Latest / Soft Ratios Research / The wrist, not the brain (on neural interfaces)

Transcript

The wrist, not the brain (on neural interfaces) · Soft Ratios Research

• 0:09
  Figure: This is Soft Ratios Radio.
• 0:12
  Ground: The voices are synthetic [pause] no act, just the thinking.
• 0:17
  Figure: Hoyd Breton, a designer at Soft Ratios Studio, teaching himself in the open.
• 0:23
  Ground: One subject at a time. [pause] Today, neural interfaces.
• 0:27
  Ground: Say the phrase "brain-computer interface" and everyone pictures the same thing. A skull. A chip. Someone in a chair moving a cursor with a thought. That's the picture that sells.
• 0:40
  Figure: And the picture isn't wrong, exactly. It's just aimed at the wrong shelf. The thing that actually shipped to ordinary people — the neural interface you can buy — went on sale September 30th, 2025. It's a wristband. It bundles with a pair of Ray-Ban glasses for seven hundred and ninety-nine dollars. It's called the Neural Band. Meta sells it. Meta did not invent it. [pause] The invention belongs to a startup called CTRL-labs. Thomas Reardon — the engineer who built Internet Explorer, then walked away from software for a doctorate in neuroscience — founded it with Patrick Kaifosh on one wager: don't read the brain at all. Read the motor neurons at the wrist, the last relay before the hand moves. Meta bought the company in 2019 — the reported price ran somewhere between five hundred million and a billion dollars — and folded that work into its Reality Labs. So the band on the shelf is CTRL-labs' machine, wearing Meta's name.
• 1:39
  Ground: A wristband called neural. That's a strange word for a thing on your arm.
• 1:46
  Figure: It's the whole trick, and it pays to get precise here. The band doesn't touch your brain. It reads muscle. Surface electromyography — sEMG. Sixteen little sensors around the wrist, listening to the electrical crackle your forearm muscles make when they're about to move your fingers.
• 2:04
  Ground: So before I've even moved a finger, it hears the order going out.
• 2:09
  Figure: That's the biology of it, yes. When a motor neuron fires, it drives the muscle fibers it owns to depolarize. That little electrical event travels out through fat and skin, and a sensor on the surface catches the smeared-together sum of it — a motor unit action potential. The command's already on the wire before the knuckle bends.
• 2:29
  Ground: Here's what nags at me. If the near-term future of "neural" input is a bracelet reading your arm muscles — why did everyone spend two decades and billions of dollars drilling into skulls? That's the gap I want to walk into. Not "is the brain implant real." It's real. The question is why the thing that reached us first went around the brain entirely.
• 2:58
  Figure: Then let's start where the field started, because the detour tells you something. The whole idea has a birthday. 1973 — a computer scientist at UCLA, Jacques Vidal, writes a paper called "Toward Direct Brain-Computer Communication." He coins the term. And he actually does something with it: he reads visual evoked potentials off the visual cortex and works out where a person's trying to move a cursor. From electrodes on the scalp.
• 3:26
  Ground: 1973. So the dream isn't new tech chasing hype. The dream's fifty years old.
• 3:34
  Figure: Older in spirit. Richard Caton found electrical currents in animal brains in 1875. Hans Berger recorded the first human EEG in 1924. But Vidal's the one who says: this could be an input. [pause] And then in 1988 you get the first genuinely practical version — Farwell and Donchin's P300 speller.
• 3:57
  Ground: Which works how, in a sentence I'd actually say out loud?
• 4:01
  Figure: A grid of letters flashes at you, row by row, column by column, randomly. You stare at the letter you want and silently count each time it lights up. When your letter flashes, your brain throws a little surprised spike about three hundred milliseconds later — the P300. It only fires for the rare thing you're waiting for. The machine watches for that spike, sees which row and which column triggered it, and where they cross is your letter.
• 4:29
  Ground: So you don't spell. You wait to be surprised, and the surprise gives you away.
