He stayed up all night playing Civilization 6 with his mind

Nathan Labenz opened the conversation by introducing Dean W. Ball as a kindred spirit: a research fellow at the Mercatus Center, author of the Hyperdimensional newsletter, and someone who had just done a deep dive on brain-computer interfaces. Both men, Labenz noted, keep a day job and an obsession. They had found each other online through Ball’s writing, and the plan for the episode was a survey — non-invasive EEG and ultrasound, invasive implants like Neuralink, and the policy questions that follow if brains can be read and written.

Ball’s résumé is not the usual neurotech one. He spent about a decade in the think-tank world on state and local economic policy, mostly housing and infrastructure. As a teenager he committed code and wrote documentation for phpBB, the open-source forum software, and the coding bug never quite left even though he never did it professionally. When ChatGPT arrived, the combination suddenly mattered: policy people were hungry for societal implications, and technical ground truth was often missing in that room.

Labenz’s own rule of thumb is that you have to figure out *what is* before you argue about *what ought to be*. Too much of the public debate, he said, jumps straight to ought. That is the job Ball was there to do — describe the devices that exist, the numbers they actually produce, and the near-term path, before anyone legislates a merge that is already, in Ball’s telling, several decades old.

Ball came to the topic as a longtime reader of J. C. R. Licklider, the Defense Department researcher who in 1960 published “Man-Computer Symbiosis.” The paper’s thesis, as Ball retold it, still sounds like a tweet from last week: AGI in five to ten years, then probably superintelligence that replaces us, “but in the meantime we’ll have a lot of fun.” Hearing that from a senior official inside the Pentagon, he said, shows how greenfield computing felt at the time. He thinks we are in a similar moment again.

The same rhetoric is back, only now the people saying it can fund labs. Sam Altman wrote a long post, “The Merge,” that treats the coupling of human and machine as an inevitability. Elon Musk has not only called it inevitable but founded Neuralink to build a technical path. In Ball’s conversations with people training frontier models, a common intuition is that if artificial superintelligence is even remotely near, the merge cannot be far behind. His question was more pedestrian: what does the hardware actually look like?

The answer he keeps returning to is that computers already are neural interfaces. You do not drive them directly with motor cortex, but they obviously interface with your brain, and you are using your brain, via touch, to manipulate them. Across the history of technology, information tools do not just change how information propagates; they change how we think. The merge, on that reading, started a long time ago. We are still on the path. It will get more vivid and more visceral. It is not a fundamentally new species of event.

Labenz had already felt a few waves of that cognitive change. Search made remembering optional; GPS made spatial maps optional. Coding used to mean jumping into the editor and scaffolding classes; language models now reward him for defining what he actually wants in conceptual terms first. His inbox followed the same arc. He used to remember keywords. With an AI email assistant he can ask questions, and he can already feel keyword search beginning to atrophy.

Ball’s version is questioning. GPT-3.5 was a hint; GPT-4 made it vivid. He would ask something, get an answer that was not quite what he wanted, and decide the model must be capable of better — so the problem was him. He refined the ask until he knew what he was really asking. In intellectual work, he said, finding the right question is 90% of the journey. A tool that makes you better at that accelerates everything downstream.

He had also noticed a split between Claude 3 Opus and GPT-4. On middle-of-the-bell-curve questions — ordinary science, history, even a lot of coding — the gap was not huge. At the frontier of a field, Opus pulled away. Researching neuroscience for his essays, he could often see GPT-4 filling in blanks; Claude 3 was simply more capable. The models had a second effect: they forced him to ask what, if anything, is distinctive about his own cognition. That kind of inquiry, he argued, compounds.

Zoom out from one user and the picture gets stranger. Labenz thinks the world’s future with AI is dramatically under-theorized: most builders still imagine the world plus their intervention, not everyone shipping at once. Ball reached for Manhattan. The island is roughly the same landmass as four hundred years ago; the difference is density. He expects a round of conceptual and cognitive densification, and he worries that people like him and Labenz are already in a bubble, extrapolating agents, while others asked ChatGPT 3.5 one basic question eighteen months ago and stopped. Neurotech, he hopes, is one way to help more people keep pace. A study that week found 30% of young people using ChatGPT for work — a lot for a little more than a year, and nowhere near everyone.

Labenz framed the rest of the hour as a jump in the dimensionality of communication. Keyboards and mice are a finite action space. Natural language opened that space; Ilya Sutskever’s line, as Labenz recalled it, was that the shocking part is “I speak to the computer and I feel I am understood.” Language is still a compressed, token-by-token stream. Brain signals would be richer still. Ball’s newsletter is called Hyperdimensional for a reason.

