Starfish’s edge is in its bet on distributed electrodes. It’s the opposite of the classic electrode long used in most academic labs: Blackrock’s Utah array. This implant records high-fidelity neural signals from its location, but it is limited to one cortex or section of the brain. It’s also bulkier and harder than newer electrodes like those of Neuralink and Axoft. The Utah array is perfect for an academic lab aiming to study one cortex of the brain in a few patients for a fixed period of a year or so. But for a consumer audience and lifespan of decades, a network of small, low-power electrodes across the brain might make more sense.

It’s worth noting that this idea is far from unique to Starfish. Dr. Maharbiz’s lab at Berkeley is making neural dust, and Dr. Etienne-Cummings and Thakor at Johns Hopkins are pioneering the microbead. Starfish itself is partnering with imec, a Belgian nanotechnology company that makes, among many other things, small probes for neurotechnology.

Let’s look at their stats compared to other options. These are from their blog.

• Low power: 1.1 mW total power consumption during normal recording.

  • Utah array: ~30 mW (source)

  • Neural dust: 7.2 mW (source)

• Physically small: 2 x 4mm (0.3mm pitch BGA).

  • Utah array: 4 x 4 mm (source)

  • Neural dust: 3 mm x 1 mm (source)

• Capable of both recording (spikes and LFP) & stimulation (biphasic pulses).

  • I believe most invasive electrodes can do this. We went over LFPs (local field potentials) in this post.

• 32 electrode sites, 16 simultaneous recording channels at 18.75kHz.

  • Utah array: Customizable, often 96 electrodes (source)

  • Neural dust: one pair of electrodes per mote (source)

Starfish has written that they expected their first chips to be ready by late 2025, but it’s currently March 2026 and I haven’t heard much from them. They also aren’t hiring to my knowledge. I hope Gabe and the neurotech community doesn’t give up on small distributed electrodes.

From their website, Merge outlines a rather unique approach to BCI. Instead of relying on electrodes implanted in the brain like most labs and companies, Merge wants to “connect with neurons using molecules instead of electrodes,” which refers ultrasound and other non-invasive imaging methods.

Pay attention to the signatories of this post:

Tyson Aflalo

Sam Altman

Alex Blania

Sandro Herbig

Sumner Norman

Mikhail Shapiro

Aside from Sam, we have leaders from two other companies. Tyson and Sumner are chief executives at Forest Neurotech, an ultrasound BCI company in San Francisco. That clues you into Merge’s non-invasive focus. More interestingly, Alex and Sandro come from World, which aims to be a blockchain-based global community providing digital identities and financial services.

From Merge’s leadership and announcement, which “demands that we think in decades rather than years,” we are seeing a totally different approach compared to a company like Phantom Neuro, which uses a more limited technology technology to solve a specific problem. In their case, they use EMG to allow amputees to control prosthetic arms. Merge clearly has a much longer time horizon and more ambitious vision. I’m excited to see what they do next.

As an interesting aside, one of the main investors in Merge is Gabe Newell, the cofounder and president of the gaming company Valve. He also dropped out of Harvard to build some of the early versions of Windows and dabbles in oceanography. Merge is also not his first intersection with BCI; Gabe also founded the company Starfish Neuroscience in 2019 to produce implantable chips. They were set to release their first product in 2025, but I haven’t heard much from them. I’m glad to see Gabe is still pushing the industry forward.

Over the weekend of October 5, g.tech is hosting their BR41N.IO hackathon. The morning is filled with presentations and demos of company projects. In the afternoon, teams start brainstorming and building using motor imagery, P300, SSVEP, and ECoG datasets. Teams must finish by around noon on the 6th for presentations and awards.

I appreciate g.tech for making it as accessible as possible, since participation is free, remote-friendly, and open to anyone. g.tech also holds a (free I think) BCI & Neurotechnology course in the spring. What an invaluable contributor to BCI education!

They initially plan to treat neurodegenerative diseases, but will eventually use their BCI to control devices and prosthetics. We explored micrometer-sized electrodes in our last few posts. This appears to be the first commercial version.

Here is their roadmap:

The founder and CEO, Tetiana, used to lead Neiry, an EEG company in Russia. Recording nanoparticles are no doubt more exciting and transcendent. If this succeeds, it should be a wake-up call to pay more attention to non-Western BCI labs and innovators.

