Every electric vehicle on the road has them. Every wind turbine spinning on a ridge in west Texas has them. Every pair of AirPods, every MRI machine, every hard drive, every guided missile in the Pentagon‘s inventory has them. Neodymium-iron-boron magnets—NdFeB if you’re reading a spec sheet, “neo magnets” if you’re not—are the strongest permanent magnets commercially available, and they are so deeply embedded in modern technology that removing them from the supply chain would be roughly equivalent to removing concrete from construction. You could technically build things without them. You just wouldn’t want to see what you’d get.

The thing is, almost nobody knows how they’re made. The manufacturing process is genuinely fascinating—part metallurgy, part materials science, part geopolitical thriller—and understanding it explains why these magnets are at the center of a supply chain crisis that involves export controls, Pentagon loans, tariff threats, and the kind of great-power competition that used to be about oil and is now about a silvery metal most people can’t pronounce.

What makes them special

A neodymium magnet is an alloy of three elements: neodymium (a rare earth element, atomic number 60), iron, and boron. The compound—Nd2Fe14B—forms a tetragonal crystal structure that was discovered independently by General Motors and Sumitomo Special Metals in 1984, which is the materials science equivalent of two people showing up to a party wearing the same outfit except the outfit happens to reshape global manufacturing for the next four decades.

What makes this crystal structure so magnetically powerful is its exceptionally high magnetocrystalline anisotropy—the atomic-level property that determines how strongly a material resists demagnetization. In plain language: the crystal lattice is shaped such that the magnetic domains align along a single preferred axis with extreme reluctance to flip. The energy density is roughly ten times higher than a standard ferrite magnet, which means a neodymium magnet the size of a quarter can do the work of a ferrite magnet the size of a coffee mug. That size-to-strength ratio is why they ended up everywhere. When you need a powerful magnetic field in a small package—an EV motor, a drone, a missile guidance system, an earbud—there is no practical substitute.

How they’re actually made

The manufacturing process is powder metallurgy, and every step matters. Screw up the particle size, the alignment pressure, or the sintering temperature by a small margin and you get a mediocre magnet instead of a great one. This is not an industry where you can wing it.

It starts with strip casting. The raw materials—neodymium (often with some praseodymium substituted in because it’s cheaper and chemically similar), iron, and boron, plus small additions of dysprosium or terbium for high-temperature applications—are melted together in a vacuum induction furnace at around 1,300°C. The molten alloy is poured onto a rapidly spinning, water-cooled copper roller, which solidifies it into thin strips. The rapid cooling is critical: it produces a fine-grained microstructure that’s optimized for the next step.

Those strips go through hydrogen decrepitation—you expose them to hydrogen gas, which diffuses into the grain boundaries and causes the alloy to crack apart into coarse chunks. This is nature doing the first stage of size reduction for you. From there, the chunks go into a jet mill operating in a nitrogen atmosphere, where high-pressure gas streams grind the material into an extremely fine powder with an average particle size of about 3 microns. That’s roughly the size of a red blood cell. The nitrogen atmosphere prevents oxidation, which would ruin the magnetic properties—neodymium is ferociously reactive with oxygen, which is also why the finished magnets need protective coatings, but we’ll get there.

Now comes the step that makes or breaks the magnet: magnetic field alignment and pressing. The powder goes into a mold, and a powerful external magnetic field—several tesla—is applied. This field physically rotates the tiny crystalline particles so their easy magnetization axes all point the same direction. The aligned powder is then compressed under enormous pressure. The alignment quality during this step directly determines the magnet’s maximum energy product—the BHmax value that shows up on the spec sheet and tells an engineer how much magnetic work the magnet can do per unit volume. A poorly aligned magnet with the exact same chemical composition will be measurably weaker than a well-aligned one. The process matters as much as the recipe.

The compressed “green compact” is then sintered in a vacuum furnace at approximately 1,050°C. Sintering fuses the powder particles together without fully melting them—it’s the difference between welding and soldering, conceptually—creating a dense, solid block with the internal crystal alignment locked in place. After sintering, the magnet goes through a two-stage annealing process at around 900°C and then 600°C, which relieves internal stresses and dissolves unstable phases that would degrade performance over time.