At bottom, Ball said, neural technology is a story of advanced sensor fusion and signal processing — the same story, from Bell Labs on, as much of the last hundred years of electronics. Thought is mediated by electricity and magnetism, so the straightforward read is to measure the fields the brain generates. The workhorse is EEG. Its cousin MEG, magnetoencephalography, needs bulky equipment and is not leaving the lab anytime soon.

The field is also older than the Neuralink demo cycle. EEG products have been shipping for more than a decade. The gap between a consumer headset and a lab rig is mostly channel count: two electrodes on the cheap end, 32 for a full helmet, 8 to 16 in the typical consumer band, and up to 256 in the lab. Sampling is usually around 1,000 Hz. An eight-channel consumer headset therefore produces about 8,000 data points per second, and millions if you wear it for any serious stretch of time.

The opponent is the skull. It exists to protect the brain, it is not particularly conductive, and it attenuates signal on the way out and the way in. Cerebrospinal fluid and other layers sit in between, but the skull is the main filter, and it is not a uniform one. Thickness and density vary across people and across a single head. An ideal reading, Ball said, has to model that geometry in real time. That is where he thinks the non-invasive stack still sits at the frontier.

Labenz wanted the units. A thousand hertz is one number every thousandth of a second; a neuron can also fire on roughly that timescale, and Ball noted that some activations are faster. EEG’s gift is temporal resolution. The spatial gold standard, fMRI, at its best takes one image every two seconds, needs superconducting hardware, and will not ride on anyone’s head. A 64-channel EEG produces more than a hundred samples in the time it takes that scanner to take one picture.

The cost is space. Even a precise electrode is looking at millions of neurons, at a resolution of centimeters or, in good cases, millimeters. Each channel really is one number per sample — the strength of the field at that spot. Labenz’s analogy was a cell tower: billions of local spikes summing into a messy broadcast, then 64 antennas trying to unmix it. Ball said the sketch is basically right. A Fourier transform takes the wobbling composite and breaks it into intensities at different frequencies. If the brain is playing a chord, the transform separates the notes.

Then you have to throw notes away. Blinks generate fields. Facial muscles do too. The headset itself is an electrical object. Traditional statistics, convolutional nets, and recurrent nets still do most of the published work. Transformer papers on EEG exist, Ball said, but they are scarce, which he finds odd given the volume of data. His first guess was academic conservatism rather than a shortage of samples.

Labenz offered two other guesses from a recent episode on Mamba. Attention is quadratic in sequence length; a thousand samples a second would swamp the context windows the public only recently saw in GPT-4. Transformers also struggle in hyper-noisy settings, where the gradient is often dominated by junk. Selective state-space models update token by token and may lock onto a rare signal instead of averaging it away. If that diagnosis is right, Ball said, the next couple of years in this literature could get interesting.

Once you are inside the literature, Ball said, you mostly admire how hard the problems are. Then you notice that EEG works anyway. The only electrical activity that really reaches the scalp is cortical, maybe one to three millimeters down; deep structures drop out. And yet lab work about eight years ago showed seizure prediction with nearly 100% accuracy up to an hour before onset, using patterns the patient cannot feel. Devices already do this for people with epilepsy. Early Parkinson’s has a distinct neural signature that can be read at something like 90-plus percent accuracy.

Consumer gadgets stay away from those medical claims, partly because of the FDA, which they would get to later. What you do see are meditation headbands and a related trick called neurofeedback: instead of writing into the brain, you show the wearer their own waves and let them train. A band pairs with a phone, a game scores focus, and the pitch is that circuits become easier to enter with practice. Ball described a sleep-mask design that plays music dynamically tuned to the EEG. He had not tried it and did not think it was shipping yet; the broader claim, that brains can learn to relax more readily, has some literature behind it.

Motor intent is the other working demo. Thinking about moving a limb already sends the electrical command; that command can be routed to a robot for someone who cannot complete the motion. Labenz sees helmet-and-arm videos constantly. The movements are often jerky or slow, which Ball chalked up to the same signal-processing limits. A lot of motor control also lives in deeper structures EEG cannot see. Pairing EEG with fMRI or PET — a sparse 3D blood-oxygenation map plus a torrent of noisy scalp voltages — is, he thinks, an underused machine-learning setup.

