De Mente Reserata: The Ultimate Power · Grey Matter

What Holds It Back, What It Will Cost, What It Will Unleash

De Mente Reserata: The Ultimate Power. What Holds It Back, What It Will Cost, What It Will Unleash

Reading the brain at a reasonable cost will do to medicine and to the mind what GPT-4 did to language, and the half of the problem that looked hardest already works. Decoders turn neural activity into sentences. What nobody has yet is a cheap way to get that activity out of a skull.

The speed of the decoding side is easy to miss because it happened in a few hospitals. In 2016 a patient with ALS, locked in, typed about two letters per minute through electrodes laid on the motor cortex and a transmitter under the skin of the chest1. In August 2023 Nature published two papers in the same issue. In one, a woman with ALS produced text at 62 words per minute through four small arrays in her speech cortex, with a 23.8% word error rate on a 125,000-word vocabulary2. In the other, a woman paralysed by a brainstem stroke reached a median of 78 words per minute through a sheet of electrodes on the cortex, with a talking avatar on top3. A year later a group at UC Davis reported 97.5% accuracy in self-paced conversation, sustained over eight months, from 256 electrodes in one gyrus4. Seven years took the field from two letters a minute to faster than most people type on a phone, which a study of 37,000 volunteers put at 36 words per minute5. GPT-4 had come out five months before the two Nature papers6, and the decoders already borrow its machinery: their last stage is a language model that turns noisy guesses into sentences.

1Vansteensel et al., Fully implanted brain-computer interface in a locked-in patient with ALS, NEJM 375, 2016. 2Willett et al., A high-performance speech neuroprosthesis, Nature 620, 2023. 3Metzger et al., A high-performance neuroprosthesis for speech decoding and avatar control, Nature 620, 2023. 4Card et al., An accurate and rapidly calibrating speech neuroprosthesis, NEJM, 2024. The rate was about 32 words per minute. 5Palin et al., How do people type on mobile devices? Observations from a study with 37,000 volunteers, MobileHCI 2019, which reports an average of 36.2 words per minute. 6OpenAI released GPT-4 on 14 March 2023; see its history.

Every one of those results needed a neurosurgeon, a hole in the skull and a patient with severe paralysis. This text is about that gap: what physics and biology put between a brain and a reader, what it will cost to close it, what comes out once it closes, and the one line that has to be drawn first. It is close enough that the line has to be drawn now.

Through the bone

The obvious move is to read from outside, and the skull answers every kind of wave badly in its own way. Near-infrared light gets in, but brain tissue scatters it every 25 to 200 µm1, so after a few millimetres a photon has forgotten where it came from, and functional near-infrared spectroscopy settles for blood oxygen 1.5 to 2 cm deep, which is the scalp, the bone and the surface of the cortex2. Ultrasound keeps its direction, but bone eats it: across human skull samples about 8 mm thick the measured attenuation was 13.3 dB per centimetre, most of it reflection and scattering3. Attenuation climbs with frequency, so transcranial work stays near or below 1 MHz4, where the wavelength in soft tissue is about 1.5 mm. Functional ultrasound resolves 100 µm in a rat at 15 MHz5, and in an adult human it reached 200 µm only through a plastic window that replaced a piece of skull after an injury6. X-rays cross bone easily and charge for it in dose, around 2 mSv for a head CT7, which rules out wearing one all day.

1Marblestone et al., Physical principles for scalable neural recording, Frontiers in Computational Neuroscience, 2013, which gives scattering lengths of 25 to 200 µm and absorption lengths around 1 mm at visible and near-infrared wavelengths. 2Pinti et al., The present and future use of functional near-infrared spectroscopy for cognitive neuroscience, Annals of the New York Academy of Sciences, 2020. 3Pinton et al., Attenuation, scattering, and absorption of ultrasound in the skull bone, Medical Physics 39, 2012. 4See the measurements collected in Evaluation of ultrasound sensors for transcranial photoacoustic sensing and imaging, 2023, where frequencies below 1 MHz are the least attenuated in frontal and occipital bone. 5Macé et al., Functional ultrasound imaging of the brain, Nature Methods 8, 2011. 6Rabut et al., Functional ultrasound imaging of human brain activity through an acoustically transparent cranial window, Science Translational Medicine 16, 2024. 7US FDA, What are the radiation risks from CT?, accessed October 2026.

