The BCI Landscape

Every brain-computer interface company sells a place to listen: outside the skull, on the brain, or inside it. The closer the sensor sits, the cleaner the signal and the higher the risk.

Learn why distance from the neuron sets what a device can read, who builds what, and how to tell a peer-reviewed result from a company claim.

Press the three places to listen, one after the other. Watch the electrode hear the target neuron, then its neighbour. Then we build, piece by piece, the physics of the skull, the companies, the rules that test them, and the habit of asking “where is the paper?”

You need no neuroscience and no electronics. We build both from zero, in plain words.

Lesson 29 of the Brain-Computer Interfaces series, a companion to Princeton NEU 560 (Jonathan Pillow). Sources are primary: the papers, the regulator’s clearances, and each company’s own statements, which this lesson always labels as claims. The full list is in the Field Guide (Chapter 9).

Choose where to listen

The traces are a toy: the weights (how strongly the electrode hears each neuron) and the noise levels are illustrative. The published results are real but come from different tasks and people, so they are not a fair race. EEG and MEG: Levy et al., arXiv 2502.17480 (preprint, typing, 35 volunteers). Surface grid: Metzger et al., Nature 2023 (speech, one participant). In the cortex: Willett et al., Nature 2023 (speech, one participant).

Chapter 0

Why You Need a Map Before You Need a Model

Sort six real headlines by the kind of evidence behind them, and learn the one habit that keeps you honest in this field

Imagine one week in the news. A start-up says it has put a chip in its twentieth patient. A university paper says a woman with paralysis spoke again through a computer. A company says its cap will let you type by thought. Another start-up raises a quarter of a billion dollars.

All four stories use the same words: brain, computer, breakthrough. They do not stand on the same ground. One has been checked by outside experts. One has been approved by a regulator. Two are a company talking about itself.

This lesson is a map of that ground. It is lesson 29 of our series, and the earlier lessons built the machinery: the spike, the recording map, recording hardware, decoders such as the Kalman decoder. Here we step back and ask who builds what, and how much we should believe each claim.

Five words you need first

A brain-computer interface (BCI) is a system that reads signals from the brain and turns them into a command for a machine, such as moving a cursor or speaking a word. Think of it as a translator standing between a neuron and a screen.

The reading part uses electrodes: small metal contacts that pick up the tiny voltages that nerve cells make. Think of an electrode as a microphone for electricity. The nerve cells it listens to are neurons, and each time one fires it makes a brief voltage blip called a spike. The first lesson built a spike from scratch.

A BCI is invasive if surgery places the electrodes under the skull, and non-invasive if they stay outside it, for example in a cap. Between the two sit in-between designs that touch the brain surface or sit inside a blood vessel. Chapter 1 explains why this choice matters so much.

Where a claim can come from

A peer-reviewed paper is one that independent experts in the field read and questioned before a journal printed it. It can still be wrong. But somebody outside the authors had to be satisfied first. Think of it as a second signature on a cheque.

A regulator is a public body that decides whether a medical device may be tested or sold. In the United States it is the Food and Drug Administration (FDA). A regulator checks safety and intended use. It does not certify that a device works as well as its maker says.

A company statement is whatever the firm says on its own website or in its own press release. It may be true. Nobody outside has checked it. Press and investor reports usually repeat the company, or report money. A funding round tells you what investors believe. It tells you nothing about what a device can do.

So we rank evidence on a ladder. A peer-reviewed human result is the top rung. A regulator action is next. A company statement is below that. A press or investor report is the lowest, because it is one step further from the data. Try the ladder yourself.

Sort the headline

Read the headline, then choose the kind of source that backs it. The six headlines are real, taken from the sources in the Field Guide. A wrong pick is fine: the reason appears below.

The headlines are paraphrased from: Willett et al., Nature 2023; FDA clearance of the Precision Layer 7 array, April 2025 (trade press); Neuralink’s January 2026 participant count (the company’s post, relayed by news sites); Bloomberg, 15 January 2026, on Merge Labs; Mitchell et al., JAMA Neurology 2023; and press coverage of Sabi’s plans. The picks are not scored anywhere.

Notice how often a number you remember from a headline had no rung at all. This is the most useful habit in the field. Before you ask “how good is it?”, ask “who says so, and who checked?”

The five questions we will keep asking

Whenever a result appears in this lesson, we put the same five questions to it. How many people were in it? What task did they do? What metric measures success, and in which units? Was it printed after peer review, or is it a preprint or a company post? And how long did the device stay in?

The answers are often modest. One participant is common. A few weeks is common. These are not failures. Early human trials are small by design. The risk is only in forgetting it, and comparing a one-person speech result with a four-person safety study as if they were the same sort of thing.

The road

  1. Three places to listen: why distance from the neuron sets how much a sensor can read.
  2. What each place has achieved: the best published results, and why they cannot be raced.
  3. The Utah array and the academic line: Blackrock, BrainGate and thirty years of results.
  4. Neuralink: threads, a robot, a chip, and what a lost thread costs.
  5. Synchron and Precision: gentler placements and what a small safety study can prove.
  6. Paradromics and the regulator’s ladder: the steps a device climbs.
  7. Sabi and Merge: the non-invasive bets and what more sensors can and cannot buy.
  8. Where the computing happens: on the implant or off it, and the cost of each.
  9. Field Guide: numbers, a checklist, the core code and every source.
A start-up announces a record funding round and says its device “reads thoughts”. What does the announcement, by itself, tell you about how well the device works?

Chapter 1

Three Places to Listen

Find out why a sensor’s distance from a neuron decides what it can tell apart, and what each distance costs in surgery

You stand outside a packed stadium. You hear a roar. You cannot tell who shouted. Now walk inside and stand beside one fan. You hear her words, and a little of the people next to her. Walk into the stands, and each voice gets clearer. Each step also costs you something: a ticket, a seat, a crowd to push through.

A brain-computer interface faces the same choice. Neurons are the fans. An electrode is your ear. The only question that matters first is how far your ear is from the voices you want.

