Can One Brain Talk to Another? Inside Project Mindlink, HORIZON's Honest Look at Brain-to-Brain Links

Picture a person sitting in a lab, wearing a cap full of sensors. They imagine moving their right hand. Far away, a second person sits with their eyes closed. A magnetic pulse touches the back of their head, and they see a small flash of light that is not really there.

Nobody spoke. Nobody typed. One “yes” went from one brain to another.

This has really been done, by serious research groups, and the results were published in peer-reviewed journals. It is also close to the whole story. After more than a decade of work, the most advanced brain-to-brain links between healthy people still carry roughly one yes-or-no decision at a time.

That gap, between one yes-or-no signal and the telepathy of films and headlines, is the subject of Project Mindlink. It is a small, exploratory study by HORIZON, the moonshot research division at LINKSPREED’s Y-LAB. This article introduces the project, explains the science behind it in plain language, and says openly why we think it is worth doing even though the chance of reaching “real” telepathy is extremely low.

If you are new to the topic, start at the top and read on. If you already know your way around brain-computer interfaces, the sections on the limits of reading and writing brain signals are where the project’s questions sit.

The short version

What exists today. Researchers have connected two or three people through computers so that one person’s mental choice reaches another person’s brain as a flash of light or a small muscle twitch. Each exchange carries about one bit: yes or no.
What does not exist. Nobody has shown that a thought, a word, or an image can be sent directly into another person’s mind. Words in these experiments are spelled out as long chains of yes/no signals.
Why it is hard. Sensors on the scalp see a blurry picture of the brain. Every brain stores meaning in its own way. And stimulating the brain from the outside is far cruder than reading it.
What implants change. People with paralysis have communicated at dozens of words per minute using electrodes implanted in the brain. That is reading only, it needs surgery, and it does not write anything into another person’s mind.
What Mindlink does. We treat the brain-to-brain link as a communication channel and ask how reliable each step is and how much information can pass per minute. The work is based on published research and public data. It involves no experiments on people.
The odds. We put the chance of general, non-invasive “semantic telepathy” far below one percent. We say that out loud, because it is part of how HORIZON works.

Where this comes from: LINKSPREED, Y-LAB and HORIZON

Most people know LINKSPREED for Web4: social networks you own and run yourself, the UIID identity system, and the tools around them. Mindlink has nothing to do with any of that, and we want to say so clearly from the start.

HORIZON is a different kind of unit. It is an internal moonshot research division inside Y-LAB that looks at high-risk scientific questions, such as parallel-universe signalling, macroscopic quantum teleportation, traversable wormholes and human metabolic suspension. Brain-to-brain communication is one of about twenty programmes in this portfolio.

HORIZON follows a few simple rules:

The honesty rule. Every project states its own probability of success out loud, even when that number is tiny. Some programmes sit at fractions of a percent or far lower.
Real calculations. If we report a number, it comes from a calculation we actually ran, and every claim about physics or biology has to rest on real, citable literature.
Decoupled from products. HORIZON projects are not tied to a product roadmap. There is no promise that any of them leads to a feature, a device or a launch.
Short public briefs. For each project we publish a brief of at most ten pages in plain language, so that negative results and tighter limits are useful to others, too.

A note on structure, since people sometimes ask. LINKSPREED LLC is based in Albuquerque, New Mexico, and serves markets outside the EU. The independent LINKSPREED UG serves Europe. Legal details are published at legal.linkspreed.com. HORIZON’s research belongs to this company family and is not an external partnership.

What would “telepathy” actually require?

“Telepathy” is a loose word, so it helps to break it into parts. For one mind to pass a thought to another through technology, three things would have to work:

Read. Something has to pick up the neural activity that carries the thought in the sender’s brain.
Carry. The information has to travel from one place to another. This is the easy part. We have the internet.
Write. Something has to place the information into the receiver’s brain in a form that the receiver experiences as meaningful.

