Elon Musk's Neuralink: AI Expert Explains

at some point when people grow up with neural links embedded in their head they'll have plenty of time through adolescence and all that kind of stuff to develop really specific neural modules in their brain that are designed to talk to this interface if you can get the language in and out directly via neural link maybe you can write a novel in 10 minutes reality isn't what we think it is so that you can just hear radio waves right so that you can hear the Bluetooth you can hear the Wi-Fi you can you can hear the television channels you you know you can hear the starlink when you're outside walking your hand you know like whatever you focus your attention on you just hear that stuff right planning for a clinic that's not something you do if you think yeah maybe we'll get to Market in 10 years I was impressed that stuff left me with a different impression of neurolink than I had after the first show and the plans for the trans Dura implantation well that is brilliant like it's a big win my biggest single takeaway from the 2022 Show and Tell was that they're they're building they're going for scale they're not timidly tinkering around with this trying to see what the possibilities are they've decided this is going to work my guess right now is AI is moving too fast and neuralink won't catch it but it could happen a lot faster you know because it's Elon time so show and tell was I looked at this and I'm like wow they're moving really fast like they're so much faster basic stuff in two years and decent stuff in five years and amazing stuff in 10 years like that could happen and that would be great if it does the beautiful thing about the neocortex is it's completely abstract in general like it's incredibly powerful learning machine my computer on my desk it doesn't want to watch it doesn't want to download anime it downloads anime because I tell it to you know it doesn't you know it doesn't care if I go to Twitter or Facebook or if I read or if I read a book for me the AI risk mitigation aspect of neurolink I think is like I said the cherry on top they're all these other things there are so many other great things that it will enable not the least of which is like in the near term it's almost dead certain to be a fix for you know this paraplegic we just know that's gonna work it's worth doing it if nothing else just for that if nothing else you know these spent a lot of time talking about upgrades right and upgrades like that's going to be a deal in some situations all you have to do is turn up the voltage right and you can overcome the insulation but sometimes you just can't do that neurolink is in a space right now where like there's so much there's fruit laying on the ground and they're just running around picking it up it'll be quite a long time before they've picked up the stuff that's laying on the ground much less have to start reaching up to branches to get it because it's just such an unexplored space it's a process and they're early in the process right now but I guarantee you there's an insane amount of Runway ahead of them they can go down that Runway and the performance is just going to get way better the craziest things that you can imagine I can imagine some pretty crazy things my guess right now is AI is moving too fast and neuralink won't catch it um so I'm you know do I think it's worth doing it's totally worth doing neural link is worth doing for all kinds of reasons that are that have nothing to do with AI alignment um so I I think it should be done but uh and it's not useless and I can see why it's an inspiration and it could turn out to the development of of neuralink it could go much faster than I'm sort of imagining because most of my expectations for how fast the technology are going to progress or from my experience over the last you know Decades of watching this back when I when I was in my 20s I I developed I designed a brain interface uh at non-invasive brain interface amplifiers right and we did medical equipment you know I mean we had to go through that whole process of like developing a medical device and whatnot and looking at that bureaucracy and how slow that thing is like that's what my expectation is based on but it could happen a lot faster you know because it's Elon time that was one of the things that I really got from the last uh show and tell was I looked at this and I'm like wow they're moving really fast like they're so much faster than my you know experience like maybe maybe we could see basic stuff in two years and decent stuff in five years and amazing stuff in 10 years like that could happen and that would be great if it does and if it does happen on that kind of time frame then then maybe it does play a role in the whole you know keeping the demon bottled up well so this whole like AI risk seems like it there's there's actually two types of risk one is like the the AI models become uh decoupled or or uh not in line with what humans are wanting but then there's also like another uh issue with AI which is like it could just be merely a tool that becomes so powerful that a bad actor wants to use yeah and then uses that against somebody so really the AI go ahead and that's the most people who think proximally like they think about like what are the threats over the next 10 years those are the threats they're most worried about right now is like this is like you know biological weapons or this is like you know nuclear weapons uh in that you know it's an incredibly powerful tool in the short run and it's not so much that it's going to go berserk and we're going to have a robot Uprising as that you know if the technology is used by I'll say Bad actors because you can have good actors in The you can have people who use these things for the best of intention and you still end up with a dystopia um um okay so Elon talks about this tertiary brain layer where we currently have our cortex and then limiting system and then neuraling could be a tertiary letter uh can you explain what he means by that and and elaborate on what you think uh the future is going to look like if they're able to make it that tertiary brand layer yeah that's a I actually I think the tertiary brain layer is a great metaphor for trying to understand the potential for what neural link can do uh you know like how do we think about the capabilities that it brings to an augmented human who's using it as a tertiary brain layer so this tertiary brain layer concept is not really applicable in the short run for people who are trying to say regain use of their limbs or move a cursor around on the screen this is the this is um this is a metaphor you can use for talking about like what the long-term plan is like what is it adding to healthy people at a point where people are voluntarily adding this so you know in in brain development this the brain is often kind of described like an onion right the oldest parts of our brain ever in evolutionary terms they're kind of like in the middle and then layers kind of got added to that so as animal skulls get bigger you know the way that Evolution would add to the brain is it would add more layers to the outside of the brain and uh so with the Advent of mammals there was a thing called uh there was a we got a cerebral cortex uh which is mostly a structure that's called the neocortex I'm mostly going to say neocortexes most of the cerebral cortex is neocortex and neocortex is interesting because of a particular because of particular traits it has so the uh the neocortex in particular it's super flexible and in fact an interesting characteristic of the the neocortex like if you were if you take a neocortex out of if you can take it out of a human brain disconnect it from all the white matter underneath it and you lay it out it's like it's a you know it's a it's a sheet that's a you know two to four millimeters thick and it's about the size of a dinner napkin right it's like 18 inches on a side or something like that is roughly the very thing and it's just a sheet of tissue now when you look at when you look inside a human skull you know the reason that you see all these folds is because it's wadded up and stuffed in your skull right so you've got this you know but it's a flat sheet of material and if you look at the structure of that flat sheet of material what you find is it's this I'll say hexagonal array of many columns like there's this structure this repeating structure that just repeats all the way through it so in evolutionary development it's a really easy thing for evolution to come up with because it you know Evolution comes up with this idea for this repeating subunit and then your brain just makes a zillion of them like they're just carbon copies of each other you know makes this sheet stuffs it in your head and then your nerves All Connect you know sort of distributed on the other side of the thing and then neocortex learns whatever it needs to learn to do whatever you need to do right which is that's a very abstract and general way of describing what the neocortex does but the beautiful thing about the neocortex is it's completely abstract in general like it's incredibly powerful learning machine and it's just like this repeating you know subunit that's repeated all throughout the neocortex okay it's kind of like a computer right uh you know computers they're arrays of ram arrays of SRAM arrays of bits um and the you know the algorithms that we run on them just they're they're built out of these simple building blocks and you just you build a lot of building blocks and you get this this super impressive behavior and because it's reprogrammable and the neocortex is totally reprogrammable this is one of its strengths it can just learn a new