Forty words a minute is not a headline number in most contexts. It is roughly what a distracted person types on a phone. But that number just came out of a human skull, over a wireless link, from an electrode array sitting on or in the motor cortex, with no cable running out the back of the head. I have sat through a decade of BCI demos that impressed me on stage and then quietly died in the appendix of a journal paper. This one is different, not because the number is huge, but because of what had to be true for the number to exist at all: a fully implanted, battery-powered, wireless system decoding intended movement into text fast enough to hold a real conversation.
The demo came from Paradromics, the Austin-based company founded by Matt Angle that has spent years positioning itself as the highest-bandwidth challenger to Neuralink. Paradromics received an FDA Breakthrough Device Designation in 2023 for its Connexus Direct Data Interface, a penetrating microelectrode system built to record from thousands of neurons at once. The company has talked for years about bandwidth as the whole game. This is the first time I have seen that bandwidth argument turn into something as mundane and as legible as words appearing on a screen, generated by a person imagining the act of writing them, with no wire attached to their head.
The Wire Was Always the Real Bottleneck
Everyone assumes the hard part of a brain-computer interface is decoding. Turning a storm of spikes into letters. That part is genuinely hard, and Stanford’s Frank Willett, working in Krishna Shenoy’s lab before Shenoy’s death in 2023, published a landmark 2023 Nature paper showing a paralyzed participant producing attempted handwriting decoded at around 90 characters per minute, later extended to attempted speech at roughly 62 words per minute. Those were extraordinary results. They were also wired. The participant sat tethered to a rack of amplifiers through a percutaneous connector, a port screwed into the skull. Every session started with technicians plugging someone into a wall.
The bottleneck was never decoding. It was always the wire.
A wired connector is a research tool, not a product. It is an infection risk sitting on an open wound in the scalp, a tether that confines the user to a lab, a piece of hardware that reminds everyone, patient included, that they are a research subject rather than a person living a life. Neuralink understood this from day one and built its whole pitch around eliminating it: the N1 implant is fully subcutaneous, charges wirelessly through the scalp, and streams data over a short-range radio link. When Noland Arbaugh became the company’s first human participant in January 2024, the story that mattered wasn’t the cursor control, which BrainGate had demonstrated years earlier with tethered systems. The story was that he could do it from his couch, no cable, no technician, no clinic visit.
Paradromics’ demo matters for the same reason, aimed at the same weak point in Neuralink’s own record. Neuralink has shown excellent cursor control and some game play. It has not, to my knowledge, published a wireless free-text typing rate that competes with what Willett’s wired Stanford system achieved. Forty words a minute, if it holds up under independent scrutiny and peer review rather than a company blog post, is the first wireless number in the same neighborhood as the best wired numbers on record.
What Is Actually Being Decoded, and Why That Distinction Still Matters
I want to be precise about language here because the phrase “thought to text” does real damage to public understanding every time it gets used loosely. Nobody has built a system that reads thoughts in the sense of reading beliefs, memories, or interior monologue. What these systems decode is intended movement. In Willett’s handwriting work, the participant imagined writing letters by hand, and the system decoded the motor cortex activity associated with that imagined movement, not the letters themselves as abstract symbols. In speech decoding work out of Edward Chang’s lab at UCSF, the signal comes from attempted articulation, the motor commands that would move a tongue and larynx that in these patients no longer respond.
Forty words a minute is attempted handwriting, decoded from motor cortex, sent over a radio.
This is not a small semantic quibble. It is the difference between a prosthetic and a surveillance device. A system that decodes attempted movement only works when the user actively tries to move, tries to write, tries to speak. It cannot eavesdrop on a wandering mind, because a wandering mind does not generate the specific motor cortex signature of intended handwriting. The privacy fear that shows up in every comment section under a BCI story, that a company will read your unspoken opinions, is not supported by how any of these systems function today. The real privacy questions are more mundane and, I’d argue, more urgent: who owns the neural data stream once it leaves the skull, what happens to that data when a startup goes bankrupt or gets acquired, and whether a decoder trained on your motor cortex can be repurposed for something you never consented to.
The Field Now Has Four Real Approaches, and They Are Not Converging
What strikes me watching this field from the outside of any one company’s PR department is how differently these teams have chosen to solve the same problem. Neuralink threads flexible electrodes deep into cortical tissue with a surgical robot. Paradromics uses a denser, more rigid microelectrode array aimed at raw channel count. Synchron, led by Tom Oxley, avoids open-brain surgery entirely, threading its Stentrode device through the jugular vein and into a blood vessel that sits against the motor cortex, trading some signal resolution for a dramatically simpler procedure that can be done by an interventional neurologist rather than a neurosurgeon. Precision Neuroscience, co-founded by former Neuralink engineer Benjamin Rapoport, sits on top of the cortex rather than penetrating it, betting that a thin, removable film of electrodes can get enough signal without the tissue damage and scarring that penetrating arrays eventually cause.
| Company | Electrode approach | Wireless | Key figure |
|---|---|---|---|
| Neuralink | Flexible penetrating threads, robot-implanted | Yes, since first human implant, Jan. 2024 | Elon Musk |
| Paradromics | Rigid, high-density penetrating array | Yes, demoed wireless typing | Matt Angle |
| Synchron | Endovascular stent electrode, no craniotomy | Yes, since early trials | Tom Oxley |
| Precision Neuroscience | Thin-film surface array, non-penetrating | Wired in current trials, wireless planned | Benjamin Rapoport |
None of these approaches has proven itself the winner, and I am skeptical anyone will win outright in the way Betamax lost to VHS. A stent electrode that avoids brain surgery will always appeal to patients and hospitals wary of a craniotomy, even if it means accepting a coarser signal. A penetrating array that gets closer to individual neurons will always win on raw bit rate, at the cost of a harder surgery and a less certain long-term safety profile as scar tissue builds around the electrodes over years. The FDA is going to end up regulating four genuinely different categories of implant that happen to solve the same downstream problem, which is its own quiet regulatory headache nobody in Washington has fully reckoned with yet.
Forty Words a Minute Is a Milestone, Not an Arrival
Natural conversational speech runs somewhere around 120 to 150 words per minute. Forty words a minute is faster than most people hunt-and-peck on a keyboard, and for someone with ALS or a high cervical spinal cord injury who currently communicates through eye-tracking software at ten or fifteen words a minute, it is transformative. It is not, however, close to giving someone their voice back in real time. The gap between 40 and 150 is the gap this entire industry is still trying to close, and every serious person working in it will tell you privately that the last third of that gap, getting from workable to natural, may be harder than everything that came before it, because it requires the kind of low-latency, high-fidelity, continuously stable signal that current electrode technology has not sustained for years at a time in a living, moving human brain.
That durability question is the one I keep coming back to and the one that gets the least attention in demo footage. A microelectrode array that reads beautifully in month three can degrade by month eighteen as the brain’s immune response encapsulates the electrode in scar tissue, a well-documented failure mode in the BrainGate literature going back over a decade. Wireless power and data transfer solve the infection and mobility problem. They do nothing for the biological one. The company that eventually wins this field will not be the one with the flashiest typing demo. It will be the one whose implant still works, unattended, five years after the surgery, in a patient who has gone back to living an ordinary life and stopped thinking about the hardware in their head at all.
That is the actual finish line, and it is a quieter, less photogenic one than 40 words per minute streamed wirelessly onto a screen. Paradromics has shown the field that the wire was never the deepest problem, only the most obvious one. The deeper problem is whether a piece of electronics can sit inside a living brain for a decade the way a pacemaker sits in a chest, boring, reliable, and beneath notice, and no demo, however impressive, has answered that yet.