Audience measurement
Cookie-free audience measurement and human/bot distinction, only if you consent. Privacy policy
The other day, I was stuck in bed, sick. I could not work on my systems. I did not have the focus to code anything. So I did what I do when I cannot do anything else: I thought.
I had just come across an article about attempts at “de-extinction” using the domestic chicken. The general idea: reactivate ancestral traits to bring out dinosaur-like features, teeth, a long tail, claws, a snout. The problem was immediately obvious to me.
Starting from a chicken to reach a raptor is like trying to turn an African elephant into a woolly mammoth. The starting point is too far away. The anatomical, behavioral, and metabolic constraints are so large that each change triggers ten more. Asking an African elephant to tolerate Siberian cold and adapt to the steppe is not an adjustment. It is a full reconstruction.
Same for the chicken. It is a domesticated animal, selected for millennia to lay eggs, gain weight fast, and meet human constraints. Its anatomy, physiology, and behavior have moved away from what we would look for in a predatory theropod morphotype.
So to me, the real problem is not only: “which genes should be reactivated?”
It comes first: which starting species should we choose?
What we need is a bird already as close as possible to the target. An organism where the required changes would be minimal. An animal able to absorb those transformations anatomically, metabolically, and behaviorally. Not a total reconstruction: a constrained shift within an already compatible morphological space.
First point: birds do not descend from dinosaurs. Birds are dinosaurs. A surviving branch of theropods.
Dromaeosaurids, the so-called “raptors”, are not the direct ancestors of modern birds. They are cousins. They coexisted with avian lineages for millions of years. Saying that birds descend from raptors is like saying humans descend from chimpanzees: it is false. Both share a common ancestor, then evolved in parallel.
That changes the approach in depth.
“Regressing a bird into a dinosaur” is a useful phrase, but it is scientifically imprecise. You do not climb back up a family tree. You do not return to an ancestor. The aim is instead to make a living organism converge toward a related morphology, now extinct, but built on shared developmental foundations.
Raptor is therefore not trying to reverse evolution. It explores what would need to be modified, preserved, or constrained to move a modern bird toward a plausible dromaeosaurid morphotype.
The truly difficult areas are elsewhere.
The first is the tail. Modern birds have a pygostyle: a fusion of the last caudal vertebrae. In 2023, Botelho showed that this fusion is driven by sterile postnatal inflammation, similar to a bone repair mechanism. Preventing that fusion may therefore be more accessible than previously thought, perhaps by intervening in the inflammatory process.
But preventing fusion is not the same as lengthening the tail. Moving from a few caudal vertebrae to a long, functional dromaeosaurid tail means prolonging axial elongation, restoring effective caudal musculature, and probably accounting for partial stiffening through tendons or equivalent structures. To date, no one has produced a modern bird with a long, functional theropod tail.
The second difficulty is the sickle claw. The hypertrophied claw on the second toe of dromaeosaurids is an emblematic trait, but its exact developmental mechanism has not been identified. Modern birds of prey have strong, curved claws, sometimes functionally comparable, but no experiment has reproduced a true dromaeosaurid sickle claw in a bird.
The third difficulty is increased mass. Moving from a 3.5 to 5 kg bird to a 15 to 20 kg animal is not just a matter of making the organism larger. It brings cascading constraints: thermoregulation, respiration, bone structure, posture, locomotion, circulation, energy expenditure. Myostatin inhibition, spectacular in some mammals, produces variable and sometimes counterintuitive results in birds. A realistic increase in mass would probably require a combination of pathways, including GH/IGF-1, musculature, bone growth, and metabolic constraints.
Finally, the horizontal posture of a theropod is not a simple angle change. It involves the pelvis, the spine, the hind limbs, balance, the tail, the position of the center of gravity, and locomotor coordination. It is a full biomechanical redesign.
These limits do not disqualify Raptor. They show where the model must stay honest.
The Magpie Goose (Anseranas semipalmata) is the most basal bird among the Anseriformes. It diverged from the other lineages in this group about 57 million years ago. This basal status does not mean it is “primitive”. It means it retains traits that are useful for a morphological exploration.