Every real experiment to date uses all three, and every one of them has been done with the simplest possible content. A “brain-computer interface” (a BCI) handles step one: it reads brain activity and turns it into a command for a computer. A “computer-brain interface” handles step three: it stimulates the brain from outside. Chain the two together through a network and you have a brain-to-brain interface.

The important point is that the computer in the middle is not a detail. It is the only reason any information crosses from one person to the other. Without it, nothing travels.

The two tools used in most studies

EEG (electroencephalography) is the cap with electrodes that sits on the scalp. It records tiny electrical signals produced by large groups of nerve cells working together. It is safe, portable and relatively cheap. It is also a coarse instrument, for reasons we come to below.

TMS (transcranial magnetic stimulation) uses a coil held against the head. A brief magnetic pulse induces a small electric current in the brain region beneath it. Over the part of the brain that controls the hand, a pulse can make the hand twitch. Over the part that processes vision, a pulse can make a person see a short flash of light called a phosphene. TMS is used in clinics and research labs under strict safety rules.

These two tools define what is possible without surgery. A coarse reader on one side, a coarse writer on the other.

What has actually been shown

The history of brain-to-brain interfaces is short, so it is easy to lay out.

Year	Study	What happened
2013	Pais-Vieira et al., rats	One rat’s brain activity was sent as electrical stimulation to another rat’s brain. The second rat learned to use it to make the matching choice. Electrodes were implanted in both animals.
2014	Grau et al.	A person imagined hand movements, and the resulting EEG pattern was sent over the internet. A receiver, separated by great distances, perceived the signal as phosphenes. Words were first converted into streams of 0s and 1s.
2014	Rao et al.	Six people in a cooperative computer game. A sender’s imagined movement, read from EEG, triggered a TMS pulse over the receiver’s motor cortex, which made the receiver’s hand press a touchpad.
2015	Stocco et al., “20 Questions”	Five pairs of people played a question-and-answer guessing game, with answers sent as visual flashes. Results were significantly better than in a sham control where the link was not real.
2019	Jiang et al., “BrainNet”	Two senders and one receiver solved a Tetris-like task together. The receiver could not see the screen. Five groups reached an average accuracy of 81.25%.

Let us look at what these studies did and did not do.

Grau and colleagues: words as bit streams

This is the study that comes closest to the popular idea. The researchers sent words between two people. But look at how: the words were encoded as binary streams, and the sender “typed” each bit through a mental action while the receiver perceived each bit as a flash or the absence of one. The words travelled, but the neural channel itself carried only a series of ones and zeros. The meaning lived in the encoding that both sides had agreed on beforehand.

Rao and colleagues: a motor signal, not a message

In this experiment, the receiver does not “receive an idea”. Their hand moves, because a magnetic pulse activated their motor cortex. The receiver did not consciously interpret anything. This is an important distinction. It is a form of remote control, closer to a reflex hammer than to a conversation.

Stocco and colleagues: where meaning really lives

The “20 Questions” experiment is the best example of how a very small channel can solve a task that looks rich. One person had a mystery item in mind. The other asked yes/no questions. Answers came back as flashes. The shared structure of the game, the list of possible items and the questions, did the heavy lifting. The brain link carried one bit at a time, and the protocol gave those bits their meaning.

BrainNet: a network of brains

BrainNet is the most elaborate demonstration so far. Two senders and one receiver worked together. Each sender decided whether a falling block should be rotated, and the decisions reached the receiver by magnetic stimulation. The receiver combined both signals and made the final call. In a second round, the senders could check the result and send corrections. When the researchers added noise to one sender’s signal on purpose, receivers learned to trust the more reliable sender, using nothing but the information delivered to their brains.

That last finding is interesting for reasons that go beyond engineering. It shows that people can learn to weigh unfamiliar, artificial signals in their heads, much like a person learns to judge which of two friends gives better advice. It does not show that anything like a thought was transmitted.