What sits between the neuron and the electrode

Start with the layers. The part of the brain that plans movement and speech is the cortex, a sheet of nerve cells a few millimetres thick on the outside of the brain. Over it lies a thin layer of clear fluid, then the hard skull, then the scalp.

When a neuron fires, small electric currents flow through the salty fluid around it. Salty fluid conducts electricity, so the current spreads in every direction, like ripples from a stone. This spreading is called volume conduction. The skull conducts poorly and smears the ripples further.

The sensor reads a voltage. That voltage is a weighted sum of the ripples from every nearby source. Sources close to the sensor get a big weight. Sources far away get a small weight, but there are many more of them. The farther the sensor, the wider the area it adds together.

This gives four common places to put a sensor, from least to most invasive.

The distances above are rounded and illustrative. The real ones vary by person and by device. The pattern is what matters: from the scalp to the cortex the distance falls by a factor of 150.

How far apart must two neurons be to be told apart?

Picture one neuron as a point that makes a bump of voltage on the sensor plane. A sensor at distance d sees that bump spread out to a width that grows with d. Call the width sigma (written σ). A point in a uniform conductor gives a potential that falls off with distance, so a reasonable first guess is that σ grows in proportion to d. We use σ = 0.4 d, a toy value.

Now put two neurons side by side, a distance s apart. Each gives a bump. The sensor sees their sum. If the bumps overlap too much, the sum has one hump and you cannot tell there were two sources. Try it.

Can the sensor tell two neurons apart?

The dashed bumps are two sources, one warm and one blue. The solid teal curve is what the sensor records: their sum. Choose a place, then drag the gap between the neurons.

3.0 mm

Toy model, not a head simulation. Each source gives a Gaussian bump with width σ = 0.4 × the distance to the sensor, and the sensor reads the sum. The distances are round illustrative values. Real heads differ by person, tissue and device; the later lessons of this series cover the real recordings.

Look at the scalp. With the neurons 3 millimetres apart, the sum has a single smooth hump. Move the gap up. At some point a dip appears in the middle. That moment is the rule we can now derive.

s
The gap between the two neurons, in millimetres.
σ
The width of the bump each neuron makes on the sensor. Two equal bumps merge into a flat top when the gap equals 2σ, and show a dip when the gap is larger.
σ = 0.4 d
The toy rule from above: the blur grows in proportion to the distance.
d
The distance from the sensor to the neurons, in millimetres.

Where does 2σ come from? The sum of two equal bumps is smooth and has one peak when the bumps are close. As they separate, the middle of the sum flattens, then sinks. The flat moment, the curve’s second slope at the centre being zero, happens exactly at a gap of two sigma. Beyond it, two peaks.

Scalp: d = 15 mm, σ = 0.4 × 15 = 6 mm, so you need a gap above 12 mm. In the cortex: d = 0.1 mm, σ = 0.04 mm, so a gap above 0.08 mm will do.

That is a factor of 150 in what can be told apart. On the scalp, neurons must be more than a centimetre apart to look different. Inside the cortex, neighbours less than a tenth of a millimetre apart can be separated. This is the whole physical case for putting things inside the head.

What the sensor can see besides

Distance also changes the kind of signal. A penetrating electrode sits close enough to hear single spikes. A surface grid sees the summed activity of a patch, a signal called the local field potential (LFP). A scalp sensor sees slow rhythms from huge crowds of neurons firing in step. The brain rhythms lesson shows those rhythms, and the recording map covers them in detail. Roughly, a spike near an electrode in the cortex is a few hundred microvolts, and scalp EEG is tens of microvolts. These are textbook orders of magnitude.

The cost side is just as plain. Surgery brings the risk of bleeding and infection, and a scar that builds up around a foreign object and slowly muffles the signal. A catheter in a vein avoids opening the skull. A cap avoids everything. Each design is a different answer to the question: how much risk will a person accept for how much signal?

The failure mode: smeared channels you cannot repair

The symptom: a scalp decoder stops improving, however much data you add. The measurement: nearby channels are almost copies of each other. Their correlation is close to 1, so 64 channels carry far fewer than 64 independent patterns. The fix: a spatial filter, which mixes channels to undo some of the smear (the spatial filters lesson builds one), or move the sensor closer. Software can sharpen a blur a little. It cannot bring back what the skull averaged away.

In the toy model, a sensor 1 mm from two neurons has σ = 0.4 mm. What is the smallest gap at which it shows two separate peaks?

Chapter 2

What Each Place Has Achieved, and Why You Cannot Race It

Read the best published speech and typing results from three placements, and learn the four things that must match before two numbers can be compared

You open three papers. One says 62 words per minute. One says 78. One says a 32% error rate. It is tempting to build a leaderboard in your head, with the winner on top. Stop. First check what each number is a number of.

This chapter reads the three results behind the hero and shows how a fair comparison works. The lesson is not that the results are bad. They are remarkable. The lesson is how easily a table of remarkable results turns into a false race.

Three results, in plain words

The first is from inside the cortex. Willett and colleagues, writing in Nature in 2023, implanted microelectrode arrays in a participant with ALS (a disease that destroys the nerves that move muscles), who can no longer speak clearly. The system read the movement commands for speech that the brain still sends. It decoded his attempted speech at 62 words per minute. The word error rate was 9.1% when the system chose from 50 words, and 23.8% when it chose from 125,000. It was one participant.

The second is from the brain surface. Metzger and colleagues, also in Nature in 2023, placed a sheet of 253 electrodes over the speech areas of a participant who could not speak. The system decoded text at a median of 78 words per minute, with a median word error rate of 25%, from a vocabulary of over 1,000 words. It also drove a talking face on screen. It was one participant.

The third is from outside the skull. Levy and colleagues at Meta, in a 2025 preprint, recorded 35 healthy volunteers while they typed sentences on a keyboard. From MEG they decoded the typed characters with an average character error rate of 32%. From EEG the rate was 67%. A preprint is a paper posted publicly before it has been peer reviewed. We could not confirm a journal version.