It has semi-webbed feet, but also functional claws. It has a stronger grasping ability than many other aquatic birds. Its neck profile is shorter than that of many geese. Its diet is mainly herbivorous, but can include an opportunistic share. Its metabolism and mass remain in a more workable range than those of a tiny or highly specialized bird.
It is not the closest morphological analog to the last toothed birds. If the only goal were proximity to archaic aquatic birds such as Hesperornithiformes, a cormorant would probably be more relevant morphometrically.
But the cormorant raises other problems: it is a protected animal, harder to breed, more specialized, and less practical for a generalizable experimental or computational approach. The Magpie Goose lives in colonies, breeds in captivity, and retains a combination of exploitable anatomical traits.
The choice of the Magpie Goose is therefore both scientific and pragmatic. It is not “the perfect bird”. It is a better starting point than the chicken for exploring a trajectory toward a predatory theropod morphotype.
Raptor is not a real de-extinction protocol. It is not a genetic editing program, and not a proposal for immediate animal experimentation.
Raptor is a computational research program.
Its goal is to explore the space of possible morphological configurations between a modern bird and a dromaeosaurid morphotype, using evolutionary algorithms constrained by developmental biology, biomechanics, and physiological viability.
The principle is not to say: “here is how to make a Velociraptor”. The principle is to ask: which combinations of traits are theoretically coherent, which ones collapse, which ones remain viable, and in what order could the transformations be explored?
What surprised me when I read the literature is that several necessary transformations have already been obtained in the lab, at least in part. Not in the Magpie Goose, but in the chicken, which remains the best-annotated and most widely used avian model in experimental embryology.
In 2015, Bhullar and his team showed that by inhibiting the FGF and Wnt pathways in the frontonasal ectodermal zone of chicken embryos, they could disrupt beak fusion. The embryos developed unfused premaxillae, with a morphology that CT scans showed to resemble Archaeopteryx and some non-avian theropods.
In 2006, Harris and his team observed that the chicken talpid2 mutant developed archosaur-type conical teeth. Direct activation of the beta-catenin pathway in normal embryos can reproduce the first stages of tooth formation. These are not complete teeth in the sense of a modern functional dentition, but it shows that some odontogenic programs remain accessible.
In 2016, Botelho and his team showed that inhibiting the IHH protein (Indian Hedgehog) prevents terminal maturation of the fibula in the chicken. The result is a longer fibula, extending all the way to the ankle, fused to the fibulare, and morphologically comparable to that of Archaeopteryx.
In 2015, Botelho also showed that paralyzing embryonic musculature can produce a non-opposable hallux, closer to the ancestral theropod state. This point matters: it is not a direct genetic manipulation, but a pharmacological one. It shows that some morphological traits do not depend only on the genome, but also on embryonic muscle activity.
These examples are not enough to create a viable raptor-like animal. But they prove one essential point: some transformations often presented as purely speculative already have an experimental basis.
Five visuals. Two of them, the genome and the constraints, call the same fitness oracle, frozen and hashed in SHA-256: the one that will judge the 50,000 evaluations of a run. Its fingerprint is shown on screen, and nothing there is illustrative. What appears is calculated. The other three calculate nothing, and do not claim to: they give the position of the species, the order of the protocol, and the exact state of the literature, publication by publication.
Seven birds and theropods, positioned by divergence time and morphological distance from the target. The two axes do not say the same thing: a taxon can be ancient and close, or recent and far.
The chicken is the farthest of all. The Magpie Goose is the best compromise among the candidates that can actually be bred. Each taxon shows what it already has and what it lacks, with references.
The model’s twenty-eight abstract morphogenetic parameters, grouped into six modules. Hover over one: its dependencies light up. Enamel depends on tooth germs, the long tail depends on its musculature, mass depends on ventilation.
Below, the measurement. Across three thousand crosses between viable parents, evaluated by the same oracle: swapping parameters one by one invalidates 27.9% of descendants; swapping whole modules invalidates 0.1%.
The protocol does not open the 28 parameters at once. It unlocks one module, lets it stabilize, then freezes it, and that module becomes a constraint for the next one. Cranial before appendicular, axial before metabolic. Move through the sequence: each step shows what it opens and what it imposes on the ones after it. Reducing the keel at the axial stage, for example, costs ventilatory capacity at the next stage, and the metabolic module pays the bill. This dependency is not a computational convenience: the hypothesis is that it keeps the intermediate forms viable. That is exactly what the runs must decide, against a control group where all modules are open from the start.
Seven sliders move the goose toward the target morphotype. With each move, the model recalculates the eight criteria and says whether the phenotype holds.
The upper target, fourteen kilos, dense plumage, predatory metabolism, is rejected: negative thermal balance. A form can look convincing and still be physiologically impossible. That is exactly what the model is meant to show.
Six transformations, ranked by what we actually know how to do. Four have been obtained in the lab and published with their DOI: the snout by Bhullar in 2015, tooth germ initiation by Harris in 2006, the complete fibula and the non-opposable digit by Botelho. The fifth, the sickle claw, rests on a plausible mechanism that no one has carried through to the phenotype. The sixth, the long tail, has no experimental basis, and that is stated as plainly as the rest. Each entry ends with the section that matters most: what the result does not prove. A dossier that showed only its successes would be worth nothing.
In Raptor, each simulated individual has an abstract genome. This genome is not a real DNA sequence. It is a vector of morphogenetic parameters organized into functional modules.
These modules represent the main transformation areas:
The important point is that the genetic algorithm’s crossover does not swap isolated “genes” at random. It exchanges whole modules, or coherent submodules. This avoids producing biologically absurd combinations: teeth without suitable cranial space, a long tail without musculature, or high mass without enough respiratory capacity.
So the model is not only looking for a visual form. It is looking for coherence between anatomy, development, physics, and behavior.
Each generated phenotype is evaluated by an eight-dimensional fitness function.
A single fitness oracle is shared by all algorithms. Its code is hashed in SHA-256 to ensure the comparison remains stable and unbiased. Each algorithm will be tested over 15 independent runs, they still have to be launched. GA and CMA-ES will be compared on morphology; NEAT on the behavioral criterion alone. The results are then compared statistically with a Friedman test and a Nemenyi post-hoc test, standard tools in evolutionary algorithm comparisons.
The goal is not to produce a beautiful chimera. The goal is to know which algorithm best explores a constrained morphogenetic space.
Raptor is designed to compare several evolutionary approaches on the same problem; the comparative runs have not yet been performed.
The first algorithm is a classic GA (Genetic Algorithm). It works with a population of simulated individuals. The best candidates, according to a shared fitness function, are selected. They reproduce, their modules are crossed, then mutated. The genome structure stays fixed, but the values evolve.
The second is CMA-ES (Covariance Matrix Adaptation Evolution Strategy). It is a more sophisticated evolutionary strategy, which automatically adapts its search distribution from previous results. It does not only search at random around the best individuals: it gradually learns which direction in the search space seems promising.
The third is NEAT (NeuroEvolution of Augmenting Topologies). Here, the interest is not only the evolution of morphological parameters, but also the evolution of decision structures. In a behavioral extension of the simulator, NEAT can evolve networks where new neurons and new connections appear progressively. This makes it possible to explore behaviors that were not explicitly programmed: orientation, approach, attack, avoidance, coordination, curiosity.
These three algorithms do not play exactly the same role. GA and CMA-ES explore the parametric morphological space very well. NEAT becomes especially interesting once embodied behavior is added: the link between body, environment, and decision.
Raptor does not transform everything at once.
The protocol calls for sequential evolution. Modules are unlocked progressively. The hypothesis is that this avoids exploding the search space and preserves the viability of intermediate forms. A control group, with all modules open from the start, will be used to test this.
