Max Hodak: The Future of Brain-Computer Interfaces

Engineering the Brain, Not Drugging It

Max Hodak co-founded Neuralink and now runs Science, a company building brain-computer interfaces from a fundamentally different angle than most of the field. Their first product, a tiny retinal implant called Prima, has already restored sight in over 40 blind patients in clinical trials. But the retinal prosthesis is just the entry point for a much larger thesis: the brain is the only organ that truly matters, and we are empirically much better at engineering it than we are at finding drugs for it.

This conversation with Garry Tan on Y Combinator’s How to Build the Future covers the spectrum from how a 2mm silicon chip restores vision, to why the brain’s representations look like AI latent spaces, to a longevity vision where the first people to live to a thousand may already be alive.

A Chip That Creates Images in the Mind’s Eye

Science’s Prima implant is a 2mm x 2mm silicon chip placed under the retina. Patients wear glasses with a camera that captures the world and a laser projector that projects the image into the eye. Wherever the laser hits the implant, it acts like a tiny solar panel, absorbing the light and exciting the bipolar cells directly above it. This bypasses the dead rods and cones to get a visual signal back into the retina.

The clinical trial across 17 European sites showed a massive effect: patients who had been unable to see faces for a decade could read every letter on an eye chart. Science is submitting for regulatory approval now, hoping for clearance later this year.

What made this work was a critical design decision about which layer of retinal cells to target. A previous company, Second Sight, stimulated the retinal ganglion cells (the 1.5 million optic nerve cells) with an electrical device. That produced flashes of light patients could identify individually, but the brain never assembled those flashes into coherent form vision.

Science targets the bipolar cells instead, one layer earlier. The retina has 150 million rods and cones connected to 100 million bipolar cells, which then compress down to 1.5 million ganglion cells. By stimulating before that 100x compression, they preserve the retina’s natural computation. The result: when you project an image onto bipolar cells, you get an image in the mind’s eye. This was an empirical discovery of their trial, the first time form vision had ever been created through a prosthetic device.

The Brain Wants to See

One of the most striking details about the patients: when blind, the brain doesn’t stop trying to see. It turns up the gain, generating phantom percepts and hallucinations. When the implant first activates, patients see a flash, but they also hear a tone played alongside it and start “seeing” the flash even when only the tone plays. The first hours of rehab are spent learning to distinguish real percepts from phantom ones, because the brain’s noise floor is so high from years of generating its own signals.

Current Prima provides black-and-white vision with a small field of view. In the next 10 years, Hodak sees a path to close to native 20/20 acuity with color and broader field of view. And because the device is agnostic to why the photoreceptors were lost, it works across macular degeneration, retinitis pigmentosa, Stargardt’s disease, and diabetic retinopathy. Geographic atrophy alone, the severe form of age-related macular degeneration, affects 1-2 million people globally, with 200 million affected by the broader condition.

Neural Engineering vs. Drug Discovery

This is the conceptual core of Hodak’s thesis. Humanity is not very good at drug discovery. Every now and then you find something like GLP-1, but far more commonly you spend a decade pursuing a lead only to run a study and get “no.” Billions have been spent trying to find drugs to slow or reverse blindness, with essentially no effect. There’s a million-dollar-per-patient gene therapy with marginal benefit for a tiny population.

Meanwhile, their retinal prosthesis took a patient who couldn’t see faces for a decade and let them read an eye chart. The neural engineering approach simply works better empirically, and Hodak sees it as a paradigm shift for medicine broadly.

The vision for Science extends beyond the retina. People need to see, hear, have balance, and have a kilobit per second of motor control. Cochlear implants already handle hearing. Motor decoding is well understood. Vision was the missing piece. Together, these capabilities speak to a fundamental reframing of what’s possible in healthcare.

The Brain Is a Computer (and It Has an API)

Hodak is direct about a claim he knows will get him yelled at by some corner of the field: the brain is a computer. A very different architecture from a von Neumann machine, but it processes information. And it has a well-defined interface.

All information flows in and out of the brain through 12 cranial nerves and 31 spinal nerves. The optic nerve is cranial nerve 2. The vestibular-cochlear nerve carrying hearing and balance is cranial nerve 8. The 31 spinal nerves carry motor commands out and sensory information in.

“You can think of that as like the API of the brain. If you can get all the signals going down those, the brain is not magically connected to the environment. Reality is whatever spikes are on the cranial and spinal nerves.”

This perspective has practical consequences. Near the brain’s inputs and outputs, the representations are “concrete” and map to intuitive things: colors, sound frequencies, muscle states. The motor cortex, just two synapses from the muscle, has representations that directly correspond to hand states and joint torques, which is why motor decoding works.

But deeper into the brain, representations quickly become abstract. In inferotemporal cortex, there’s a map of “object space,” a manifold where each point corresponds to the percept of a possible object: a vase, the Eiffel Tower, a car, a zebra. Millions of neurons represent this space.

The Brain’s Latent Spaces Are Real

The biggest surprise from the convergence of AI and neuroscience: when you train image models or language models, the internal representations that emerge look a lot like the representations found in the brain.

“It is a latent space. Exactly. There’s this huge unification going on between AI and neuroscience.”

Hodak notes this was unexpected. Ten years ago, the assumption was that AI researchers would learn from neuroscience. It’s been the other way around. Many neuroscientists have migrated to AI research because they’re essentially still doing neuroscience, just with systems that are far easier to experiment on.

For Science, this is directly actionable. Neural activity recorded from the brain is “just another latent.” If you can translate between the brain’s latent representations and a model’s latent representations, you unlock extraordinary capabilities. This is the theoretical foundation for their biohybrid approach.

