The first Neuralink patient beat a world record with most of his electrodes gone

Lex Fridman recorded this conversation at Neuralink, walking from Elon Musk through co-founder DJ Seo, head neurosurgeon Matthew MacDougall, brain-interface software lead Bliss Chapman, and finally Noland Arbaugh, the first human to receive the device. Musk opens with the second implant already in a person and more than 400 electrodes returning signal. How fast the trial scales, he says, is a regulatory question: the company is hoping for ten participants by the end of 2024.

The surprising early result is bandwidth. Even with only roughly 10 to 15 percent of Noland’s electrodes working, the team reached a bit rate twice the previous world record. Musk’s near-term forecast is not modest. He talks about hundreds, then thousands of bits per second, and maybe a megabit within five years — faster than anyone can type or speak. At that point, he says, the bottleneck is the other person, unless they have an implant too.

He likes to run the numbers on an ordinary day. There are 86,400 seconds in 24 hours, and almost nobody emits 86,400 tokens in that window, so the average human output rate is less than one bit per second. Language is a lossy codec: you model the other mind, compress a concept into syllables, and hope they decompress it into something close. Memes, he adds, are denser compression still, because they ride on a template already sitting in the listener’s head.

The long-term aspiration is not a better mouse; it is to keep humans in the loop with AI by raising communication bandwidth. Even in a benign scenario, Musk says, an intelligence that talks in terabits will get bored waiting for a few words. Fridman asks the obvious follow-up — what use are humans then — and Musk’s answer is will, or purpose. Cortex already serves the limbic system, phones and laptops are a tertiary compute layer, and the implant is meant to stop that stack from leaving the biological layer behind.

Asked whether a high-channel implant will mostly talk to AI, Musk insists on the medical stack first. The opening job is bridging neuron damage in the spinal cord, neck, or brain, as with the first two patients. A second product, Blindsight, would skip ruined eyes and optic nerves and stimulate visual cortex directly. Schizophrenia, seizures, and memory access sit further up what he calls a tech tree: you need letters before you get *Lord of the Rings*.

He is blunt about risk. A new device cannot be made perfectly safe, so you start where the reward is highest: someone who has almost no path from brain to body. Stephen Hawking is his example of how profound faster communication could have been. Once thousands of people have used the implant for years and the risk looks minimal, then you can talk about augmentation. Until then, he says, the sensible order is neuron damage, then more ambitious uses.

That does not mean aiming only for parity with able-bodied typing. If you are already in the motor cortex of a person with quadriplegia, Musk’s instinct is to give them a data rate that *exceeds* a typical human. The same logic applies to vision. Early Blindsight would be low resolution, but electric fields among even 10,000 electrodes might be patterned into something closer to a megapixel. Over time he wants higher-than-eye resolution, plus wavelengths eyes cannot see: radar, ultraviolet, infrared, “eagle vision,” Geordi La Forge from *Star Trek*.

Fridman asks whether Neuralink is Musk’s best current approach to AI safety. Musk will not call it a panacea. Years ago, he says, he asked what would inhibit aligning collective human will with artificial intelligence, and the answer was the low data rate, especially the slow output. If you want to make a plant happy and the plant never speaks, you do not know what a tree wants. Raise intake and output by three to six orders of magnitude and the link gets less absurd.

He does think hundreds of millions of people could have implants in a couple of decades, if the device is extremely safe and offers superhuman abilities, including memory upload so you would not lose memories. At that point it would supersede the phone, whose main job is guessing what you want between painfully slow keystrokes. A computer doing trillions of instructions per second experiences a whole second as an eternity. High bit rate, he admits, would change the human experience in ways that are hard to imagine, and it is not around the corner: maybe 10 or 15 years. Fridman could perhaps get one in less than ten, depending on the use. Musk is more confident that within a year or two someone with an implant will outperform a pro gamer on reaction time.

The rest of Musk’s hour ranges into xAI, Optimus, and first-principles engineering. Training compute is the engine of the AI race; Tesla video and future humanoid robots are, in his view, the scarce real-world data. He recites a five-step mantra: question the requirements, delete the step, simplify, speed up, then automate, in that order, because the common failure of smart engineers is optimizing a thing that should not exist. Digital superintelligence, he adds later, might be a great filter. Bandwidth is one attempted hedge. Being multiplanetary is another.