skill or it can add to a skill that it that that it knows now the older part of the brain is is different from that the older part of the brain take much longer to develop and it's much closer to you know an analog microwave oven or a television or something like that like it's it's built the purpose you don't Pro you don't program it so you have these two things mammals have these two things in their head we have this part of our brain that that that we've had for as long as the brains have been around um that does all the the basic housekeeping of keeping us alive you know it makes uh it makes you know it talks your your your heart and your lungs talk to it and you know it makes you hungry it regulates your body temperature like all of the super basic simple control stuff it's basically got so why do we why what was the point of a neocortex well a neocortex is basically what makes you smart right it gives you this flexibility where you can go out and learn stuff you can learn really fast stuff right so all of the all of the oh it's almost like there's there's a human and then there's an animal right like inside your body like you're a human being but your body's an animal and you cohabit your body with this animal right and the animal has all its needs and and all this stuff that it does you know but you're you know you're in this body with it using it to be your actor in the world the animal itself and so the old brain is kind of the animal part of it and the neocortex that's kind of the human part of it right although and Elon you know he points us out that like all of your desires they come from your old the older part of your brain not your neocortex your neocortex really is it's kind of like a computer right like my computer on my desk it doesn't want to watch it doesn't want to download anime it downloads anime because I tell it to you know it doesn't you know it doesn't care if I go to Twitter or Facebook or if I read or if I read a book or watch CNBC like I want to do things and it does my bidding right and that's the relationship that your neocortex has to the older part of your brain your older part of your brain wants stuff and your neocortex is a tool that the old part of your brain uses to get what it wants out of the world and it's an incredibly powerful tool like in in most of the you know aspects of things that we would consider intellectual that's all your neocortex right you're like the old part of your brain is just it's like this brute it just wants things right I mean in certain ways you know it's it's a brute that that evolved over a billion years and it's got a lot of common sense about what it takes to get along in the world right so I don't want to completely dismiss it it's got a lot of value but it's not flexible it's not super adaptable right it's like an alligator right and it is kind of like an alligator like the what an alligator has in a brain like the whole part of the alligator's brain that's kind of like what the old part of the human brain is like right so okay so now we've got this model it and you know and Elon rightly points out like you know nobody wants to get rid of their limbic system say that's what he's referring to as the old brain and nobody wants to get rid of their neocortex right I mean you want to you want to enjoy Thanksgiving dinner right uh and uh but you know you also don't want to be an idiot right you know and it's like if you want to enjoy Thanksgiving dinner need Olympics if you want to enjoy anything you need one big system because you know a computer doesn't enjoy anything it doesn't have any of that stuff built into the algorithms that you know that it's running it's a tool that a human uses and a neocortex is a tool that the old part of your brain uses to get what it wants out of stuff okay so if we add to the brain and we're going to add a thing that's even more like a neocortex than a neocortex is it's even more programmable it's got even more because your neocortex has so much IO compared to the rest of your brain like its i o rate it's just phenomenal like your vision goes into your neocortex right doesn't go into the old part of your brain you're you know or it's like in mammal brain's Vision just goes straight into the neocortex and it does all of the processing that goes uh that goes into that kind of stuff although the the visual cortex is kind of a slightly different part of the neocortex it has a slightly different structure but to a first approximation you know it's it's like everything else it is the visual cortex is a little bit specialized but it's the only part of the neocortex that has significant amount of specialization to it um so so if you add a machine another layer on top of this you kind of do to the neocortex what the neocortex did to the old brain right which is you give the neocortex an even more powerful tool for processing information for accessing the outside world right it's even more flexible and it's got unlimited capacity because it doesn't have to fit in your head the problem the limitation on human intelligence is we have to fit our skull we have to hit our fit our brain inside our skull right and you probably know that that skull size is the limiter on it's been the the uh evolutionarily limited uh in human beings for the last I don't know a million years or something like that like women's pelvises have evolved to be able to uh to have a baby with the largest possible head like the head is the constraint when you're born your half of you is head right your head's like half your body when you're born right like that's how important it is and the first five years of your life when you're growing up that's almost all brain development like you're you're totally helpless like somebody has to take care certainly for the first couple of years you're you know it's a humans like our gestation period wasn't long enough for our brain to get as big as Evolution wanted it to get your head doubles in size you know and after after you're born and that's all you know that's all development that goes on you know but we can't get any smarter inside this case what we can do is we can take the lid off and we can connect you to the whole outside world to Giant data centers or you know whatever other Hardware there is out there and on and now in human intelligence is no longer limited by the confines of our scholar you know the size of our mother's pelvises so we can we can move on from that um so uh so the tertiary cortex is basically you can think of it as a lever you know the the neocortex is secondary the primary thing is the brain you know the brain that keeps your heart beating you know the brain that where makes you hungry and keeps your body temperature at 98 Degrees and that kind of stuff and then there's another layer on top of that that has all of these other wonderful capabilities that's a secondary layer and the tertiary layer is what you get when we when you take the human brain and you plug it into the wire world sure um okay so there's another reason for people to care about what neurlink is doing it's to cure brain problems and central nervous system problems yeah um are you excited about that more so than the AI risk mitigation or the other way around no definitely I feel like AI risk mitigation is a cherry on top like that's the thing that we might get if the timing works out um and and if it turns out I mean we don't know how AI is going to develop right now we don't know that the threats I I'm actually an AI Optimist like in the spectrum of people who worry about stuff I think and and Elon has made comments along these lines too that he thinks you know that AI going bad is like a one percent probability and I would generally agree with that I don't think AI is likely to go bad I think it's unlikely to go bad in the same way that I can feel like we are unlikely to get hit by a comment but you know you plan for it because the consequences are bad if you know if it does happen to go that way so I think personally and I've been thinking about this a really long time we could talk about it for hours if you want to uh I think that AI is very unlikely and 99 is probably a good thing I think AI is 99 likely to develop the way we want it to to just be a tool that mostly just gets you I mean like you know you've always every every tool like somebody will do something bad with it but by and large I expect it to be you know a very substantial Improvement to The Human Condition you know I think that's overwhelmingly likely it but the probability the possibility that goes bad is not zero and so you definitely want to spend time thinking about it because of the potential consequences so for me the AI risk mitigation aspect of neurolink I think is like I said the cherry on top there are all these other things that there are so many other great things that it will enable not the least of which is like in the near term it's almost dead certain to be a fix for you know this paraplegia quadriplegia sort of phenomenon where you've got a significant number of people who have serious brain or spinal cord injuries like you know that we just know that's going to work right and that it's worth doing it if nothing else just for that if nothing else sure so kind of this like AI potential risk although it has a one percent chance or or you perceive it to have a one percent chance uh that one percent chance is is a huge issue if that happens yeah the the brain illnesses and spine problems like those are definitely a hundred percent they exist and 100 chance that they are soluble yes uh given enough time so and enough uh and not even a lot at this point right I would say you know neuralink has already demonstrated the core you know capabilities for being