• 4:37
  Figure: It's the oddball paradigm — the target's the low-probability event, one in six on a six-by-six grid, and the brain can't help flinching at it. Clever. And slow. That's the ceiling that haunts the whole non-invasive side: you can read the brain from outside, but you're reading through the skull, and the signal comes back faint and smeared.
• 4:59
  Ground: So the logic writes itself. Faint signal from outside — go inside. Get the electrode next to the neuron.
• 5:06
  Figure: And they did. The enabling piece of hardware was the Utah array — Richard Normann, University of Utah, around 1992. A little bed of stiff silicon needles you press into the cortex. It became the standard for twenty years. It's what let tetraplegic patients in the BrainGate program move cursors and robotic arms. Single-neuron resolution. The real thing.
• 5:29
  Ground: And this is where I want to slow you down, because "the real thing" is exactly the phrase that hides the problem. You put a stiff bed of needles into living tissue. The tissue is not a circuit board. It's wet, it moves, it defends itself. What happens after you close the skull?
• 5:50
  Figure: It fights back. That's not a metaphor — it's a documented cellular sequence, and it has a name: gliosis. Within minutes of insertion, you've torn cells and broken the blood-brain barrier. Immune cells called microglia reach toward the wound. They release signals that wake up the astrocytes —
• 6:08
  Ground: The brain's repair crew.
• 6:11
  Figure: More like the brain's scar-builders. The astrocytes swell, multiply, and weave a dense sheath around the electrode. In the dish you can watch it form in layers — microglia coating the shaft, astrocytes wrapping that in a mesh. And that scar does two things at once. It's electrically resistant, so it muffles the signal. And it physically shoves the living neurons back, away from the tip that's trying to hear them.
• 6:38
  Ground: So the body's response to the listening device is to build a wall around it and evict the neighbors.
• 6:45
  Figure: And now the numbers get sobering, which is where the myth of the permanent implant starts to crack. The largest longevity study looked at over six thousand recording sessions, across almost nine years, fifty-five Utah arrays. Average lifespan of usable recordings: six hundred and twenty-two days. Some lasted past a thousand. One went nine years.
• 7:08
  Ground: That doesn't sound like decay. That sounds durable.
• 7:11
  Figure: The average flatters it. Look at individual neurons and it's harsher. An earlier study tracked single units: sixty-seven percent stable through fifteen days. Thirty-one percent through forty-seven days. Twenty-one percent through a hundred and six. Eight percent stable past nine months. In rat motor cortex the yield fell from about fifty-three percent in week one to thirteen percent by week twenty-four.
• 7:37
  Ground: So the array survives. The specific voices it was hearing don't. It's like keeping the microphone but losing the people who were talking into it, one by one.
• 7:48
  Figure: And that reframes the engineering problem entirely. You can't build a consumer product on a signal that quietly walks away from you over months. For medicine — for someone with no other way to speak — you accept it, you re-tune, you fight the drift. For a mass-market gadget, that fight is the whole reason it can't be the brain. Not yet. Maybe not this decade.
• 8:11
  Ground: Which sends us back to the wrist. But I don't want to let Meta off easy just because their thing's non-invasive and clever. Let me push on the word again. They call it a neural interface. It is not reading your nervous system's thoughts. It's reading the exhaust from your muscles. Isn't "neural" here basically marketing?
• 8:33
  Figure: Partly. And partly it's technically defensible, which is what makes it slippery. The signal does originate in motor neurons — the alpha motor neurons firing is what starts the whole chain. So the electrical event has a neural cause. But you're catching it after it's driven a muscle, out at the periphery. Their own paper's honest about this in a way the branding isn't: it says brain interfaces have shown high-bandwidth communication only with invasive electrodes and custom decoders built for one person at a time. The wrist, by contrast —
• 9:06
  Ground: Generalizes.
• 9:07
  Figure: Out of the box. That's the July 2025 Nature paper — Kaifosh, Reardon, and colleagues. They trained on a hundred-plus hours of data from a hundred people per task, and the decoder works on a new person without being retrained for their body. That's the thing the brain implant can't do. Every implant so far is a bespoke suit tailored to one cortex.