The Avatar Queue: Growing New Nerves

Science’s biohybrid neural interface is the most ambitious part of their pipeline. The intuition starts from a simple observation: your brain’s two hemispheres are connected by the corpus callosum, about 200 million fibers, and despite processing different halves of the world separately, you experience one integrated moment.

If nature wanted to build an ultra-high-bandwidth brain-to-brain connection, or a new cranial nerve, an “internet nerve,” it would grow a new nerve bundle with a connector at the end. That’s exactly what Science is building.

They seed their implant with living neurons, heavily engineered stem cells derived from a single cell line that has been hidden from the immune system (hypoimmunogenic). When grafted onto the brain, these neurons grow throughout the brain and form biological connections, at least in animal models. No wires are placed into the brain. No genetic modification of the patient’s neurons is required.

Hodak compares it to the ponytails in James Cameron’s Avatar: a big new cranial nerve with a connector at the end.

The advantage over other approaches: optogenetics requires genetically modifying the patient’s brain neurons, which is a one-way door. If it goes wrong, it can go really wrong. With biohybrid, the only edited cells are the graft cells. If they die, the patient is essentially no worse off than before.

The trade-off: biohybrid is harder to deploy and will likely be “backloaded” relative to simpler approaches for near-term medical needs. But for the highest-bandwidth applications, the kind of interface that might eventually let two brains share conscious experience, biology is the right medium.

Conjoined Twins and the Physics of Consciousness

Hodak raises a natural case study that hints at what high-bandwidth brain-to-brain interfaces might feel like. There’s a pair of conjoined twins in Canada whose brains are connected through a biological cable between their thalami, visible on MRI. Over this cable, they can share meaningful elements of conscious experience. They can see through each other’s eyes to some degree. They can coordinate tasks without speaking. And critically, they don’t confuse the other’s signals for their own, unlike schizophrenic misattributed monologue.

The open question: are they sending classical information, or is there phenomenal binding happening over this cable, the way your two hemispheres bind into one moment of consciousness? Do they have three or four image modes instead of the normal two (eyes-open vision and imagination)?

Hodak’s long-term vision for BCI connects to consciousness research itself. If the only thing you can truly know about consciousness is your own, then brain-computer interfaces may be the only way to study it directly. And once you understand whatever the brain is taking advantage of that the universe supports, you can eventually build super-intelligent conscious machines that humans can be part of through ultra-high-bandwidth connections.

The Vessel: A Backpack Instead of an ICU

Science’s third project, Vessel, emerged from a case study Hodak read in The Lancet a decade ago. A 17-year-old in Boston was being kept alive on ECMO (a heart-lung machine) while waiting for a lung transplant. When complications made him ineligible, the doctors faced an ethical dilemma. He was alive, playing video games, doing homework. But he was consuming a half-million-dollar-a-month ICU suite. Eventually they stopped changing the oxygenator filter, it clotted, and he died.

When Hodak searched PubMed for “ECMO ethical dilemma,” there were multiple pages of results. Doctors were actively discouraging families from pursuing ECMO in critical care to avoid creating a “bridge to nowhere.” When he suggested treating it as a “destination therapy” rather than a bridge, the response from the medical community was, in his words, “shouting and throwing things.”

The same fundamental technology has already transformed organ transplantation: over 75% of US liver transplants now use machine perfusion. But the systems cost $500,000 and can only be moved by private jet. Science’s vision: refine it to the point where you could check a kidney as luggage on a United flight, or where that 17-year-old could have gone home with a backpack.

The Origin Story: An Email from Sam Altman

Hodak got into Neuralink through an email from Sam Altman in early 2016 with the subject line “crazy question”: Elon is starting a brain-computer interface company, who should run it? Hodak’s first reaction was to recommend his friends at MIT. An hour later, he emailed back asking if he could be considered.

Elon had already come up with the name Neuralink and had a clear motivation: he saw what was coming in AI far earlier than most, and the implication that humanity needed to merge with machines was obvious to him. Over the second half of 2016, a shifting group met weekly in the evenings, eventually coalescing around the thin-film polymer thread approach.

What Hodak took away was less about specific technology and more about execution. Neuralink was “the ultimate startup PhD,” teaching him how to run a technically complex, multidisciplinary company. His key career advice to young engineers: figure out what you want, then be very high agency about getting there, but also consider working for someone extraordinary first. Running a startup is an oral tradition passed down from a handful of Silicon Valley cultures, and getting that education at 20 instead of 28 can make a huge difference.

Afterthoughts

Max Hodak occupies a rare position: deep enough in neuroscience to have done primate neural decoding since 2008, deep enough in engineering to have co-founded Neuralink, and bold enough to bet his company on the idea that biology is the right medium for the highest-bandwidth brain interfaces. A few things worth noting:

• The layer targeting insight is a textbook example of first-principles engineering in biology. Second Sight stimulated past the retina’s compression layer and got flashes. Science stimulates before it and gets images. The difference in outcomes is enormous, and it came from methodically exploring all four quadrants of a 2x2 design matrix.
• The AI-neuroscience convergence is not a metaphor. Brain representations literally look like model latent spaces. This has gone from a poetic analogy to a practical engineering tool. The brain’s “API” through cranial and spinal nerves is a well-defined interface that we’re only now learning to speak fluently.
• The 2035 event horizon is Hodak’s way of saying the compound effects of BCI plus AI plus bioengineering become unforeseeable within a decade. His claim that the first people to live to a thousand may already be alive is not flippant; it follows directly from the neural engineering approach to medicine, where you support the brain and treat everything else as replaceable.
• The biohybrid bet is genuinely distinct. Most BCI companies are doing electronics miniaturization. Science is growing living neurons that form biological connections with the brain. This is either the most important approach in the field or a dead end. The conjoined twins case study suggests the biology might actually support things that electronics cannot: phenomenal binding between separate neural systems.