A BCI like this would eliminate most of the downsides of invasive electrodes without sacrificing the recording and stimulating resolution. It would be by far the easiest BCI to use, not requiring surgery or an annoying EEG headset digging into your scalp.

Read more on their website.

To understand how Elemind works, we need to understand “communication through coherence.” This is a concept by neuroscientist Pascal Fries that says that neurons that oscillate in phase with other neurons transmit information more easily with each other. When I say a neuron oscillates, I mean that it fires off spikes in voltage at a certain frequency.

Elemind’s goals, outlined in their white paper, are simply to estimate the phase of these neuronal networks in real time and stimulate them based on that phase.

The first part of that involves their proprietary algorithm: the end-corrected Hilbert Transform (ecHT). Estimating the phase of neural oscillations is difficult, because they are so dynamic. This algorithm is a step up from previous techniques in how fast and efficient it is.

Both Elemind and Somnee are closed-loop, meaning that they continuously record neural signals in the form of EEG and use that to tailor their stimulation to the patient. The key difference between Elemind and Somnee is that Somnee measures the strongest frequency in the alpha and theta bands and stimulates according to that. Elemind also looks at their data in the frequency domain, but they calculate the phase, or the precise timing of the wave. That is necessary to stimulate in a way that is in phase or out of phase.

In the case of Elemind, they use their headband to deliver auditory stimulation out of phase with the alpha waves (7.5-12.5 Hz) for thirty minutes. Why alpha waves? There is an association between them and arousal before and during sleep. I haven’t seen anyone use auditory stimulation to affect neural signals in this way before. I wouldn’t expect that listening to a sound of a certain frequency and phase would reproduce that frequency and phase in someone’s neuronal networks.

So how well does it actually work?

According to their study, the stimulation decreased sleep onset latency (SOL) by about 10 minutes, with a rather larger variance. These charts from Figure 5 show the difference between the Sham group (EEG recording with no stimulation) and the Stim group (EEG recording with tailored stimulation).

They decreased the average SOL from 36 minutes for Sham to 25 minutes for Stim, a decrease of 31%. In Somnee’s trial, they decreased the sleep onset time from 22 minutes on average to 16 minutes, a 27% decrease. However, the 22 minute baseline for the Somnee trial is not from a sham group, but a group with open-loop non-personalized stimulation (I couldn’t find the sleep onset time for their sham group). You would expect the open-loop group to sleep at least slightly better than their sham group, which they did in terms of sleep duration, so we can hypothesize that the 27% decrease in onset time is an underestimation if you compare it to the sham group.

Which one is cheaper? Somnee’s headband is $429 with a 45-day trial, while Elemind’s is $349 with a 30-day trial plus a subscription fee for the app. I don’t see why they can’t apply both auditory and electrical stimulation to make people fall asleep even faster. It’ll be like X.com and Confinity merging to create PayPal.

That’s where Somnee comes in. This Berkeley company makes the Smart Headband to help you fall asleep faster and sleep more efficiently. It includes an electroencephalography (EEG) sensor to record your brain state at bedtime and an electrode for transcranial electrical stimulation (tES), which will stimulate the brain for 15 minutes to induce a more restful state. Because the precise way your brain is stimulated depends on how the headband interprets its signals, this makes it one of the few closed-loop BCIs on the market.

As outlined in their research paper, the Berkeley researchers combine two technologies in a novel way. tES (more specifically, transcranial alternating current stimulation or tACS) is a non-invasive way of stimulating the brain with no harmful side effects. Previous studies have found that stimulating the frontal cortex at 5 Hz can make patients drowsy. EEG, which records a jumble of neuronal activity, can provide biomarkers for sleep quality. Everyone has slightly different baseline activity levels recorded by EEG, so they tried both a fixed stimulatory pattern and one tailored for each patient based on their EEG.

EEG signals are typically transformed from time-domain into frequency domain using a Fourier Transform. So, instead of looking at the signals over time, you see how strong the signal is at 1 Hz, 5 Hz, 10 Hz, 60 Hz, etc. Neuroscientists usually separate the frequencies into bands that act as biomarkers, like the alpha band (8 - 13 Hz) and theta band (4 - 8 Hz).