DJ Seo, Neuralink’s president and COO, came to brains through electrical engineering, MEMS, millimeter-wave circuits, and a Berkeley PhD on “neural dust”: ultrasound-powered motes meant to sit beside neurons in what he calls a 37°C bag of salt water. Electromagnetic waves attenuate badly in tissue; ultrasound does not, which is why hospitals already image with it. The field’s longer arc, as he tells it, runs from Galvani’s frog legs in the 1790s through Berger’s EEG, Hodgkin and Huxley’s membrane model, Eb Fetz’s 1969 closed-loop monkey work, and Apostolos Georgopoulos’s motor tuning curves in the 1980s.

Invasive versus non-invasive, he says, is a question of what you want. EEG and ECoG can do a lot. His own interest is high-resolution local activity. The stadium analogy is the one he reaches for: outside the arena you hear cheers and know if the home crowd is happy, but not the score or the next play. Drop the microphone into the huddle and you get the chatter. Put an electrode more than about a hundred microns from a cell, roughly the width of a human hair, and you lose that neuron. A hundred-micron cube of cortex holds on the order of 40 neurons. Most of them, he notes, are quiet; the field even talks about “dark neurons.”

The Link itself is three pieces: the N1 implant, the R1 surgical robot, and a B1 app running a relatively simple model that turns spikes into cursor motion. Sixty-four flexible threads, each with 16 electrodes spaced along three to four millimeters, give 1,024 channels in motor cortex, only the last three to five millimeters into tissue. A custom ASIC amplifies, digitizes, and detects spikes on-device because a thousand channels at just under 20 kHz and 10 bits is about 200 megabits, too much heat to stream raw. What leaves the skull over Bluetooth is mostly whether a spike happened. The implant is about the size of a US quarter and about nine millimeters thick. Most of the volume is a rechargeable cell, inductively charged under a two-degree Celsius tissue-heating limit.

Each thread starts around 16 microns wide and tapers to about 84, against a typical hair of 80 to 100 microns, and is less than five microns thick: two microns of polyimide, a 400-nanometer metal stack, another two microns of polyimide. The conductor is a “tiramisu” of titanium, platinum, and gold, two microns across. Seo says you will not see the 16 electrode sites with the naked eye. Flexibility is the point. Rigid Utah arrays are a bed of silicon needles, often 64 to 128 shanks, recorded only at the tip, with a percutaneous port that is a classic infection path. Neuralink’s threads record along their depth and never leave a hole through the skin.

You cannot sew 64 of those by hand. R1 is a multi-axis gantry on a granite slab that weighs about a ton, with optics and a needle that threads a loop at the end of each filament. A 405-nanometer light makes the polyimide fluoresce so the robot can find the loop. The needle tip is 10 to 12 microns, “only slightly larger than the diameter of a red blood cell.” Computer vision maps vessels; a human still approves targets; the robot inserts one thread at a time. Seo’s goal is one click and a surgery that takes minutes. Noland’s operation ran about three and a half hours end to end, with robot insertion just over 30 minutes. An hour after he woke, he could modulate spikes by thinking about crunching a fist.

The company is vertically integrated because the supply chain does not exist. They run their own microfab, a custom femtosecond laser mill for needle geometry, accelerated-lifetime salt-water chambers that age implants four times faster, and a mock operating room with a 3D-printed skull and hydrogel brain that copies the patient’s vasculature. They practiced Noland’s anatomy until the craniectomy looked familiar. Histology from animals implanted for seven months, Seo says, shows neurons abutting threads and almost no collagen scar — the usual killer of rigid arrays. Avoiding vessels with R1 is part of why the immune response looks muted.