What an error rate measures

The word error rate (WER) counts the mistakes a decoder makes in a sentence and divides by the number of words in the true sentence. A mistake is a wrong word, a missing word or an extra word. Suppose the true sentence has 20 words. The decoder gets 2 wrong, drops 1 and adds 1. That is 4 mistakes in 20 words, a WER of 4 / 20 = 20%.

The character error rate (CER) does the same thing one letter at a time. It is a different ruler. A 32% CER and a 32% WER are not the same quality. A sentence with a few wrong letters can still be read by a person, so letter errors can look worse on paper than they feel.

The size of the list the decoder chooses from matters most of all. Picking one of 50 words is easy. Even random guessing is right 1 time in 50. Picking one of 125,000 words is a huge task. Willett’s own system scores 9.1% on the small list and 23.8% on the large one. Same person, same implant, and the error more than doubles. So 25% on about 1,000 words and 23.8% on 125,000 words are not the same kind of number, even though both are word error rates.

Compare two results below. Pick any two and read the rows.

Can these two be compared?

Choose result A, then result B. The bars show the headline error figure of each, with its unit. The table shows what differs. A fair race needs the same task, the same measure and the same size of word list.

Result A

Result B

Numbers are from the abstracts: Willett et al., Nature 2023 (23.8% is the 125,000-word figure); Metzger et al., Nature 2023 (median 25%; vocabulary and 253 channels from the paper’s text); Levy et al., arXiv 2502.17480 (CER 32% MEG, 67% EEG). The Metzger and Willett papers each report one participant; Levy reports 35 volunteers. The comparison rules are this lesson’s own.

Try the pair MEG typing and EEG typing. They match on every row, because they come from the same preprint, the same volunteers and the same task. That comparison is fair, and it is a real finding: for typing, the same decoder works about twice as well on MEG as on EEG. Now try Cortex speech against MEG typing. Every row differs. Putting a 23.8% next to a 32% there tells you nothing.

Two more cautions

First, the people differ in kind. The cortex and surface results come from people who cannot speak, who want the device and can try to speak on demand. The non-invasive result comes from healthy volunteers who typed with their fingers. A decoder may be reading finger-movement signals, not an inner voice. The result is useful, but it answers a different question.

Second, MEG (magnetoencephalography) senses the tiny magnetic fields of the brain with very sensitive detectors. It normally needs a shielded room and a large machine. It is non-invasive but not wearable. Do not picture a cap when you read a MEG number.

The failure mode: the leaderboard that is not one

The symptom: a slide ranks “invasive 62, surface 78, non-invasive 32%” as if they were points on one axis. The measurement: list, for each number, the task, the measure, the word list, the people and the review status. Here none of the four match. The fix: compare only within a row that matches, and say so out loud. A fair statement is “within one preprint, MEG beat EEG at typing”. An unfair one is “implants beat caps”. Implants may well do better. These papers do not show it.

Which comparison is closest to a fair one?

Chapter 3

The Utah Array, Blackrock and the Academic Line

Follow the one electrode that carried almost every published human result for twenty years, and see what the academic labs added on top of it

Before any start-up put a chip in a person, a small square of silicon needles did the job. It is called the Utah array. Picture a tiny bed of nails, ten by ten, each nail a hair-thin electrode. A surgeon presses it into the surface of the motor cortex. Each nail listens near a few neurons.

Maynard, Nordhausen and Normann described it in 1997, in a paper whose title says what they hoped for: a recording structure “for potential brain-computer interfaces”. Of the hundred sites, 96 are wired out as recording channels. The array is made by Blackrock Neurotech of Salt Lake City, formerly Cyberkinetics. Blackrock sells the hardware to research groups. Most of the famous names in this chapter did not build the array. They used it.

From a cell’s preference to a pointing arrow

The oldest idea in the field is older than the array. In 1986 Georgopoulos, Schwartz and Kettner recorded from motor-cortex neurons in monkeys that were reaching in different directions. Each neuron fired fastest for one direction. Call that its preferred direction. It fires less when the arm moves away from it. This smooth pattern is called tuning.

A single cell is a poor compass. A crowd is a good one. The population vector adds up one arrow per neuron. Each arrow points along the neuron’s preferred direction, and its length is how much that neuron fires above its resting rate. The sum points where the arm is heading.

Take three neurons with preferred directions of 0, 120 and 240 degrees. In a moment of movement they fire at +8, −2 and −6 hertz relative to rest. The three arrows are 8 times (1, 0), then −2 times (−0.5, 0.866), then −6 times (−0.5, −0.866). Add the x parts: 8 + 1 + 3 = 12. Add the y parts: −1.732 + 5.196 = 3.464. The arrow is (12, 3.464), which points at the angle whose tangent is 3.464 / 12 = 0.289, about 16 degrees. Three cells already give a decent heading. A hundred give a sharp one.

That is the seed of every decoder in this series. The linear decoders lesson turns it into a fitted formula and the Kalman decoder adds a model of how the arm moves from one moment to the next. The array is the ear. The decoder is the interpreter. Most of the progress in the table below came from better interpreters.

Thirty years in one timeline

Here are the milestones of the academic line, each with its journal. Drag the year. Green dots are peer-reviewed human or animal results, and the warm dot is hardware.

Scrub the years

Drag the year from 1986 to 2025. Each dot lights when its paper appears. The read-out names the latest milestone and counts how many are behind you.

2006

Milestones: Georgopoulos et al., Science 1986; Maynard et al., 1997; Serruya et al., Nature 2002; Hochberg et al., Nature 2006 and 2012; Willett et al., Nature 2021 and 2023; Rubin et al., Neurology 2023; Card et al., NEJM 2024. The full references are in the Field Guide. The choice of milestones is this lesson’s, not a complete history.