• 9:31
  Ground: And that's the real reveal, isn't it. Not invasive versus non-invasive. It's bespoke versus generic. The brain gives you the richest signal and demands you refit it to every single skull. The wrist gives you a coarser signal that fits everyone off the rack. And "fits everyone" is what a product is.
• 9:54
  Figure: Let me put a hard edge on "coarser," though, so we don't romanticize the band either. The CTRL-labs team's own paper — published from inside Meta — concedes the boring old inputs still win. A trackpad beats it. A Joy-Con beats it. Physical handwriting beats it. Their sEMG handwriting clocks about twenty-one words a minute; typing on your phone is around thirty-six.
• 10:17
  Ground: So it's slower than the thing already in my pocket.
• 10:20
  Figure: Today, yes. Its case isn't speed. It's that it works with your hand at your side, in the dark, with no camera watching you, all day, on a wristband's battery. It's not trying to beat the phone at the desk. It's trying to be the input for glasses — for the moment when there's no screen to touch and no keyboard to find.
• 10:40
  Ground: Let me sit inside the hardware for a second, because the ordinariness of it is the whole point and it's easy to skate past. Sixteen sensing pairs around the wrist. Little gold-plated discs, six and a half millimeters across, twenty millimeters apart. Sampling two thousand times a second. And here's the humbling part — the physics says only the motor units within about ten to twelve millimeters of a sensor contribute any real signal. Everything deeper is lost in the tissue. [pause] So this thing that feels like magic on your arm is actually straining to hear a whisper through fat and skin, catching only the muscles closest to the surface. It's not omniscient. It's a very good listener pressed against a wall, and it only hears the room nearest to it. [pause] And that reframes the achievement. The magic isn't that it hears everything. It's that they built something useful out of hearing almost nothing cleanly — a faint, smeared, surface signal — and let a decoder guess the rest. Which, come to think of it, is the same trick the brain implant pulls. Both of them are reading a poor signal and betting on a smart guess.
• 11:54
  Figure: That's the thread that ties the whole field together, and it's the piece people miss when they argue about electrodes. The decoder is the breakthrough. Not the sensor. The last few years of genuine progress are really the years the AI got good enough to turn a sparse, noisy signal into fluent output.
• 12:13
  Ground: Give me the medical version of that, because that's where the stakes are real. Nobody's paralysis is cured by a wristband scrolling a menu.
• 12:22
  Figure: Speech restoration. This is where the science is genuinely staggering, and I want to be careful to keep the proven separate from the promised. Willett and colleagues, Nature 2023: a man with ALS, intracortical arrays. On a fifty-word vocabulary, a nine-percent word error rate. On a hundred-and-twenty-five-thousand-word vocabulary — essentially open English — twenty-three point eight percent error. Decoding attempted speech at sixty-two words a minute. Natural conversation is around a hundred and sixty.
• 12:55
  Ground: So a third of natural speed, and one word in four wrong on the big vocabulary. That's not a finished product.
• 13:03
  Figure: It wasn't. But hold the dates. The very next year — Card and colleagues, New England Journal of Medicine, the BrainGate2 study. Another man with ALS, severe dysarthria, four arrays, two hundred and fifty-six electrodes in the precentral gyrus. Day one, after thirty minutes of training: ninety-nine point six percent accuracy on fifty words. Day two, they open it to a hundred and twenty-five thousand words, an hour and a half more of training: ninety point two percent. And past eight months, it held ninety-seven and a half percent. Word error rate down around two and a half percent, self-paced, about thirty-one words a minute.
• 13:44
  Ground: Wait. From a quarter of the words wrong to two and a half percent, in roughly a year? That's not an incremental step. What changed — did the electrodes get that much better in twelve months?
• 13:58
  Figure: No. The electrodes are the same generation of idea. The language model got better. The decoder learned to lean on what a plausible English sentence looks like — strong priors filling the gaps in a weak signal. Same physical listening problem. Vastly smarter guess.
• 14:16
  Ground: So the honest headline isn't "we can read the brain now." It's "the machine got better at finishing our sentences." And that should unsettle us a little, shouldn't it? Because a system leaning that hard on priors — how much of the output is the person, and how much is the model's confident guess about what a person like them would probably say?