EEG and MEG listen to the brain's own fields instead, and their trouble is signal to noise. The magnetic field of the brain at the sensor is between 10 and 1,000 femtotesla; the Earth's is about 50 µT, roughly a billion times larger1. The sources are current dipoles, whose fields fall with the square or the cube of the distance2, and a helium-cooled sensor already sits about 2 cm off the scalp for the insulation it needs3. The skull conducts poorly and smears the electric potential sideways, so EEG sees only large patches firing together. Epilepsy surgery measured how large: in simultaneous scalp and intracranial recordings, cortical spikes covering less than 6 square centimetres never showed up on the scalp, and those over 10 showed up 90% of the time4. A few square centimetres of cortex hold tens of millions of neurons5. EEG and MEG see what is synchronous and superficial, and almost everything that carries the content of a thought is neither.

1NYU Neuroscience of Language Lab, MEG FAQ. 2Hämäläinen et al., Magnetoencephalography: theory, instrumentation, and applications to noninvasive studies of the working human brain, Reviews of Modern Physics 65, 1993, the standard reference for the physics of both methods. 3Brookes et al., Magnetoencephalography with optically pumped magnetometers (OPM-MEG): the next generation of functional neuroimaging, Trends in Neurosciences 45, 2022. 4Tao et al., Intracranial EEG substrates of scalp EEG interictal spikes, Epilepsia 46, 2005. 5Azevedo et al. (2009, cited below) count about 16 billion neurons in the cerebral cortex, which spreads over roughly 1,500 to 3,000 square centimetres according to Nunez in Scholarpedia's EEG article; that is several million per square centimetre.

Under the engineering sits a theorem. Helmholtz showed in 1853 that infinitely many current distributions inside a conductor produce the same field outside it, so the inverse problem has no unique answer1. Some configurations are silent outright: in a spherical head, radial currents leave no magnetic field outside at all. Algorithms pick one answer among the infinite ones by adding priors, and the answer then reports the priors as much as the data. Activity that leaves no trace outside the head cannot be recovered by any amount of computation, and fine structure at the scale of columns and cells is exactly what cancels before it leaves. That ceiling holds for every reader that stays outside, and it is why I expect the first cheap readers to sit under the bone.

1The result and its consequences for MEG are set out in Hämäläinen et al., 1993, cited above.

Under the bone

Going in fixes the physics and swaps it for biology and plumbing. An electrode in the cortex hears few neurons beyond about 100 µm and none beyond 160 µm1, so every channel is a local window onto a brain of about 86 billion neurons2 whose neocortex alone holds around 1.5×10141.5 \times 10^{14}1.5×1014 synapses3. Nobody needs all of them, and even a modest window is heavy: a thousand channels sampled at 30 kHz and 10 bits produce 300 megabits per second of raw data, and pushing that through a radio costs more power than an implant can spend, which is why designs now compress on the chip before transmitting4. Power is where bandwidth turns into a medical limit. The whole brain runs on about 15 W and tolerates local warming of around 2 °C over hours without serious damage5, and every extra bit read and sent is heat deposited next to neurons, in an organ that cannot sweat it away.

1Marblestone et al., 2013, citing recordings with multi-site silicon probes. 2Azevedo et al., Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain, Journal of Comparative Neurology 513, 2009. 3Pakkenberg et al., Aging and the human neocortex, Experimental Gerontology 38, 2003. 4He et al., An event-based neural compressive telemetry with >11× loss-less data reduction for high-bandwidth intracortical brain computer interfaces, IEEE Transactions on Biomedical Circuits and Systems, 2024, which puts the serializer for hundreds of megabits per second above 80 mW, too much for a cortical implant. 5Marblestone et al., 2013.

Then the tissue fights back. Neuralink's N1 puts 1,024 electrodes on 64 threads thinner than a hair1, and in its first participant a number of threads retracted in the weeks after surgery, leaving fewer working electrodes2. The Utah array, the kind used in the Stanford and UC Davis speech studies, carries up to 96 electrodes on a 4 by 4 mm square; astrocytes start walling it off within four to six weeks and the scar can last years3. A device that degrades and needs a craniotomy to replace is a research instrument, whatever its channel count.

1Neuralink, PRIME study brochure. 2CNBC, Neuralink's first in-human brain implant has experienced a problem, 8 May 2024. 3Blackrock Neurotech, Utah array; on the glial response, Salatino et al., Glial responses to implanted electrodes in the brain, Nature Biomedical Engineering 1, 2017.