You can already buy this. A 32-channel headset plus software will, after calibration, let an enthusiast slide a block around a screen. It is nothing like Neuralink’s cursor, and it is not cheap, but it is not vaporware. Headbands and over-ear designs typically land somewhere between $500 and $1,000. The company Ball named as furthest along on a true consumer interface is EMOTIV. Apple has even explored a patent for putting electrodes on AirPods to watch for medical conditions. Labenz’s ladder of resolution stuck: two channels for sleep versus wake, four for stress, eight for a coarse emotion, sixteen and thirty-two for the clumsier motor tricks.

Labenz quoted Musk from a Neuralink event: the best thing to decode a neural net is another neural net. Ball has watched problem after problem fall to the “unprincipled” application of neural nets, each time after specialists said it would never work. His humility now is to experiment in that direction. Academic labs, he thinks, stay with older pipelines partly so new papers remain comparable to old ones. Startups and corporate R&D will not care. Apple holds patents and has an exploratory design group; Meta has been more public. In his review of the Vision Pro, Mark Zuckerberg treated eye tracking as fine “until we get the neural interface hooked up,” as if that step were inevitable.

The missing ingredient is labeled data. The brain is a black box you cannot ablate at will. The field spent years on tiny fMRI samples, mapping regions and frequencies in a low-data regime that naturally produced principled, small-n methods. You cannot run the bitter lesson until you have many people, many sessions, and a record of what they were trying to do with a cursor or an arm. Population models of skull variation barely exist. High-dimensional thought looks personal: Ball’s associations with Mozart or Claude 3 Opus are probably not Labenz’s. Low-dimensional thought — “I am scared of this tiger, I am trying to run” — already has a common, readable signature.

The previous Cognitive Revolution episode had been with Paul Scotti, author of MindEye2. The project reconstructs what a person was looking at from fMRI. Version one trained a custom model per subject and needed 30 to 40 hours in the scanner. Version two trained a shared latent space on seven of eight people and brought the eighth online with about an hour of data. Visual-cortex voxel counts ran from roughly 12,000 to nearly 18,000, a 40% spread, so each person still needs a small adapter. The images barely overlapped, so the study cannot say how similar two people’s “Claude 3” is — and the shared space worked anyway. Later in this conversation Labenz added the detail that still sits at the center of the Chinese writeup: those voxels are not just edges and colors. Scotti’s model could predict a *caption* from the brain state, because visual cortex already carries semantics.

Neuralink, Ball said, is the invasive version of the same pipeline: dense electrodes in cortex, some of the ML on-device for latency. The first patient is a young quadriplegic, paralyzed in a diving accident, with no use of his arms. After the implant he drove a standard Windows machine — not special software — moving a cursor, playing chess. The first night the system was hooked up, he stayed up playing Civilization 6. Typing has been shown, including with non-invasive EEG, though not in a form a consumer would want. None of this, Ball said, should be a galloping shock to anyone who has been watching the lab demos. What may be underestimated, in classic Musk fashion, is the robotic surgeon that can place the threads with little human intervention. That is a cost-structure move, aimed at scale. On the non-invasive side, fNIRS — functional near-infrared spectroscopy — is a wearable cousin of fMRI, reading blood oxygenation without the superconducting magnet, at EEG-like cost and capability.

The next generalization Ball can actually picture is mind-reading in the narrow sense: decoding imagined images, then words, then using those as prompts for a language model. Plenty of people cannot formulate the question they can think. Even before full inner speech, basic motives look reachable. He pointed at the Rabbit R1’s “large action model” and at Apple Shortcuts: see today’s calendar, check the weather, turn the bedroom lights on. If a headset can read those intents and hand them to automation that already exists, it will feel shockingly like telepathy. The stone-cold reality, he said, is still crude. Everything is converging on a qualitative leap anyway. The brain-computer interface people actually want still needs write access, not just read.

Labenz put the field between GPT-1 and GPT-2. Fine-tuned GPT-2 could classify and complete sentences; it did not reason, and it did not do few-shot learning. Ball had tried, in his housing-policy days, to feed it municipal zoning codes and get a thumbs-up or thumbs-down on restrictiveness. The answers were primitive. About eight consumer neural devices have shipped far enough that someone is building a company on them. MindEye2’s entire public dataset is eight people and a couple of hundred hours in an MRI. Once headbands are in the wild, even among enthusiasts, the data regime goes vertical. The architectures already work in other domains. Unless this one is strangely different, Labenz said, image decoding, verbalizing thought, and usable control all get better mostly by collecting more signal.