One possible sequence would be:
Each stabilized module becomes a constraint for the next one. This should limit impossible forms and progressively build global coherence, this is what the comparison with the control group must establish.
The respiratory system cannot be treated as a detail.
Birds already have a highly advanced system of air sacs and unidirectional ventilation. Non-avian theropods probably had important elements of this system as well. This makes the Magpie Goose interesting: it already starts from an efficient avian respiratory architecture.
But increasing mass, modifying the rib cage, reducing the keel, or changing posture can compromise respiratory balance. A viable raptor-like morphology must therefore maintain enough respiratory capacity to support an active metabolism.
In the simulator, a phenotype whose respiratory capacity does not keep up with mass and metabolism must be strongly penalized. This is exactly the kind of constraint that prevents the model from producing a spectacular but biologically absurd creature.
Thermoregulation is the other major constraint.
A modern bird operates with a high internal temperature, often around 40–42°C. As mass increases sharply, the surface-area-to-volume ratio decreases: the animal dissipates heat less effectively. A heavier individual, covered in feathers and carrying a high metabolism, could quickly become nonviable in a hot climate.
The simulator must therefore integrate internal temperature, mass, plumage, exposed surface area, vascularization, and heat dissipation strategies. Some phenotypes will have to be rejected not because they are morphologically far from the target, but because they would overheat.
This is one reason Raptor cannot be a simple silhouette generator. A form can be visually convincing and physiologically impossible.
The question of teeth also calls for caution.
Activation of beta-catenin can initiate dental structures in chickens. But forming first-generation teeth is not the same as restoring a complete, functional dentition, with mature enamel and continuous replacement.
Birds lost their teeth about 116 million years ago. Several genes linked to enamel and dentin have become pseudogenes: ENAM, AMELX, AMBN, AMTN, MMP20, DSPP. Restoring a fully functional dentition would therefore probably require far more than activating a single pathway.
In Raptor, this means classifying teeth by levels:
This distinction avoids overstating what the literature actually supports.
Raptor does not claim to recreate a Velociraptor.
It does not claim to produce an extinct species.
It does not claim that a Magpie Goose could become a real dromaeosaurid.
What Raptor proposes is something else: exploring a biologically constrained morphological space between a modern bird and a related theropod form. It tests how far known, theoretical, or unknown mechanisms can be combined without breaking the coherence of the body.
The result, if such a program were ever transposed to the living, would not be a dinosaur. It would be a new organism, a chimera in the morphological sense: an animal that never existed, combining ancient, modern, and artificially selected traits.
That is precisely why the simulator is essential. Before any real biological speculation, the possible, the impossible, the dead ends, and the dangerous zones must be mapped.
Raptor is as much a project about biology as it is a project about artificial intelligence.
It asks a simple question: what happens when an evolutionary system is given a body, a world, physical constraints, developmental constraints, and selection pressure?
Intelligence then does not come from a dataset. It does not come from imitation. It comes from the interaction between an organism and its environment.
That is embodied cognition.
In such a framework, behavior is not entirely programmed. It emerges from morphology, constraints, available energy, risks, opportunities, and survival goals. A medium-sized predator, too small to dominate by brute force alone but large enough to be dangerous, is pushed toward strategies: anticipation, coordination, exploration, learning, communication.
This does not mean such an animal would automatically become “intelligent” in the human sense. But it creates a selective terrain where complex behaviors have strong reasons to emerge.
This is the direct link with genetic algorithms: give an agent a body, an environment, and constraints, and you do not only get an optimized solution. Sometimes you get behaviors you did not explicitly write.
Raptor is not a promise of de-extinction.
It is a computational research instrument.
It combines:
Its purpose is to explore, compare, falsify, and understand.
The question is not: “can we build a raptor?”
The question is: which morphogenetic trajectories remain coherent when imagination is constrained by biology?
And if some trajectories hold, then Raptor becomes more than a speculative experiment. It becomes a method for exploring the boundary between morphology, development, artificial evolution, and embodied cognition.
Genetic algorithms + developmental biology + physical simulation = emergent morphogenic evolution.