The price of a channel

Put together, the obstacles become a price, and the price is lopsided. In animal research the channel is already cheap: a Neuropixels 1.0 probe costs about $1,400 for nearly a thousand recording sites1, a dollar and a half per site, and Neuropixels 2.0 put 5,120 sites on a four-shank probe2. In a human the probe is the cheap part, and the operating room, the surgeon, the follow-up and the regulatory file are the rest. Outside the skull the bill is just as steep: a conventional SQUID MEG costs about $3.5 million up front and $200,000 a year to run, mostly helium3. Optically pumped magnetometers need no cryogenics and sit on the scalp in a helmet the subject wears while moving4, which removes the 2 cm gap and brings the system to about $1.4 million and $70,000 a year5. Cheaper, and still a hospital purchase.

1Simons Foundation, A new era in neural recording, 2021. 2Steinmetz et al., Neuropixels 2.0: a miniaturized high-density probe for stable, long-term brain recordings, Science 372, 2021. 3Pedersen et al., Wearable OPM-MEG: a changing landscape for epilepsy, Epilepsia, 2022. 4Boto et al., Moving magnetoencephalography towards real-world applications with a wearable system, Nature 555, 2018. 5Pedersen et al., 2022, cited above.

The companies are attacking the operation itself. Precision Neuroscience got FDA 510(k) clearance in March 2025 for a thin-film cortical interface, for implantation of up to 30 days1. Synchron avoids opening the skull altogether: its Stentrode travels up the jugular vein on a catheter and lodges in a vessel beside the motor cortex, and its six-patient COMMAND study reported no device-related serious adverse events at one year2. It pays in resolution, since the vessel wall keeps the electrodes farther from the neurons and they record field potentials instead of single cells. Neuralink reported 21 participants in January 20263, Paradromics got approval in November 2025 to start a trial with a 421-electrode implant4, and both send their data out without a cable through the skin5. The whole field still counts its human users in tens.

1MassDevice, Precision Neuroscience wins FDA clearance for BCI cortical interface, 2025. 2Synchron, Positive results from the U.S. COMMAND study, 30 September 2024. 3Reuters via US News, Neuralink says it has 21 participants enrolled in trials, 28 January 2026. 4Paradromics, FDA approval for the Connect-One study, 20 November 2025. 5University of Michigan, first-in-human Paradromics wireless brain-computer interface; Neuralink's brochure, cited above.

The scale to aim at is the cochlear implant, which passed one million devices worldwide in 20221: a neural interface with a modest electrode count, a routine operation and an established path to payment. Brain reading gets there when three things fall. The operation has to become an outpatient procedure (endovascular, a thin film through a slit, or no hole at all). The channel has to cost in a human roughly what it costs in a mouse, with electrodes that outlast the scar by a decade. And someone has to finish the first full approval of a permanent implant, after which the road is cheaper for everyone behind.

1Zeng, Celebrating the one millionth cochlear implant, JASA Express Letters 2, 2022.

The decoders lower the bar for all three, because they squeeze more out of poor signals every year. From three seconds of MEG, a model trained on 175 volunteers picked the right speech segment among more than a thousand candidates 41% of the time on average1. Meta's Brain2Qwerty decoded typed sentences with a 32% character error rate from MEG and 67% from EEG2. An fMRI decoder at UT Austin reconstructed the gist of stories a person heard or imagined, after 16 hours of scanning each subject3. The realistic near-term target is tens of micrometres and milliseconds over a few chosen regions, read well enough to decode one channel of intent such as speech or movement. The speech results came from a few hundred electrodes on one strip of cortex, and the hard part is making those few hundred cheap, durable and safe.

1Défossez et al., Decoding speech perception from non-invasive brain recordings, Nature Machine Intelligence 5, 2023. 2Lévy et al., Brain-to-text decoding: a non-invasive approach via typing, 2025. 3Tang et al., Semantic reconstruction of continuous language from non-invasive brain recordings, Nature Neuroscience 26, 2023.

What a cheap reader unleashes

Medicine goes first, because it already has the patients and the payers. Speech for people who lost it has been shown. Walking followed: a brain-spine interface let a man with chronic tetraplegia stand and walk in his community, and he kept part of that walking with crutches when the device was switched off1. Epilepsy shows the scale of what is left. The WHO counts around 50 million people with it and estimates that up to 70% could live seizure-free if properly diagnosed and treated2, which leaves millions whose seizures still arrive without warning. In 2013 an implanted advisory system in 15 patients in Melbourne learned to flag periods of high seizure likelihood, with sensitivities between 65% and 100% in the 11 whose algorithms qualified3. A reader that cheap and that permanent gives every one of those patients a forecast.

