55% of People Have Had a Lucid Dream. Prophetic AI Wants the Rest to Have One Every Night

Nathan Labenz introduces this episode as a crossover. Prophetic AI had come up on Cognitive Revolution in a conversation with Dean W. Ball about brain-computer interfaces, and Labenz already had the company on his invite list. Ate-a-Pi, host of the new podcast Emergent Behavior — built in the Czech Republic by Infinite.cz, with an AI-generated voice — had already sat down with the co-founders. The conversation that follows is theirs. Eric Wollberg and Wesley Berry are building Halo, a consumer headband that reads brain activity and stimulates specific regions. The first product is nightly lucid dreaming. The ambition, they keep saying, is a company for consciousness experiences, not a dreaming gadget.

Wollberg’s capsule version: two core technologies, transcranial focused ultrasound as the stimulation modality and neural transformer architectures tuned for neural data. The model is trained on simultaneous EEG-fMRI. It outputs steering instructions — targets — for the transducers. EEG on the headband then does reinforcement learning to improve the model over time. When they showed the first model, Morpheus 1, a couple of weeks before the taping, Wollberg closed by insisting they should be known for taking any discrete, universal brain state focused in the prefrontal cortex and putting it on the same piece of hardware. Halo is the prototype.

Ate-a-Pi’s own experiments were hypnagogic: lying down, almost asleep, trying to move objects behind closed eyelids, then waking unsure it happened. A night with Halo is more boring at the front end. You pair the band with an app, calibrate once, and go to sleep. It will not knock you out if you drank too much coffee. Consumer neurotech, Wollberg jokes, has been stuck in a Muse-and-Mendi neural-feedback paradigm for a decade — an EEG band plus content, asking you to work. His analogy is Ozempic flying off shelves while gym memberships gather dust. Halo is supposed to do the work. You just sleep.

The band watches for the EEG signature of REM. That prompt wakes the transformer, the transducers turn on, and the first spatial instructions go out. From then on the model is hunting a gamma-frequency spike during REM, the signature of lucidity. Wollberg narrates it as a loop: is she lucid, is she lucid, great, keep her lucid. When REM ends, stimulation stops and you continue the night. A REM cycle is generally about 20 minutes, so that is the ceiling of a session, though *Inception* got time dilation right — 10 or 20 lucid minutes feel longer. Berry’s definition of success is a self-sustaining sequence: prefrontal cortex active without assistance.

Do all people have lucid states in REM? About 55% self-report at least one lucid dream in their life. A smaller share has a natural proclivity. Wollberg is in that second group. He had his first lucid dream at about 12, woke equally terrified and excited, found early internet forums, and later the work of Stephen LaBerge, whose 1980 Stanford PhD kicked off the modern study of the state. LaBerge and others built training methods — reality checking, mnemonic induction — so people without the knack can still improve. Wollberg is 30. He has been doing this for 18 years.

The neural correlate is simple: prefrontal-cortex activation during REM. REM is when you dream. The prefrontal cortex is where conscious awareness, working memory, and decision-making live. It is, Wollberg says, where *you* are, which is why you feel so present in this conversation. Nothing neurobiological or neuroanatomical prevents the other 45% from having that combination. If you have had a lucid dream, he insists, you will remember it, because working memory was on. Vivid non-lucid dreams fade by the first cup of coffee. He therefore believes those 45% have not had it, and that neurostimulation can still give them an induced, stabilized version.

Ate-a-Pi’s falling-asleep experiments were a different state. Hypnagogia sits between waking and sleep; the phenomenal space is a bit more malleable, but you are not in REM. Lucidity itself is a scale. You can know you are in a dream and still be along for the ride with no control. Wollberg will allow that to be called a lucid dream. What they are driving for is the 95th percentile: awareness, the ability to pivot between first and third person, dissociation, and control. Control is what people want, and it is a function of how active the prefrontal cortex is relative to how active it is right now, in the waking conversation.

At height, being in a lucid dream is being as present as you are in this room, able to change phenomenological components of the dream at will. The lucid-dreaming subreddit has more than half a million people. Wollberg sorts what they do into three buckets. Recreation is the ultimate VR: fly, make buildings appear out of the ground, talk to dream characters. Those characters, he noticed only after using ChatGPT, behave a little like large language models — sometimes striking, sometimes absolute gibberish, almost as if they are hallucinating.

The second bucket is productive. They have investors who write code in lucid dreams and run it in the morning. A friend, Case Findley, who helped with branding, did the first hardware render of Halo as a designer who works in his lucid dreams. The history is long: Srinivasa Ramanujan and infinite series, Niels Bohr, Salvador Dalí. The third bucket is metaphysical, closer to a spiritual or psychedelic experience — simply becoming aware inside your own consciousness. Concretely, people who have lost someone go looking. Two years earlier Wollberg lost his 98-year-old grandmother; in the first lucid dream after, he went to talk to her. You carry a mental model of the people you know, so she responds the way you remember, not like a generic dream extra. Closure after breakups is in the same file.