able to do this there's the question I mean efficacy has been demonstrated right like you will be able to provide useful capabilities to these people safety needs to be refined and it needs to be proven and so that is just going to take time right and you know they've got all this lifetime testing they have all this bio compatibility testing they need to refine the surgical procedures the surgical procedures really it's as safe as it possibly can be right and the the you know they spend a lot of time talking about upgrades right and upgrades like that's going to be a deal that's that's going to be a significant issue not the least uh because the you know the the tech that they're working on for the implant right now it's not the kind of thing that's going to last 50 years like even if you put it in somebody it's going to stop working in five years or something like that and so you you know you need a robust follow-on like what do you do after five years is up and they need to figure it out figure that out and that's a critical part of the safety aspect of it right so uh so there's work to be done on safety I feel like for the for the you know the worst the people in the worst case scenarios right now the efficacy is already sufficient I mean you've um is it the Utah array uh the brain gate stuff right they've already demonstrated you've got utility right it just it needs to be refined so it's safe it's repeatable so that the lifetime is long enough that the risk reward um you know the cost benefit of doing it can be improved tremendously and it will be improved tremendously but you know we know the values there the question is just like how much value can we get out of it can we get a million dollars out of it can we get a billion dollars out you know like because the sky is the limit you can go really high if you if you get the reward up high enough and you get the risk down low enough then it does become like Lasik and anybody can get it and uh and there are all kinds of you know as it gets safer and as it becomes more capable more and more there's so many you know there's so many people in the world who have whose lives could be improved if you could Tinker around in there a little bit right like we know there's a problem we kind of know how to solve it um but we just don't have access and so for a lot of people like depression right is is one I mean depression is you know it's a pretty broad spectrum of there are many many different things that cause depression but there's a space of things that cause depression which are addressable by you know this kind of implant and you know so it'll be and depression is a serious problem right not only do lots and lots of people have it there are lots of people who really suffer from it and where it it makes it just a massive difference in their ability to lead their lives so I I personally have a decent understanding of how conceptually neuralink works but can you explain like just at a very high level the concept of like putting the probes next to neurons and and all these very simple things to somebody who's never heard of a Utah array deep brain simulator or a neural link yeah so I mean uh we've known for a long time your your brain is it's neural tissue it's mostly neurons and glial cells right so there's there's some structural tissue in there that keeps everything in place and there's um there's a vascular Network which delivers uh nutrients and oxygen and it takes away waste and whatnot but aside from that most of your brain is just neurons it's a lot of neurons a neuron is a cell it's got a cell body it has inputs and it has an output and it makes a decision um the uh neural tissue so neurons are extremely Dynamic cells they uh they are able if you get a lot of neurons together they can express a complicated Behavior by connecting to other neurons so like a single neuron it's got some inputs and some outputs and does some simple processing if you take a bunch of neurons and you connect them together and you connect them in the right way you can get much more sophisticated Vapors the more neurons you have the more complex a behavior you can get and essentially all of the behavior expressed by a human being or a mammal or most most creatures that have brains it's their its consequences of just the connections the connections that these neurons make right the neuron the networks of neurons in your brain they can learn right and so that's the thing that that um that is a really critical characteristic of nerve tissue right nerve tissue has got this adaptive capability it's got this very high level information processing capability that most of the rest of your body doesn't have I mean you can talk about whether you know gut biomes and that kind of stuff they have these complicated feedback networks and that kind of stuff but for the in your immune system is a fantastically sophisticated information processing system but most of the behavior you get from you know mammals humans is is is nerve tissue now individual neurons um this is a cell that has electrical activity like it it uh the way that it propagates a signal through it is it is is this it's essentially electrical the the the a neuron sets up a charge difference across its cellular membrane um that charge difference is kind of like stacking dominoes in that you can have an effect you can have something happen in one part of the cell and they're they're kind of tree shaped you know you've got a cell body you've got like these wires that go out the input wires are called dendrites and then you have one and out typically one output wire called an axon and uh and so the cell body essentially what it does is it collects inputs from these dendrites makes a decision it does its processing and then uh and then it fires the axon in When Its Behavior is triggered right so and the way these signals travel along the wires is they're a lot they're electric potential chemical gradients across the cellular membrane which is kind of stacked like a bunch of dominoes so you and the reason that you want to do this is because uh is because nerve cells can be quite large and they want to they want to make a decision at one point and send the result of that to another part of the cell really fast compared to other things and the mechanism that nerve cells came up with is they trigger this gradient that propagates down one of the either an acts either a dendritic uh the dendritic tree into the cell or out of the or out of the you know the output goes along the axon so it's electrical like there's rather it's these aren't like wires in a in a in a car where you know or a house where you just have a piece of copper you know and you put an electrical potential on it and you know electron shift and so there's like a current running through it but it is electrical in the sense that that it's this it's information transfer which is mediated by this electric potential across it so we have this tissue that's made up of on you know 80 billion 100 billion neurons inside your head and uh the information like what your brain does is essentially the behavior that it expresses is captured in the interconnection between all these neurons but the individual neurons we can externally trigger them we can trigger an input to a neuron or an output to a neuron and we can tell when they're triggered because we can see this electrical potential propagating along say an axon or even at the or even at the neuron body so uh the way you stimulate it is if you have an electrode if you embed an electrode in the tissue and it's relatively close to a neuron if you pull a current through the electrode you will you'll generate an electric field local to the tip of that and if the electric field density exceeds a threshold it can trigger an ion channel in a neuron to trick and then that set of dominoes Cascades goes on so essentially you can run a current into some nerve tissue it can you can cause a neuron to fire so when you cause a neuron to fire all the other neurons that take it is an input get it get a signal and so that signal can Cascade on through the brain and it becomes it's just a signal and you've injected it into the information Network similarly when when a neuron fires it creates this thing called an action potential if you've seen the plot in fact this is what the neural link the neural Link logo is right it's an action potential right well if you look at this particular point on you know a nerve fiber and one of those nerve signals propagates long well that's what you see you see this wave it propagates down and a probe that is next to that wave it'll see that action potential shape you know as that wave propagates past a certain point so the pro so a probe will let you read a signal from a neuron and it'll also let you trigger a neuron right and so these are the two things we get we we you know we can monitor neurons and we can trigger neurons uh and so you essentially a probe is you're taking you're basically sticking the tip of a small metal wire into the brain tissue when the way that the probes get installed right now um you know they basically take a wire each individual wire has um 16 probes on it so it's not the the you know a probe that goes into the nerve tissue right now it's not a single probe it's actually got 16 wires embedded in a common insulating sheath and each one of those wires comes to essentially if you look at the tip of the probe you get these 16 little dots along the edge of it and each one of them is the Terminus of a different wire so when they put a single probe in you actually get uh 16 different it with the current probe design you get 16 different locations where the tip of a wire is protruding into the brain tissue and at that