Yuanming Hu, Luke Anderson, Tzu-Mao Li, et al. DiffTaichi: Differentiable Programming for Physical Simulation. International Conference on Learning Representations, 2020. https://doi.org/10.48550/arXiv.1910.00935 ↩

14. Josh Tobin, Rachel Fong, Alex Ray, et al. Domain randomization for transferring deep neural networks from simulation to the real world. IEEE/RSJ International Conference on Intelligent Robots and Systems, 2017. https://doi.org/10.1109/IROS.2017.8202133 ↩

15. Jonathan R. Karr, Jayodita C. Sanghvi, Derek N. Macklin, et al. A Whole-Cell Computational Model Predicts Phenotype from Genotype. Cell, 2012. https://doi.org/10.1016/j.cell.2012.05.044 ↩

16. John Jumper, Richard Evans, Alexander Pritzel, et al. Highly accurate protein structure prediction with AlphaFold. Nature, 2021. https://doi.org/10.1038/s41586-021-03819-2 ↩