From Restoring Sight to Reimagining the Brain, with Max Hodak
In a Nutshell
Max Hodak's company Science has achieved European regulatory approval for Prima, a retinal prosthesis that restores functional vision to blind patients via a subretinal implant paired with laser-projecting glasses—patients can now read books and complete puzzles. The company is building on this proof-of-concept to pursue broader brain-computer interfaces and biohybrid neural technologies aimed at achieving "substrate independence" for the brain. Hodak argues the brain functions literally as a computer, making direct engineering approaches more effective than traditional drug discovery for treating neurological conditions.
These notes were generated by AI and may contain inaccuracies.
The brain can be understood very literally and plainly as a computer. Computational problems can be solved by arranging matter in a certain way and then letting the process unfold. The skull functions as a brain in a vat, with the brain connected to the environment through a small number of wires - the cranial and spinal nerves. Getting visual signals, auditory signals, balance, and motor input and output into and out of the brain is an end in itself and represents the central object of focus.
The retinal prosthesis serves as a great proof of concept that the approach is on the right track. Nobody had previously been able to restore a form vision image in the mind's eye of a blind patient in this way. Current capabilities need to add depth of grayscale, and there is a path to get at least red and green color vision. The technology can compound upon this path through an engineering process to make increasingly better products.
Max Hodak is the founder and CEO of Science, formerly of Neuralink. The company recently received regulatory approval in Europe for Prima, an implant that helps people who have gone blind see again. The company's quest involves sustaining the human experience and achieving substrate independence for the brain. The discussion covers alignable representations between AI models and the future of neuroscience.
Science is fundamentally a medical device company. The mission is to use a differentiated understanding of the universe to improve the human condition. The company specifically uses an understanding of how to work with the brain to get big effect sizes that are not often achieved in medicine. The main product is a retinal prosthesis, which can be thought of as a cochlear implant for the eye. Cochlear implants represent some of the biggest impacts in all of medicine.
The Prima device is a tiny chip implanted under the retina in the back of the eye for patients who have gone blind due to loss of light-sensitive cells. The clinical trial focused on macular degeneration, with upcoming studies planned for retinitis pigmentosa, Stargardt's, and other diseases. The patient wears glasses with a laser projector that projects an image onto the implant, which then stimulates the retina to bypass dead rods and cones and directly stimulate the retina to get a visual signal back into the brain.
When starting the company, multiple ideas were considered. One direction was the biohybrid neural interface, where living neurons are engrafted that grow in and form new biological connections. This remains a big research project that needed to be paired with a near-term business. The team asked what was the most valuable thing they could do and concluded they could restore vision to the blind with the resources available in early 2021.
To restore vision requires understanding how the brain gets vision. The first stop of the optic nerve into the brain is the lateral geniculate nucleus in the thalamus. Options for restoring vision include going into the retina, the thalamus, or V1 (visual cortex containing half a billion cells at the back of the brain). Scientific and technical reasons led to the conclusion that if an optic nerve exists, stimulation should occur in the retina.
All variants of retinal stimulation were explored early on. An in-house gene therapy was developed that affected the retina in one direction. A survey of electrical stimulators was conducted, and ultrasound was also examined. The team developed a state-of-the-art gene therapy indigenously, which is probably going to humans next year, and identified state-of-the-art electrical stimulation work.
A company in France called Pixium had the state-of-the-art work in electrically stimulating the retina, originally developed by an inventor at Stanford and licensed to the small French company. After getting to know them over several years, Science acquired them in late 2022 when they saw something that nobody else really saw at the time.
It took about 2 years post-acquisition to get Prima to the place where marketing approval was possible. In July, marketing approval was received in Europe for Prima to start commercially selling. This means the product is really commercially available, with first sales expected in the coming weeks.
People forget that the moonshot worked - footprints were left on the moon. The term "moonshot" has gotten used in Silicon Valley to mean things with extremely long odds that are unlikely to work. Historically, the success rate of moonshots is higher than people give them credit for. Having a real business is the most important thing, and this represents the start of that.
The pipeline contains three elements: work in vision, biohybrid neural interfaces, and a perfusion program called Vessel. These three things together form the minimum set that if successful on the 10-15 year timescale could drive a significant revolution in medicine broadly.
