AJ Keller

The through-line

Everything AJ Keller says traces back to a question he heard his father ask over a keyboard: "why does the computer not know that I want to save when I'm done typing" [1]. That became a lifelong preoccupation with the asymmetry between humans and machines, and the conviction that "computers are able to learn from humans faster than humans are able to learn from other humans" is the thing worth building for the next twenty-five or thirty years [1]. The ambition is bandwidth. Two people in conversation exchange a great deal of meaning, and he wants a channel that beats it: "imagine the world where I could exchange even more information about myself or my desires with a computer uh almost instantaneously at the speed of thought" [1]. Neurosity is described as the vehicle for that, a company he wants to make "the most trusted and loved neurotechnology company in the world" [1], building computers controlled by thought [3].

The shift over time is in what he thinks the near-term product actually is. Early on he framed brain-computer interfaces as an instantaneous translation machine between brain and computer, the equivalent of speaking English to a friend who relays it in French [1]. What shipped instead is a measurement instrument for mental states, "essentially a speedometer for your brain" that can tell "when you're over speeding or when you're under speeding" and help you get up to speed and hold it [1][2]. He also concedes the field is early in a specific way: the phenomena are real and the physics allows it, but the engineering has not caught up. "the things that we're trying to do are physically possible just technologically not feasible yet," and his job is to close that gap through hardware innovation [2]. He compares the moment to the Wright brothers, since air always existed and what changed was the understanding of the technology [1].

On why the hardware had to come first

He tried the software-only path and says plainly that he failed at it: he could not "produce um reliable enough software that was useful on a daily basis for uh consumers for the mass Market" because the underlying hardware did not exist [2]. That failure produced the company's founding logic, which he states as a rule rather than an ambition: "the future of software in this space is led by great hardware," and therefore "neurosity exists because it has to" [2]. He left Boeing at the end of 2015 to work on this full time, was building intellectual property for the headset by 2017, and the Crown is the third generation of hardware [1][2].

Boeing supplied both the trigger and the discipline. Watching robots assemble three hundred million dollar airplanes convinced him the future is robots working alongside humans, which demands ever higher bandwidth between the factory worker and the machine: a single "mental tap" registered as an error potential in milliseconds could stop a collision faster than a person could yell stop or reach a button, saving hundreds of millions in damage [1]. The company also taught him "how you can put quality over everything you can put quality over deadlines you know and it's okay to push back the deadline if it means the quality is going to be uh higher" [2]. What it did not give him was meaning; measuring robot-drilled holes by hand to check for drill bit wear was not work he found significant enough, and he was too junior to have built a plane he later flew on [2].

On designing for people who have to want to wear it

Keller is blunt that most brain-computer interfaces look bad for structural reasons: to fit every head shape and hairstyle and get sensors through hair, the best systems articulate each electrode and "it ends up looking like spiders" [1]. His counter-position is that adoption is a design problem before it is a technical one. "I often will say this is a 51 branding 49 tech company" [1]. The scaling test he uses is conversational: it is far easier when people say "can I try it" than when the company has to ask "will you let me put this on your head" [1]. He refuses headbands and anything crossing the face, because "the face is like uh the most beautiful part of any human" and obscuring it without adding something is "a big No-No" in his design philosophy [1]. The goal is that wearers "feel empowered that they're being crowned" [1], and that the device amplifies rather than subtracts from personal brand, something you would be proud to wear at the office or leave on your desk as part of your identity while programming [2].

The form came from constraining the problem and then letting nature answer it. The requirement was whole-head scanning covering the motor cortex, which runs ear to ear, so sensors had to pass through the hair on top of the head [1]. His co-founder Alex Castillo, who has a design degree from Parsons University, found the answer at a steakhouse dinner with his girlfriend's parents, looking at mounted trophy heads and seeing how a ram's horns brace and wrap the skull the way rams wrap each other's heads when fighting [1][2]. There was no alternative to Alex doing it: "there is no one else to come up with the design it's just you and me were you know we're just co-founders like we couldn't have any money to pay anybody" [1]. For newer work, sources of inspiration have widened, including typing prompts into DALL·E and DALL·E 2 to see what comes back, and deliberately not looking at competitors so as to lean on instinct and user feedback instead of getting distracted by what already exists [2].