• 14:38
  Figure: That's the right worry, and it lives at the seam of every one of these results. When a UC Davis team pushed to real-time brain-to-voice in 2025 — synthesizing actual voice, not text, at about a fortieth of a second of latency, catching intonation, catching the rise of a question — the thing that made it feel like a phone call instead of a text message was the same thing that makes you nervous. The model's intervening, fast, between intention and utterance.
• 15:08
  Ground: And I'd add: these are single people. One participant. One cortex.
• 15:14
  Figure: Which is the caveat that has to ride along with every one of these numbers. Most of the speech results are a single participant. Ninety-seven percent accuracy in one man's motor cortex tells you the approach can work. It does not tell you it generalizes. That word — generalize — is exactly the one the wristband gets to claim and the implant does not.
• 15:37
  Ground: Let me re-ask the question we started with, now that we've been inside both. It isn't really "brain or wrist." It's: what kind of wrongness can you live with? Because none of these read a clean thought. They all read a smudge and guess. So the design question underneath everything is — when the machine is wrong, how wrong, and who pays for it?
• 16:02
  Figure: And that's a formalized field, not a vibe. The person who named it cleanly is Bliss Chapman, who leads the brain-interface software at Neuralink. His framing: there are different cost functions for different errors, and part of the design is building something that, when it's wrong, is still useful to the person using it. A wrong letter in a menu is a shrug. A wrong word in the sentence "do not" — that's a different kind of wrong.
• 16:29
  Ground: So you don't design for the machine being right. You design for the shape of its mistakes.
• 16:35
  Figure: And there's older theory underneath that, which is the part I find genuinely beautiful. Merel and Paninski, around 2013, formalized what everyone had been calling "the user gets used to it" into actual mathematics. They modeled the human and the decoder as two agents in one control problem, both learning at once. The decoder adapts to your signal; you adapt your signal to the decoder. Co-adaptation.
• 17:01
  Ground: Two learners circling each other. That's not a tool you pick up. That's a relationship.
• 17:09
  Figure: It's coordinate descent, in their language — you optimize, then it optimizes, back and forth toward a shared minimum. And they proved something almost deflating and completely honest: under their assumptions, all that mutual dancing can't beat what you'd get from one optimal fixed decoder plus the user simply learning well. The synergy people romanticize has a mathematical ceiling.
• 17:32
  Ground: I want to resist collapsing that into "so co-adaptation is pointless," because that's not what it says. It says the ceiling exists in an idealized world. In the messy one — where the signal drifts, where the neurons walk away over months — the adaptation isn't a bonus. It's how you keep standing on ground that's moving under you. The dance isn't for extra performance. It's for staying upright.
• 18:00
  Figure: And there's a concrete number that makes it feel real rather than theoretical. The CTRL-labs scientists found that even a little personalization — a small amount of tuning to the individual — improved their handwriting recognition by up to sixteen percent. That's co-adaptation earning its keep on a shipping product, not a whiteboard.
• 18:20
  Ground: Sixteen percent from the machine bending toward one person. Which loops us right back: the wristband's whole boast was that it works without that. Generic, off the rack.
• 18:34
  Figure: Both are true, and the tension is the honest state of it. It generalizes enough to work on a stranger's arm out of the box — and it still gets meaningfully better the moment it stops treating you like a stranger. Generic to get in the door. Bespoke to get good.
• 18:50
  Ground: There's a third thing we haven't touched, and it's the one that outlasts all the hardware arguments. What happens to the signal after it's read? Because whether it comes off your cortex or your wrist, it's data now. It's leaving your body and going somewhere.
• 19:07
  Figure: And the law's started to notice, unusually fast. Colorado went first — HB 24-1058, in force August 2024, the first US state to specifically protect neural data. California followed with SB 1223, operative January 2025. Both fold neural data into their strongest "sensitive" tier.
• 19:30
  Ground: Define neural data, though. Because if it means "brain," the wristband walks right through the gap.
