Science Corporation says it expects to keep improving PRIMA’s vision capabilities with a new chip and a reworked form factor for the glasses, which currently need a battery pack to run. That is the sentence I keep returning to, because it is not a marketing line. It is an architecture statement. The company is telling you, in plain language, which parts of this system it considers finished and which parts it considers drafts.
The device itself: a chip implanted in the back of the eye during an hour-long outpatient procedure, paired with camera-equipped glasses that transmit visual information to it. PRIMA has now launched in Europe following CE marking for adults with severe central vision loss caused by geographic atrophy from dry age-related macular degeneration. In the United States it holds both Breakthrough Device and Humanitarian Use Device designations. Those are the verified contours of the thing. What interests me is the shape of the system underneath them.
Perception split across a surgical boundary
If you sketch PRIMA as a pipeline, you get something that looks uncomfortably familiar to anyone who has built an embodied agent. Sensor at the front: a camera on a pair of glasses. Compute and power: on the head, tethered to a battery. Actuator at the far end: a chip sitting against retinal tissue, writing a signal into a substrate it does not own and cannot inspect. And then a decoder that nobody designed and nobody can retrain directly, which is the patient’s visual cortex.
That last stage is the part most engineering discussions skip. In a conventional perception stack, the model’s output is consumed by code you wrote. Here the output is consumed by biology with its own priors, its own plasticity schedule, and its own tolerance for noise. You cannot push a fix to it. You can only change what you send it and wait for it to adapt. Every design choice upstream of the retina is therefore a choice about what a biological decoder can learn to use, not what a benchmark says is accurate.
Why the upgrade path lives outside the body
Here is where the company’s own framing becomes instructive. The implanted chip requires surgery to place and, presumably, to replace. The glasses do not. So the system’s mutable logic — framing, exposure, whatever processing happens between photons and the signal sent inward — sits on the removable side of the boundary. The expensive, permanent, high-risk component sits on the other.
Anyone who has shipped agent systems into hardware will recognize this as the correct instinct and will also recognize how often it gets violated. The pattern is simple to state and hard to honor:
- Put the parts you expect to revise where revision is cheap.
- Put the parts you expect to keep where cost is paid once.
- Make the interface between them narrow enough that improving one side does not invalidate the other.
Science Corp saying it will improve capability with a new chip and a new glasses form factor tells you both sides are still moving. That is honest about where the engineering stands. It also means the interface between camera and implant is doing real work holding the design together while both ends evolve.
The battery pack is the specification
The glasses currently require a battery pack. Treat that as a constraint statement rather than an inconvenience. A wearable power budget sets a ceiling on how much processing can happen between the camera and the implant. Every operation you want to run on the incoming frame — stabilization, contrast handling, any kind of scene interpretation — competes with the hours the user gets before the thing goes dark.
This is the same tradeoff curve that governs on-device agents everywhere, with the stakes changed. If you move processing off the head, you buy compute and pay in latency, and latency in a visual prosthesis is not a quality metric, it is whether the world moves when you move. If you keep it local, you pay in power and thermal budget against a device sitting on a face. The form factor work the company is pointing toward is, in effect, work on that curve.
What this actually demonstrates
I want to be careful about what the verified record supports. The CE mark covers adults with severe central vision loss from geographic atrophy due to dry AMD. The stated aim is restoring some vision. Not sight restored, not a solved problem — a regulated device cleared for a specific population, now commercially available in Europe, with the company openly describing the next hardware revision.
That is a more interesting result than a cleaner story would be. It shows a perception system shipping while two of its three layers are still being redesigned, held together by a boundary chosen for reversibility rather than performance. For those of us building agents whose sensors and actuators touch the physical world, the design discipline on display here is worth studying closely. The hardest constraint in this system is not the camera or the chip. It is that one component cannot be patched, so everything that can be was moved to where it can.
🕒 Published: