What this covers

Bing Brunton discusses how simulating the fruit-fly connectome—the neural wiring diagram—becomes scientifically generative only when the simulation is embodied in a physically realistic body rather than abstracted away. The episode centers on a concrete case: her team simulated the ~4,000-neuron front-leg subnetwork of the fly ventral nerve cord, then computationally pruned it down to a minimal three-neuron circuit capable of generating walking rhythms. This reduction was not fitted to existing observations but produced genuine predictions—such as identifying a neck-descending neuron that, when later activated with a laser in real flies, produced the predicted leg tap. The core argument is that behavioral vs biological fidelity are not the same thing: a worm's 300-neuron connectome, when plugged into a reinforcement-learning algorithm to drive a simulated fly body, can produce convincing fly-like walking without capturing anything biologically meaningful about how flies actually work.

Brunton argues that because nervous systems evolved to control particular bodies with specific limbs, joints, and muscles, understanding a brain requires understanding it in context—a claim she connects to broader ideas about embodied cognition and why human intuition fails for recurrence and feedback loops that dominate neural circuits. She contrasts reductionist pruning (numerical simulation with no machine learning) against deep-learning approaches, arguing that the former preserves interpretability while the latter obscures mechanism. The discussion also addresses an older debate in neuroscience—whether the connectivity matrix alone suffices to understand function—and moves toward the possibility that digital twin models could illuminate rehabilitation and motor recovery, where biological understanding remains sparse. Brunton touches on why the larger fruit-fly connectome may paradoxically be easier to decode than the smaller, simpler C. elegans, and acknowledges that bio-inspiration vs biological detail are separable: engineering does not require copying biology's implementation to harness its ideas.

Sharpest takeaway

Brunton argues that the fruit-fly connectome, when simulated and embodied in a physically realistic body, can actually generate novel, testable predictions about how neural circuits control behavior—but only if the interfaces between brain and body are engineered meaningfully rather than papered over with deep learning, since behavioral fidelity does not imply biological fidelity.

  • Simulating the fly ventral nerve cord connectome and pruning it identified a minimal three-neuron circuit sufficient to generate walking rhythms—a model-first prediction.
  • Brains evolved to control specific bodies, so understanding the nervous system requires embodying it in a realistic body and environment ('the brain does not live in a jar').
  • A worm connectome can be trained to drive a fly body and walk convincingly, proving that behavioral realism alone is not evidence of biological meaning.

The argument · threads7 threads · 28 claims
0.67

Nervous systems evolved to control specific bodies and generate behavior through physical interaction with the world.

4 pointscentrality 5/5
  • Essentially all biological movement is cyclic—walking, running, swimming, slithering, crawling, even bacteria spinning their flagella—so the nervous system must have a way of generating rhythmic instructions for muscles, which is fundamental to locomotion as well as breathing and digestion.

    “all animal movements are rhythmic. Actually, not just animals, even bacteria move by spinning their flagellum”

  • The nervous system always evolved to control a specific body with particular limbs, muscles, joints, and sensors in order to move through the world, so the brain must be understood in the context of the body it interacts with to produce behavior—motivating building biologically interpretable, physically realistic 'digital twins' of behaving animals embodied in a physics-engine virtual environment.

    “It's so obvious to me that the brain does not live in a jar. It always controlled a body, and it always controlled a specific body with these limbs and these muscles and these joints and these sensors”

  • The longest cell in the human body is about as tall as the person: a single sensory neuron runs from the big toe all the way up to the brainstem, and having one cell span this distance—rather than handing the message off through many cells—allows the signal to travel very fast, which matters because the brain needs rapid information (e.g., to keep you from falling over when you kick a rock).

    “one end of it is at your big toe... And the other end of it, it goes all the way up to your brain stem... The advantage of having one cell do it is that you can do it really fast.”

  • The only agents we all agree are intelligent and conscious are embodied ones, and since all nervous systems evolved ~500 million years ago to control bodies for sensorimotor tasks—seeking food and mates and moving through the world—everything we call reasoning and consciousness was built on top of the neural machinery for sensorimotor control, so understanding that substrate is likely important and a strong constraint on how higher capabilities could have arisen.

    “We have no examples that we all agree on of agents that are intelligent and conscious except the ones that are embodied.”

0.67

Connectome data alone is insufficient; understanding requires integrating chemical signaling, cell identity, and mechanical computation with connectivity.

4 pointscentrality 5/5
  • The C. elegans connectome of ~300 neurons has been fully mapped for decades, yet we still do not understand it, because the worm performs much of its computation outside the neural connectivity matrix—through chemical communication (neurotransmitters squirted at each other) and mechanical computation (reflex loops coupled to its squishy body)—so having the wiring diagram alone was insufficient.

    “they do a lot of computation not using that connectivity matrix. There's a ton of chemical communication... There's a lot of mechanical computation.”

  • Unlike the genome, which everyone agrees should be at base-pair resolution, the term 'connectome' has no agreed-upon scale of description—some use it for cell-by-cell wiring while others use it for coarse-grained area-by-area maps (e.g., the human connectome), and experts disagree on what resolution is required.

