A Path to BCI for All
What technology do we need to bring human-AI collaboration to nine billion people?
What is a brain-computer interface, really? And what are they for?
For some, brain-computer interfaces are a medical tool, surgically implanted devices justified by their enormous benefit to patients locked behind paralysis and neurodegenerative disease. For others, brain-computer interfaces are an opportunity, a way to take advantage of increasingly cheap sensors to gather health data that consumers might want to use.
We have a third perspective. For us, BCI isn’t just a tool to help a comparatively small number of patients, or a way to tempt the public with cute applications of current technology. It’s a path to the future, to our place in the future, for all of us. It’s the most promising way forward, to break free of the limitations of our evolutionary past and become something more. To keep up with artificial intelligence, to collaborate with it, to make it a part of us. It’s something that we want to benefit everyone, not merely a few, and something we want to be transformative, not merely nifty.
That perspective comes with fundamentally different priorities. It means we’re looking for a technology that can develop like cell phones: something that may begin bulky, expensive, and impractical, but with development and iteration can grow cheaper and lighter. It means working on the fundamentals, finding key technologies, like modern semiconductor chips and batteries were for cell phones, that make ambitious goals radically more feasible. And it means building a device not for one application or another, but something that can cover most of the use-cases we can imagine, and uses we can’t yet begin to predict. In fifty years, cell phones went from limited toys for the wealthy to essential work tools for some of the poorest people in the world. As technology moves faster, we expect BCI to follow an even quicker path.
We don’t expect this to be easy. Taking these priorities seriously means pushing for a kind of holy grail, a combination of traits companies have dreamed of and none yet have managed. Over the last year, we’ve sifted through the approaches on offer, and thought through what we’ll need. The technical demands are stringent, we won’t deny that. But they’re worth pursuing. After all, the future is at stake.
Evading the Surgical Bottleneck
In science fiction, brain-computer interfaces tend to be implants. Cyberpunk novels are filled with hackers with ports on the side of their heads, ready to plug in to a 1980’s vision of the internet. Right now, the most prominent players in brain-computer interfaces are following suit. Neuralink inserts electrodes surgically, close to key neurons in patients’ brains, while Synchron uses blood vessels to gain entry in a less invasive approach.
Neuralink and Synchron are building medical devices, and in that context, this approach makes sense. While researchers are still discussing how best to measure the benefit of these technologies, it is clear that, for patients unable to communicate or navigate the world on their own, the technology could be life-transforming. With benefits like that on the table, both patients and doctors are more than willing to accept the risks and costs of surgical intervention.
Those risks, and costs, are much less reasonable for the healthy. Surgery takes expertise, even with surgical robots. It takes time to heal, even for the simplest procedures. And it takes a risk, even for less delicate operations than modifying a human brain. There are roughly 1.5 million plastic surgeries in the US each year, and a bit over half a million LASIK surgeries. It is hard to imagine any procedure involving brain surgery to be anywhere near as common, and hard to imagine even a lighter surgery for minimally implantable tech to cover the thousandfold gap to universal adoption.
At e184, we are working to preserve our voice in the future, and we view brain-computer interfaces as essential to that work. We need a solution that will not just be accessible to a few, but that has the potential to benefit everyone in the world. That means we cannot let ourselves be bottlenecked by surgical techniques. We need a non-implantable technology.
Our First Goal: A Mental Smartphone
We won’t achieve the symbiosis we aim for in one shot. On the other hand, we don’t want to get sidetracked. We want to build foundational technology, technology that doesn’t just address a current need but paves the way for what we aim to build in future. That means we need to make sure our system has the right capabilities to be that foundation.
What are those capabilities?
Right now, the most general AI systems available interact via text and images. As a baseline, then, we want an interface that lets a user, with a thought, communicate a text or an image. We want users to be able to think through a message word by word, and send that message as a prompt to an AI, or a message to another user. We want users to be able to imagine an image, guide an AI through fleshing it out, then print it or post it to social media. In short, we want a device much like a smartphone, operated by the user’s mind. If we can manage that, we will have a solid foundation to go farther, and meaningfully enhance human cognition.