Read the dots from left to right. In 1997 the hardware appears. In 2002 monkeys use a handful of neurons to move a cursor. In 2006 comes the first human result of the BrainGate team: a participant with tetraplegia moved a cursor and operated a simple robot hand from a 96-channel array. By 2012 two participants reached and grasped with a robot arm. In 2021 the decoder reads imagined handwriting, 90 characters a minute with an offline language model correcting errors. In 2023 and 2024 it reads attempted speech, as in Chapter 2.

The array hardly changed between 1997 and 2024. The decoding changed enormously, from a straight line fitted through the neuron counts to recurrent neural networks and language models. The academic line shows that much of the progress came from software that listens better.

How many people, and how safe?

Counting the people is harder than it sounds. Secondary sources put the total for the Utah array in research somewhere around thirty to forty, with BrainGate studies accounting for most of them. We found no authoritative count, so treat it as an order of magnitude: a few dozen people in all, over about two decades. A BrainGate interim safety report (Rubin and colleagues, Neurology, 2023) is the best peer-reviewed source on safety. It tracks the first participants over the long term and is the paper to read before trusting any claim about implant safety.

The failure mode: the signal fades

The symptom: a decoder that worked on Monday works worse by Friday, and worse again months later. The measurement: the mean firing rate and the number of detectable neurons on each channel drift, so a fixed decoder maps the same intention to a different input. The fix: recalibrate often, or use decoders that adapt. The drift lesson covers both. It is also why Card and colleagues in 2024 stressed rapid calibration in the title of their paper.

Between 1997 and 2024 the Utah array changed little, yet decoded speech went from impossible to 62 words per minute. What changed most?

Chapter 4

Neuralink: Threads, a Robot, a Chip, and What a Lost Thread Costs

Count the electrodes and the data they make, and see why on-implant computing and honest reporting both matter

The Utah array is a rigid bed of nails. A moving brain is soft. Every heartbeat and every turn of the head makes the brain shift a little inside the skull. A stiff needle in a soft, moving organ irritates it. Neuralink, founded in 2016, bet that thinner and softer would be better, and that a robot should do the delicate work of placing it.

This chapter separates what Neuralink has published from what it has said. The difference is large, and it is the main lesson.

What is on paper

In 2019 the company put out a white paper in the Journal of Medical Internet Research (Musk and Neuralink, An integrated brain-machine interface platform with thousands of channels). It describes flexible polymer threads much thinner than a hair, a robot that sews them into the brain, and custom chips that read them. The headline numbers are 3,072 electrodes spread over 96 threads in a recording system, tested in animals. It is a company paper in a journal. It shows the hardware, and it contains no result from a person.

The device now in people is called N1. According to the company it has 64 threads with 16 electrodes each, so 64 × 16 = 1,024 electrodes. It sits flush in a hole in the skull and sends data by radio, with no wire through the skin. A surgical robot inserts the threads. The human trial is called PRIME. The first participant received the implant in January 2024.

What is only said

We searched for a peer-reviewed paper reporting PRIME results from people. We did not find one. What exists is the company’s blog and social posts, and news reports about them. In January 2026 the company said 21 participants were enrolled in trials worldwide. That is a company statement, relayed by news sites.

Cursor speed is reported in bits per second (BPS). Here is how a simplified version works. A grid has 16 cells. Picking the right one among 16 equally likely cells carries log base 2 of 16 = 4 bits. If the participant picks one correctly every half a second, that is 4 ÷ 0.5 = 8 bits per second. Published figures for Neuralink range from about 4.6 to about 10 depending on the source and the date. All come from the company or the press.

What happened to the threads

In the weeks after the first surgery, some threads pulled back out of the brain. The company acknowledged this publicly: a number of threads retracted and the number of effective electrodes fell. The participant told the Wall Street Journal that about 85 percent of the threads had come loose. That figure is his account, and we could not check it. The company said it changed its decoding algorithm to compensate and that performance recovered and later rose above its earlier level. A report on company documents said the team considered inserting future threads deeper.

Treat this as a case study in what a design change does downstream. A chip had 1,024 inputs. Some went quiet. The decoder had to work with what remained. Try the arithmetic.

Count the threads and the data

Drag the number of threads still in place. The bars, on a logarithmic scale, compare the raw stream, the stream if only spike events are sent, and the budget of a low-power implant radio.

64

Illustrative: 20,000 samples per second and 10 bits per sample for the raw stream; 10 detected spikes per second per channel and 16 bits per event for the spike-only stream; a 1 megabit-per-second radio budget. The thread and electrode counts (64, 16, 1,024) are the company’s own description of N1. The real chip settings are not public in this detail.

Each thread carries 16 electrodes, so the electrode count is threads times 16. The raw data rate follows from three numbers.

R
The raw data rate in bits per second.
n
The number of channels, here threads times 16.
f
How many samples per second each channel takes. Spikes last about a millisecond, so f must be tens of thousands.
b
The bits stored in each sample.

1,024 × 20,000 × 10 = 204,800,000 bits per second, or 204.8 megabits per second.

That is about 200 times the 1 megabit-per-second budget in the device above. A radio that fits in a skull, runs off a small battery and must not heat the brain cannot send that. The way out is to send only what the decoder needs: the moments when a spike happened. Sending 10 events per second per channel at 16 bits each is 1,024 × 10 × 16 = 163,840 bits per second, about 0.16 megabits. That fits. This is why spike detection is done on the implant, as the implant compression lesson develops and the analog front end lesson sets up.

The failure mode: a quiet channel is not a bug you can see

The symptom: cursor speed drops after surgery and the team cannot tell why. The measurement: count channels with detected spikes per day. If the count falls from hundreds to dozens, threads have moved or scar has formed. The fix is a decoder that reweights its inputs each day. That kind of repair is why this field depends on the recalibration methods from the last chapter and on honest reporting of the electrode count over time, not just the best day.

Why must an implant detect spikes on the chip instead of sending every raw sample over its radio?