location you can monitor the neurons that are in the environment and you can trigger them by pushing the current you can potentially trigger them so right now the thing is brain tissue is so dense that you can just take a probe and you can stick it into the brain tissue and you will be close to a whole bunch of neurons now some some of the tips will be really close to them some of them will be farther away because there is all this structural tissue there's other tissue in there aside from just neurons and it's possible when you put it in that you might be close to the axon of a neuron but not close to the nerve body you know I mean there's different parts of a neuron that you might be close to so the characteristics will vary a little bit and you might be close to two or three neurons and so what your probe is actually reading is some some you know uh it you know these essentially these different signals are interfering at the probe tip and similarly when you trigger uh when you send a triggering input current into the thing you might be triggering multiple neurons but it's a a characteristic of brain tissue is that neurons in order to keep the wiring short functions that are closely related are very closely positioned inside the tissue so um right now the probes that get inserted into the brain are um they they go they you know so they're they're plugging into the neocortex right now and the mammals that they plug into as we mentioned before the neocortex is like this it's a sheet of tissue it's right at the outside of your brain so it's really easy to access and you've got it's it's mostly made up of these repeating units that are called mini columns or micro columns where each one of those is like um it they're they're functionally very similar they're all very very similar um they once they've got installed and you train the brain up how they connect will make every mini column unique but the the original structure of the like when you're born right they're all about the same and uh they're about the diameter of a probe right now they're 30 microns or 40 microns wide that's approximately the width of and then the the you know they're they're called columns because they're like I said 40 microns wide but then they go through two three four millimeters thick layer they every mini column is the full height of the of the of the layer of tissue so when a probe goes in right now it sticks into a mini column right because that's because that's what we have and you've got these 16 you know surface things that that are sort of they're bumping into different parts of the of the mini column as they as it works its way through the tissue and so uh when you know a particular probe is talking to one or a small number of many columns where each mini column is kind of a functional unit right it's got like a thing it does and so you can trigger that mini column to cause it to go and so when the other mini column that it talks to or that depend on it they they're like oh that guy fired right and they take it as an input and they run with it does that answer your question it does uh I'm gonna show a video the monkey mind pong demonstration back in 2021 included this blog post and then they show the specific area in the cortex that they implanted these Electro threads and then they they show like exactly what they're corresponding to um so this like basically neuralink is building on top of a lot of Prior research they're basically doing similar to what the Utah array and deeper in stimulator as functions were but then the Utah ran deeper in simulator nearly as information dense or or able to provide you with this level of detail as what neurolink is doing is that correct yes and there's a couple of reasons for that one of them like a Utah array is um it's a you know it's a hundred probes it's a 10 by 10 grid that's fabricated as a single device that gets inserted it's physically it's relatively large so the individual probes are they're separated substantially in space because it's just physically larger and then they're all fixed it's a rigid array so when you put the when you put it in you know the probes go in on this grid and uh so you you don't get to avoid vasculature right you get one sensor at each one of those point and they're they've got to lay exactly on a rectangular grid because it's a because it's a rigid thing and I think the probe tips are actually larger too so you know so one advantage of this is is you can get a lot more probes in um they use a 64 threads with 16 like the current in version that they were showing us in the 2022 Show and Tell is um you know 16 probes on 64 threads for a total of 1024 total probe so like obviously a thousand is better than a hundred also because the because the individual probes um you know the robot basically can look at the brain tissue and you can pick better in worse places so of the probes they put in more of them are going to be useful because because you're choosing the most fruitful places to put them so that's another big advantage over a Utah Ray um the the so another big problem with anytime you introduce anything into the body the bot you know depending on the degree of biocompatibility the body will have a response to it like your immune system will respond uh your you have a healing system also that that you know it fills in air gaps and that kind of stuff and and uh so a probe that you inject into any tissue in your body your body will have a response to and the brain is no exception to that so the brain will form a fibrous isolating sheath around anything that it just that it you know kind of decides is a foreign object embedded into it and the problem with once it forms that sheath is that becomes like an insulator it becomes an obstacle to the probe doing its job so a not uncommon situation that you have is you stick a probe in the brain and on day one it works great but you know a month later two three months later once the brain's healing response is like it's put in insulating sheath around your probe now so your probe is still in there but you know you're not getting as good uh interface to the address like to in some situations all you have to do is turn up the voltage right and you can overcome the insulation but sometimes you just can't do that and because what will happen as you change that voltage the shape of the electrical field around the end of the probe and how that affects neighboring neurons will be a thing so like ideally you want to put you want to have a really good connection what they call lowsey or low interfer you don't want to be a short but you want a low impedance connection so that when you're so that you're very sensitive when a nearby neuron Fires for the sensing thing and when you want to trigger a neuron you can use a very small current the more current you have to use the more tissue you're going to activate so instead of activating one or two or three neurons you might be activating 500 or something you know which it's a much more confusing signal that's harder for the brain to deal with and then similarly when you're trying to read stuff it might be much harder to read a single neuron's Behavior right so you so essentially it's just cruder if you if the if the brain doesn't accept it so as the now biocompatible materials are materials that are less likely to provoke that healing response I mean the ideal material you stick it in the tissue and tissue has no chemical reaction to it at all Now with an electrical probe there's some challenge there because you will get electrochemical reactions if you just run electricity into tissue so it may be that getting perfect acceptance maybe that won't maybe that won't happen ever but certainly the smaller the object is because one of the things is your you know your immune system it notices things because it has an immune sensor cell bump into some kind of surface and that triggers a response in that tissue the smaller the surface of the probe is that you put in the less likely you are to trigger a response and so at some level like if you could make like if you could get the wire all the way down to like a single molecule of width you'd be you know almost none of them would trigger any response from tissue at all and so right now they're actually I don't know what their current thread diameters and other needles they said their needles were 40 microns the thread must be significantly small but if your thread is like 10 microns or something you're getting you're getting really far down and uh now if it's 10 microns it has to be flexible because it's you know for long rid long things you can make something out of tungsten which is an incredibly hard rigid material but if you make it 10 microns you know if you make it 10 microns wide and you know a half inch long it's going to be really flexible right because everything is flexible when it has that kind of aspect ratio so so that's great it means on the one hand you can use almost any material you want you can even use crystalline materials as long as you the bend radius isn't very tight it can still be a fine wire because it'd be flexible enough if you made it thin enough right as like they were talking about using amorphous silicon carbide as an insulator right I don't know how they're going to deposit it on the wires but silicon carpets really hard right like silicon carbide it's like diamond you know like you silicon carbide drill bits they like cut through steel like with butter it's incredibly stiff and brittle stuff yeah and yet but you can use it as a insulating material and a flexible wire if you make it thin enough right sure um okay so Elon talks about how a deep brain stimulator is kind of like uh how like he he used the metaphor of back