Here is an example of a frequency-domain graph with the bands identified, from this paper.

For the fixed tACS stimulation, the researchers applied two in-phase 10 Hz and 5 Hz sinusoidal waves to target the alpha and theta bands respectively. They chose them because alpha band activity (8–13 Hz) is linked to stage I sleep and theta band activity (4–8 Hz) is linked to the transition to stage II sleep, according to another study. The amplitude of the waves was 0.6 mA peak-to-peak.

For the personalized tACS stimulation, the researchers took EEG samples of patients in a relaxed, closed-eyed state (similar to someone laying down to sleep), and they identified the peak closest to 5 Hz within 4-6 Hz and the peak closest to 10 Hz within 9-11 Hz. They then applied similar sinusoidal waves as the fixed stimulation, just with the slightly-changed frequencies for the two sine waves. They started in phase, but didn’t remain that way unless the 5 Hz-adjacent wave was half the frequency of the 10 Hz-adjacent one.

They also tested a sham session without stimulation and measured sleep quality with a Fitbit. For either one, you don’t actually have to wear the headband to sleep. You can just wear it for the 15-minute stimulation and take it off to sleep. If it works, you would expect the personalized stimulation to outperform the fixed one, which would outperform the sham. If it doesn’t work, would I be writing this blog post?

Let’s check the results.

P1 refers to fixed stimulation, and P2 refers to tailored stimulation. I wish they would change their graphs from the Excel default, but other than that, they show that people who use the personalized stimulation fall asleep faster (B) and sleep longer (A) than people who even use the fixed stimulation. Why don’t they have the sleep onset bar for the sham? My guess it that the number is not comparable, since those people don’t stay up for 15 minutes receiving the stimulation.

I’m surprised that the P1 patients slept for essentially as long as the sham ones. The only difference between P1 and P2 is at most a 1 Hz change in the frequency of both of the sinusoids. Maybe the fact that the fixed stimulation waves are harmonics of each other masks their effect somehow, so if they applied 4.9 Hz and 10.1 Hz instead of 5 Hz and 10 Hz it would remove that problem. That’s just pure speculation, though. This paper points to other papers that show open-loop (non-personalized) stimulation can improve sleep, but this paper seems to refute that idea. The other papers use either transcranial magnetic stimulation (TMS), auditory stimulation, or transcranial direct current stimulation (tDCS) during sleep, while this paper uses alternating current tACS before sleep. Maybe there’s something about tACS or applying it before sleep that doesn’t improve sleep quality, or maybe those papers are measuring sleep quality using a metric other than duration.

The authors actually wonder the same thing in their conclusion, and they have the same idea about the harmonic nature of their fixed stimulation. They should just hire me at this point. They also point to their small sample size.

Either way, let’s not let this mystery and the default Excel graphs take away from the result: that there is a way to apply a current to your brain for 15 minutes that helps you sleep longer. Imagine if one day, we can have implanted electrodes that can stimulate the right neurons to trigger your brain to sleep at the push of a button. Or even ones that can inhibit it. Is there a way to apply tACS in a way that makes you less drowsy? You can use that to lock in at critical moments, like driving home after a long night or staying awake for an 8 AM meeting. I can’t wait for the future.

The Phantom X shares little in common with the implanted motor cortex arrays of many clinical studies, like this one. They use an implantable EMG (electromyography) sensor around the patient’s forearm to read the signals emitted by muscle contractions. The device then decodes the signals and use them to control a prosthetic hand or leg.

The band looks like this.

And this image shows how it is implanted and used.

Even more impressive, in order to test the device before all the necessary safety issues are worked out, Phantom ran their first study without implanting the band. They simply wrapped the 32-electrode array over the skin of the forearm, almost like how you would wear a watch. So whatever the results of the study show, you would expect an implanted band to record higher quality signals and produce better results.

From there, the signals are passed through an amplifier and bandpass filter of 5 Hz-4 kHz. Finally, they are transmitted wirelessly to a computer running a model to translate them into movements. That involves filtering out noise in the data coming from the patient moving by filtering out signals less than 200 Hz and then extracting features from the time-domain and frequency-domain data. They trained the model using able-bodied subjects instructed to perform various gestures with their hands such as a thumbs-up, a peace sign, a fist, and pointing.

The Austin team measured their success in three different ways. They first graded each gesture as an A, B, or C based on how easy the patient could reproduce the activity. They then calculated a Gesture Accuracy (GA) score as the percent of gestures that received an A. They also calculated a Modified Gesture Accuracy (MGA) like a GPA, where the percentages of gestures that received B and C were added to the GA after being weighted by 0.75 and 0.5. The MGA will always be greater than the GA, because the GA only includes gestures that received an A. Finally, they calculated the Classifier Gesture Accuracy (CGA) as the percent of gestures that the machine learning model correctly identified. The GA, MGA, and CGA were all highly correlated anyway.

So how did they do?

This graph clearly depicts very high accuracies for all the tasks they tested, except for index point and index tap (bending your index finger down to touch the palm). The patients also experienced fairly low latencies of less than 300 ms, which this histogram shows.

While the data suffers from some limitations, such as a controlled office environment and small sample size (n = 11), it is still extremely promising. This study leaves a lot more accuracy on the table that can be achieved once the device is implanted, trained on more subjects, and further optimized.

You can find their study here.

The conclusion I draw from this study is that Phantom could find success in focusing on a non-invasive band altogether. It would avoid a lot of safety risks and complications that would crop up when developing an implanted band. On the other hand, a non-invasive band has to be taken off and put on, is vulnerable to movement and signal degradation, and might not be good enough for extremely fine motor skills like playing a musical instrument or typing. I’m eager to see what the Phantom X can do when it reaches its full potential.

San Francisco’s own Prophetic is making it possible with BCI. Their aim is to use non-invasive ultrasound to activate the dorsolateral prefrontal cortex (DLPFC). This is the part of the brain that is responsible for higher-order cognitive functions like planning, decision making, and emotional regulation. This is the target.

The modality is ultrasound, specifically transcranial focused ultrasound stimulation (tFUS). This involves sending high-frequency sound wave waves on the order of a megahertz, or a million oscillations per second. This is much higher pitched than anything we or even our pets can hear, but it does cause small vibrations at the target. So unlike other BCI stimulation, tFUS doesn’t use electricity or magnetism. It uses physical, mechanical vibrations and small movements in the brain. tFUS is also one of the only neuromodulation techniques that can stimulate deep in the brain with high spatial precision, all without requiring an implant.

How do vibrations inhibit or stimulate neurons? The exact mechanism is unclear. It is possible that there are ion channels in the neurons that can open or close in response to mechanical force. It’s also possible that the movements can change the capacitance or voltage across a neuron’s membrane, which then causes ion channels to open or close. Or something else entirely.

Ultrasound is typically delivered in pulses that last about a millisecond, and with hundreds of these pulses per second. Higher frequencies stimulate the neurons and make them more likely to fire, while lower frequencies interestingly enough do the opposite, inhibiting that part of the brain.

You can also play with the frequency of the actual sound. Higher frequencies let you target parts of the brain more precisely, but they attenuate through the skull more, which makes it harder to penetrate the brain. Conversely, lower frequencies are less precise but maintain their strength as they reach deeper parts of the brain.

We will explore tFUS more in depth in a paper spotlight.

When I was putting together focused ultrasound and the prefrontal cortex, the first idea that came to mind was a tool to make you smarter or more mature. In theory, using tFUS to stimulate that part of the brain could make you better at planning, decision-making, and emotional regulation. Maybe you come up with more creative ideas or foolproof plans. This would be great to have on the job.

My second idea was the opposite. Maybe you leave work after a long day of serious thinking and you want to unwind. You go to a bar to have a drink or two to loosen up. But why spend the money and risk a DUI or a hangover? Instead, you could flip on your handy tFUS device at a lower pulse frequency, weakening the sober prefrontal cortex. Seriously, who’s building this???

The founders of Prophetic are certainly much more creative than I. They noticed that the DLPFC is dormant during dreaming, unless it’s lucid dreaming. I’m not sure how that was discovered, but it makes sense. You can probably guess where this is going—Prophetic is developing the Halo to stimulate your DLPFC while you dream to allow you to control your somnic destiny. You can place a refundable $100 deposit on the headband, which is estimated to cost $2,000. If it works, not a bad price to pay to live the life of your dreams for a third of the time (or subjectively more if your perception of time is distorted). Hopefully you don’t go crazy from waking up to your alarm after that every night.