At this point you have a block of sintered NdFeB that is extremely hard, extremely brittle, and not yet the shape anyone needs. Machining comes next—diamond-tipped saws and grinding wheels cut the blocks into the precise geometries required for specific applications: arcs for motors, discs for speakers, rings for sensors. This is delicate work because the material shatters like ceramic if you look at it wrong. The kerf loss (material wasted in cutting) is a meaningful cost factor, especially when neodymium oxide costs upward of $70 per kilogram.

Then: coating. Unprotected NdFeB corrodes aggressively. The neodymium-rich grain boundary phase reacts with moisture and oxygen, forming hydroxides that literally cause the magnet to disintegrate over time—structural failure from the inside out. The standard solution is a multi-layer nickel-copper-nickel electroplating, though epoxy coatings, zinc plating, and parylene are used depending on the application environment.

Finally, the magnet is magnetized. A pulse magnetizer blasts it with a field of approximately 5 tesla, which saturates the aligned domains and produces a permanent magnet ready for installation. The whole process, from raw oxide to finished magnet, involves dozens of precision-controlled steps, and the yield at each stage matters. This isn’t assembling a product. It’s growing one.

Why this is a geopolitical problem

China produces roughly 85% of the world’s NdFeB magnets. Not 85% of the neodymium—85% of the finished magnets. Japan and Vietnam account for most of the rest. The United States, as of early 2025, produced approximately zero sintered NdFeB magnets at commercial scale. That’s a supply chain that has the resilience of a house of cards in a wind tunnel, and everyone involved knows it.

The concentration isn’t accidental. China made a strategic bet on rare earth processing decades ago—Deng Xiaoping reportedly said in 1992 that the Middle East has oil and China has rare earths—and then spent thirty years building out the mining, separation, refining, alloying, and magnet manufacturing infrastructure while everyone else was content to buy cheap finished products. The result is a vertically integrated supply chain that’s extraordinarily difficult to replicate quickly, because it’s not just the magnet factory you need. It’s the solvent extraction plant, the oxide separation facility, the metal reduction furnace, the alloy production line, and the workforce that knows how to run all of it. Each step has its own chemistry, its own equipment, its own failure modes.

The consequences of this concentration became concrete in April 2025, when China imposed export licensing requirements on dysprosium, terbium, and finished magnets. Export volumes reportedly dropped roughly 74% in May compared to the prior year. These aren’t abstract tariff games—dysprosium is the element that gives NdFeB magnets their high-temperature performance, and without it, the magnets in an F-35’s flight control actuators, an MQ-9 Reaper drone’s guidance system, or a Virginia-class submarine’s propulsion motor would lose their magnetic properties at operating temperatures. The Pentagon noticed.

The U.S. response has been a scramble. MP Materials—which operates the Mountain Pass mine in California, the only active rare earth mining operation of scale in the country—opened a magnet manufacturing facility in Fort Worth, Texas, in 2025. They began trial production of automotive-grade sintered NdFeB magnets late that year, with a target capacity of about 1,000 metric tons per year. For context, global production is somewhere around 220,000 to 240,000 metric tons annually. So the Fort Worth facility, at full capacity, would represent roughly 0.4% of global supply. It’s a start. It’s not a solution.

The Pentagon, meanwhile, awarded a conditional $620 million loan to Vulcan Elements and ReElement Technologies to scale domestic magnet production for defense applications. President Trump publicly threatened 200% tariffs on Chinese goods if Beijing restricted rare earth magnet shipments. The EU passed the Critical Raw Materials Act. Everyone is suddenly very interested in a supply chain they ignored for thirty years.

Why there’s no easy substitute

The reason this matters so much is that there is no drop-in replacement for NdFeB in most high-performance applications. Ferrite magnets are cheap and abundant, but they’re roughly one-tenth the energy density—you’d need a motor ten times the size to produce the same torque, which defeats the purpose of using permanent magnets in the first place. Samarium cobalt magnets handle high temperatures better but cost significantly more and use cobalt, which has its own supply chain problems centered on the Democratic Republic of Congo. Researchers have explored ferrite-based alternatives, iron nitride, and manganese-based compounds, but none have come close to NdFeB’s combination of magnetic strength, manufacturability, and cost at scale.