Already, researchers have made progress towards some of these baseline capabilities. Using magnetoencephalography in a magnetically shielded room, a team at Meta have managed to predict what volunteers are typing based on their brain activity with reasonable accuracy. Invasive approaches have gone farther, and using brain implants researchers are able not only to measure direct attempts at speech, but even to pick up on speech that their patients only imagine. Decoding images is harder, but there has been steady progress in the last few decades, going from identifying small black and white grids to using diffusion models to generate appropriate images from fMRI and even EEG data. These methods focus on decoding images that are seen, not imagined. Reconstructing imagined images will be more challenging, but there has some progress in reproducing imagined images from fMRI data.
In contrast, research on interfacing in the “other direction” – that is, stimulating a user’s brain to convey speech or images – is still in its infancy. While there are researchers and companies that stimulate patients’ brains with brain-computer interfaces, they generally focus on specific medical outcomes, like treating epilepsy. At the moment, it is unclear how this side of a general-use brain-computer interface would work, and this is a question we postpone for the future. Initially, our mental smartphone will be one-way: a tool to send messages and images, not an additional way to receive them.
That still adds up to a substantial challenge, as we are well aware. It will mean reading signals from multiple regions of the brain, to decode both imagined speech and images: ideally, we want to get as close to whole cortical coverage as possible, to decode signals from the entire cerebral cortex. And it will mean doing so without using an implant, to an unprecedented resolution.
The Measure of a Brain
In the ideal case, we would want to detect the activity of individual neurons in the cerebral cortex. This would require distinguishing objects a fraction of a millimeter in size, corresponding to the length of dendrites of human pyramidal cells. If that proves infeasible, then we expect we will at least need to distinguish the activity of small groups of cells, on the scales that invasive techniques probe via implanted electrodes, meaning resolution of a few millimeters at worst. Currently, this is the scale that medical BCI systems have needed to probe in order to achieve communication rates on the same order as those of speech with a realistically large vocabulary and acceptable error rate. Non-surgical techniques have yet to achieve this threshold, and we suspect this is due in part to the absence of any technique that probes the same resolution as implanted BCI. If we can achieve that few-millimeter resolution, we have a chance.
In addition to achieving good spatial resolution, we will need to achieve good temporal resolution. Neuron voltage spikes can be spaced hundreds or even tens of milliseconds apart, with peaks often lasting only a few milliseconds. That means we will need to resolve and distinguish signals on those timescales. If we measured with a lower temporal resolution, we would not be able to see activity of individual neurons, and would be restricted to collective measurements.
How do existing non-implanted methods compare?
Some methods detect brain activity indirectly, via its impact on blood flow. This includes functional Magnetic Resonance Imaging (fMRI), which detects movement of oxygenated blood in the brain via inducing resonant effects in hydrogen nuclei in a strong magnetic field, functional Near-Infrared Spectroscopy (fNIRS), which distinguishes oxygenated blood in the outer brain through its absorption of near-infrared light, and functional Ultrasound (fUS) which uses the Doppler effect to measure the relative speed of blood to an ultrasound probe.
All of these methods rely on neurovascular coupling to track brain activity via its dependence on blood oxygen. The enormous magnets required by fMRI make it purely useful for medical and research applications. Portable devices need to be based on fNIRS or fUS, though both are still quite motion-sensitive, with fNIRS having additional sensitivity to scalp details due to only probing a fairly superficial layer and fUS faces challenges getting a high-resolution signal through the skull. More fundamentally, all three methods, to different degrees, suffer from a shared disadvantage: timing. Blood flow in the brain responds to signals on time scales of seconds, at least ten times too slow for the kind of real-time communication we are aiming for.