About four weeks after surgery, Noland noticed his control getting worse. Impedance and spike-rate plots, plus the electrodes stacked along each thread, showed the rest being pulled out. A human brain, Seo notes, is about ten times a monkey or sheep brain and moved more than they expected. That is why you run a clinical trial. The team did not immediately operate again. They changed what the implant sent: not only the BOSS spike-sorter output but spike-band power, the energy in a frequency band Noland could still modulate even from farther away. Retrained on those features, the model recovered. Performance, Seo says, ended up better than before. Noland had just set 8.5 bits per second, against a prior human record of 4.6, with a personal goal of 10 — roughly median Neuralink staff using a hand on a mouse.

Channel count is still the hardware roadmap. Seo wants 3,000 to 6,000 channels by the end of the year and 16,000 the year after, which means narrower traces, lower-power circuits, denser bonds, and a hermetic barrier in salt water. Multiple implants are the other scaling path; monkeys have already worn two, one per hemisphere. The second named product is vision: stimulate visual cortex to paint phosphenes, those yellowish dots you can also get by pressing on your eyelids. Short term, object-detection overlays so you do not walk into furniture; long term, denser pixels and cameras that are not limited to visible light. A million people in the US are legally blind. Retinal prostheses help only a slice of them. Cortex stimulation is the bypass when the optic nerve is gone.

Seo would implant himself, “maybe not right now,” and would get jealous if participants hit 15, 20, or 100 bits per second. He can already imagine eight billion people with a Link once psychiatric uses — depression, anxiety, appetite — and then phone-replacement join motor and visual restoration. Mind-reading in the science-fiction sense he will not rule out, but he does not think BCI alone gets you there. The corpus callosum’s couple hundred million axons is his tongue-in-cheek existence proof for how many channels a true mind-meld might take. For now the study’s job is safety and efficacy on a wildly varying population of people with tetraplegia, then a larger pivotal trial before anyone markets the device.

Matthew MacDougall, Neuralink’s head neurosurgeon, calls his portion of the procedure one of the most basic operations in the specialty, kin to trephinations with healed bone edges from ancient Peru. He cuts the scalp over the hand knob, the omega-shaped fold of motor cortex that lights up on fMRI when even a person with quadriplegia imagines moving their fingers. He opens a perfectly round one-inch hole, reflects dura, and shows the cortex to R1. The robot inserts the threads to a precise depth while avoiding surface vessels. He then seats the implant in the skull hole, screws it down, and sews the skin. The whole case is a few hours, and he argues the risk is far below tumor or aneurysm surgery because nothing dives deep.

Noland’s January case was done with partners at Barrow Neurological Institute under a pressure MacDougall does not feel in ordinary production neurosurgery: conference rooms watching a live stream of something never done in a human. Unknown unknowns included extra brain motion or sag. He was extremely nervous. After exposing the brain he unscrubbed, placed insertion targets with a lasso tool that auto-avoids vessels, and watched electrodes go in without bleeding. The team had already done hundreds of animal surgeries and practiced on a pulsating hydrogel brain in a 3D-printed skull that matched Noland’s face, wheeled through mock CT and timeout checks like a real patient.

Deep-brain stimulation already reaches the bottom of the brain with about a one-in-a-hundred bleed risk from a blind pass. That is not an acceptable profile for Neuralink. MacDougall wants two or three orders of magnitude safer, “safe enough really that you or I, without a profound medical problem, might on our lunch break someday” take an upgrade. They have not solved arbitrarily deep targets. They have, in animals, crude spinal-cord electrodes that turn motor intention into grasp and stepping. Digital telepathy, he says, already goes surprisingly far because so much of life is on a screen. He would implant himself on safety grounds tomorrow, but the current use case is a mouse, and he can already use one of those.

Utah arrays, MacDougall says, taught Neuralink what not to copy. Rigid spikes hammered in with an air gun provoke scar that walls the electrode off from neurons; some labs even tried chemotherapy to slow that scar. Flexible, tiny threads avoid much of the bleeding and immune response, and in animal models the tissue around them stays healthy for years. He would like the operation simple enough that any neurosurgeon, and someday perhaps non-neurosurgeons, could do it. Standing shoulder to shoulder with R1 in the middle of a case, he reads the body language as a brother in arms, not a threat to his job in his own career. He might tell his kids to think twice in twenty years.