Chapter 5

Synchron and Precision: Gentler Placements, and What a Small Safety Study Can Prove

Compare a vessel route and a brain-surface film, then compute how little “no problems in six people” really tells you

Not everyone wants a needle in the cortex. Two companies chose a gentler route. They accept a blurrier signal in return for an easier operation. Their stories also teach the most useful statistics lesson in the field: what it means when a trial reports that nothing went wrong.

Synchron: through the vein

Synchron avoids opening the skull. Its device, the Stentrode, is a stent (a small mesh tube of the kind doctors use to hold blood vessels open) covered in electrodes. A surgeon feeds it on a catheter through a vein in the neck, up to a large vein that runs beside the motor cortex. The mesh opens and presses against the vessel wall. The electrodes then listen to the cortex through the wall, a few millimetres away. In Chapter 1 that is the “in a vessel” setting: less sharp than the cortex, with no brain surgery.

Two results exist. The first is peer reviewed. Mitchell and colleagues reported the SWITCH study in JAMA Neurology in 2023: four patients with severe paralysis, followed for 12 months. They reported no serious device-related adverse events and no blocked vessel or device movement. The second is the COMMAND study in the United States, with six patients. Synchron announced in 2024 that all six met the primary safety goal over one year. We found that result in a company release and press coverage, presented at a surgeons’ meeting. We did not find a peer-reviewed paper for it. It is interim data from the company.

Precision: a film on the surface

Precision Neuroscience makes a very thin film that lies on the surface of the brain under the skull. Its array is called Layer 7. The array holds 1,024 electrodes on a film about 1.5 square centimetres in area, according to trade press. In April 2025 the FDA cleared it through a 510(k) for implantation up to 30 days. A 510(k) is a route for devices that are substantially like an already-cleared device. It allows use, and in this case the use is short and the purpose is mapping and recording. It is not a clearance to restore speech.

The company says it has tested the film in 37 patients during surgeries done for other reasons. That is a company statement, and we found no peer-reviewed paper for it. This surface route is the “on the brain” setting of the hero. Its research cousin is the 253-electrode grid in Metzger’s speech paper from Chapter 2.

What “no serious problems in 6 people” means

Suppose a trial of 6 people reports zero serious events. Is the device safe? It is encouraging. It cannot show that the risk is small. Here is why.

Say the true chance of a serious event for each person is p. The chance a person has no event is 1 − p. For n people, all independent, the chance that nobody has an event is (1 − p)n. If the real p is high, seeing zero events is unlikely. If p is low, it is likely. So zero events lets us rule out high values of p, but not low ones.

Statisticians pick a threshold. Ask: what is the largest p that would still produce zero events in n people at least 5 percent of the time? Set (1 − p)n = 0.05 and solve. That value is the 95 percent upper bound for the event rate. Any p above it is hard to believe, because zero events would be rarer than 1 time in 20.

pmax
The highest true event rate that is still consistent with seeing zero events, at the 95 percent level.
n
The number of people in the trial. Each one is assumed to be independent of the others.

n = 4 (SWITCH): 0.051/4 = 0.473, so pmax = 52.7%.   n = 6 (COMMAND): 0.051/6 = 0.607, so pmax = 39.3%.

Read that again. After four people with no serious events, the true risk could still be as high as one in two. After six, as high as about two in five. The rule of thumb called the rule of three gives a quick check: pmax is about 3 / n. For n = 6 that says 50%, close enough as a rough guide though a little high. Drag n below to see how fast the bound falls.

How many people must you watch?

The curve is the 95% upper bound on the true event rate after zero events. The two marked points are the two studies above. Drag the number of people and read the bound.

6

Exact one-sided binomial bound, assuming independent people and zero events. It is a statistics exercise, not a claim about any device: SWITCH and COMMAND also report follow-up details that this single number ignores. The study sizes (4 and 6) are from Mitchell et al. 2023 and Synchron’s 2024 announcement.

To be 95 percent sure the true rate is under 10 percent, you need 29 people with no events, because ln 0.05 / ln 0.9 = 28.4. To get under 5 percent you need 59. Early feasibility studies are never that large. That is not a flaw. It is why early studies speak of “feasibility” and why the next studies must be bigger.

The failure mode: reading silence as safety

The symptom: a headline says “no serious adverse events” and the reader hears “safe”. The measurement: compute pmax from the number of people. The fix: report the bound next to the count, and ask how long the follow-up was. The honest reading of SWITCH is “in four people, over a year, nothing serious was seen”, which is a good start and nothing more.

A trial of 6 people sees no serious adverse events. What is the best statement about the true event rate?

Chapter 6

Paradromics and the Regulator’s Ladder

Climb the steps a medical device must take, and place each company on the step its own records support

A news site writes “FDA approves brain implant”. It sounds like the end of a long road. Often it is the first step. The word “approval” covers several very different decisions, and mixing them up is the most common error in coverage of this field.

This chapter lays out the steps, then uses Paradromics as a worked case, because its record is recent and clear.

The steps, in plain words

Think of a driving test. First you practise on a closed track. Then you may drive with an instructor beside you. Then you drive more, under rules, to prove you can do it safely. Last you get a licence, for one kind of vehicle only. A medical device follows the same shape.

  1. Lab and animal work. The company shows the device records and survives. Results here are the company’s own, often with a paper.
  2. An Investigational Device Exemption (IDE). The FDA allows the company to test the device in people, under conditions. It does not say the device works.
  3. An early feasibility study (EFS). A few participants, often two to ten. The question is: is it safe enough to keep going, and does it do anything useful?
  4. A pivotal trial. A larger, planned study that tests whether the device does what the maker says it does, for a stated use.
  5. Marketing authorization. The FDA allows sale for that stated use. Routes include a 510(k) for devices like one already cleared, and a premarket approval (PMA) for high-risk devices.

Each step proves less than it sounds like. Step 2 is permission to try. Step 3 is a small trial, as Chapter 5 showed. Only step 4 asks whether the device works well enough to justify the risk, and only step 5 lets anyone buy it.

Paradromics, step by step

Paradromics builds Connexus, a device with micro-electrodes that go into the brain and a receiver in the chest that sends the data through the skin by radio. It aims at speech: a person who cannot talk thinks the words, and the device types or speaks them.