The limiting factor, he says, is imagination: what you imagine becomes. Engraved on the inside of Halo is the company slogan: Prometheus stole fire from the gods; we will steal dreams from the prophets. Fire was not invented. Lightning already burned dry fields. Humans learned control, and then built industrial society. Animals dream; primates dream; octopuses dream, and you can watch it, because their colors and textures change as they move through an environment. What they are trying to control is, in their minds, potentially as powerful as fire. High prefrontal activation also means you remember the night. A social-plus-journal app will let you write it down and, with generative models, wrap it in video or stills — both as a memory aid and as something followers can see, if you want them to.

Ate-a-Pi asks about the fMRI work that reconstructs images a person is seeing or imagining. Could you generate what someone saw in a lucid dream? Wollberg is excited and cold at once. Those experiments train on a handful of images — on the order of 20 — and then have people imagine the same 20. Labeling the entirety of phenomenological experience is a different problem. He puts useful reconstruction years if not decades out.

The second piece they ever published on the blog was on noetic sovereignty, an offshoot of the cognitive-autonomy movement. These are profound technologies. Society, he says, has to talk about whether it wants that level of so-called mind reading. Conscious experience is the most intimate thing you will ever have. Prophetic’s stated posture is to bring you to brain states and let you fill in the content. If you do not want to post what you dreamed, that is completely up to you. The capability, on a medium-to-long horizon, he treats as inevitable. The conversation about how far to enable it should not wait for the demo.

They closed a pre-seed in June. Wollberg’s itch is older. A natural lucid dreamer for 18 years, in 2018 he was in Jerusalem working for Israel’s Innovation Authority, the government’s venture arm, a self-described hopeless generalist. Living there pushed ultimate questions; the one that would hold his attention was consciousness. He talks about lucid dreaming as a particle accelerator for it: a conscious experience with little to zero sensory input, the way the Large Hadron Collider uses magnets to make exotic collisions. Landmark 2013 and 2014 studies had used electrical stimulation to induce lucidity, statistically significant on only two of seven variables on the scale — awareness that you are dreaming, and dissociation.

Electrical and electromagnetic stimulation are, internally, the vacuum tubes of the field. Three limits: depth, because the skull evolved to keep things out of the brain; spread, because current that does get through fans across the surface with no precision; and no ability to steer. It worked at all only because lucidity is REM plus prefrontal surface, and the electrodes could tickle that surface. Deep-tech investing had taught him a filter: is this in R or in D? Venture capital was not designed to underwrite research; universities and governments are. In 2018 he tabled it.

In 2022 he found transcranial focused ultrasound and the neural transformer and decided they had entered D. Against the three old limits: depth is centimeters into the brain, noninvasively; precision is millimeters, which he calls a paradigm shift rather than an order of magnitude; and you can steer those millimeter pulses in three-dimensional space, which matters because the brain fires in 3D patterns. Berry was already applying neural transformers in neurotech and had started down the ultrasound rabbit hole. They found each other. TFUS for neuromodulation dates to around 2004. Element counts on piezoelectric transducers have been rising, size and cost falling, on curves Wollberg named Wes’s Law because Berry would not. In hardware, he says, you always want to be riding some kind of curve.

Version one of Morpheus was sourced mostly from open-source simultaneous EEG-fMRI, plus some lucid-dreaming data from Donders. The model’s job is not to classify REM. That problem was solved a long time ago and they leave it to other techniques. Given a brain state, Morpheus is supposed to output TFUS instructions continuously until the brain responds. They mask the prefrontal cortex in the fMRI, find voxels of heightened activity, and at the same timestamp line up the EEG sequence. The transformer learns spatial and temporal patterns between the two, approximating what a highly active prefrontal looks like, then how to emit ultrasound instructions that pull a sleeper up toward that state.

A confusion Ate-a-Pi has to have walked back: the showcased model is not trained on people being ultrasonically stimulated, and it is not primarily trained on people in lucid dreams. It is waking-state data, where the prefrontal is already on, supplemented with EEG of lucidity. The simultaneous EEG-fMRI lucid corpus is what they are aggregating now. They collaborate with the Donders Institute, which Wollberg calls probably the top lucid-dreaming lab in the world, led by Martin Dresler, whose 2012–2014 work helped establish the neural correlates. Prophetic is doing the largest neuroimaging aggregation of lucid-dreaming data ever, about four simultaneous EEG-fMRI sessions a week. Wollberg had been on the phone with Dresler that day; the first new dataset was due perhaps that day or the next.

The closed loop on-device is the obvious one: EEG from the headband prompts the model, TFUS fires, the next EEG is compared, targets update. EEG currently samples at 1,000 Hz, downsample-able if there is redundancy. Ultrasound carrier is on the order of 500 kHz; pulses aim at gamma, around 40 per second. The main adjustment each loop is spatial target more than timing. Because the band sits slightly differently every night, real-time EEG is also how they correct aim. The datasets are dense: an fMRI spatial reading every 2.1 seconds sits next to 2,100 EEG samples in the same window.