People have spent huge amounts of time and money looking for drugs to restore vision, hearing, stop Parkinson's, or help paralyzed people move again. Understanding the biology and molecular detail required to make a drug has been very difficult. However, the brain is a computer, and when dealt with as a computer, things work very well. Demonstration effects in medicine like a cochlear implant being turned on or a deep brain stimulator being turned on simply do not occur with most drugs.
Small molecules are super hard. Even with highly engineered patient-specific CAR-T therapies, giant immune overreactions can occur. In contrast, if electrodes are placed in M1, the patient will probably be using a computer in an hour. The retinal prosthesis is a great proof of concept, though current limitations include a small field of view like looking through a straw and black and white only vision.
In the clinical trial, the main achievement was the existence proof that success was an impossible outcome. Patients were filling in Sudoku puzzles and crossword puzzles, and some patients were reading books. Videos of these patients were viewed, one patient was met with, and surgeons were talked to. This is described as one of those things that seems too good to be true.
Telling people that the brain is a computer tends to make them angry. This view starts from a defensive place. The brain very literally, very clearly, plainly is a computer. The universe generally can be viewed as a computer - computational problems can be solved by arranging matter in a certain way and letting the process unfold. The brain is the same thing.
There is nothing special about transistors. Computers can be understood in an idealized way as Turing machines - abstract computers that go from state to state in ways subject to laws that make transformations interesting and meaningful.
The view that the brain is a computer was fairly contrarian both in the sense that BCI has a broader interpretation than motor decoding, and whether a retinal prosthesis counts as a BCI. If retinal prostheses are considered BCIs, this opens up interpretations of many areas of medicine that could be accessible to this approach.
Science comes from a different culture than conventional biotech industry. There has always been an East Coast, West Coast divide in biotech. Science is more of a tech company than a conventional biotech company. The device view of historical biology problems makes the company even more of a tech company by biotech standards. Most funding came from tech investors rather than biotech investors.
BCI is a category like how pharma is a category. The comparison to making "one bet on drugs" illustrates the breadth of possibilities within the field. Everything from silent speech devices to radar through the face represents different approaches that may be BCI in greater or lesser degree.
Silent speech devices are basically hand substitutes. Hands work really well and do not require much conscious thought. The concern with spontaneous actions like summoning an Uber just by thinking about it highlights the need for explicit communication. Extra hands would be useful, but getting vision, hearing, balance, and a kilobit per second of motor control gets halfway to the Matrix.
There is a distinction between communicating with a thing and redrawing the border around the brain. The way the two hemispheres of the brain are used as one integrated bound thing is different than the way one talks to another person. All communication is about creating correlations between brains. When speaking, big correlations are driven between brains because communication is premised on pre-shared language or common education.
One mode involves pre-shared structure between two brains, whether an AI model or biological brain, communicating over the channel. The other mode involves adding some new structural capability. Figuring out where the transition between these modes happens is a compelling area of research.
If the end of the artificial intelligence quest is care, the question becomes whether one might want their brain to be a better computer or a richer one in terms of understanding other people's experiences. If a brain is scanned into a computer creating a software simulation, does that count as the original person? Would that make someone feel better about dying of cancer if the software replica could continue their venture investing job?
General anesthesia produces a break in consciousness, but people accept this and realize it is fine. The question is why talking to a software replica and then going away feels different. There is an asymmetry in the operators that actually change things in physics - creation or annihilation operators. Life involves creating new life or new minds, and there are times when they can be annihilated or destroyed, and ways they can change while intact.
A conscious moment involves experiencing a bunch of things in parallel - seeing, hearing, feeling, and smelling things simultaneously. The question is how the brain constructs each of those and causes them to be perceived together to the exclusion of other things. One never gets someone else's vision and hearing - there needs to be a fundamental explanation for how that partitioning happens.
Continuity is greatly important. People accept significant drift in their identity over time as long as they have continuity. However, if a software simulation answers exactly like the original would but lacks phenomenal continuity, that is less satisfying. Dramatic morph but continuous experience would be preferable to significantly degraded IQ, though degraded IQ might be acceptable for some period if it then got backfilled.