• 19:37
  Figure: They saw the gap and closed it. Both statutes define neural data as information generated by measuring the activity of the central or peripheral nervous system. Peripheral. Written in. The wrist is explicitly covered. California even carves out the loophole — it says the data has to come from the nervous system itself, not be inferred from non-neural signals. So a heart-rate tracker guessing your stress isn't neural data. The band reading your motor neurons is.
• 20:06
  Ground: So the law drew the line exactly where Meta's marketing drew it — and used it against the marketing. "You called it neural. Fine. Then it's neural data, and here are the rules."
• 20:18
  Figure: And it matters because the industry's baseline is alarming. The Neurorights Foundation, with Perseus Strategies, went through thirty consumer neurotech companies in April 2024. Twenty-nine of the thirty — ninety-six point six seven percent — appear to have access to the user's neural data with no meaningful limit on it. Only one company offered any real restriction on selling it onward. Fewer than half encrypt and de-identify. And the medical shield you'd assume exists — HIPAA — doesn't apply to a gadget you bought online.
• 20:52
  Ground: So the same signal that would be locked behind medical consent in a hospital is, on a consumer wristband, just... a data stream with a privacy policy nobody reads.
• 21:04
  Figure: Which is the sting hidden in the whole detour to the periphery. The reason the wrist won on convenience — no surgery, all day, works on anyone — is the same reason the data problem gets bigger, not smaller. Millions of unregulated, always-on, generic devices generate far more nervous-system data than a handful of clinical implants ever could. The scale that makes it a product makes it a surveillance question.
• 21:29
  Ground: And it moves the human where the machine is going. Apple did this in May 2025 — a protocol they call BCI HID. It makes neural input a native, first-class category in the operating system. Not a hack bolted onto an accessibility feature. Alongside touch, alongside voice, alongside typing.
• 21:52
  Figure: With Synchron as the first partner — their Stentrode, the sixteen-electrode implant that goes in through the jugular vein, no craniotomy. And the interesting part is the closed loop: the iPad shares what's on screen back to the decoder, so the decoder knows the context it's decoding into. The device tells the brain-reader what the choices are, to make the reading better.
• 22:14
  Ground: So it's co-adaptation again, wearing an operating system. The screen and the signal, each shaping the other in real time.
• 22:23
  Figure: Once it's a native input type, the whole lineage clicks into place. The mouse waited for the graphical desktop. Multitouch waited for the phone. Each new way of pointing arrived when a new kind of computer needed it. Neural input — brain or wrist — is being positioned as the pointer for the thing with no surface to touch. Glasses. Ambient computing. The screen you don't hold.
• 22:51
  Ground: Then here's where I land, and I want to keep the split honest rather than tie a bow on it. We opened with the picture everyone has — the skull, the chip, the mind being read. And almost none of that picture is what reached us. What reached us reads muscle, not mind. It guesses more than it knows. It works because a language model finishes our sentences, and it's governed by laws that had to reach past the brain to catch it.
• 23:17
  Figure: And the piece that should stay uncomfortable is the honesty buried in the word itself. Every one of these — the wrist, the Stentrode, the arrays in a man's motor cortex — decodes attempted movement or attempted speech. The intention to act. Not the private thought behind it. The mind-reading fear is aimed at a capability nobody has demonstrated, while the real thing quietly redraws what counts as your data.
• 23:44
  Ground: So the thing we were afraid of and the thing that actually arrived aren't the same thing at all. We braced for a machine that reads the mind. What showed up is a machine that reads the body's outbox — the order already on its way out — and a very confident guesser standing behind it, and a data pipe running off your arm that the law only just learned to name.
• 24:07
  Figure: The brain stayed hard. It fights the needle, it walks away from the electrode, it insists on a bespoke fit for every skull. And the body, at the wrist, turned out to be the easy door — generic, all-day, already halfway to the machine. The future got in through the door nobody was watching.
• 24:27
  Ground: If the interface that reached us reads the command leaving the muscle rather than the thought behind it — the body's outbox, not the mind — then when it acts a half-second before we're aware of deciding, whose intention are we actually watching: ours, or the machine's guess about ours?