Andrew Huberman on Sleep, AI, and Neurotech

A mouse learns to press a lever for a reward. Researchers tag the exact neurons that fire, in the exact order, while it does this. Later, they reactivate those same neurons in the same sequence, and the mouse performs the behavior perfectly. Then they try something else: they fire the same neurons in reverse, or all at once, like banging every key on a piano instead of playing the tune. The mouse does the behavior anyway, just as well (35:52). This is the finding that, by Andrew Huberman's own admission, unsettles a lot of neuroscientists, because it suggests that the order in which memory neurons fire might not matter much at all (36:11).

Huberman is a Stanford neuroscientist and the host of the Huberman Lab podcast. On this episode of Invest Like the Best, host Patrick O'Shaughnessy gets him talking across a wide range: how to actually fall asleep, why he thinks the biggest AI companies are secretly biotech companies, and the strange afterlife of the scientist who edited human embryos. The through-line is a field that can build stimulation hardware faster than it can explain what the hardware is stimulating.

Three doors into sleep

Huberman's starting point is personal: most high performers already know their own sleep number. His is seven hours. Eight leaves him groggy, six is fine, and two nights of four hours or less and he is, in his word, cratered (03:58). The harder problem, he says, is not knowing how much sleep you need. It's getting your mind to stop running long enough to use the time you have.

He breaks the skill of falling asleep into three parts, and says the order you do them in doesn't matter (08:00). First, stop thinking by shifting attention away from planning and memory and into pure sensation: the feel of the sheets, the sound of breathing, the sense of your own heartbeat. Second, slow the heart rate with long, complete exhales. Exhaling triggers a real physiological mechanism called respiratory sinus arrhythmia, in which the vagus nerve signals the heart to slow down as the lungs empty (10:36). Third, and strangest, forget where your limbs are in space. Huberman points to studies showing that rocking beds, the kind that put babies to sleep, work on adults too, because side-to-side motion generates compensatory eye movements under closed eyelids that scramble the brain's sense of body position (11:42). A team from MIT has built an eye mask that mimics this directly, using mild stimulation behind the ears to move the eyes and help people fall asleep faster, with commercial release expected in roughly seven to twelve months (13:56).

"The eyes are two pieces of brain, they're just not in the cranial vault, which makes them unusual in terms of parts of brain." — Andrew Huberman [13:15]

It's a small, strange fact that does a lot of work in this conversation: vision is not a window onto the brain, it is brain tissue, sitting outside the skull, which is part of why eye movement can be used as a lever to change brain state.

The real AI arms race

From sleep hardware, the conversation widens into Huberman's actual thesis for where neurotechnology is going. His claim is blunt: companies that look like language-model companies are not really language-model companies.

"Every single one of those is going to be a biotech company. They're all interested in the brain." — Andrew Huberman [22:52]

He means Meta, OpenAI, and Anthropic, alongside Elon Musk's Neuralink, a company building brain implants, and the broader push toward what he calls reading from and writing to the nervous system. Reading, in current practice, mostly means blunt proxies: a smartwatch tracking heart rate and heart rate variability (HRV), a measure of the variation in time between heartbeats that is often used as a stress indicator. Huberman's former postdoctoral researcher Malisse Ylmaz, a neuroscientist and biotech founder, has built a wearable cuff that reads autonomic nervous system activity directly instead of inferring it from heart rate, and has already piloted it with police departments and first responders (41:03).

Writing to the brain today mostly means pharmacology, which is effective but imprecise: caffeine, Adderall, and sedatives all raise or lower neural firing thresholds across the whole brain rather than in one targeted spot (27:27). Huberman's forecast for what replaces this is specific and strange. A person gets injected with a harmless viral vector, which is simply a carrier that delivers a gene into neurons rather than causing illness. That gene makes targeted neurons sensitive to light or ultrasound. The person then wears a cap or headband that beams nonspecific light or sound at the brain; only the gene-tagged neurons respond, and an AI system calibrates the intensity automatically rather than a person turning a dial by hand (44:24). He expects built-in safeguards, comparing them to the off-switches already used on vagal nerve stimulators, implanted devices that use mild electrical pulses on the vagus nerve to treat depression (46:04). He thinks something close to this, even if just a basic vagal stimulator under the skin, will be a mainstream curiosity within about a year, floating a figure of 20 percent of entrepreneurs wanting one (30:07). He's similarly confident that whole-body MRI scanning becomes common within twenty months, despite recent pushback from parts of the medical establishment, including a Wall Street Journal piece arguing against widespread use (47:26).

The ethical edge of this story has a name: He Jiankui, a scientist who trained as a postdoctoral researcher at Stanford before running his own lab in China. He announced he had used CRISPR, a gene-editing tool, to delete the HIV receptor in twin babies, aiming to make the children immune to HIV (53:22). The scientific community reacted with alarm, the Chinese government said his lab would be shut down, and he disappeared from public view. Huberman says He reportedly now runs a laboratory in Austin, Texas, though he's careful to flag that he can't confirm this (54:49). The detail that sharpens the story: the same HIV receptor He deleted is also linked to neural circuit formation and memory, which raises an unresolved question about whether the goal was only disease prevention (56:47).

Huberman is unapologetically more excited than alarmed by all of this, which puts him at odds with a public health media environment he describes without much restraint.

"The public health discourse around health and longevity is a fucking disaster. It is a bunch of clowns." — Andrew Huberman [18:21]

His complaint isn't that people are wrong, it's that the bar for an audience has dropped to showing abs and admitting to steroids, crowding out people with actual training in science or medicine (19:02). He predicts a correction: a downturn in performative optimization content, followed by a wave of credentialed educators doing the slower, less theatrical work of explaining what's actually known.

The episode closes on a story that quietly answers its own opening puzzle. In graduate school, Huberman abandoned the lab he'd formally joined to sneak into a different one at night, chasing a research question he couldn't stop thinking about. His eventual graduate advisor, Barbara Chapman, told him to simply join her lab instead of the one he'd committed to, because, in his telling, what matters is wanting to know the answer badly enough.

"Success is determined by how precisely you match your genuine curiosity to the work that you do." — Andrew Huberman [67:58]

It's a tidy echo of the mouse and the scrambled neurons. The sequence you're supposed to follow, in a career or in a circuit, may matter less than people assume. What produces the result, in both cases, is something closer to intensity and fit than order.