Neuralink and the bandwidth argument
Elon Musk's new company wants a high-bandwidth link between brains and computers. The argument is about output speed. I think it's partly right, and too optimistic on timing.
Wait But Why published its long explainer on Neuralink this week, Elon Musk’s new company building a brain-computer interface. The core argument is about bandwidth. Our brains take in information quickly through vision, but we get it out very slowly, through fingers on a keyboard or speech. Musk’s claim is that this output bottleneck is the thing that will keep humans from keeping up with AI, and that a direct neural interface could widen it by orders of magnitude.
I think the bandwidth framing is partly right. When I explain a technical idea to someone, the idea in my head is a picture: a mesh deforming, a system diagram, a sense of where the risk is. I compress it into a sequence of words and they decompress it into their own picture, which is never quite the same. Anyone who manages engineers knows how much time goes into that compression and how much gets lost in it.
But I’m less sure the bottleneck is the wire. A lot of the slowness is that the thought isn’t fully formed until I try to put it into words. Writing isn’t only output. It’s also how I find out what I actually think. A faster link that sends a half-formed picture might not help as much as it seems.
On the engineering side, the obstacles are big. Current implanted interfaces, like the Utah array used in research with paralyzed patients, have around a hundred electrodes and let people move a cursor or a robotic arm. That’s an incredible achievement for those patients, and it’s also very far from “download a thought.” To go beyond it you need many thousands of channels, implanted safely, that last years in the body without scar tissue degrading the signal, and you need to decode what they mean. The decoding part is where machine learning helps, but you can’t decode signals you’re not recording.
Musk says the first products will be medical, and that makes sense. Here’s my guess on timing: something consumers can buy for non-medical use within ten years, with much higher bandwidth than today’s research implants. A decade is a long time, and there’s a lot of money and talent going into this.
Even if that’s too fast, I like that someone is working on the interface between people and computers from the other end. For fifty years we’ve improved the computer side, faster chips and better screens, and done almost nothing about the human side.