The open question, for Ball, is how consistent the language of thought is across skulls. If you have to get an fMRI to make a headset useful, the product changes; he joked about putting a scanner in the back of an Apple Store, then admitted that even a 30-minute calibration is a different consumer motion than a Vision Pro demo. Still, translating impulses and basic desires into actions taken by existing AI systems “just seems super possible in the near term.” He would be shocked if it did not happen in five to ten years. Maybe a lot less.

The parallel is agents. GPT-4 can look expert at medical diagnosis and still fail to click the right button, probably because mid-length computer-use data was not in the original mix. Hardware is shipping in glasses, helmets, and headbands that do not quite work, because the right brain data was not there either — but unlike a web agent, you cannot fake that data without actual brains. A 32-channel band may already be enough for many intents once decoding improves; you may not need a hardware refresh. Neuralink’s implant, they looked up live, has 1,024 electrodes, dense and local, down to single neurons in a small patch, not a whole-brain read. Connecting that implant to the internet is, Ball said, “absolutely terrifying” as a cybersecurity problem. Almost no serious compute has been thrown at the non-invasive version. He sees low-hanging fruit.

Writing is the same physics with worse geometry. The skull, Ball said, is less a lid than a shield: it is there to keep things out. Transcranial direct and alternating current, and transcranial magnetic stimulation, have shown lab effects, but the hardware is expensive, the field diffuses, focus is poor, and people vary. Electrodes and ultrasonic transducers are on a cheap-and-dense curve; TMS is not. Some researchers still bet on magnetism if costs fall. Ball is interested in what could happen soon, so he mostly set those methods aside.

The bet is transcranial focused ultrasound, a century-old idea that came back in the last ten or twenty years. You fire sound far above human hearing — dogs and cats neither; rats and bats can hear it — at a frequency, amplitude, and pulse that land a few millimeters to about a centimeter into the brain, at millimeter spatial resolution. You cannot write the hippocampus. You can reach tissue no other non-invasive tool can, and the transducers are getting cheaper. The FDA treats doses below a threshold as safe. Nobody fully knows why it works. Theories include changing how readily tissue carries current, or kicking off neurotransmitter release. High doses heat tissue. Daily, years-long use has not been modeled, because the studies are one-off lab sessions.

The finding that most excited Ball in the whole literature was tactile. Blindfolded subjects feel one pin, then two close together, and try to tell them apart. Aimed at somatosensory cortex, TFUS improved that discrimination while it was on, and forty minutes later the gain was still there, only slightly faded. To him that looked like a bump in the brain’s channel bandwidth to the hand. Other groups have reported shifts in mood, focus, visual discrimination, auditory discrimination, and the ability to tell close vibration frequencies apart. You can induce phosphenes, the blobs you see when you press on your eyelids. Inhibitory protocols show up mainly in pain attenuation, with a side effect of slower motor reaction times. Both of the leading TFUS authors Ball spoke with were confident the dose can still be raised safely. Writing a thought, he said, is not remotely close.

A New York company, Prophetic AI, is trying to close the loop. Its “ultrasonic transformer,” trained on EEG and fMRI, takes a headband’s EEG as input and emits commands to ultrasonic transducers on the same band. The first target is lucid dreaming: sleep is an easy EEG signature even at low channel counts, a gamma spike in prefrontal cortex flags the lucid moment, and the transducers keep firing to hold the state. Focus and a positive mood are the next advertised rungs; higher-dimensional conscious experience is the ambition. Labenz’s forecast, which Ball mostly shared, is that consumer devices will get fairly good at reading in the next few years, remain crude at writing — valence, not “think about Claude 3 Opus” — and need implants for high-bandwidth two-way traffic. The Matrix moment where Trinity downloads a helicopter is, Ball thinks, hard to encode as ultrasound instructions across several regions. He would not rule it out. He would also not wear a lucid-dream band every night: his guess is that it would feel less like a pill and more like a sudden inner shift, and that daily use might resemble taking LSD every day, a Syd Barrett doom loop. Resolution and data, not a mysterious new physics, are what separate “happy” from “exactly this moment of my life.”

Labenz has been asking frontier AI people a hypothetical: if Neuralink had a million patients and was considered safe, would you get one? Some answers came back game-theoretic — how else would I keep up? Ball reached for interior design. Victorian American houses had dark red and green walls because kerosene lamps stained plaster; white walls are a luxury that electricity made possible. The point is not wallpaper. It is that the effects of a technology at scale are almost impossible to predict. You can try to get the mechanisms clear. The rest will surprise you.