Here is the record, from the company’s own releases, which match press reports. In 2025 the company placed the device temporarily in a patient already having brain surgery for epilepsy, to show it could record safely. In November 2025 the FDA granted an IDE for a study called Connect-One, an early feasibility study starting with two participants who have impaired speech and limited movement. The sites are UC Davis, Massachusetts General Hospital and the University of Michigan. In June 2026 the company announced the first implantation in the study, at University of Michigan Health.

The company also reports more than 200 bits per second of information transfer, but says the figure comes from preclinical models, which means animal experiments. It is a claim about the hardware’s capacity, not a result in a person, and it uses a different measure from the speech results of Chapter 2. We found no peer-reviewed human paper from the company.

So Paradromics is on step 3 and has just started. That is a fine place to be. It is not a failure to be there. It is where the evidence stops.

A worked reading. Take the headline “Paradromics implants first patient in FDA-approved study”. Ask the three questions. Which step? An IDE, so step 2 became step 3 when the first person was implanted. For which use? An early feasibility study of safety and of speech and computer control, with two participants first. For how long? The company says the participant will be followed for six years. What does the headline not tell you? Any decoding result. The first implant happened in June 2026, and as of our sources no outcome had been published. Now compare a different headline: “FDA clears brain interface”. Ask the same three questions, and you find a 510(k), a 30 day limit and a mapping role. Two headlines with the same verb describe different rungs, and only the three questions separate them.

Where does the record put them?

The five steps rise from left to right. Choose a company. The marker sits on the step its sources support, and the read-out names the evidence.

Placements use the dated sources in the Field Guide and may be out of date: the field moves quickly. A marker shows the furthest step we could document, not a ranking of quality. Precision’s 510(k) covers the thin film for up to 30 days, not a permanent speech system.

Step on each company in turn. Notice that every one sits on step 3 or on a narrow step 5. As far as our sources show, no permanent communication implant has passed a pivotal trial. Sources age fast, so check the FDA’s own databases and the trial registry before you rely on this.

The failure mode: approval means different things

The symptom: a headline says “FDA approves”. The measurement: ask three things. Which step? For which use? For how long? “IDE approved” means permission to test. “510(k) cleared for 30 days” means a short stay in a limited role. “PMA approved” means sale for a stated use. The fix is to write the step next to the word, every time.

A company announces that the FDA granted an IDE for its brain implant. What does that tell you?

Chapter 7

Sabi and Merge: The Non-Invasive Bets

Learn what these two companies say, what is and is not on the record, and what extra sensors can and cannot buy through a skull

The most attractive BCI is one you never operate on. You put on a cap and think. Two young companies are betting that the right sensors plus the right AI can make that real. Both are well funded. Neither has, so far as we could find, a peer-reviewed human result. This chapter keeps those facts straight.

Sabi

Sabi sells the line “Just think”. Its website describes three steps. First, brain imaging with “novel neuroimaging sensors powered by custom ASICs”, sensors with a one-millimetre footprint. (An ASIC is a chip designed for one job.) Second, collecting what it calls the world’s largest neural decoding dataset. Third, mapping brain signals to thoughts with what it calls the world’s most capable Brain Foundation Model. The page invites visitors to join a waitlist. It gives no date and no accuracy figure.

The site does not name the type of signal. Press coverage describes a cap of EEG sensors with a very large sensor count, up to about 100,000 according to some outlets, a funding round of tens of millions of dollars, and a shipping target of the end of 2026. All of this is company claim or press report. We found no paper, no registry entry and no independent test.

A foundation model is a large model trained once on a very broad pile of data, then adapted to many tasks. Language models are the famous example. For brain signals the hope is the same: train on many hours from many people, then adapt to a new person with a few minutes of data. Peer-reviewed work exists on this idea for spikes (for example the POYO model and the NDT2 pretraining paper cited in the Field Guide). The idea is real research. Whether Sabi’s version works is unproven in public.

Merge Labs

Merge Labs came out of stealth in January 2026 with a $252 million seed round that Bloomberg reported, at a valuation of $850 million (secondary coverage differs on whether this is a final or a targeted figure). Press names OpenAI and Bain Capital among the investors and Sam Altman among the backers. It also names scientific co-founders from the Caltech Chen Brain-Machine Interface Center. Reports describe a non-invasive approach that uses ultrasound, a sound wave above human hearing, to read and influence the brain. One analysis calls the company early in its research.

Again, no public human result that we could find. For this chapter the useful fact is the size of the bet. The money shows how many serious investors think the problem is worth attacking. It does not show that the physics will cooperate.

What more sensors can buy

Chapter 1 said the skull smooths the signal. Now ask what happens if you add sensors without moving them closer. The answer comes from sampling theory, the same idea as in the DFT lesson. Imagine a pattern across the scalp that varies smoothly. If the pattern has no detail finer than a certain size, then sensors spaced about half that size apart capture all of it. Putting in more sensors adds no new detail. Each extra sensor can only help average out noise.

Take a toy line of scalp 20 centimetres long with a blur of 6 millimetres. The finest pattern is about 0.3 cycles per centimetre. The sampling rule says you need at least twice that, 0.6 samples per centimetre, so 12 sensors across 20 centimetres. Past that, the signal is fully captured. Noise is different. If each sensor has independent noise, averaging N of them cuts the noise by the square root of N. Try it with the slider.

More sensors, same skull

The warm curve is the true smooth pattern. The teal curve is rebuilt from noisy sensors spread along the line. Drag the number of sensors from 4 to 100,000 and watch the error.

84

Toy model. A 20 cm line; the true pattern is three smooth waves (0.1, 0.2 and 0.3 cycles per cm) because the skull removes finer detail; every sensor adds independent noise of the same size as the signal; the rebuild averages the sensors that fall in each 0.5 cm cell and joins the cells with straight lines. It illustrates sampling and averaging. It says nothing about any company’s sensors.