At the point of achieving substrate independence, almost anything becomes possible. One of the big missing pieces is connectomics, which is getting close to the point where the project could be done. While still relatively far from a human connectome, a mouse connectome is not that far away and would be enormously useful for facilitating research. Basic questions remain about the overall architecture of the brain.
There is increased interest in BCIs given the progress of AI model research because representations should be shared or empirically seem to be. This is the platonic representation hypothesis, which is controversial but appears to have something real happening. Mathematical objects inside big AI models look a lot like things seen in neuroscience, with very similar geometry in how concepts are represented.
Alignments can be achieved between animal brain neural recordings and AI model internal representations. This was a big clue that AI was on the right track and not hitting a wall. There is some fact about the universe where these systems as they learn are grabbing onto some true underlying data manifold. If enough compute is applied to matter, intelligence emerges.
Some faction of people do not want the representation hypothesis to be true for reasons that are not totally clear. It is also not fully understood whether the phenomenon is global or local in nature. Relational structures between ideas can be recovered, but it might work locally where disconnected things might be placed differently. There is a bunch of technical detail that remains unknown.
Working on AI is considered one of the most fertile ways to study neuroscience today. Several neuroscience researchers have moved to AI companies like OpenAI and Anthropic, finding it easier to conduct neuroscience research on AI models than traditional methods.
Restoring vision to the blind represents a viable business opportunity. Age-related macular degeneration affects one in two people with early stage by age 80, and one in ten by age 85. The total addressable market for the current version is hundreds of thousands of patients in the US and Europe, expanding to millions with the next version.
Second Sight's retinal prosthesis from a decade ago, which provided different functionality than current approaches, received FDA approval despite limited performance. It generated revenue of approximately $150,000 per patient in the mid-2010s. A gene therapy for a narrow 5% patient population with one specific indication reimburses at nearly half a million dollars per eye, despite achieving only 0.1 lines of improvement and merely slowing degeneration rates.
Vision restoration therapies command high prices due to expensive development costs and historically high failure rates. Vision represents a dominant sense, and even minimal restoration provides significant value. The first version targets hundreds of thousands of patients, while the next generation undergoing animal studies should reach millions.
The brain is the only organ that cannot be transplanted in principle. Other organs like the heart, pancreas, liver, and lungs serve as support systems to maintain brain function. The brain functions as a computer generating our experience of the world, connected to the environment through cranial and spinal nerves.
The skull functions similarly to a brain in a vat scenario, with the brain connected to the world through a small number of wires including the optic nerve (nerve two) and vestibulocochlear nerve (nerve eight). The world is generated by the brain, making the ability to input and output visual, auditory, balance, motor, and somatic sensory signals an end in itself.
Through combination of brain-computer interfaces and perfusion medicine programs, significant improvements in both lifespan and healthspan are possible. This approach aims to create better quality of life through what feels like lateral moves rather than solving anticipated problems.
Success in this field will reduce the fragility inherent in the human condition. Neurodegeneration remains difficult and requires substantial investment. Cardiovascular disease and cancer that metastasizes to the brain represent the two leading causes of death that could be addressed through this work.
Human bodies are designed for Earth conditions. Serious space exploration requires adapting humans to environments like hard vacuum rather than exporting Earth conditions. Preserving oneself and adapting oneself represent the same fundamental project of substrate independence.
The simplest premise for BCI companies involves non-invasively communicating with AI models at high bandwidth. However, the focus is not on brain keyboard applications but on generating vision, hearing, and achieving substrate independence. Language represents thinking, and the concept of special latent states that aren't language is considered unlikely.
A cognitive bottleneck exists at approximately 10 bits per second, supported by multiple independent lines of evidence. This includes studies where individuals with perfect memory viewed Manhattan from helicopters and their recall accuracy was measured. This bottleneck appears to relate to language processing.
Brain keyboard applications might be possible through EEG and MEG combinations, but represent different product categories than vision or hearing generation. The comparison is made to AR glasses where attention was already fully occupied, suggesting brain keyboard might not fundamentally change human capability in the same way as sensory restoration.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