He starts with courts. The idea of being under oath changes if we want it to, once a headset can add a biological read to spoken testimony. One way to think about the next twenty years of technology, he said, is whether we will impose artificial constraints on social life. Nick Bostrom’s “solved world” — every good in abundance, godlike control over atoms — is not a future Ball expects, but if it arrived he suspects it would “quickly devolve into hell.” A society that can know at the biological level whether you are lying is a different society. He is pretty confident that some degree of deception is an important part of life, and of judicial life. Courts already assume the truth will be bent.

Warrants and subpoenas get strange when varying degrees of thought are on disk. Ball’s instinct is that access is useful up to a point and dangerous at the extreme: if everyone has a Neuralink and every thought is recorded at high resolution, that corpus should not be available for advertising. Europe will regulate; at least half of America will want to. The present FDA picture is looser than the sci-fi. Most of these gadgets sit in a mid-level risk bucket, then slip out of the heavy process via the “general wellness” exemption — so long as you do not claim to diagnose or treat a specific condition. When the Apple Watch added a blood-oxygen sensor in 2020, Apple was careful to say it had nothing to do with COVID. Neurotech companies have generally taken the same exemption. Asked to say clearly whether it applies, the FDA has refused. Ball recognized the move from crypto: uncertainty itself is a regulatory tactic. It preserves the option to pull every wellness device off the market in a heartbeat. Until then, if you stay under the TFUS safety line and do not promise to cure a disease, you can sell.

He also declined the story that evolutionary advantage forces universal adoption. There is already an incentive to jog after the podcast, eat as Andrew Huberman recommends, and skip alcohol. Most people do not. The AI-safety community, he thinks, overrates how far a legible Darwinian algorithm drives what societies actually take up. The other panic — if there is an incentive, everyone will have to use it; what about those who refuse? — he takes seriously. He thinks about the Amish. There will be forms of digital Amish, and they need to be designed for. People who want to enhance cognition should have the liberty to do that too. “We should want there to be more cognition in the world, especially human cognition.” Wanting less human thought with GPT-5 around the corner is a world model he is not sure he can defend. Down the hall at Mercatus sits Robin Hanson, who argues we are in a strange “dream time” before evolutionary constraints snap back, and who has a near-term story about Amish fertility and technological stagnation. Ball’s narrower prediction is political valence. The Vision Pro already feels to him like a statement. Using a brain-computer interface, or refusing to, will get coded the same way. Issue polls that ask whether you want HAL 9000 to kill you next year and then report 98% opposition are, in his view, not data.

Outside the AI bubble, Labenz still finds the default gut reaction to this stack is negative — shaped by fiction, the Terminator, and a lot of secondhand media. He does not think that makes polling worthless, only that it is downstream of ignorance. His advice to business people who have never really used the tools is simple: spend time with the technology, give it data that is personal, and watch what it does. The surface area is too large to outsource to articles. Ball agreed that pessimism is real. He thinks techno-optimists should not pretend otherwise; the polls exaggerate a tendency that is already there.

Asked how listeners should acclimate to the merge, Ball started with a book: Nita Farahany’s *The Battle for Your Brain*. Then he recommended a Vision Pro demo. Gaze control, he said, is already shockingly close to a neural interface, and the natural form factor for EEG or ultrasound is probably something that wraps the head the way the Quest and the Vision Pro do. Add eye tracking and hand tracking, and the neural signal may not need to be that good to become interesting. That is sensor fusion again, from another angle. The demo floored him in the way GPT-4 did after GPT-3: you can be told the resolution, or you can sit in it. He has an older Quest and still has not bought the Apple headset, mostly because he is unsure how much he would use it. The experience itself, he said, is definitely going to be a thing — not entirely a neural interface, and not *not* one either.

The rest of the advice is deliberately boring. Proceed from empirically demonstrated ground truth, not from the surrounding narrative, which he expects to be toxic and misleading in several directions. Model the technology yourself, including against his model. In the technology-scouting business, he said, you have to keep a high degree of epistemic humility. Before anyone buys an EMOTIV-class band, the episode also leaves a quieter checklist: know where the brain data goes; watch the Neuralink chess and Civilization footage and skim MindEye2 so the ceiling and the floor are visible; practice asking language models better questions on real work; and write down, now, which kinds of cognition you are willing to outsource and which you will keep training yourself. That last list is easier to draft before the hardware is everywhere.