So, Neuralink is coming up with a press release this week. A commentary thread on brain machine interfaces. 1/

In the 1990s, I joined a team to make brain implants to study cortical plasticity. Initially, as an engineer, I was asked to make the neural processing end and leave the biological interface to my colleague. I disagreed. Limitations on function were the biological interface.2/

I worked on that, and we made it work. How well did it work? Put in 64 microelectrodes. Spikes on a quarter of them, good fields on half. And of the electrodes with good spikes, perhaps half relevant to behavior. 3/

This was in primary sensory cortex. Holding spikes from single neurons was possible across days but unlikely. Why? 4/

The stiffness of the neuropil is 10^6 lower than that of the metal in the electrode. Small amounts of movement leads to shearing movement in the tissue, and the electrode would move. 5/

But with good materials, and good tip positioning, you could do some science, and we did. It was difficult slow work, and I added different scientific directions as a result. 6/

The field of brain machine interfaces is moving towards eCoG grids. Why? These are sheets of electrodes you lay over the surface of the brain. Good signals on each lead every day. 7/

My friend Steve Hsiao showed the high frequency energy signals come from local action potentials underneath, and my former colleague Eddie Chang, and others, are showing you can do a lot of science with these signals. 8/

OK, back to the Neuralink approach (probably why you are reading this in the first place). They randomly insert a huge number of leads into the brain and hope to couple this to a brain machine interface. I think these electrodes ARE better than what I used. 9/

But the problems of density of neural signals remain. Random insertion has a needle in a haystack problem. Only a small fraction of those inserted will have relevant signals. 10/

And they will still be less capable than eCoG grids in the amount of information they can convey from inside the cranium to outside (a favorite measure of mine, Krish Shenoy uses it well). 11/

You can make a better brain machine interface than the Utah probe approach with enough money. You might be able to control more than two scalars continuously. You might get three. But in the meantime, the much cheaper, much easier, eCoG grids will be stealing the show. 12

They have a ready patient population to test for nearly no cost - epilepsy patients who need eCoG grids anyway. Minimal additional risk to use informed consent and study. And, already ahead of the microelectrode approach. 13/

I am not sure the microelectrode approach will ever work well. I think there are two real problems. The important one is targeting relevant neural populations with each electrode. That is - find high information targets, then implant. The second is tissue reaction. 14/

You cannot just use a huge number of electrodes. When their density gets too high in the brain, the tissue reacts (remember the shearing problem). So I look forward to some of my colleague pursuing the next gen eCoG grids and brain machine interfaces. 15/

And I am leery of colleagues only pursuing microelectrode work in animals. Not horribly concerned, but pointing your career in a direction that is not growing is always an issue for a scientist. In another few years, as the eCoG grids take off.... 16/

There will be less tolerance for the microelectrode approach. Lower cost, lower risk, easier surgery, and ready to deploy very soon in large numbers. It will be the future. 17/17