A bit is a bit: why one signal is so far from a sentence

A bit is the smallest unit of information: a choice between two possibilities. Yes or no. Left or right. Flash or no flash.

Three bits are enough to pick one item out of eight. About four and a bit are enough to single out one item from a list of twenty. That is how the “20 Questions” game works.

The trouble is time. In published brain-to-brain experiments, each exchange takes seconds, and some of those seconds are lost to errors and re-tries. Our own modelling of a realistic non-invasive link, using plausible accuracy ranges for each step, gives an order of magnitude of a few bits per minute. For scale, natural human conversation runs at around 160 words per minute, according to the speech-decoding literature (Willett et al., 2023). A single word is worth many bits.

Even the best-established one-way brain-computer interfaces, those that only read a sender’s brain and do not write anything into another, have historically been described as reaching somewhere in the range of ten to twenty-five bits per minute at best (Wolpaw et al., 2002). A link that has to pass through a second brain cannot be faster than the slower of its two halves.

That last point deserves a plain-language explanation, because it is the reason why “just build better hardware” is not a complete answer.

The chain can only lose information

Think of a message passed along a line of people in a noisy room. Each person can drop or garble parts of the message. Nobody along the line can restore something that the first person never said clearly. The message can only get worse, or at best stay the same.

A brain-to-brain interface is such a chain: the sender’s brain and sensors, the software that decodes the signal, the network, the stimulation device, and finally the receiver’s brain. Information theory has a precise version of this idea (it is called the data processing inequality), but the intuition is enough: a chain can never carry more than its weakest link. If the sender’s side produces a noisy single bit, the most refined stimulation in the world cannot turn it into a clean sentence.

Why reading the brain from outside is blurry

Now to the first of the main obstacles. Why can’t we simply put better sensors on the head and read thoughts?

The skull smears the signal

The electrical activity of nerve cells has to pass through the skull and the scalp before an electrode on the head sees it. The skull conducts electricity poorly, and it spreads the signal out. The result is that each electrode on the scalp reflects activity from a region several centimetres wide, and neighbouring electrodes see very similar signals.

A useful picture: listening to a stadium from outside the walls. You can tell when the crowd cheers, and you can tell roughly which side of the stadium is louder. You cannot make out what any individual fan is shouting.

Only large, synchronised activity shows up

EEG picks up the combined activity of very large groups of cells that fire in step. That works well for broad rhythms, such as the drop in a certain brain wave over the left or right motor area when you imagine moving a hand. It works poorly for fine-grained patterns that distinguish one word or idea from another.

Sources cannot be uniquely recovered

There is also a mathematical problem. Many different patterns of activity inside the head can produce exactly the same voltages on the scalp. Working backwards from the scalp to the sources is what is known as an ill-posed inverse problem. Clever assumptions can give you a plausible answer, but not the answer. This is not a limit of today’s equipment. It is a limit of the problem itself.

Other ways of looking at the brain

Method	What it is good at	Why it falls short for reading thoughts
EEG	Fast timing, portable, low cost	Blurry in space, noisy, sensitive to muscle and eye movement
MEG (magnetoencephalography)	Fast timing, less distortion by the skull	Needs a shielded room and a very sensitive scanner, still limited in detail
fMRI (functional magnetic resonance imaging)	Good spatial detail	Slow, because it follows blood flow rather than nerve signals, and it needs a large scanner

Each method trades something away. None of them gives you fine detail, fast timing and portability at the same time.

Why every brain has its own “language”

The second obstacle is less about physics and more about biology.

Decoders are trained per person

Look at the strongest non-invasive language results. In a 2023 study in Nature Neuroscience, Jerry Tang, Alexander Huth and colleagues used fMRI to reconstruct the gist of language from brain activity. Their system could produce word sequences that recovered the meaning of speech a person was hearing, speech they imagined, and even silent videos. That is a remarkable result.