Years of high-stakes surgery have given him a visceral sense that death is inevitable. He does not believe his generation will outrun entropy. He is comfortable with the certainty, if not with the tragedy of imagining his kids without him. Asked whether he has ever seen consciousness in the field, he says he never found it and never removed it, “like a dementor in Harry Potter.” His analog is touch: you know what it means for skin to be contacted. Consciousness, he thinks, is that same sensory mapping applied to thought. You feel the part of the brain that thinks of red things or coffee the way you feel a palm being touched. It is not a warp of space-time or a quantum field effect, however often people reach for the latest physics to make it magical.

The lever he wants to give people is optional. A dial on happiness makes audiences uneasy until you mention major depression and suicide. He would give them that knob. Politics, addiction to outrage, even fertility and blood pressure have legitimate targets in the brain. “We gotta suffer a little, just not so much that you lose hope.” Musk, earlier, had put death more information-theoretically: what are we but our memories, and what is death but the loss of information? A painless disintegration and reintegration with no information loss would, in that thought experiment, not be death. Neuralink’s nearer job is not uploading a copy. It is widening the pipe while you are still here.

Bliss Chapman has spoken with hundreds of people with spinal cord injury or ALS. What he hears, over and over, is a request for independence: communicate without a stick in the mouth, work a computer well enough to hold a job, answer a child before the child wanders off. A mouth stick or eye tracker can move a cursor, but Noland only has the stick when someone puts it in front of him, and spasms make cameras and oral styli dangerous. A wireless implant that is literally in his body can be used in bed at 2 a.m. without a caretaker. Chapman was in the operating room watching live spikes on a phone while threads went in, then in the ICU the next morning for body mapping. Noland found a single channel that fired when he moved one finger, counted it down, and proved it on command.

The engineering target is “the Tesla Roadster version of a mouse,” good enough that people with paralysis dominate e-sports. End-to-end latency from brain spike to cursor is about 22 milliseconds, versus roughly 75 milliseconds for a neuron to move a hand and about five for a top gaming mouse. Noland told them the cursor moved before he felt himself intending it. Decoding is a mapping from spike trains to HID events, trained in an awkward open-loop game because you cannot see a paralyzed user move. Closed-loop then lets user and model co-adapt, which can lock in a local minimum of tricks that vanish the next morning. Chapman is clear that this labeling problem is unsolved, will not be solved just by adding channels, and does not need to be solved for useful control. It does need to be solved if you want the world’s best mouse.

Bits per second, a metric he credits to Krishna Shenoy, is how they keep score: log of the number of targets, times correct minus incorrect, over a time window. Academic and internal Neuralink records sat around 4.2 to 4.6 BPS. Noland’s record is 8.5; median staff on a hand mouse is about 10. Chapman’s own Webgrid ritual — late night, peanut butter, sitting on the floor — is 17. By March, Noland had selected 89,285 Webgrid targets, sometimes from 2 a.m. to 6 a.m. without being asked. When threads retracted, an over-the-air firmware update to spike-band power, plus a dwell cursor, put him back on an independent path. Magnetic targets, “quick scroll” for manga, and a mixer for gain, smoothing, and friction are the unglamorous details of making errors cheap enough to live with.

Noland Arbaugh’s accident was in 2016, a dive into a lake that he describes as a freak knock to the left side of the head. He was conscious, face-down, and understood within seconds that he could not move and would drown when he ran out of breath. Two friends who were lifeguards pulled him out. He cracked jokes on the way to surgery so his mother would not be called until there were answers. Nerve pain later made him scream that his legs were on fire. At the two-year mark, June 30, 2018, when textbooks say you get back as much as you will get, he was sad. He does not think he could be a husband and father as a quadriplegic. He misses the smell and texture of turning a paper book. What he did not do, he says, is wallow. Faith, family, and a stubborn refusal to treat stress as useful got him through.