1Lorach et al., Walking naturally after spinal cord injury using a brain-spine interface, Nature 618, 2023. 2World Health Organization, Epilepsy fact sheet. 3Cook et al., Prediction of seizure likelihood with a long-term, implanted seizure advisory system in patients with drug-resistant epilepsy: a first-in-man study, The Lancet Neurology 12, 2013.

The science of the mind would get its first real instrument. Psychiatry still diagnoses by the number and type of symptoms, and the NIMH itself calls the clinical criteria for a disorder, built by expert consensus, somewhat arbitrary1. The best map of working cortex anyone has is a cubic millimetre of mouse visual cortex, over 200,000 cells and half a billion synapses, assembled by a consortium over years2. How frustration, sustained effort or stress becomes a response of the brain is a question about people living ordinary days, and today almost nobody outside an epilepsy ward is recorded for longer than a scan. Cheap reading would let that question be asked of thousands of people for months.

1National Institute of Mental Health, About RDoC. 2MICrONS Consortium, Functional connectomics spanning multiple areas of mouse visual cortex, Nature 640, 2025.

Then communication and work, which are the same problem seen from two sides. The decoders above already beat thumbs. Since the keyboard, everything a person wants from a machine has gone out through the fingers, and the language models have made the machine side so fast that the slow side is now the person: a model answers in seconds what took minutes to type. A man who cannot move a muscle and an engineer at a desk hit the same wall at different heights, and the first device that reads intent well enough for the first will be studied hard by the second.

The same signal carries what its owner never meant to send. In 2012 researchers from Oxford, Berkeley and Geneva showed that a consumer EEG headset, watching the brain's responses to images of bank cards, PIN digits and places, reduced the uncertainty about a user's private information by roughly 15 to 40% compared with guessing1. When the Neurorights Foundation reviewed the policies of thirty companies selling consumer neurotechnology in 2024, 29 appeared to have access to the user's neural data with no meaningful limit, and 29 could transfer data to third parties2.

1Martinovic et al., On the feasibility of side-channel attacks with brain-computer interfaces, USENIX Security 2012. 2Genser, Damianos and Yuste, Safeguarding brain data: assessing the privacy practices of consumer neurotechnology companies, Neurorights Foundation, April 2024.

The line comes first

Today's non-invasive decoders carry a safeguard inside their weakness. The fMRI decoder needed the subject's cooperation both to train and to work, and decoders trained on other people performed barely above chance1. That protection is an accident of the technology, and it shrinks with every gain in decoding. The data is already moving through consumer channels: in 2023 Chile's Supreme Court ordered Emotiv to delete the brain data its headset had collected from a former senator, who had found that its terms granted the company a perpetual licence over it2.

1Tang et al., 2023, cited above. 2Chilean Supreme Court, ruling of 9 August 2023, analysed in Chilean Supreme Court ruling on the protection of brain activity.

Law has started to move. Chile amended its constitution in October 2021 so that scientific and technological development must respect mental integrity and the law must specially protect brain activity and the information derived from it1. Colorado added neural data to the sensitive data of its privacy act in April 2024, the first US state to do so2, California followed in September 20243 and Montana in May 20254. In November 2025 UNESCO adopted the first global recommendation on the ethics of neurotechnology, which binds nobody5.

1Law 21.383, published 14 October 2021, amending article 19 number 1 of the Constitution; see Carey's summary. 2HB 24-1058, signed 17 April 2024; see Hunton's analysis. 3SB 1223, signed 28 September 2024, in force 1 January 2025; see McDermott's summary. 4SB 163, signed 1 May 2025, in force 1 October 2025; see Cooley's note. 5UNESCO, Recommendation on the Ethics of Neurotechnology, adopted 11 November 2025.

These are good starts with a narrow scope. Most of them treat neural data as one more kind of personal data, to be consented to and deleted, which assumes a person can understand what a recording of their brain will reveal once tomorrow's decoder reads it. Nobody can, and the distance between a company that holds years of recordings and a patient who signed a consent form to walk again is very large. The rules on ownership, inference and refusal have to exist before reading a brain becomes cheap, because after that point the data is already collected, and a recording cannot be unmade. I would take the revolution two years late with that line drawn over the revolution on time without it.

Grey Matter is built on that condition. It follows the frontier as it happens (the companies, the trials, the approvals and the patents, each entry dated and sourced), it explains the tissue the devices are trying to read, from the ion channel to the oscillation, and every card about a technology links to the ethical question it raises, so the two are never read apart.