It also comes with conditions. The decoder had to be trained on many hours of data from each individual person, and the researchers reported that the participant’s cooperation was needed both to train the decoder and to apply it. The result captures the gist, not the exact words, and it works with a large scanner. It reads. It does not write.

A different 2023 study, by Alexandre Défossez and colleagues, used contrastive learning on recordings from 175 volunteers and decoded what people were listening to. From three seconds of MEG signal, the system could identify the matching speech segment with up to 41% accuracy among more than a thousand possibilities on average, and up to 80% for the best participants. Again, a real result, on perceived speech rather than spontaneous inner thought, and with the more sensitive MEG scanner.

Meaning is spread out, and personal

The brain does not keep the meaning of a word in one tidy spot. Semantic information is distributed across wide areas, and the details differ between people. That is why a decoder trained on one person does not simply work on another.

For a brain-to-brain link this creates a double problem. It is hard enough to read a person’s code. To make a second person experience the same meaning, you would have to write into their code, which is different and which nobody has measured in a way that would allow it.

Perceived is not the same as thought

Most of the best non-invasive results decode what the person hears or sees, because outside stimuli drive the brain in a strong, time-locked way. Spontaneous inner thought has no outside clock and no clear ground truth to train against. That is one more reason why the results from decoding studies should not be read as “the machine can read your mind”.

Writing is harder than reading

If reading from the outside is blurry, writing from the outside is cruder still.

A TMS coil affects a patch of brain several centimetres wide. What it can do is make a hand twitch if it sits over the motor area, or produce a flash of light if it sits over the visual area. What it cannot do, as far as anyone has shown, is place a specific word, a face or a concept into a person’s experience.

In the experiments above, the receiver’s vocabulary consists of two things: flash or no flash, twitch or no twitch. To go beyond that, researchers would have to find ways to create reliably different, recognisable percepts on demand, one for each symbol. Whether that is possible in healthy volunteers is an open question that, from the literature we have reviewed, nobody has settled.

What implants can do, and why that is not telepathy

Things look very different when electrodes sit directly on or in the brain.

In 2023, Francis Willett and colleagues at Stanford reported a speech decoder for a participant with ALS who can no longer speak clearly. Electrode arrays implanted in the brain recorded nerve-cell activity while the participant tried to speak. The system turned this into text at 62 words per minute, with a word error rate of 9.1% on a 50-word vocabulary and 23.8% on a vocabulary of 125,000 words. That is 3.4 times faster than the previous record at the time.

In 2024, Nicholas Card and colleagues, working with a 45-year-old man with ALS, used four implanted arrays with 256 electrodes in total. On the first day of use, after 30 minutes of data collection, the system reached 99.6% accuracy on a 50-word vocabulary. On the second day it reached 90.2% on a 125,000-word vocabulary. Over more than eight months and 248 hours of self-paced conversation, accuracy stayed at 97.5%.

These are life-changing results for people who have lost the ability to speak. They are also worth reading carefully:

The signal comes from the part of the brain that controls the mouth and tongue. The system decodes the movements the person is trying to make, not abstract thoughts.
It requires brain surgery and is aimed at patients.
It goes in one direction. Nothing is written into anyone else’s brain.
Each decoder is trained for the individual who received the implant.

So there are really two different worlds. In the invasive world, signals are rich and reading works well, but there is no writing and no brain-to-brain link. In the non-invasive world, a brain-to-brain link exists but carries a single bit. The gap between the two is not just a matter of resolution. It is a gap in what the systems can do at all.

What Project Mindlink actually does

With this background, here is what the project is. Mindlink is not an attempt to build telepathy. We treat the brain-to-brain link as a communication channel, and we ask the questions that engineers ask about any channel: how reliable is each step, how much information gets through per minute, and what limits it?

The work has three strands, described here at a high level.

1. The sender side, in the open

We are building a reproducible analysis pipeline that detects a very simple mental signal in public EEG recordings: imagining a left-hand versus a right-hand movement. The point is to see how reliable the first link in the chain is across many different people, using data that anyone can download.