The substitution problem is especially acute in two sectors: electric vehicles and wind turbines. A typical EV traction motor uses 1 to 2 kilograms of NdFeB magnets. A direct-drive offshore wind turbine—the kind being deployed at scale in the North Sea and off the U.S. Atlantic coast—uses roughly 600 kilograms per megawatt of capacity. If you’re planning to electrify the global vehicle fleet and simultaneously triple offshore wind capacity by 2040, you need a lot more neodymium than currently exists in the processing pipeline. The bottleneck isn’t the ore. It’s the processing, the separation, and the magnet manufacturing—and those bottlenecks are sitting in a country that has demonstrated a willingness to use them as leverage.

This is the kind of constraint that doesn’t show up in the optimistic energy transition models, and it’s the kind of thing that makes the difference between a plan that works on a slide deck and a plan that works on a random Tuesday when the supply ship doesn’t arrive.

We cover neodymium magnets—along with 35 other critical elements and minerals, from lithium to uranium to gallium nitride—across 36 lectures in our Rare Earth Elements & Critical Minerals course. If you want the full supply chain story, from the Bayan Obo mine to the inside of an F-35 actuator, that’s where it lives.

When I read headlines about brain-computer interfaces—and there’s a new one roughly every 72 hours, each seemingly announcing that the future has arrived—I’m reading them with the same part of my brain that reads MRI reports and discharge summaries. I’m looking for the mechanism, the sample size, the follow-up period, and the part of the press release that got quietly omitted. And what I keep finding is a field where the actual science is genuinely remarkable, the engineering is legitimately impressive, and the gap between what’s been demonstrated and what’s being promised is roughly the width of the Grand Canyon.

So here’s where brain-computer interfaces actually stand in March 2026. Not the press release version. Not the “we’re five years from The Matrix” version. The clinical reality, company by company, with the caveats attached.

What a BCI actually does (the 30-second version)

Your motor cortex generates electrical signals when you intend to move. In a healthy nervous system, those signals travel down your spinal cord, through peripheral nerves, and to your muscles. In someone with a spinal cord injury or ALS, the signals still fire at the top—the brain is still doing its job—but the wiring downstream is broken. A brain-computer interface picks up those electrical signals directly from the cortex and routes them to an external device instead of to muscles. You think about moving your hand, the electrodes record the neural activity, a decoder translates it into a digital command, and a cursor moves on a screen or a robotic arm reaches for a cup. That’s it. That’s the core mechanism. Everything else is engineering.

The engineering, of course, is where it gets complicated—and where the companies diverge in ways that matter enormously for which patients actually benefit, when, and at what risk.

Neuralink: The one you’ve heard of

Neuralink gets roughly 95% of the media coverage in this space despite being neither the first nor the furthest along clinically. What they have is Elon Musk, which in the attention economy is worth more than a decade of peer-reviewed publications. The N1 implant is a coin-sized device with 1,024 electrodes distributed across 64 ultra-thin polymer threads, inserted into the motor cortex by a custom surgical robot. The threads are thinner than a human hair—about 5 microns—and that’s genuinely impressive from a materials science standpoint. The pitch is high electrode count plus wireless transmission plus a cosmetically invisible implant that sits flush with the skull.

As of early 2026, Neuralink has 21 participants enrolled in its global clinical trials, up from 12 in September 2025. The first patient, Noland Arbaugh—quadriplegic from a diving accident—demonstrated the ability to control a cursor, play video games, browse the internet, and post on social media using the implant. The third patient, Brad Smith, who has ALS and is ventilator-dependent, can type using the device. These are real outcomes. They matter. A person who cannot move anything below the neck using thought alone to navigate a computer is not a small thing.

But—and this is where my press release detector starts going off—Musk announced on December 31, 2025, that Neuralink would begin “high-volume production” of BCI devices and move to “almost entirely automated surgical procedures” in 2026. He also announced the Blindsight implant for restoring vision in the blind would begin its first patient trial this year. If you’ve followed Musk’s timeline promises across any of his companies—Tesla Full Self-Driving, the Cybertruck, the Boring Company’s Vegas Loop—you know that announced timelines and delivered timelines have a relationship best described as aspirational. “High-volume production” in 2026 from a company with 21 trial participants is a claim that requires a lot of intermediate steps that haven’t been publicly demonstrated: manufacturing consistency, surgical standardization, long-term safety data, and FDA clearance for anything beyond an investigational device. None of those things have happened yet.

The first patient also had a documented issue: some electrode threads retracted from the cortex after implantation, reducing the number of functional electrodes. Neuralink adjusted its approach for subsequent patients, but this is exactly the kind of biocompatibility challenge that doesn’t show up in the demo reel. Brain tissue is not a circuit board. It’s alive, it moves, it forms scar tissue around foreign objects, and it does not appreciate being punctured by 64 threads, no matter how thin they are.

Synchron: The one that doesn’t require brain surgery

Synchron is doing something fundamentally different, and I think the approach deserves more attention than it gets. Their Stentrode device is an endovascular BCI—it’s threaded up through the jugular vein and deployed inside a blood vessel on the surface of the motor cortex, using the same catheter techniques that interventional neurologists (my people) use every day to treat strokes and aneurysms. No craniotomy. No opening the skull. No penetrating brain tissue. The median procedure time in their COMMAND trial was 20 minutes.

The tradeoff is resolution. The Stentrode has 16 electrodes compared to Neuralink’s 1,024. It’s sitting inside a blood vessel, not directly in cortical tissue, so the signals it picks up are less granular—think of it as the difference between sitting in the front row at a concert versus listening from the parking lot. You can still hear the music, but you’re not picking up the individual instruments. For the current application—point-and-click cursor control, typing, navigating apps—that’s sufficient. An ALS patient in their trial became the first person to control an iPad with a BCI, using Apple’s native Switch Control accessibility feature. He later connected to an Apple Vision Pro and Amazon Alexa using only his thoughts. These are consumer devices, unmodified, working with a brain implant through standard accessibility protocols. That’s a very different value proposition than a research demo in a lab.

Synchron’s COMMAND trial—six patients, 12-month follow-up—met its primary safety endpoint with no device-related serious adverse events to the brain or vasculature. They raised $200 million in Series D funding in late 2025, backed by Bezos, Gates, and the Qatar Investment Authority, and are preparing pivotal trials for 2026 ahead of commercial approval. They’ve also announced a next-generation whole-brain interface with significantly higher channel counts, though details won’t arrive until later this year.

The strategic bet Synchron is making is that a lower-resolution device implanted through an existing, well-understood medical procedure will reach more patients faster than a higher-resolution device that requires brain surgery. From a regulatory and clinical adoption standpoint, that bet has a lot of logic behind it. Interventional neuroradiologists already know how to navigate catheters through cerebral vasculature. The training curve is short. The risk profile maps onto procedures we’ve been doing for decades.

BrainGate and Blackrock Neurotech: The ones that were here first

The BrainGate consortium—a collaboration across Mass General Brigham, Brown University, Stanford, and several other institutions—has been running clinical trials of implantable BCIs since 2004, which is two decades before Neuralink implanted its first patient. Their technology is built on Blackrock Neurotech’s Utah Array, a grid of 96 silicon microelectrodes that penetrates the cortex. It’s the workhorse of the field. Nathan Copeland, implanted in 2015, holds the record for longest continuous use of a brain-computer interface. In 2016, he used a robotic arm to fist-bump Barack Obama, which remains the single best piece of BCI marketing ever produced despite not being marketing at all.

In 2025, a BrainGate team at UC Davis demonstrated that a man with ALS could speak through a BCI-driven voice synthesizer reconstructed from recordings of his pre-disease voice. More recently, a BrainGate study published in Nature Neuroscience showed two paralyzed patients typing on a standard QWERTY keyboard layout using attempted finger movements—not imagined cursor control, actual attempted typing—at speeds approaching 90 characters per minute. That’s not far from the average typing speed of a non-disabled person on a phone. The decoder is getting better because the algorithms are getting better, and the algorithms are getting better because we have more data from more patients over longer periods. This is the boring, incremental, deeply important work that doesn’t generate Musk-level headlines but is actually what moves the field forward.