At 4 sensors the rebuild misses most of the shape. At about a dozen it finds the shape, noisily. Above a few dozen, the error falls slowly: each tenfold increase in sensors cuts it by about three, the square root of ten, until the 0.5 cm cells set a floor. At 100,000 sensors you have a very clean picture of a very blurry pattern.

That leaves two ways extra sensors and clever software can still help. Averaging lowers noise. And a model that knows what brain patterns look like can fill in what the skull removed, by using prior knowledge. That is the case for a brain foundation model. It is a case, not a result.

The failure mode: adding channels that add nothing

The symptom: accuracy stops improving as channels go from 64 to 640. The measurement: check how many independent patterns the channels carry, for example by the number of components needed to explain most of the variance (see the EEG preprocessing lesson and PCA). If a few dozen explain nearly all of it, the extra channels are copies. The fix: use the extra channels to average noise, move sensors closer, or lean on a model with strong priors.

In the toy line, the signal has no detail finer than about 0.3 cycles per centimetre. What do sensors beyond the sampling minimum of 12 still do?

Chapter 8

Where the Computing Happens: On the Implant or Off It

Choose the point where the data leaves the head, and see how each choice trades radio power against compute power

Every company in this lesson answers one more design question, and it does not show up in headlines. Where does the computer sit? A BCI is a pipeline: sense, clean, find the spikes, count them, decode them, send a command. Somewhere along that line the data has to leave the head. Whatever happens before that point must run on a chip inside the body, powered by a small battery and kept cool.

Why cool? Brain tissue is easily harmed by heat. An implant that warms its surroundings by about a degree risks damage. So an implant has a strict power budget, and every operation and every bit sent over the air spends it.

The pipeline, stage by stage

Here are the five stages, using Chapter 4’s numbers for 1,024 channels.

  1. Amplify and digitize. Convert tiny voltages into numbers, 20,000 times a second, 10 bits each. This must be on the implant. Output: 204.8 megabits per second. (The analog front end lesson builds this stage, and the quantization lesson explains the 10 bits.)
  2. Detect spikes. Mark the moments when a spike happened. Output: about 0.16 megabits per second in the toy count.
  3. Count. Add up the spikes in short windows, say 20 milliseconds. Output: about 0.2 megabits per second.
  4. Decode. Turn the counts into a cursor velocity or a word, with a Kalman filter or a neural network. Output: a few numbers, almost nothing.
  5. Act. Move the cursor or speak. This happens outside.

You can cut the pipeline anywhere after stage 1. Cut early and the implant does little computing but must radio a lot of data. Cut late and the radio is nearly silent but the implant must run the decoder.

The trade in one equation

P
The power the implant draws in total, in milliwatts.
C
The power of every stage that runs on the implant. It grows as you move the cut later.
E
The energy the radio spends per bit sent. A toy value of 50 nanojoules per bit is used here.
R
The data rate that crosses the radio, in bits per second. It falls as you move the cut later.

Work the first cut. Sending the raw stream: 50 × 10−9 joules per bit times 204.8 × 106 bits per second = 10.24 watts. That is 10,240 milliwatts for the radio alone, hundreds of times above any safe budget. Now cut after spike detection: 50 × 10−9 × 163,840 = 8.2 milliwatts. The radio drops by a factor of 1,250. The cost is a few milliwatts of detection logic. Try each cut.

Where do you cut the pipeline?

Choose the last stage that runs on the implant. The bars show compute power and radio power against a toy budget of 40 milliwatts.

Illustrative numbers for 1,024 channels: compute power of 20, 5, 1 and 25 milliwatts for the four stages; data rates of 204.8 Mbit/s, 0.164 Mbit/s, 0.2 Mbit/s and about 0.001 Mbit/s; a radio cost of 50 nanojoules per bit; a 40 milliwatt budget. They are a teaching sketch, not any company’s chip.

The best cuts here are detect and count. Decoding on the implant removes the radio but costs the most compute and, just as important, freezes the decoder inside the body. Yet drift means decoders need frequent retraining (Chapter 3). Moving the cut late makes updates harder. This is why many systems cut early and decode on a computer outside.

Where the people and the methods sit

This pipeline is also a map of where different skills enter. Chip designers and signal-processing engineers own stages 1 and 2. Estimation and machine learning own stages 3 and 4: the Kalman filter, its adaptive versions, recurrent networks and transformers. Calibration and drift sit across stage 4 and the recalibration lesson. Safety and regulation wrap the whole thing, as Chapters 5 and 6 showed.

A real example shows the pipeline at work with the decoder outside the body. Wairagkar and colleagues (Nature, 2025) used 256 microelectrodes in a participant with ALS and decoded his voice causally, which means each output uses only past and present activity. The paper reports synthesis within 10 milliseconds of the neural activity. IEEE Spectrum, reporting on the same work, quotes a delay of 25 milliseconds. Both can be right, because they may measure different spans of the pipeline. It is a reminder to ask which latency a number measures.

The failure mode: the radio eats the battery

The symptom: an implant that works on the bench runs hot or empties its battery in hours once it streams. The measurement: log the radio duty cycle and the bits sent per second. The fix: run more stages on the implant, so the cut falls later in the pipeline and less crosses the radio. Detect and count before sending.

With a radio that costs 50 nanojoules per bit, sending a 204.8 megabit-per-second raw stream costs about 10 watts. What is the usual fix?

Chapter 9

Field Guide: The Landscape on One Page

Review the map, keep the numbers and the checklist, read the core as code, and find every source

You can now read BCI news the way an engineer does. You know that distance from the neuron sets what a sensor can tell apart. You know the best published results come from small groups and cannot be raced against each other. You know the companies by what each placed in the body and by the step each has reached. And you know the habit that matters most: ask who says it, and who checked.