The third generation also changed how he holds his own judgment. He describes learning "that it's okay for me to be extremely detail-oriented when I'm designing a device," that when intuition says a part might break in production later he can push until it is fixed, and that trusting that instinct took a while: "I do have good intuition I have listened to my customers I do know what we want I do know what will break" [1]. Everything counts, from how someone lifts the device out of the box to the pamphlet inside it, and the aesthetic references are Soho streetwear and the way Tiffany presents its boxes, down to gold-plated metal on the Crown [1].

On putting the computer on your head

The technical decision he considers most consequential is moving the compute onto the device. Laboratory BCIs are big nets wired to machines inside shielded rooms, which is excellent science under very specific circumstances [1]. Neurosity packaged the scanner, the translator and the computer into something that sits on the back of the head and does pattern recognition locally, "a majority of the compute on the edge," producing a stable system you can actually build applications against [1]. He reaches for the obvious analogy: it is the same move as Oculus going from tethered to untethered by putting enough compute on the headset [1].

Software came before silicon in the literal sense. "we wrote the SDK and the API for this before we ever made the hardware," specifying a brain-computer interface where a few lines of code get you securely authenticated data, with the heavy processing done on the device so clients do not have to write it [2]. That specification then drove the internal requirements of the first hardware [2]. The payoff is JavaScript-based, works across React Native and the web, and has no native drivers or dependencies to compile against, which turns weeks of integration work into minutes and lets the programmer spend their time on user value instead [2]. The underlying objective is speed: an ecosystem that lets the company "move at the speed of software" while discovering which applications actually deliver value [2]. Keller notes he was the first person to build a JavaScript SDK for a brain-computer interface, and that this was how he met Alex, who came from GrubHub and understood the web as a democratizing technology from growing up in the Dominican Republic with internet access but no console games, a conviction the two of them have always shared [2].

On flow, and treating the brain as a state machine

The Crown "measures EEG and essentially creates like a meter for concentration," a speedometer for getting up to thinking speed the way a car speedometer helps you reach highway speed [2]. The signal is actionable: a particular brain wave is either increasing, decreasing or sustaining, and "if you're not able to increase this particular brain wave you're going to have a really hard time completing that task and getting into flow" [2]. The shipped ritual is five minutes of custom music with your eyes closed before you start work [1]. Keller treats the brain the way a game developer treats states, moving users from unfocused to focused, with the finding that the meditative state is the necessary way station: "if you can't be in a meditative State then you're probably not going to be able to shift into that focused state" [1].

What makes it work is the loop rather than the sound. Stimulus alone is "only about 80 percent effective"; the system needs closed-loop feedback, giving the brain a stimulus, reading the reaction, and doing it again [1]. A reinforcement learning algorithm takes brain data as input, predicts what will drive focus up, and is rewarded when activity increases and punished when it drops, getting more accurate with every session [1]. He compares it to a parent learning to soothe a baby: you figure it out, then you get good at it, and the answer changes day to day, twinkle twinkle little star today and itsy bitsy spider tomorrow [1]. On Cal Newport's model of deep work as five rooms you pass through in sequence, he says the headset gets you to room five faster because "we know what your brain looks like when you're in five," and when your brain is not producing that wave, the curated music causes it to "start to organize itself into that room room level five," with the embodied feeling of flow arriving during that organizing process [2].

He is careful about what generalizes. What maintains your flow state is personal, whether that is lo-fi, movie soundtracks or silence, and most people already know their own answer; the most common finding is that "what works for you doesn't work for me" [2]. He prefers house electronic around 128 beats per minute and some cinematic music, while Alex leans to post rock, alternative rock and metal [2]. What appears universal is the organizing step itself, because brains naturally synchronize to an outside stimulus, and the trick is to have them synchronize to something chosen rather than something random [2]. A project with UC Berkeley students is testing whether a twenty second scan can suggest the music that shifts a given person into flow [2].