In most clinical BCI trials, researchers build a narrow program from scratch to collect the data they want or perform the task they want. If they want to explore cursor control, they build a program that lets the user move a cursor over some targets and not much else.

Naturally, Apple sees an opportunity to dominate a field of many disjoint competitors with a smooth, unified ecosystem. By partnering with Synchron to build a pipeline for controlling consumer devices, Apple is bringing neurotechnology out of the clinic and into the home. Many people severely limit their thinking of neurotechnology as purely a clinical invention, whether due to ethics, marketing, or a simple lack of imagination. But the way I see it, limiting BCI just to paralyzed people is like limiting airplanes to people who can’t drive. It’s like only allowing people who can’t write by hand to use keyboards. Sure, disabled and paralyzed people have the most to gain from BCI, but if we stop there, we would miss out on a world-changing invention like the airplane or the keyboard.

Altman is joining the BCI race just as it heats up. Neuralink has implanted at least seven patients with its thread electrode. Other companies are progressing rapidly, too. The FDA has cleared Precision Neuroscience’s interface, Paradromics has launched clinical trials, and Synchron is collaborating with Apple to integrate into their ecosystem.

Merge Labs joins a cadre of promising startups, including Axoft, Motif, and Neurosoft. Perhaps Merge Labs has an advantage by being motivated by personal competition.

Sam Altman and Alex Bania

There are a few differences between in silico and in vitro. An artificial neural network has defined input and output neurons, and information travels in one direction. A petri dish of neurons is pure chaos, at least initially. These researchers segmented the chip into two halves, one for input and one for output. They placed electrodes to read neuronal spikes from the output half. High spike activity on one side moves the paddle up, and spikes on the other side move the paddle down. On the input side, they applied electric current in specific patterns according to the ball’s location and paddle’s location. These currents stimulated the neurons to fire.

Unless you somehow train the neurons to play, the signals you read from the output neurons will be pure nonsense. With an artificial network, you can manually adjust the weights of each connection based on their derivative with respect to an error function. This strategy guarantees that each weight converges onto a value that minimizes the error function, at least to a local minimum. You can’t do that with actual cells. How would a neuron know that it should try to return the ball?

Herein lies the genius of Dr. Brett Kagan and his Cortical Labs team. They operated under the theory of active inference via the free energy principle. In simple terms, they believe the plate of neurons seeks to minimize its variational free energy by minimizing its surprise, or the difference between its predictions and observations. Theoretically, the neurons should learn to fire in a way that keeps their input predictable. The researchers provided feedback using electrical stimulation: a predictable, uniform pattern for good behavior, like when the neurons intercepted the ball. Conversely, they shocked them with unpredictable, random stimulations for bad behavior, like missing the ball. In that way, teaching a palm-sized plate of neurons does not seem too different from training a dog.

And just like a dog, they learned.

This chart demonstrates an improvement in average rally length (hit-to-miss ratio) in the mouse cortical cells (MCCs) and human cortical cells (HCCs) after five minutes of training. In contrast, the media-only control (CTL), in silico control (IS)—which moved the paddle randomly—and rest session control (RST)—where neurons received no sensory input—all failed to improve as expected.

That a plate of neurons can be taught to behave a certain way is nothing short of profound. If one small plate can learn so quickly with such basic commands, imagine what a large-scale, long-term, optimized ensemble of neurons can do. It’s not hard to conceive of a world where OpenAI and Microsoft, instead of building colossal server farms and data centers, construct three-dimensional scaffolds for neurons to grow on, with gallons of media pumped in to feed the developing brain.

Cortical Labs in Australia is already selling the personal computer of neurotechnology: the CL1. It’s an enclosed environment to grow neurons over a silicon chip that can stimulate and record to your liking. I can’t wait to see how people make use of this.

If you want to dip your toes into the life of one of these researchers, you can take a swing at this year’s Brain-to-Text challenge. UC Davis has provided a dataset of 10,948 words to analyze and their current strategy, which achieves only a 6.70% word error rate.

Competitors have until December 31, 2025 to submit a solution. You could win up to $5,000 from Blackrock Neurotech. Read more here, and good luck.