in the day if the TV wasn't working and you had this big box top TV that you could just yeah you just smacked yeah and a deep brain simulator is like that right that is kind of what's going on it's it I mean you've got like I'm not an expert in this space right my understanding of a deep brain stimulator is you've got you have a you have a a you have a set of neurons that is misbehaving and so what you do is you smack it to get it to shut up right and then it shuts up for a while and so like the the the problem that it's solving is that it is that it's it's triggering the this this set of neurons out of cycle to to just break the bad behavior break the cycle of this bad behavior and like you know that's not ideal it it would be you know obviously you'd like to have a better intervention or you just figured out what the root cause was and uh and solve that but but for people who have problems that are addressed by deep brain stimulation it's way better than nothing sure uh and then likewise with the Utah or both of them uh have like this device protruding out of the head and and neuralink does not it's wireless and that's like gigantically forward yeah so it's kind of amazing that the wireless stuff has taken so long to come to that space because you know I mean so this is a thing like you you need microelectronics both because you need to make it small and you need to get the power requirement down really low if you want to go Wireless and microelectronics they have to be customized it's going to be a custom silicon design and that kind of stuff and building even a single chip can be pretty expensive and pretty complicated and so my guess is that these you know DBS systems and uh that DBS systems and Utah arrays they just like the volume has been too low to justify the development and it's nice that you know neuralink basically started with that I mean I know they they had some earlier stage stuff where they they built a couple of things that were hardwired and whatnot but going to custom silicon so that you can go Wireless like that I mean it you have to do it there's you just can't you know no no reasonable long-term therapy is going to have people with like a box glued to their skull and a wire hanging off of it that they have to be Tethered to so I guess one of the things that I've found striking is that with Tesla like in the early days of Tesla Elon in his interviews was often talking about how like they're gonna be able to get to like batteries that are much denser than current batteries and um like the manufacturing of electric vehicles or just vehicles in general will improve over time but it was never like or I had never heard him say like oh here's the current state of the art version of what we're working on and we're gonna just 10x or 100x easily versus at the end of this 2022 update when there were questions about like okay how does this compare to like a deep brain stimulator well the head neurosurgeon Dr Matt McDougall was talking about how you're just sticking in this gigantic uh probe into the brain and hoping you don't really know important yeah yeah and so elon's like hey this is a really low bar for us to clear yeah and moving forward like they're going to clear it way way further like the goal I think is to get to nanoscale uh threat threads and electrodes I mean I think that's worth considering when you compare um you know for instance your battery example is like if you start with a with a relatively mature technology that already has a lot of development you know it's harder there's not as much low hanging fruit you know in terms of things that you can do that will make you know radical improvements neurolink is in a space right now where like there's so much there's fruit laying on the ground and they're just running around picking it up right and it'll be it'll be quite a long time before they've picked up the stuff that's laying on the ground much less have to start reaching up to branches to get it because it's just such an unexplored space yes so yeah the the pace of improvement if you look at the Core specs like the channel count sensitivity like they they didn't spend a lot of time talking about you know the the chip itself right now it's you know it's an eight it's a analog the digital converter it's a digital to analog converter it's got one of those on each of the channels they Multiplex I mean they don't I mean they've got a set of moxes they've got a set of adcs and dacs that they actually generate the signals with and then they have some really really simple onboard electronics that basically you know interprets the signals filters them you know tries to look for Action potentials or spikes right and then log that with some very with some very simple uh um you know algorithms and uh to to sort of extract this stuff I mean they're super super such early days because like I guarantee you like there's layers and layers of better that you could do as these algorithms go but it's going to take a while to get there like you know you you um it's a process and they're early in the process right now but I guarantee you there's an insane amount of Runway ahead of them that that you know they can they can go down that Runway and the performance is just going to get way better and that that's not even they're looking at a 16 000 Channel EA one simple way to just get a lot better is just have a lot more channels um the the odds that you're that you're talking to exactly the right neuron they get higher the more neurons you're talking to um you know you get so much more statistics from from your uh so you're I mean your ability to extract subtle patterns out of neural activity like the more probes you have like that that's just going to get a lot better and you know I I S I see no reason why they won't get to millions of of probes and and that's probably where it starts getting real where um you know you don't need you're not focused on really simple heuristics and algorithms and stuff to get your signals out to inject signals in you can start looking at it you know in a big data kind of format going for much more abstract much more valuable and information dense interfaces to the brain like so right now just you know moving a cursor around on the screen that's a it's a it's a lot to somebody who can't do it but but it you know what would be a lot better is if you could you know Express an abstract concept right and the neural network outside just takes that you know pipes it into uh into a universal mimetics base so that a computer can work on it or it can be handed to another person and you don't even have to explain something to them they just know you know you form the concept in your mind it gets transferred to an outside thing you if you have enough electrodes and you have enough processing like that's a thing that becomes possible okay I guess I wanted to ask you about like the longest term things that are like the craziest things that you can imagine For What knurling could enable in the future are I can imagine some pretty crazy things but I'd say like the the uh so here's like this is a really simple one but this is a thing that I think you get to at some point like way down the road which is uh it becomes a way for us to slip our mortal shells and move into another substrate foreign we're on a biological substrate right now right you're there in your body right with all of its limitations i o limitations like immortality right you know uh your your your intellectual like once you get outside the meat once you can step away from that you know now you know all of the constraints that your physical body uh impose on you they potentially Fall Away lots of them do anyway so like if you want to travel Interstellar distances like that's a problem for human beings today because you got to be in this can for like 200 years or 10 000 years to get to the nearest star depending on what the technology is that gets you but you know if you can just turn your clock rate down what's 10 000 years you know you can just do that wait it out um so like to me like that's the that's kind of the craziest obvious thing right is is it you know it it opens this pathway for people to be able to to get to to get outside so there's a couple of other ones that I'm running out on a list that some of them we talked about um one of them you talked with Dave Lee about um but I get I get so I guess the first one is you said like you could slip your your current substrate there's kind of this debate whether or not it makes more sense to terraform Mars to be more livable for humans like they are on earth or like adapt humans or modify or have like I don't know just change ourselves to be better suited for Mars which one do you think is more appropriate like I think that changing People for Mars is going to happen way faster than the changing it the and the one real I think I can't remember if I mentioned this before but uh there are Technologies that are physically possible that would allow you to terraform Mars really fast um it's just uh we don't know how fast those things are going to come to fruition so we could get really surprised and it could be that you know in the next 50 years or so 100 years like we develop significant capacities in those Technologies these are like you know self-replicating Micro Machines kind of the gray goo as a tool or um you know vacuum phase manufacturing tools type stuff where you can build arbitrarily complicated molecules um completely from scratch to the point where you can build molecule size robots that can self-replicate and so a relatively small number of those things can have an arbitrarily big impact because they just grow exponentially once you dump them onto a substrate if they can