One thing to be clear about: we will not report accuracy figures on human data until we have run the pipeline on real, public recordings and checked the results with proper statistics. Small datasets can look impressive purely by chance, and we do not want to add to the pile of overconfident brain-signal claims.

2. The whole chain

Next, we combine what we learn about the sender side with what published studies report about the receiver side. The aim is an honest estimate of how many bits per minute a realistic non-invasive brain-to-brain link can carry, and what happens if each exchange carries more than one bit, for instance a choice from a short list of agreed symbols, in the spirit of “20 Questions”.

3. A paper design, not an experiment

Finally, we sketch what a slightly richer link could look like on paper, and we look for the weakest points in our own design. We also try to argue as hard as we can against our own idea, and we write down how serious each objection is.

Mindlink does not involve experiments on people, and it does not involve any brain stimulation. If a human study ever made sense, it would have to happen in partnership with an established, ethics-approved laboratory. That would be a separate decision with its own review.

Not everyone can send even one bit

One finding from the BCI literature shapes the whole design. In studies of brain-computer interfaces based on imagined movement, a sizeable share of healthy people, often described as roughly one in five, cannot produce a signal that is strong enough for reliable control. Researchers call this “BCI illiteracy” (Blankertz et al., 2010), and it is a poor name, because it is not a matter of skill or intelligence.

It matters for brain-to-brain links because it means that the sender’s brain sets part of the limit. Not only the hardware does.

The odds, said out loud

HORIZON’s honesty rule asks us to state a probability, so here it is.

We estimate the chance that general, non-invasive semantic telepathy becomes possible in the foreseeable future as extremely low, far below one percent.

The reasoning is straightforward. For such a system to exist, several things would all have to become true at once: non-invasive sensors would have to read spontaneous meaning at useful speed; some form of non-invasive stimulation would have to write specific content, not just flashes and twitches; there would have to be a way to translate between two people’s different neural codes; and all of it would have to be safe for repeated use. Each of those is a major open problem, and today there is no known route for any of them. Multiplying even generous odds for each step gives a very small number.

On the other hand, the chance that we can measure and explain what the existing binary links can and cannot do, using published results and public data, is high. That is the realistic part of the project, and it is the part we are committing to.

What would change our mind

We try to be clear about what would make us revise the estimate upwards. For example:

a convincing, reproducible demonstration of non-invasive stimulation that creates several clearly distinguishable percepts on demand
evidence of shared neural coding across people that is strong enough to translate between individuals
non-invasive reading of spontaneous content, rather than perceived stimuli, that works without hours of per-person training and without the person’s active cooperation

If any of these appear in the literature, we will say so.

So why do it?

If the odds of the big goal are so small, why spend time on it? There are four reasons, and none of them depend on the big goal succeeding.

Clear facts in a noisy debate. Brain technology attracts big promises, in both directions. A neutral, open estimate of what non-invasive brain-to-brain links can carry helps journalists, students and curious readers separate what has been demonstrated from what is speculation.

Methods that carry over. Building clean analysis pipelines for noisy biological signals, using proper statistical checks, and being strict about bit-rate accounting are skills and tools that apply to many other fields.

Teaching value. Mindlink is a good way to explain how physics, information theory and neuroscience meet. We plan to turn parts of the material into accessible educational content.

Responsible framing. Mental privacy matters. The research we cite shows that today’s decoders need a person’s cooperation, and being exact about those limits protects against both hype and unnecessary fear.

And one more point, which we think matters. If the project mainly confirms how large the gap is, that is a valid result. A well-measured “no” is useful, because it stops other people from wasting time on the same dead end and tells them where the real bottlenecks are.

Questions people ask

Is this telepathy?

Not in the popular sense. Telepathy in films means thoughts, words or images passing directly from one mind to another. The experiments that exist carry a single yes-or-no signal through a chain of computers and stimulation devices, under laboratory conditions.

