Brain-machine interfaces (BMIs) are devices that read from or write to neural tissue with sufficient fidelity to substitute for the body's native input-output channels. Neuralink, Synchron, Precision Neuroscience, and a growing cohort of academic labs have moved the field from proof-of-concept to implanted human trials. The short-term value is medical — restoring movement, sight, and speech to people with severe impairment — but the long-term trajectory points at a redefinition of what "communication bandwidth" means.
Three generations of capability
The first generation restores function: cursor control, limb prosthetics, and speech decoding for people with paralysis or ALS. The second generation augments healthy users: memory prosthetics, mood regulation, and closed-loop therapy for depression or chronic pain. The third generation — speculative but not implausible — compresses the latency between minds and machines, or between minds and minds, to something closer to thought than typing. Each generation crosses a different ethical line.
Bandwidth as the binding constraint
Human language is a remarkably slow channel: roughly forty bits per second. Most of the economic value of Open-Source AGI bottlenecks on how quickly humans can state intent and review output. BMIs are one of the few credible routes to widening that channel. If they succeed, the collaboration pattern between humans and models shifts from prompt-and-review to something more like co-cognition, with profound implications for Education Abundance and software authorship.
Risks and open questions
The failure modes are severe and unfamiliar: surgical complications, long-term biocompatibility, covert influence on cognition or mood, and the concentration of neural data in private hands. Consent doctrines designed for medical devices do not yet cover cognitive augmentation at scale. Whether BMIs arrive inside frameworks of Verifiable Identity and strong data rights, or outside them, will shape whether they become an abundance technology or a surveillance one.