Science Corp's Max Hodak on restoring vision with a retinal chip — and the biohybrid brain interface beyond it

Summary

Science Corp's retinal prosthesis has cleared a major clinical milestone. A trial in age-related macular degeneration patients, completed last summer and published in the New England Journal of Medicine, produced results that Max Hodak describes as unprecedented: the best-performing patients went from reading zero lines on an eye chart to reading the entire chart. The device, called the Science Prima chip, is a 2mm wireless implant placed under the retina in a procedure comparable in complexity to LASIK, requiring only a local injection to numb the eye. There is no implanted battery or cable. Glasses worn by the patient project an infrared laser image onto the chip, which acts as a solar panel, stimulating the 100 million bipolar cells in the retina directly above each pixel.

The technical rationale for targeting the retina over cortex is specific. Stimulating deeper in the visual pathway, at the optic nerve or thalamus, produces only diffuse, uninterpretable flashes because the signal has already been compressed. The Prima trial is the first time patients have achieved what Hodak calls form vision, the ability to intuitively read words rather than laboriously decode individual line segments. The chip has only 400 electrodes, but because the eye moves continuously, it feeds an updating stream of data into the brain's world model rather than projecting a static 400-pixel image, producing substantially better perceptual outcomes than electrode count alone would suggest.

Science Corp submitted the device for marketing approval in Europe last summer. That filing ran to tens of thousands of pages and a 65 GB PDF. European approval is expected to come before U.S. clearance, with a target market launch of next summer. The FDA process is running slower, which Hodak acknowledges is unusual. Revenue from the retinal product is intended to fund the company's longer-term program.

That second program is a biohybrid neural interface for cortical applications. Rather than inserting electrodes into brain tissue or using genetic modification via optogenetics, Science Corp grows engineered biological cells, cloaked from the immune system, and places them on the brain's surface. The cells then grow inward and form biological connections autonomously, potentially in the billions. Hodak frames this as nature's answer to ultra-high-bandwidth neural interfacing, a new cranial nerve with a biological USB port. He is explicit that this remains a research program, paid for by the retinal device business.

AI is already contributing to the science. Science Corp recently published a paper on a new optogenetic protein, discovered using AI models, that is sensitive enough to respond to ordinary indoor office lighting rather than requiring bright laser illumination. Reducing the optical power needed matters directly for safety: brain implants are thermally limited, and dimmer LEDs allow more of them, a potential path from thousands to hundreds of thousands of stimulation points. The same protein work opens a roadmap for a future product, potentially available within five to seven years, in which patients receive only an injection that makes their own bipolar cells light-sensitive, eliminating the chip and possibly the glasses entirely. The other area where AI has had measurable internal impact is regulatory documentation, navigating the hundreds of overlapping standards embedded in a European approval filing.

On the broader BCI landscape, Hodak argues the field is better understood as a longevity story than an AI adjacency story. Near-term, cortical implants remain serious brain surgeries irrelevant to healthy users with functioning limbs. The value proposition scales as bodies deteriorate. His stated marker for the technology's maturity: by 2035, offering a terminal patient the option of full neural migration rather than death from pancreatic cancer should be technically feasible, even if not yet widespread.

On AI risk, Hodak flags a failure mode he considers underrated and more probable than an autonomous termination scenario: humans progressively delegating decisions to models because the models demonstrably make better calls. He cites a reported case of a U.S. military combatant commander running personnel decisions through ChatGPT as an early signal. The concern is not a single catastrophic event but a gradual erosion of human agency through accumulated deference, with AI-influenced self-harm rates as one concrete metric worth monitoring over time.