Precision Neuroscience: The one designed to come back out

Founded by Neuralink co-founder Benjamin Rapoport, Precision takes yet another approach. Their Layer 7 Cortical Interface is an ultra-thin film—one-fifth the thickness of an eyelash—studded with 1,024 platinum microelectrodes. It sits on the brain surface without penetrating tissue, slipped through a small slit in the skull. The critical difference: it’s designed to be safely removable. Every other invasive BCI is essentially permanent. Precision’s device can come out without damaging the cortex, which is a significant advantage for a field where the long-term effects of brain implants are still being studied.

Precision received FDA 510(k) clearance—the first full regulatory clearance for any of the new commercial BCI technologies—for implants lasting up to 30 days, and they’ve performed 38 human implant procedures. Their current application is short-term brain mapping and neural data collection, not yet chronic implantation for daily use. But the data they’re collecting is building the training sets for neural decoding algorithms that could eventually power long-term devices.

What’s actually hard (the part nobody puts on the slide deck)

Here’s what I think about when I read BCI press releases, and what I wish the coverage spent more time on:

Signal degradation over time. The brain forms glial scar tissue around penetrating electrodes. The signal quality starts high and degrades over months to years as the immune response walls off the foreign object. This is the single biggest unsolved problem in chronic invasive BCIs, and no company has publicly demonstrated a solution that works at scale over five-plus years.

The decoder problem. Current BCIs require calibration—sometimes daily—because the neural signals drift. The relationship between a specific pattern of neural activity and an intended movement isn’t fixed. It shifts as neurons adapt to the implant, as the user’s neural strategies change, and as electrodes move or degrade. Making decoders that stay accurate without constant recalibration is an active area of research, and AI is helping, but it’s not solved.

Surgical risk. Any procedure involving the brain carries risk. Infection, hemorrhage, device failure requiring revision surgery—these are not theoretical concerns. They’re the everyday calculus of neurosurgery. Neuralink’s thread retraction in their first patient is a concrete example. Synchron’s endovascular approach has a lower risk profile precisely because it doesn’t penetrate the brain, but even catheter-based procedures have complication rates.

The regulatory path. These devices are in early feasibility trials. The road from there to FDA-approved commercial products typically takes years—pivotal trials with larger patient populations, long-term follow-up data, manufacturing quality controls, post-market surveillance plans. Musk saying “high-volume production in 2026” doesn’t change the regulatory timeline. The FDA moves at the speed of evidence, not the speed of X posts.

Where this goes

The honest assessment: BCIs for people with severe paralysis will almost certainly become commercially available within the next three to five years. Not as consumer products—as medical devices, implanted by neurosurgeons, prescribed for specific clinical indications, and covered (eventually) by insurance. The first approved products will probably do what the current trials are demonstrating: cursor control, typing, basic device navigation. Not telepathy. Not memory enhancement. Not uploading your consciousness to the cloud.

The longer-term applications—speech decoding for people who’ve lost the ability to talk, sensory feedback for prosthetic limbs, treatment of neuropsychiatric conditions—are genuinely possible but further out. They require higher-resolution interfaces, better algorithms, and a much deeper understanding of neural coding than we currently have.

What I want people to take away from this is that the technology is real, the progress is meaningful, and the patients benefiting from these devices are experiencing something that would have been science fiction 20 years ago. But the field is in its early clinical phase, and the distance between “a paralyzed person can move a cursor” and “BCIs are a consumer product” is enormous—measured not in engineering breakthroughs but in safety data, regulatory milestones, and the slow, grinding, essential work of proving that these devices help more than they harm over the long term.

We cover the full history, science, and engineering of brain-computer interfaces—from the earliest EEG experiments to every company and approach described above—across 48 lectures in our Neuroprosthetics & Brain-Computer Interfaces course. If this piece made you want the granular version, that’s where it lives.

In 1992, a neuroscientist at the University of Parma named Giacomo Rizzolatti was studying the premotor cortex of macaque monkeys—specifically, the neurons that fired when a monkey reached for a peanut. Standard motor mapping stuff. Electrode in the brain, monkey grabs food, neuron fires, graduate student logs it, everybody goes home. Except one afternoon, a researcher reached for his own lunch in front of the monkey, and the same neuron fired. The monkey wasn’t moving. It was watching someone else move. And the cell lit up like it couldn’t tell the difference.