    “we don't quite agree on the scale of description of like, what is a... Like, do you need to have every single neuron in that spider for that to be the connectome of a spider?”

  • The connectome is not just the connectivity matrix of which neurons connect to which; cell identity matters because a dopamine cell and a serotonin cell firing the same action potential convey completely different messages, and the message received depends on the receiving cell's identity—the same signal from cell A produces different effects in cells B and C.

    “when cell A speaks to cell B and cell A says exactly the same thing to cell C, depending on the identities of B and C, they could hear very different messages and do very different things with it.”

  • The longstanding debate over connectomes—articulated in essays by figures like Eve Marder and Cory Bargmann—was never about whether unobservable details (channels, biophysical properties) could be derived from the connectome (everyone agreed they couldn't); the real disagreement was whether the measurable connectivity matrix alone is sufficient to understand function, versus being utterly useless because all the other details are needed.

    “the disagreement was whether or not the stuff that you can measure, effectively these connectivity matrices, is that sufficient?”

0.53

Computational models enable discovery by predicting unknown phenomena and revealing minimal circuit architectures that human intuition cannot access.

4 pointscentrality 4/5
  • Unlike most computational modeling that fits equations to already-known observations, this work is a rare case where the model came before the experiments: starting agnostic from the connectome dataset, the team generated genuine predictions of things not previously known, including a neck-descending neuron predicted to make the leg tap when activated—which was then confirmed by ordering a pre-existing fly line, activating the neuron with a laser, and observing the leg tap.

    “This is one of the few instances where I feel like the model actually came before the experiments. We were agnostic going in. We had this giant dataset. We're like, 'Let's just simulate it.' And then we made a prediction of things we didn't know before.”

  • Humans can reason well through linear forward paths (A to B to C to D) but have very poor intuition for what happens with feedback circuits and recurrence (when D feeds back to B and C feeds back to A); since nervous systems are full of feedback and recurrence, this human cognitive limitation is a key justification for needing complex computational models that humans cannot replace with intuition.

    “As soon as there's recurrence, when D goes back to B and then C goes back to A, our intuition for what's going to happen is really poor.”

  • Rather than building up from individual experimentally-tractable components, the team took a reductionist approach: they simulated the full ~4,000-neuron front-leg subnetwork of the fly ventral nerve cord, confirmed it could generate motor rhythms, then computationally pruned cells one at a time until reaching the minimal circuit that still produced the rhythm—using brute-force numerical simulation with no machine learning or deep learning.

    “we took the reductionist approach. We're like, it's in here somewhere. We got it down to a network of 4,000 cells or so.”

  • The minimal circuit needed to generate the fly's leg-walking rhythm turned out to be just three identified neurons—two excitatory (E1, E2) and one inhibitory (I1)—connected in a specific, understandable architecture motif that explains why this tiny circuit can generate cycles; the team had expected dozens of cells would be needed.

    “The answer was three. Three cells. That's the minimum you need. And they have names. We know who they are in the fly nervous system.”

0.47

Embodied digital twins simulating nervous system and body together can illuminate motor control, plasticity, and therapeutic rehabilitation across timescales.

2 pointscentrality 4/5
  • Getting biologically meaningful results from connectome simulation depends on actually engineering the interfaces between nervous system and body—analogous to how the same software can run on a Mac or PC via an emulator if the interfaces (like HDMI) are right; if you don't care about meaningful interfaces, you can get many things to plug together without biological significance.

    “the real trick to get something meaningful out of it is to actually engineer those interfaces.”

  • Embodied digital-twin models could illuminate poorly-understood holistic interactions between the nervous system and musculoskeletal system across multiple timescales—how the body compensates for injury (limping, adopting a new gait, strengthening one side) via plasticity, growing new muscles, tendons or neural pathways—potentially yielding new therapeutic guidance for rehabilitation, where even physical therapists disagree on best practices.

    “What happens when we break it in different ways? How does it compensate? What is the role of plasticity?... if we understand that, perhaps we'll have some new clues about how to design therapeutics to help that process work better”

0.44

Biological economy and scale-dependent circuit organization suggest larger connectomes with specialized cell types may be more interpretable than tiny highly-multiplexed ones.

10 pointscentrality 2/5
  • Just as a civil engineer designing a roof beam doesn't need to solve Schrodinger's equations for the atoms in the beam, neuroscientists building holistic models of an animal's nervous system don't need every biophysical detail of each neuron—but the difficulty is that we know we don't need all details, yet we don't know which specific details are actually crucial.

    “But we are not solving Schrodinger's equations in order to design a roof. It's just way too much.”

  • Roughly half of the cells in the brain are not neurons but glia, which were historically dismissed as mere custodial support but are now understood to be involved in vital functions and have their own dynamics, making them an exciting emerging field in neuroscience.

    “the rough estimate, in my understanding, is that half of the cells in your brain are not neurons. Our word for not neurons is just glia”

  • Biology is not wasteful with neurons—because neurons are among the most expensive cells in the body to maintain, if a cell exists there is probably a good reason for it; over a long enough timescale the body would eliminate truly unnecessary cells, making vestigial neurons less plausible than vestigial organs (which persist because evolution got them 'stuck').