Personalization is the next layer, reinforcement learning with two feedback channels. Explicit: you wake up and rate the sequence, one to ten, or even rate it awake. Neural: ping the brain with TFUS and listen to the echo — gamma spikes, patterns across electrode placements — as a continuous reward or penalty. Berry’s analogy is a TikTok or Twitter feed. A like is explicit; dwell time is continuous. The model learns which tokens work for this person. Wollberg’s lay version: they only train on states that are discrete and universal. Flow is a counterexample, not discrete — a surfer’s flow is motor cortex, a chess player’s is spatial reasoning — so a general model would need much more data. Lucid dreaming is the same thing in everyone: prefrontal on during REM. Ate-a-Pi might live in one region of the vector space of sequences, Berry and Wollberg in others; RL walks each person toward their neighborhood.

Open-source simultaneous recordings can get you terabytes. Hundreds of terabytes or petabytes come when people sleep in the device every night, contributing gigabytes of neuroimaging plus the TFUS sequences that produced the echoes. That is when the reinforcement layer really works. The day after they showed Morpheus they published a piece on the Qualia Factory, a marketing name for a purpose-built simultaneous EEG-fMRI lab. Bring in people who are extraordinary at lucid dreaming, meditation, focus, positive mood; train successive models that load more experiences onto the same hardware. At launch they hope for more than lucid dreaming: focus and positive mood, both already induced with TFUS in the literature. Lucid dreaming is the consumer killer app. Extra states expand TAM — wear it while jamming on a spreadsheet, or when you want to feel better.

Consumer-hardware founders told Wollberg to de-risk demand and show an order book to manufacturers, because small-batch runs are death by a thousand cuts: 500 units, 250 dead, send more money, 250 dead. The reservation program has $2.4 million of booked revenue in a separate account, fully refundable, not used for development. Anyone who put money down when they were a website and a dream should get a device. They have spent zero dollars on marketing. A beta program launched with the model; about 3,500 people signed up in five days. The first manufacturing run will track the reservation list. Qualia, he notes, is Charles Peirce’s term for the units of conscious experience — the color red, the smell of coffee — which is why the lab got the name.

If they had to write a first-run tutorial, Wollberg would start with grokking the environment, talking to a dream character if one is nearby, and trying to make a flower or a building appear. Because what you imagine becomes, go in knowing the first place you want to be. An in-dream vocal walkthrough, the way some lab studies have had real-time dialogue with dreamers, is not on the roadmap. Sensory input can activate the thalamus, which regulates sleep, and you wake up. A researcher who did a PhD on auditory induction had to map each subject’s waking threshold; light cues have the same problem — the body decides something is happening outside. Music in the background is the same channel. The point of Halo is to circumvent ears, nose, and eyes and only push prefrontal cortex. Education happens before bed, as video, not as a voice in the dark.

Is it a medical device? They make no medical claims. The ultrasound regulatory threshold is 720 milliwatts per cubic centimeter; Halo is in the 100 to 200 range, well below, in Wollberg’s phrase less than a sonogram on a mother’s womb. You can already buy recreational electrical stimulators that are far less precise, and ultrasound’s safety record is better. Almost everyone else in TFUS is targeting medical applications, which is fine and slow. They are targeting recreational consumer experiences, which is how they stay out of FDA device review. V1 is the only hardware they will talk about in detail. Card79, which did Neuralink’s hardware, built the prototype. Comfort is the constraint: side sleepers, stomach sleepers, back sleepers, all night. They built a 3D scanline hydrophone tank so they can measure pressure fields and match them to simulation, because they cannot cut a skull open to watch the wave propagate. Neurostimulation validation was slated for that spring, with a hope of the first ultrasonically induced lucid dream soon after. V2, someday, rides Wes’s Law further — more elements, less power, more experiences as subscriptions — and maybe loses the side electronics pack.

Purists show up in the comments: lucidity should be earned through meditation, not handed out. Wollberg is unsurprised. He spent a stretch in psychedelics and watched the same split, ethnobotanists calling compounds sacred versus labs at University College London and Johns Hopkins trying to move the science. His reply is the fire line again: only lightning gets to make fire, because we might burn the village. The world, he says, is moved by people who ask what if we gave an extraordinary thing to everyone. They had not yet stimulated beta users; friends at the Institute for Advanced Consciousness Studies have run ultrasound into meditation and positive mood, including a video of a woman who had never meditated dropped into a deep state and becoming emotional. Paraphrasing Sam Altman at the end, Wollberg says these are the worst these neural transformers will ever be, and the worst TFUS will ever be. He does not have a good picture of the world when the stack is available at mass. It will be strange. He wants to explore it alongside everyone else.