That’s the origin story of mirror neurons, and it’s one of those moments in neuroscience where a single observation cracks open a door that everyone then spends thirty years arguing about the size of. The finding was replicated, published in 1996, and promptly became one of the most overhyped discoveries in the history of brain science—V.S. Ramachandran called them “the driving force behind the great leap forward in human evolution,” which is the neuroscience equivalent of calling a rookie quarterback the next Tom Brady after one preseason game. The actual data, as usual, is more interesting than the hype, and considerably more complicated.

So what do mirror neurons actually do? The basic mechanism is straightforward: these are neurons in the premotor and parietal cortex that fire both when an animal performs an action and when it observes another individual performing the same action. Grab a peanut, the cell fires. Watch someone else grab a peanut, the same cell fires. The neuron doesn’t distinguish between doing and seeing—or more precisely, it encodes both, which is a meaningfully different claim than the pop-science version where your brain “simulates” everything it sees like some kind of empathy PlayStation.

The pop-science version went roughly like this: mirror neurons are the biological basis of empathy, imitation, language, theory of mind, and possibly the entire foundation of human civilization. You can still find TED talks making this argument. The actual neuroscience community has, over the past two decades, walked most of that back—not because mirror neurons aren’t real or important, but because the leap from “this neuron fires during observation and execution” to “this neuron explains human culture” requires about fourteen intermediate steps that nobody has convincingly demonstrated.

Here’s what we actually know, species by species.

Macaques remain the best-studied case because you can do single-neuron recordings in them, which you generally cannot do in humans for obvious ethical reasons involving the part where you stick an electrode into someone’s brain. Rizzolatti’s lab and subsequent groups have mapped mirror neurons primarily in area F5 of the ventral premotor cortex and in the inferior parietal lobule. These neurons are action-specific—they respond to hand grasping, mouth actions, tool use—and they’re modulated by context. A macaque mirror neuron that fires when it watches another monkey grasp a peanut to eat it may not fire when the same monkey grasps the same peanut to place it in a container. The neuron isn’t just mirroring movement. It’s encoding the goal of the action, which is a much more interesting finding than the simple mirror story.

The caveat—and this matters—is that macaques are actually terrible imitators. They don’t readily copy novel behaviors from observation. So if mirror neurons are supposedly the neural substrate of imitation, we have a problem, because the species in which they were discovered doesn’t really imitate. This is the kind of inconvenient fact that tends to get footnoted rather than headlined.

Great apes are a different story. Chimpanzees, bonobos, gorillas, and orangutans all demonstrate genuine imitation—learning novel motor sequences by watching others perform them. The problem is that single-neuron recordings in great apes are extremely rare for ethical and practical reasons, so the direct electrophysiological evidence for mirror neurons in apes is thin. What we have instead is a lot of fMRI and behavioral data suggesting that homologous brain regions (the ape equivalents of F5 and the inferior parietal cortex) are active during action observation. The inference is reasonable—these are our closest relatives, the anatomy is conserved, the behavior is consistent—but it’s still an inference, not a measurement. We’re reading the box score, not watching the game.

Humans are where the story gets both more exciting and more contentious. You can’t ethically do single-neuron recordings in healthy humans, but a handful of studies in epilepsy patients with implanted electrodes (who were being monitored for seizure localization, not mirror neuron research) have found neurons in the supplementary motor area and medial temporal lobe that respond to both observed and executed actions. Iacoboni’s UCLA group published some of this work in the 2010s. The broader human evidence comes from fMRI, EEG mu-suppression studies, and transcranial magnetic stimulation—all of which point to a “mirror neuron system” distributed across premotor cortex, inferior parietal lobule, and the superior temporal sulcus. The system is real. The question is what it actually does versus what we’d like it to do.

The honest answer, as of 2026: mirror neurons in humans are probably involved in action understanding—recognizing what someone is doing and predicting what they’ll do next. There’s decent evidence they contribute to motor learning through observation. The link to empathy is much weaker than the popular narrative suggests, and the link to language is speculative at best. Gregory Hickok’s 2014 book The Myth of Mirror Neurons did a pretty thorough job of separating the signal from the noise here, and the field has been more careful since.

Now, here’s where it gets genuinely weird. Because mirror neurons—or at least mirror-like neural systems—aren’t limited to primates.