Can a machine read my thoughts?

Not from a cap on your head, and not against your will. The best non-invasive language decoders need hours of data from the person, a large scanner and the person’s cooperation, and they capture the gist of what someone hears or imagines. Implanted systems decode attempted speech for people who chose to have surgery.

Could someone send a thought into my head?

No such capability has been shown. The best that stimulation has done in healthy volunteers is to produce a flash of light or a muscle twitch.

Is LINKSPREED building a brain product?

No. Mindlink is exploratory research. It is not connected to UIID or to any other LINKSPREED product, and it comes with no promise of a future feature or device.

Will you test this on people?

Not in this project. Mindlink works with published research, public data and calculations.

What happens next, and how to follow it

The next phase of Mindlink is built around public data and simulation: running the sender-side pipeline on real recordings, checking the results with proper statistics, and refining the whole-chain estimate.

The public documents for the project, including the short public brief, will be available for download soon. You will find them, together with the rest of the HORIZON portfolio, at horizon.linkspreed.com.

If you would like more information, want to discuss the work, or have a suggestion for a research group, a dataset or a critique we should consider, write to us at [email protected]. Critical feedback is especially welcome. A project that tries to be honest about its odds should expect to be challenged.

References

Studies discussed in this article, in order of appearance:

Pais-Vieira M, Lebedev M, Kunicki C, Wang J, Nicolelis MAL (2013). A brain-to-brain interface for real-time sharing of sensorimotor information. Scientific Reports 3:1319.
Grau C, Ginhoux R, Riera A, et al. (2014). Conscious brain-to-brain communication in humans using non-invasive technologies. PLoS ONE 9(8):e105225. doi:10.1371/journal.pone.0105225
Rao RPN, Stocco A, Bryan M, et al. (2014). A direct brain-to-brain interface in humans. PLoS ONE 9(11):e111332. doi:10.1371/journal.pone.0111332
Stocco A, Prat CS, Losey DM, et al. (2015). Playing 20 Questions with the mind: collaborative problem solving by humans using a brain-to-brain interface. PLoS ONE 10(9):e0137303. doi:10.1371/journal.pone.0137303
Jiang L, Stocco A, Losey DM, Abernethy JA, Prat CS, Rao RPN (2019). BrainNet: a multi-person brain-to-brain interface for direct collaboration between brains. Scientific Reports 9:6115. doi:10.1038/s41598-019-41895-7
Wolpaw JR, Birbaumer N, McFarland DJ, Pfurtscheller G, Vaughan TM (2002). Brain-computer interfaces for communication and control. Clinical Neurophysiology 113:767-791.
Blankertz B, Sannelli C, Halder S, et al. (2010). Neurophysiological predictor of SMR-based BCI performance. NeuroImage 51:1303-1309.
Tang J, LeBel A, Jain S, Huth AG (2023). Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience 26:858-866. doi:10.1038/s41593-023-01304-9
Défossez A, Caucheteux C, Rapin J, Kabeli O, King J-R (2023). Decoding speech perception from non-invasive brain recordings. Nature Machine Intelligence. doi:10.1038/s42256-023-00714-5
Willett FR, Kunz EM, Fan C, et al. (2023). A high-performance speech neuroprosthesis. Nature 620:1031-1036. doi:10.1038/s41586-023-06377-x
Card NS, Wairagkar M, Iacobacci C, et al. (2024). An accurate and rapidly calibrating speech neuroprosthesis. New England Journal of Medicine 391:609-618. doi:10.1056/NEJMoa2314132

About LINKSPREED and HORIZON

LINKSPREED is building Web4, an internet where people and organisations own their data, identity and digital relationships. HORIZON is the exploratory research division at LINKSPREED’s Y-LAB. It studies high-risk scientific questions under an explicit honesty rule: every project states its real chance of success, and the results, including negative ones, are shared in short public briefs. Learn more at horizon.linkspreed.com. For questions, contact [email protected].