    “Because the cells are so expensive to maintain. Neurons are some of the most expensive cells to maintain in your body. I think my hypothesis would be that if a cell is actually not necessary, the body would find a way for it not to be there over a longer time frame.”

  • Paradoxically, the larger fruit-fly connectome (150k brain + 22k ventral nerve cord neurons) may be easier to understand from the connectivity matrix than the tiny C. elegans (300 neurons), because the fly has jointed limbs, enough cells to form distinct cell types rather than every cell being a unique snowflake, and relies less on multiplexed multimodal cells.

    “the connectivity matrix of the fruit fly, because of, that it is a little bit bigger, it has jointed limbs just like humans do, and it has enough cells that there're actual cell types. Like, not every single cell is just its own little snowflake.”

  • Reconstructing a connectome at sufficient scale is technically extraordinarily difficult, involving running a transmission electron microscope for six months straight while making zero mistakes, which fueled early skepticism that it was even possible.

    “it involves, I don't know, running a transmission electron microscope for six months straight, making zero mistakes.”

  • C. elegans neurons are so unspecialized that, due to extreme compression in such a tiny nervous system, single cells handle multiple sensory modalities at once (multiplexing), unlike larger animals with dedicated visual and olfactory systems—a computational feature of how the worm nervous system works.

    “They have single cells that have multiple sensory modalities going into it because it's just so tiny, it's so compressed... They've had to multiplex in that way.”

  • Building artificial intelligence does not require understanding the details of how biology implements intelligence; bio-inspired engineering repeatedly shows that the mere concept that biology does something is enough to inspire good solutions—imitating bird flight failed and had to be discarded for fixed-wing aircraft, and roboticists implemented central pattern generators just from the concept of a spinal rhythm generator without needing the actual cells.

    “imitating the way that birds fly was an utter failure. We had to throw it all out the window and start over. Fixed-wing aircraft, that's where it's at”

  • Even seemingly structural body parts like bones are not inert like a stainless-steel beam: while their material properties come largely from a calcium matrix, bones are intricate meshy structures threaded with blood vessels because they must be vascularized to receive sugar and oxygen, and only cells can keep them alive.

    “even something that you think is structural, it's not like a stainless steel beam in a building. It's alive... And so there's cells all inside it”

  • An artificial intelligent system may need to be embedded in something interactive with a physics—not necessarily our physics, but with rules, constraints, some notion of energy and conservation laws rather than limitless everything—though Brunton flags this as speculation rather than established fact.

    “the concept that it may be necessary for that agent, for that intelligent system to be actually embedded in something that's interactive, that has a physics, so to speak. It doesn't have to be our physics, but I think it should have maybe some rules and constraints, some notion of energy, some notion of conservation laws”

  • All living organisms as we know them are composed of cells, though there is also non-cellular material in the extracellular space, and viruses are a debatable edge case.

    “all I meant was that all of life as we know it is made out of cells.”

0.36

Central pattern generators, the best-understood neural circuits, establish that rhythmic locomotion arises from circuit architecture, not reflexes alone.

3 pointscentrality 3/5
  • The three-neuron circuit is hypothesized to be sufficient to generate the basic walking rhythm but not to actually control walking, because legs have many more muscles than degrees of freedom that must be coordinated to move without being clumsy, so additional cells are required to translate the rhythm into coordinated movement.

    “we believe these three neurons, our hypothesis is at least three neurons generates the basic rhythm. And then there's other cells involved to make it actually walk.”

  • Since experiments in the 1910s, it has been known that rhythmic movement is generated not by reflex alone but by the central nervous system itself via central pattern generator (CPG) circuits, which also drive breathing and digestion—and the best-understood CPG is the crab digestive (stomatogastric) circuit, arguably the only neural circuit truly understood in all of neurobiology.

    “Ever since the 1910s, some of the first experiments demonstrated that the generation of these rhythms is not by reflex only”

  • Unlike all other flying animals—where bird wings and bat wings are modified arms—insect wings are evolutionarily novel limbs that did not arise from pre-existing legs or arms, and this novelty is reflected in dedicated clumps of cells in the ventral nerve cord that correspond to the wings.

    “The insect wings are actually novel limbs... They're not modified legs. There's theories about exactly how they evolved, but they're actually novel structures.”

0.31

Behavioral matching between mismatched connectomes and bodies proves fidelity without biological correctness, requiring careful interface engineering to avoid spurious results.

1 pointcentrality 3/5
  • Matching an animal's observed behavior is not evidence of biological correctness: a worm (C. elegans) connectome of only 300 cells can be plugged into a reinforcement-learning algorithm to control a biomechanically realistic fly body and produce convincing fly-like walking, demonstrating that if enough deep learning is used to tune unconstrained parameters, the connectome and the realistic body become biologically meaningless—you get behavioral fidelity without any biological fidelity.

    “You can get behavior fidelity without any biological fidelity.”