Songbirds have what might be the most compelling mirror system outside of mammals. In zebra finches and other oscine songbirds, neurons in a region called the HVC (used to stand for “High Vocal Center” but now it’s just HVC because the original name was anatomically inaccurate, which is the neuroscience version of a company rebranding after a scandal) fire both when the bird sings a specific note sequence and when it hears the same sequence sung by another bird. These aren’t just auditory neurons responding to sound—they’re sensorimotor neurons that link production and perception of the same vocalization. The parallel to primate mirror neurons is striking, and it evolved completely independently, which tells you something about how useful this computational architecture must be.

The songbird mirror system is deeply involved in vocal learning—young birds learn their species’ song by listening to a tutor and gradually matching their own output to the template, and the mirror-like neurons in HVC are a critical part of that error-correction loop. This is arguably a cleaner example of mirror neurons supporting imitation than anything in the primate literature, which is both fascinating and slightly embarrassing for the people who spent two decades claiming mirror neurons were a uniquely primate innovation.

Parrots are the other avian case worth knowing. Alex the African Grey—Irene Pepperberg’s famous research subject—could label objects, understand concepts like “same” and “different,” and produce novel combinations of learned words. Parrots are vocal learners like songbirds, but they’re not closely related to them—vocal learning evolved independently in parrots, songbirds, and hummingbirds, which means the mirror-like neural circuitry that supports it likely evolved independently too. Parrot neuroscience is less developed than songbird work (partly because parrots are harder to work with and live approximately forever), but the behavioral evidence for action-perception coupling is strong. A parrot that watches you wave and then waves back is doing something that macaques—the species where we actually found mirror neurons—basically can’t do.

Dolphins present maybe the most interesting case because they combine vocal learning, complex social cognition, and a brain that is anatomically very different from a primate brain. Dolphins can imitate novel motor behaviors on command (the “do this” paradigm developed by Louis Herman’s lab at the University of Hawaii in the 1990s), and they engage in vocal mimicry—copying signature whistles of other dolphins, which functions as something like calling someone by name. The neural basis is largely unknown because, to state the obvious, you cannot put a dolphin in an fMRI scanner with any meaningful cooperation, and single-neuron recordings in cetaceans are essentially nonexistent. What we have is behavioral evidence that strongly implies a mirror-like system, layered on top of a brain with a completely different cortical organization—dolphins have an insular cortex that may serve some of the functions that premotor cortex serves in primates, but honestly, cetacean neuroanatomy is still more question marks than answers.

The pattern that emerges across all these species is that mirror-like neural mechanisms seem to pop up wherever you find sophisticated social learning—whether that’s vocal imitation in songbirds, motor imitation in apes, or behavioral mimicry in dolphins. And these systems evolved independently in lineages that diverged hundreds of millions of years ago, which suggests that coupling action perception to action production is such a useful computational trick that evolution keeps reinventing it. It’s convergent evolution at the neural architecture level, which is roughly as cool as neuroscience gets.

What the pop-science narrative got wrong was the specificity of the claim. Mirror neurons aren’t the secret to human empathy or the origin of language or the biological basis of civilization. They’re a neural mechanism for linking what you see to what you do—one piece of a much larger puzzle that includes prefrontal cortex, temporal lobe social cognition networks, and a dozen other systems that we’re still mapping. But what the pop-science narrative got right, even if accidentally, was the intuition that something deep is happening when one brain watches another brain act and encodes that observation in the language of its own motor system. That’s not empathy, exactly. But it’s the scaffolding that makes empathy—and imitation, and social learning, and maybe culture—mechanistically possible.

The fact that an octopus, which diverged from our lineage over 500 million years ago, can watch another octopus open a jar and then do it themselves raises the question of whether mirror-like computation might be even more widespread than we currently think. We genuinely don’t know. The electrophysiology hasn’t been done. But the behavioral signatures keep showing up in species we didn’t expect, and every time they do, the story gets bigger.

We cover mirror neurons—and the broader neuroscience of social cognition across the animal kingdom—in depth across several lectures in our Neurozoology course, which traces the evolution of cognition from mycelial networks to primate brains across 48 lectures and 69 hours of audio. If the octopus jar thing made you want to know more, that’s a good sign.