The Neuralink press release: a thread. 1/

The big ticket items were two-fold. They demonstrated recordings in real-time from a pig, and they demonstrated prediction of limb position from activity. These are both solid advances/milestones for them. 2/

They are not breakthroughs, except perhaps that they are doing this wireless - but that is a completely straightforward advance that was technologically achievable two decades ago. I even wrote a white paper on it....3/

The movement tracking seems impressive, but groups led by Krish Shenoy and John Donoghue did as much in the late 1990s. And Andy Schwartz does this far better, even today. However, those implants failed in translating to humans. Sort of. 4/

The implants could be demonstrated efficacious in an anecdotal sense, but total communication bandwidth was not adequate, and cost/east of use were not where they needed to be. 5/

The larger problems were not in the anecdotal displays, they were in the consistency of function. While those prior implants worked great in some subjects, the ones you see in videos, they ceased functioning completely, or mostly, in many others. The word in the field was... 6/

Plan to get all your data in the first 4-6 weeks. Sometimes they work longer than that, but that is not the rule. If you want to publish with these methods, get data right after surgery. 7/

And that is a CRITICAL aspect of function necessary to translate to humans. It is awesome if it can work demonstrably in one human. Can you make it work that well in 10 out of 10? 8/

And the next critical milestone is total bandwidth. An anecdotally good implant could control two scalars (think of this as being able to move and click a cursor). It could not move three. To be able to lift a glass of water and take a drink requires control of seven. 9/

And the reasons are pretty straightforward. What portion of the motor strip can be easily implanted in one surgery? Perhaps that can be increased, which should result in more bandwidth, but I am dubious 1024 is enough. 10/

The next bottleneck is the ability to introduce signals into the nervous system. The available bandwidth there is truly impoverished - like a few bits per second per electrode. And, introducing current causes neurotrophic effects making it tricky to perform over long periods 11/

They demonstrated they are trying to solve this problem. No advances were shown. 12/

So where is Neuralink? They've made a better implant. It is fully subcutaneous, and has demonstrated the bandwidth of 4-6 Utah arrays already, working in an ambulating pig. That's very good! But the reasons prior strategies failed have not been approached yet. 13/

What is your worst bandwidth like in, say, five attempts? That's a critical translation question, and one that sunk prior efforts. 14/

There are two parts to that - what is the actual bandwidth, and can you do it repeatedly, and both are relevant. 15/

And the other place they are working, but even all of neuroscience is stuck, is the ability to send signals into the nervous system electrically. The cochlear implant is still the best we can do, and we cannot even do one tenth as well elsewhere in the nervous system. 16/

Not even in the retina where my long-time colleague Robert Greenberg spent a huge fraction of his career (and he is much MUCH smarter and more capable than Musk). 17/

So I did not see anything that I considered a significant advance that would make this approach viable where prior approaches have failed. The reasons prior approaches failed remain ahead for Neuralink. I hope they continue to make progress - they have made a lot so far...18/

But the next steps are the ones that are make or break. Can you do it consistently across subjects? What is the bandwidth? Can you stimulate to introduce signals? 19/