The one-paragraph summary

A BCI reads the brain with electrodes placed on the scalp, in a vessel, on the brain or in the cortex. Closer sensors tell neurons apart better (the 2σ rule) but need surgery. Blackrock’s Utah array carried two decades of academic human results, with progress coming mainly from decoders. Neuralink uses 1,024 electrodes on threads and a robot, and its human results are company statements. Synchron and Precision choose gentler routes, and Paradromics has just begun its first study. Sabi and Merge bet on non-invasive sensing plus AI, with funding and claims but no public human result we could find. Safety studies of four to six people cannot show a risk is small. The regulator’s ladder has five steps, and most companies stand on the third.

The numbers that matter

WhatNumberSource and label
Gap two neurons need to look separate (toy)12 mm on the scalp; 0.08 mm in the cortexChapter 1 toy model, σ = 0.4 d
Attempted speech, inside the cortex62 words per minute; 9.1% WER (50 words), 23.8% (125,000)Willett 2023, Nature, one person: peer reviewed
Attempted speech, brain surface78 words per minute; median 25% WER; 253 channelsMetzger 2023, Nature, one person: peer reviewed
Typing decode, outside the skullCER 32% with MEG, 67% with EEG; 35 volunteersLevy 2025, arXiv: preprint
Stent-electrode safety, SWITCH4 patients, 12 months, no serious device-related eventsMitchell 2023, JAMA Neurology: peer reviewed
Stent-electrode safety, COMMAND6 patients, primary safety goal metSynchron 2024: company, interim
Bound after zero events52.7% (n = 4); 39.3% (n = 6); 29 people for 10%, 59 for 5%Chapter 5 arithmetic
N1 hardware64 threads × 16 = 1,024 electrodesNeuralink description: company claim
Neuralink enrolment21 participants worldwide, January 2026Neuralink post, relayed by news: company claim
Layer 7 film1,024 electrodes, 510(k) for up to 30 days, April 2025Trade press on the FDA clearance: regulator action
Connect-OneIDE November 2025; two participants first; first implant June 2026Paradromics releases: company claim
Raw data rate, 1,024 channels204.8 Mbit/s raw; 0.16 Mbit/s spike events onlyChapter 4, illustrative settings
Radio power at 50 nJ per bit10.2 W raw; 8.2 mW spikes onlyChapter 8, toy radio cost
Merge Labs seed round$252 million, valuation reported as $850 million, January 2026Bloomberg: press

The checklist: before you believe a BCI claim

The core as code

Four small functions carry this lesson’s arithmetic. Each one answers a question from a chapter.

pythonimport math

def resolvable(gap_mm, dist_mm, k=0.4):            # Ch 1: can a sensor tell two neurons apart?
    return gap_mm > 2 * k * dist_mm                 # two equal bumps of width sigma = k * d split when gap > 2 sigma

def upper_bound(n, conf=0.95):                 # Ch 5: zero events in n people
    return 1 - (1 - conf) ** (1 / n)         # solve (1 - p)^n = 1 - conf for p

def people_needed(p_target, conf=0.95):
    return math.ceil(math.log(1 - conf) / math.log(1 - p_target))

def power_mw(compute_mw, channels, fs=20_000, bits=10, keep=1.0, e_bit=50e-9):
    rate = channels * fs * bits * keep           # Ch 4, 8: R = n * f * b, times the share kept by on-implant processing
    return compute_mw + e_bit * rate * 1000     # P = C + E * R, in milliwatts

print(resolvable(3, 15), resolvable(3, 0.1))               # False True
print(round(upper_bound(4), 3), round(upper_bound(6), 3))  # 0.527 0.393
print(people_needed(0.10), people_needed(0.05))         # 29 59

Read it line by line and name the chapter. The first function is the physics of placement. The second and third are the honesty about small trials. The fourth is the budget that forces computing onto the implant. Chapter 2 and Chapter 6 have no code because they are habits: compare like with like, and write the step next to the word.

Where this sits

This lesson is the map for the second half of the series. The earlier lessons built the pieces: the spike, the recording map, recording hardware, implant compression, event-related potentials and the Kalman decoder. Here we put names and evidence on top of them. The lessons that follow in this series read the individual papers behind the table above.

The Princeton course NEU 560 (Jonathan Pillow) teaches the statistical models behind these decoders. This lesson draws on its approach to neural data but takes no slides from it. Everything above is re-taught from the primary sources listed below.

Keep going

Sources

Peer-reviewed or preprint papers. Georgopoulos et al., Science 1986. Maynard et al., 1997. Serruya et al., Nature 2002. Hochberg et al., Nature 2006 and 2012. Willett et al., Nature 2021 and 2023. Metzger et al., Nature 2023. Card et al., NEJM 2024. Wairagkar et al., Nature 2025. Rubin et al., Neurology 2023. Mitchell et al., JAMA Neurology 2023. Musk and Neuralink, JMIR 2019. Levy et al., arXiv 2502.17480 (preprint). Azabou et al., POYO, NeurIPS 2023. Ye et al., NDT2, NeurIPS 2023.

Company statements and press, labelled as claims. Neuralink PRIME update (May 2024; we could not retrieve its text and cite it for existence only); Neuralink’s January 2026 participant count, as relayed by Benzinga; the thread retraction as reported by Popular Science; Synchron’s COMMAND announcement; Precision’s clearance, in Fierce Biotech; Paradromics’ IDE release and first-implant release; Sabi’s site; Bloomberg on Merge Labs; Blackrock Neurotech. Check each against newer reports: the field moves fast.

The takeaway. Distance from the neuron decides what a sensor can read, and the evidence behind a claim decides how far to trust it. Learn both rulers, and the landscape stops being a wall of headlines and becomes a map.
A press story says a company’s cap with 100,000 sensors “reads thoughts” and has raised $50 million. What is the most useful next step?

Now press Teach and explain the landscape out loud from memory: why distance matters, how the two ends of the skull trade, what a small safety study can and cannot show, and why a funding round is not a result. Then go back to the three places to listen at the top and press them again.

The BCI Landscape, Brain-Computer Interfaces lesson 29
Princeton NEU 560 · Willett 2023 · Mitchell 2023 · Back to Gleams