On mental health before medication

The therapeutic ambition is explicit: products that "start helping mental health without having to really jump first to medication," digital therapeutics rather than a first prescription [1]. He speaks from his own position, saying he is on Vyvanse at the time of the conversation and would love a future without dependence on stimulants to be productive, that he has not been able to drop it yet, but that when he runs out he can lean on the headset more [2]. He connects this to ADHD in children and students building better habits without drugs, and calls it the reason he is fighting for this [2]. Among the use cases he finds most compelling is a preventative mechanism for rumination, helping someone stuck on a ruminating thought push into a mental state they would rather be in, ideally covered by insurance or an HSA [1].

He divides the field cleanly. Invasive interfaces are for control, bypass and stimulation: routing signals around a spinal cord injury to restore walking, the Neuralink category, or eye prosthetics of the kind a company called science is working on, and these are the true instantaneous communication machines [1]. Non-invasive work lives at the level of major mental states, attention and adherence: how meditative or focused someone is, signs of depression, anxiety and PTSD, and Alzheimer's progression, all of which can be read from outside the head without going inside [1].

On AI as the natural language of brain data

His starting premise is that "the brain is AI," which makes machine learning the obvious tool for reverse engineering the electrical activity coming out of the head [1]. Present use is the reinforcement learning loop that learns more about each user with every session [1]. Beyond that he sees three generative directions. The first is data augmentation: with roughly ten thousand opted-in research datasets, train a network to predict the next EEG sample given the last minute of data and use it to expand the corpus for the generalized model [1]. The second is music composed in real time rather than drawn from the current catalog of neuro musicians, generating the exact genre and instrumentation predicted to move a brain between states as fast as possible [1]. The third is analysis. After uploading sixteen of his own EEG datasets to the advanced data analytics component of ChatGPT-4, he found it doing work "way more hardcore than I could using state-of-the-art python toolboxes," even writing an email that pulled a key insight from the latest dataset by comparing it against the previous fifteen, to the surprise of his own neuroscientist [1]. He sees a paying monthly product in it, a "Fleet of neuroscientists" continuously reviewing a user's data for insights, on an opt-in basis [1].

On staying consumer, and letting developers do the marketing

Neurosity tried B2B early and abandoned it for two reasons: sales lead times were too long, and "you end up creating a device that's designed to meet the requirements of the purchaser the business instead of the consumer" [2]. He is categorical about the orientation, "definitely always consumer focused never B2B," on the grounds that people are the ones using it in the end [2]. The word B2B has been replaced by third party. Third-party developers build on the platform, but their users still onboard through the Neurosity app, maintain the device there and shift into focus there, then use OAuth to authorize the third-party application to access data from the Crown [2].

That structure doubles as distribution. Early marketing targeted people who wanted to build BCI applications, people like Alex and himself, and only later did the products become useful to non-programmers [2]. He credits builders such as a lean and farza, who made things with the headset and posted about them, and says the company owes a lot of its start to them [2], a relationship with the initial customer base that has opened up further possibilities around productivity and thought control [3]. On the commercial side, the Crown ships for ninety-nine dollars a month for as long as you want it, the company has passed two million dollars in total revenue, and demand has outstripped supply for three years, which is why he was in Toronto working with the contract manufacturer to speed up the assembly line [1][2].

Takeaways

• The software-only route failed first, and that failure became the thesis: "the future of software in this space is led by great hardware," so "neurosity exists because it has to" [2].
• Put the compute on the device. Local pattern recognition on the edge is what turns lab equipment into a stable platform developers can build on, the same unlock as Oculus going untethered [1].
• Write the SDK before the hardware. Specifying a few-lines-of-code JavaScript API with no native dependencies set the internal hardware requirements and cut integration from weeks to minutes [2].
• Adoption is a branding problem: "this is a 51 branding 49 tech company," and the test is whether people ask to try it rather than being asked to wear it [1].
• Stimulus alone is roughly 80 percent effective; the closed loop is what matters, with reinforcement learning rewarded when brain activity rises and punished when it falls [1].
• Flow has a personal layer and a universal one: the music that maintains your state is idiosyncratic, but brains universally synchronize to an outside stimulus, so choose the stimulus deliberately [2].
• Avoid B2B if you want a consumer product, because building to a purchaser's requirements produces a device designed for the buyer rather than the user [2].
• Non-invasive BCI reads mental states from outside the head, including focus, meditation, depression, anxiety, PTSD and Alzheimer's progression, while invasive BCI is the category for direct control, bypass and stimulation [1].