He almost declined to go first. The first implant might be the worst version; he could wait for number two or three. Being first felt like a chance he had always wanted. He FaceTimed Musk the morning of surgery, told the room he hoped Musk was not too starstruck, prayed, then faked amnesia at his mother when he woke — a planned prank, anesthesia-aided, that she has not forgiven. Hours later Neuralink staff put an iPad of eight spike channels in front of him. He scanned by trying to wiggle fingers and caught a yellow spike on the top row. Everyone clapped. He thought that was unnecessary. That was what was supposed to happen.

For years after the accident he had stared at body parts and ordered them to move, sometimes for hours, a fatiguing attempted movement that hospitals encourage in hope of new pathways. That practice made Neuralink’s body-mapping feel native. Attempted movement — physically trying to shrug or swing a hand, even when nothing moves — gave him cursor control within a week or two and did not surprise him. Imagined movement, visualizing without trying, did not click at first. Then came a Tuesday on Webgrid when he looked at the next target before attempting anything and the cursor shot over. He swore. For the first time it felt like controlling a computer with his mind rather than with a phantom arm. He now mixes both, and he experiments without telling the team until afterward.

They told him about the retraction on the day of his tour of Neuralink’s Fremont facility. His first instinct was to go back to surgery the next day. The first operation had been painless; he did not want a month of new doors slammed shut. On the short drive over he prayed, decided not to ruin the tour, and had one of the best days of his life anyway. For a few days he did not know if cursor control would return. He decided that even if it never did, he would keep doing body mapping and data collection for a year if that helped everyone after him. “The whole reason that I did this was to help people.”

Within a couple of weeks they switched how they measured spikes. Spike-band power felt better immediately. Dwell clicking at 0.3 seconds was a downgrade from left-and-right click, but it was a path. He is back to 8.5 BPS on a 35-by-35 grid, hunting 9 and then 10, racing implant battery percentages at 2 a.m. with a giant BPS counter behind the targets that he asked them to put there because zen mode did not show the number. Accidental dwell clicks have cost him chess games. He wants on-demand clicks again, phone and console support, every hidden cursor parameter, and a way to type that might start as finger spelling and end as thinking a letter. He would take an upgraded implant whenever they will let him. No regrets about the surgery.

Independence is the word he keeps using. Sitting at a computer all night without someone propping an iPad and waiting up is a load taken off his family. He plays *Civilization VI* as Korea, rushing science, occasionally furious when a diplomatic victory ends the game by accident. He would like to drive an Optimus robot: 99 percent of what he cannot do needs a caretaker’s hands, and a robot that held a book under his nose might give him the smell of paper again. Touch is what he asks God for. The first thing he would do with a working hand is squeeze his mother’s. Until then, the implant is not a cure. It is a way to put some choices back in his own head.

What Neuralink has shown in this recording is specific. A person with no sensation below the shoulders can calibrate a model, move a cursor, play games, read, and text without a mouth stick. The first human surgery was boring to people who had rehearsed it hundreds of times, which is how they wanted it. When biology pulled the wires back, a vertically integrated stack — custom threads, custom ASIC, custom app, over-the-air firmware — recovered a world-record bit rate. None of that is the same as solving decade-scale stability. Seo’s priority remains keeping threads in cortex. MacDougall’s safety bar remains lunch-break elective. Chapman’s decoder still cannot label millisecond intent. Noland still dwell-clicks.

Musk’s merge story and the clinic’s paralysis story share a device and not much else. One is a claim about AI timelines and human relevance. The other is a man who wants to take a load off his parents and someday hold a book. Psychiatric indications, obesity, depression, memory replay, and healthy-person augmentation sit behind a wall of risk, ethics, and FDA benefit-risk math that nobody in the room pretends to have climbed. DJ Seo can imagine eight billion users. Matthew MacDougall would implant tomorrow and still says there is no value proposition for him yet. Noland would upgrade tomorrow and still wants a hand.

Fridman closes with Aldous Huxley on the privacy of experience: we can pool information about what we feel, but never the experiences themselves. That is the poetic version of Musk’s compression complaint and of Chapman’s remark that some thoughts are most precisely described in poetry. Neuralink’s near work is cruder and more moving. It is a wireless mouse for people who cannot hold one, a robot that refuses to hit a vessel, and a first patient who, told his threads had retracted, went on a factory tour anyway and then came back to beat his own record.