replicate in the environment right so these are like micro scale Von Neumann machines so in principle like you know the you know that the stuff has been worked out from first principle and it's possible to do that we know from an information theoretic standpoint we know from we know that thermodynamics allows it we know that chemistry allows it uh we know that you can build these molecules we just don't know how to do it right now we know they can exist so that's a technology like we don't know when we're going to get it and if we did get it suddenly then terraforming Mars over a span of you know 50 years it suddenly becomes something that that becomes possible but absent that you know they're you know terraforming Mars in less than I don't know a thousand years or something like that is pretty challenging and it's pretty hard for me to believe that the human race survives and yet doesn't uh develop the ability to extend our biological you step outside our biological elements like it shouldn't take nearly that long to do that that kind of thing and the the deal is Mars is just really big right so if you don't use a self-replicating technology of some sort to tackle it then you have to build all the individual things to modify this planet size mass and that's just an insane amount of industrial of capacity it people have talked about you know what if we do you know macro scale Von Neumann type machines to do that and that's another possibility but that that you know that's also a pretty does macro scale Von Neumann machines you know you uh you need to be able to build a robot that you can drop on a planet that can make a semiconductor fabric you know Fab and uh like the there are a lot of core technologies that go into robots right now that are already pretty much down at the molecular scale like you know five nanometer and you know long before we're doing this there will be one denominator ICS and then there will be 0.1 nanometer ICS and now you're starting to get you know chip feature sizes which are getting down on the order of a molecule and the way that we build that kind of stuff right now is just has an incredible amount of overhead Building self-replicating Systems that can build semiconductor Fabs it's certainly possible it's just not going to happen soon um whereas you know being able to modify our biology whether we do it directly or whether we step outside the biology by um you know interfacing biology to machines and using those machines to extend to extend our capabilities um you know my guess is like like the runway on that is more visible like that's a 50 100 Year kind of thing whereas I don't know if the self-replicating um semiconductor fabric kit uh Fab is is a 50-year thing or not might be I see um okay so in your conversation with Dave you talked about like what it would be like to watch a movie without an erlink and then watch a movie with the or basically like get all of the same feelings that you would have gotten without watching with an earlink but you just get it downloaded or you just get those feelings streamed to your brain you don't have to watch the movie you just uh you just it so this is a little complicated right so um you know the the whole Neo downloading Kung Fu kind of thing unfortunately that doesn't actually work um because the thing is um to get Kung Fu in your biological brain the neurons in your brain have to rearrange and there's a learning algorithm that they use to do that and it takes time and information exchange and feedback right so being able to reach inside someone's head and just rearrange all the neurons so they use Kung so they know Kung Fu like that's not something interface to the brain now if you have a silicon substrate inside your skull or external to your skull that you use in part to manage your body well you can download Kung Fu to the external thing pretty quickly like that's a thing we can do with computers is you can just dump a bunch of stuff into their memory and they can reconfigure themselves like really really fast unfortunately biological neurons to the extent that the biology itself is doing the learning that's going to be rate limited by the Learning algorithm that your neural tissue uses so that I mean that's a critical limitation on it straightforward implications of what we get from neuralink so um now let's say you want to download French right uh okay reconfiguring the neurons in your head in your Language Center or actually it turns out when you learn a second language it doesn't end up being confined to your Language Center it ends up distributed all over your brain but uh reconfiguring like whatever spare neural tissue you have that isn't desperately needed for something else right now so that you can load French into it like there's a learning process you know you can go learn French as an adult you can learn it really well if you spend a lot of time on it um but you've got all these neurons in your head that reconfigure to imprint French onto them all the rules and that and and you know um in your language center right at the bottom of your Language Center the input you know you've got sound coming in and then you know there's stuff that takes sound and breaks it down uh you know into bits and pieces that get turned into words and all the other Concepts that you have and as you work your way up that stack there's a point in your language stack where the idea is represented right in other words you start with French and the sentence is said and you work your way up the stack and at some point you get to the concept which is being expressed in this leads and then you go you can you know that becomes an input to other parts of your brain that like take the idea manipulate the idea regurgitate it form its own idea maybe and stick that in the language stack where it gets all the way down to the motors that move your lips and that kind of stuff so that you can say the thing that you want to say in response but there is this layer where the concept exists and it is it's not entirely divorced from French but it's largely you know now it's a concept it's not bound to that kind of stuff if you have a neural link that plugs into that layer right then the input is a concept and the output is a concept so if you have that layer and you have an external box that knows French right then you can you know there's this potential This is complicated right there are things that you have to do that might not be easy or might not even be possible in order to like to do a simple you know implementation of this thing that I'm describing but this is an approach right you plug into the conceptual Center so the outside so the outside system here's French converts it into the concept and then you get the concept but the thing is your brain already has the concept layer already because we use it for English so you have a pre-existing structure in there that can understand kind of an arbitrary language once it gets converted to the concept phase and so so the way potentially you could get French really fast is to uh is to is to have all of the specifics of the of you know it's essentially skip the non-french part of your brain you you know do that in an external system and then have that system plug into your brain at the point where whether it's French or English or German or Hindi or whatever doesn't matter right the experience of the person using the system is they know French right they hear French they know what it means right and they want to say French they imagine what they want to say and the external system converts that into French and says it so like that's doable like wait from a first principle standpoint that's doable and it doesn't require much adaptation in the in the biology so it's not going to be nearly as rate limited by the Learning apparatus and and it's more General right so this isn't like because you're getting a concept input directly into the brain you know if you decide you want to switch to Russian you know you just the external system flips over to Russian because the interface to the concept layer is the same like once you've developed that interface then uh then you get all the languages you know or you can skip the language like if you if you have an interface directly to the concept thing you can also take the concept out so you can take the concept out of your head and inject that into somebody else's head after it goes through translation layer and stuff like because everybody's concept you know Hardware is going to be unique to that individual they they have a common mapping space but um but they're not identical so you have to do a translation thing but you could work at the concept level having a conversation with another person so that's kind of like that gets to what Tim Urban and Elon talk about where communication and words are lossy like if I have an idea and I have to compress it down into words and then spew it out to you and then you process it and then re redisemble it then we lose so much information yeah and there's you know there are other challenges too the the the language is it it's highly compressed it is error corrected right like there are all these aspects of language are designed to compensate for noise and the signal and there's these other feedback channels like you look at someone's lips and that affects your interpretation of the phonemes like if you can see their face you look at their expression you get a sense of like how they're like people respond to what you're saying and that's actually part of the give and take of face-to-face conversations then you have a dictionary compression system like there's a part of your brain when you're having a conversation with somebody that is like it has an idea of what