Electroencephalography, or EEG, instead detects electrical signals from voltage jumps in the cell membranes of groups of pyramidal cells (along with electrical signals from muscles in the head, eye blinks, and saccades, which must be disambiguated). This is a favored method for many non-implanted approaches at the moment, from devices that gather health data from a small number of electrodes, like ear-EEG, to measurements based on hundreds of electrodes across the scalp.
While EEG is well-suited to current approaches, we do not expect it to achieve the kind of precision we need. The voltage signals detected by EEG are transmitted nonuniformly through the brain and skull, leading to volume conduction effects that make them challenging to interpret with real precision. EEG is best at detecting radially-oriented currents towards the outside of the brain, demanding more powerful interpretive techniques to access the remaining two-thirds of cortical currents. Actually localizing signals often depends on models for the heads of individual patients, an added source of potential error. Even if signals for EEG could be perfectly disambiguated and traced, the technology has a more fundamental limitation: in order for a signal to be detectable at all on the scalp, it must be of sufficient strength. Potentials in the scalp act as a low-pass filter, restricting the frequencies that are able to pass through. This rules out detecting signals from individual neurons or small groups, instead demanding large synchronous patches of the cortex to fire. As such, even the most optimistic estimates expect a resolution of centimeters, likely too low to reach the accuracy we need to decode speech and mental imagery.
(It may be possible to improve on this resolution limit by coupling the brain’s electrical activity with induced ultrasound waves, creating a mixed signal that can pass through the skull and scalp. While this idea is intriguing, it is still too early to tell whether it can overcome EEG’s disadvantages to a sufficient extent.)
Finally, magnetoencephalography, or MEG, detects magnetic signals generated by electric currents within the brain. As currents from axon and synaptic sources cancel out, MEG only detects signals from dendritic currents in pyramidal neurons, which are conveniently exactly the brain activity that is most relevant for this kind of BCI. Unlike electrical signals, magnetic fields pass through the head almost entirely undistorted, probing much deeper and giving a much cleaner interpretation than EEG, and can localize sources to a 2-3 millimeter scale.
MEG does, however, have one enormous limitation: the Earth’s magnetic field. While brain signals lead to magnetic fields outside the scalp in the range of hundreds of femto-Tesla, the Earth’s magnetic field is around fifty micro-Tesla, a difference of eight orders of magnitude. In practice, this limits MEG to clinical and research environments with either passive or active magnetic shielding. Conventional MEG uses superconductors, which necessitates cryogenic cooling, leading to extremely bulky setups. It is also extraordinarily sensitive to motion, so a subject’s head must be kept as motionless as possible.
Optically pumped magnetometers have recently seen use for MEG (referred to as OPM-MEG), and their requirements are much lighter. They still require shielding, but less than conventional MEG, and are able to be used for moving subjects. They do have additional issues with cross-talk between sensors and field drift, and heat and power transmission makes it challenging to pack enough of them next to the head to achieve high resolution. Still, they represent a significant step towards our goals.
The Upshot
We are at a pivotal time in history. It matters enormously what we choose to build. If we are to enhance human intelligence to catch up, and collaborate, with machine intelligence, we need brain-computer interfaces that can be accessible to all. That will mean non-surgical technology, and it will mean technology powerful enough, with high spatial and temporal resolution, to pick up what we want to communicate.
That in turn, means that none of the technologies available today will suffice. Hemodynamic methods are simply too slow, and EEG has too low resolution, while current MEG cannot be used outside of magnetically shielded rooms.
However, while the limitations of hemodynamic and electrical measurements are biological, the limitations of MEG are a matter of engineering. Already, OPM-MEG represents a step in the right direction. A MEG sensor that can do what we need of it, achieving high resolution in everyday life, does not contradict any physical law. It simply does not exist…yet.
Thus, a brain-computer interface company that aims for true human-AI symbiosis…must first be a sensor company.
In an upcoming post, we’ll explain what that means.
We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application: