The acuity ceiling
Start with the number that matters. The best reported acuities from visual prostheses remain far below the threshold for legal blindness; the phosphene image is sparse, low-contrast, and unstable. Simulation studies suggest that on the order of several hundred to about a thousand well-behaved phosphenes would support tasks like reading or face recognition — but that word, well-behaved, hides the whole problem, because real phosphenes are anything but. Raw electrode count is therefore an optimistic proxy that overpromises.
A sampling bound on prosthetic acuity: spread N phosphenes over a field of view of solid angle Omega and their average angular pitch is about p; by a Nyquist argument the finest resolvable detail (minimum angle of resolution, MAR) cannot beat roughly twice that pitch. This is the optimistic ceiling — it assumes the phosphenes tile the field regularly.
Reality falls short of even this optimistic bound. Current-spread overlap merges neighbouring phosphenes, dead electrodes punch holes, axon-bundle streaks smear points into lines, and the warped retinotopic map scatters phosphenes irregularly. Each effect drives the effective N far below the electrode count, so the true acuity ceiling sits well under the equation's promise. This is why prosthetic acuity limits are a property of perception and physics, not merely of manufacturing.
Perceptual learning does half the work
The brain is not a passive screen. Over weeks and months, users improve markedly as they learn to interpret an unnatural phosphene code — perceptual learning with a prosthesis. Measured performance is therefore a moving target of device plus trained brain, echoing the co-adaptation theme that runs through Volume III: two learners, the person and the decoder, converging together. This is hopeful — the brain is a willing partner — and a methodological trap, because early and late performance differ so much that training silently confounds any comparison between devices.
The read/write asymmetry
Step back to the deepest reason vision is hard. The read/write asymmetry says decoding tolerates a statistical black box — you can predict intent without understanding the code — but writing a percept demands that a given stimulus reliably produce a given experience. That requires controlling which cells fire, how they interact, and how downstream cortex interprets the result. We can read the brain far more richly than we can write to it, and vision is the sharpest illustration of that gap in the whole field.
This reframes what progress means. The lever is not channel count but precision and interpretability of the write — which is exactly why the encoding models of Guide 3 and the cell-type-specific optical methods of Guide 4 matter more than sheer electrode density. The frontier of visual neuroprosthetics is control over what you write, not bandwidth into the brain.
An honest reading of the next decade
Where does an honest reading point? Realistic near-term: better subretinal photovoltaic reading in macular degeneration, richer cortical phosphene control through steering and encoders, and the first optogenetic products maturing toward the clinic. Unlikely near-term: high-resolution naturalistic vision, or reliable full-field face recognition. The field's honest promise is meaningful partial vision — navigation, object localization, large-text reading — that materially improves a blind person's life, rather than a restored normal visual world.