they know and that help that helps inform how you want to explain something like the way you talk to your mom who doesn't know how to use a computer versus your little brother who knows how to use one really great about a particular topic related to computers might be really different because the dictionary that your brain is using to like figure out what set of Concepts it can easily convey and which ones are going to be harder it's different for different people and for people in different cultures the dictionary can be really different right so our language apparatus has to use that dictionary as part of the thing and the because if you use the dictionary it greatly reduces the amount of communication you have to do to convey a concept right it's it's a Dictionary lookup of an internal state for your for the person that you're that you're interacting with and it lets you know like what's the most compressed way to express a particular thing so it's a huge Advantage but if you're having a conversation with somebody who you don't know their dictionary it's a big problem now all of a sudden your language is incredibly sort of cumbersome um the thing is like one of the things is if you can get to the concept level you might be able to skip the dictionary right so you can have a meaningful nuanced conversation with somebody who has nothing in common with you like no common language no common cultural expectations right so there's another idea that you had shared with me that um is like what if you pair GPT for or open ai's chat GPT with a neuralink then what what does that lead to what implications does that have yeah I think it's I mean the that Tech is going to be so Good by the time neuralink is available the the big limiter there is so you know there's moving a cursor around on the screen but ideally to get to higher bandwidth and yet plug into a symbol system that human beings already that we know is present in there because at some point when people grow up with neural links embedded in their head they'll have plenty of time through adolescence and all that kind of stuff to develop really specific neural modules in their brain that are designed to talk to this interface and can be pretty rich but when we're implanting these things in fully formed adults who want to be able to use it in less than 20 years um what you need to do is have the system adapt to structures they already have present so that they can use that so for instance like I words are a concept that I have that I can work with right now we know they're in there someplace right and so if you can plug into the part of my brain which can express words then so so that essentially you plug a neural Link in and so it you know if I'm if I sub vocalize a word because I'm going to say something or I think of a word that I want to say I think of a sentence right I'm like my brain is making the language and then at some point as it works its way down the stack before I say it or write it um you know it's present in I'm going to call them words right or a set of tokens you could plug that into a model that understands stuff at that level and the bandwidth goes way up as soon as you do that because now instead of moving a cursor laboriously across the screen to select one of 64 boxes I now have 8 000 or 10 000 or 50 000 unique tokens that I can instantaneously sample and feed one right after the other so that the bandwidth goes re up a lot so if you'd asked you know five years ago well what would you do with that you could say well I could type a letter really fast right or you know but one of the things you might not have thought is like well you know I could give really quick instructions to my Optimus robot so fast that I'm almost not even thinking about it it's just doing what I want well what do you what you need for that is a language interface to an optimist robot or I mean chat jpt is pretty interesting we were talking about you know how amazing Chad gbt is the interface is just words right so if you can generate words quickly and efficiently and you plug into something like chat GPT which is you know it's an it's it's an amazingly capable system for taking natural language and and using you know you know taking that out into the world manipulating in really sophisticated ways and then providing you feedback also in the form of words to respond to your request right like you know it's like I was writing uh I was writing some code with it the other day where I was asking it to write this various numpy programs for me and it's amazing you know you're just like I want to write this program and I wanted to kind of do this and that another thing boom you get this block of code and it's not perfect but it's pretty close like it's so much now so if that if that's a direct high bandwidth interface that you can respond to really quickly like that starts to become really interesting and and that's something that's almost for sure doable like uh you know it'll neural Link's not at the point right now where you're going to be able to plug it in get words out of somebody's brain or inject words into somebody's brain but the promise is there it's very likely to be possible to do something like that it might take a significant I mean in the beginning it'll certainly there will be a long learning curve for people who are trying to use that and as the software gets more sophisticated there will be less and less of a learning curve um but you know they're like there's this this potential to have a language kind of interface because the thing is we know language is present in the brain we know about we know approximately where it is so we even know approximately where to put an interface to get access to language in the brain and we have these AIS we're building that process language and they do amazing stuff with it so you know it seems really obvious to me that that you would you would take the language out of the brain and stick it into this language-based AI one of the ways that I've been thinking about knurling uh for for these longer term aspirations is it's basically like you could you could have all of the internet like the cloud immediately accessible yeah in your brain so then like with this whole language thing like if I'm going to a place that I've never been before I don't know the local language let's say I'm going to China I had never spoken Chinese then I could like pre-download the Chinese language package and then and then be able to reference any of the words that I'm thinking in English like it's just pulling the Google translated version immediately and out and outputting does that sound feasible is that yeah I mean you don't need neuralink for you know for for that and we'll we'll you know we'll have that without neurolink it's a simultaneous uh Peril you know parallel translation is is uh that's a technology that exists and it's getting better really fast um you know I mean it your phone can do it right it's uh people it's been demonstrated a bunch of times as the As It Gets a bit more refined and a bit easier to use I think you're going to see that as a universal capability so yeah you could do that with neural link um the things that I think are more fun to think about are the things that like you can't do them without neuralink now the the thing that I was just describing with chat GPT you know say you know uh or like whatever future version of that technology exists by the time neurolink for uh you know neural link for language centers is uh is a is a product or is is a technology that's available to people uh that like you can obviously do that with your phone too right because you know you can open chat GPT and use a speech interface to talk to chat GPT and you can get the responses back from the the interesting thing is like how much tighter that Loop could potentially get with uh you know if you have a direct neural interface and the and the potential is you know kind of impressive for you know I mean you could I a human being using a voice interface say say you wanted to use chat GPT to write a novel and people are doing this already you know where they write a novel they write a paper or something like that they're tricks to using it you know you get the output from it you decide to make a change you have a slightly different request and you stick the stuff together you go through a process of editing and whatnot but you can you know you can generate um sizable you know books in really short period of time with some level of quality right now and in the future you'll be able to do a lot better than that right but if you have to do all of the in the back and forth all happens with typing on a keyboard you know it you know if you wanted to write the Great American novel it's it might still take you days or something like that of refinement that kind of stuff but you know if you can get the language in and out uh directly via neural link maybe you can write a novel in 10 minutes right maybe maybe it takes an hour like that would be that's kind of a really impressed like a human being actually writing a novel in the span of an hour right like faster than you can read a novel I'll say yeah will you be like somebody else could have the the feelings that they got from reading the novel yeah there's that too that requires that like there's some really cool things about going higher in the stack when you asked earlier I think I digressed away from the whole the conversation with Dave about uh about you know at what level do you want to receive the movie right do you want to receive the movie by watching it you know shall we figure out what the photons are and then stimulate the receptor receptive centers in your visual cortex so you see the movie or you know because the thing is your visual cortex converts that into a set of perceptions now if we plug into the perceptual Parts maybe we just give you the perceptions directly so like you feel like you're saying like all the concepts then it's kind of like a dream like dreams kind of have this very non-specific thing like you know what's going on but nothing really exists until you look at it closely in a dream right because it's happening at kind of a conceptual level inside or you can go even higher than that what's your emotional response you know and that that's actually that's kind of deeper at that if you want to really do the emotional response it might be necessary to go down into the limbic system and stimulate some of that that stuff directly I'm not sure so certainly like if you want to smell food you have to go down into the limbic center right because your olfactory senses go straight into your animal brain but uh yeah you know it's reality isn't what we think it is right reality is the set of perceptions the way there's a bunch of Hardware in our brain that takes uh it takes what our senses detect from The Real World it builds a set of perceptions and then our experience our you know qual our qualitative uh experience of reality is inside that illusion which is being constructed from our senses right and so you can plug straight into that and skip the whole lower levels of the i o so we're kind of limited by our human biological Hardware to sense and perceive the world but in the future like we could have advanced sensors that are superhuman and we could have senses that we don't currently have and then process those with the neuralink right so like I think you had mentioned infrared vision and I don't know like hearing hearing a decibel ranges that we don't currently hear yeah and for sound ultrasonic yeah are there other things I mean one of the coolest things to be able to hear like if you wanted to add something is be able to hear if like have have uh software-defined radio so that you can just hear radio waves right so that you can hear the Bluetooth you can hear the Wi-Fi you can you can hear the television channels you you know you can hear the starlink when you're outside walking your hand you know like whatever you focus your attention on you just hear that stuff right and how about like perceiving uh elect or magnetic fields sure yeah I mean people have done that one that that was uh the um people have done it a couple of different ways but there was one that I thought was kind of interesting where somebody they built a a bracelet is actually to go around your ankle and it had um it has a set of 16 little vibration uh thing you know kind of like you know how your iPhone can vibrate or whatever the deal is so it's got 16 of those arranged equally around and you strap it on your ankle and what it does is you get a slow tapping sensation that's aligned with the magnetic field wherever you're at right so you just you put this on your ankle and you walk around with it right and what happens after a while is your brain starts to integrate its understanding of the magnetic field uh you know where you're at like you get in the car you can sense the magnetic field change or you walk through your house and the the bracelet you know it notices the little electronic devices in the various rooms and that kind of stuff and so you you can like you can find your way around the house in the dark right because your little magnetic sensor thing is telling you things about the environment that you can't see normally so that's a it's an interesting experiment in learning because because your brain is really good at taking almost any kind of input and then integrating it into its world view and then learning to use it well for its in this case it's the sense of presence or you know or understanding things about the environment that you don't get but there's a learning process with that right so neural link will be like that on steroids like almost anything that you decide to plug into the brain which is a new sense you know the brain will rapidly get to work figuring out how to make use of that to understand the world better sure there were a lot of things that I I made this quick little list of uh stuff that I thought you know so what what did we see that was new at this neural link Show and Tell right and I was trying to think of what those were and I'd so like my biggest takeaway was the manufacturing scale up right and what I wrote down is like you can tell they're manufacturing scaling you can tell they mentioned focus on particular things that you would associate with manufacturing like scaling up to make a product like you know lifetime testing they've got these new facilities in the work you know they've got a dual operating theater now and it looks like one of them is cranial and the other is spinal you know so they've got both like uh so I thought that was significant they mentioned building a clinic right or you know they're building plans for a clinic so like planning for a clinic that's not something you do if you think yeah maybe we'll get to Market in 10 years right planning for a clinic is something you do like when you're expecting to move pretty fast so I was I had like I was impressed that that stuff left me with a different impression of neurolink than I had after the first show and tell so the other let's see they walked us through robot upgrades I thought the custom Optical path was an interesting optimization and that's so on the on the robot they need you know they need to get visibility down inside the same space in this really small location where they're also got probes going around and they have a couple of different things that they need they need to be able to you know illuminate the tissue in a particular way and they need to be able to take a visual signal back out of it and they've got this really small space right so you develop a custom Optical pathway that separates the channels with polarization right that's a lot of optimization right like that's not just fooling around like that's a pretty complicated piece of very specialized work to make the robot better right so the the takeaway that I get from that thing is is the robot it's they have really high expectations for it and it's getting relatively mature and they're throwing their like it's not the they're not developing in certain respects the neural link that they're making right now is kind of definitely a gen one and you know Gen 3 is going to look really different and Gen 5 is going to look really different from gen 3 and so on their points along a path um but the robot it doesn't they it they didn't do a quick and dirty robot right to just like so they could get the implants basically working right they're really thinking about like what does the ideal robot look like and developing the tech for like you know the high volume robot that you're really going to want when you're uh when you know he's like yeah I need a certain level of accuracy I need a certain level of performance I need a certain cost point that I'm that I've got to get to like my impression is that they're they're thinking about it in those terms this is like not like a research project this is a we're scaling up to mass manufacture this thing which also says like if they're scaling up to mass manufacturers thing it's like they have no sir they have no real doubt that they are going to mass manufacture like they're making those Investments right now um the plans for the trans Dura implantation well that is brilliant like that is really good like if they can not break the dura when they put the probes in that solves a whole bunch of potential issues like it's a big win so I and and it's not easy so it's going to be interesting to see how they do that like how do you seeing the vasculature through the dura so that you don't have to break the dirt when you do the implant when you put the needles I mean putting the needles through the Durham you know they're working on coming up with a better needle needs to be stronger because the door is pretty tough uh they'll figure that out right it's not an easy problem uh it I I thought it was kind of interesting that there that that apparently their approach is to try lots and lots and lots of needle designs uh and then see the strengths and weaknesses uh you know that they're uh of these different approaches before they sort of you know dial down to a particular one which like that's interesting like most people they sit down and they try to understand why the why a particular system didn't work and then they designed a slightly different one maybe they are doing that and they're just doing it really fast um so they talked about 16k channels in the next you know they're gonna do another chip that's 4K per chip and a new version of the implant that's four chips 16 channels 16 channels is a lot more than a thousand channels like that's gonna be a really big step even 4 000 channels is a lot um so that was another thing that really struck me about what was going on um the fact that they developed brain proxies I thought was kind of interesting I mean and to a certain extent it makes sense that you would do that but the amount of effort that they're putting into like developing good brain proxies so that you know so they can do their you know they don't they do less in Vivo testing and More in vitro testing but also like building a brain proxy that has the right sort of kind of like uh that's that's a good proxy for like you know this whole thing where like you take a needle and you stick a thread in to the tissue and then you want to pull the needle out and leave the thread in place uh so think about the properties of the brain of brain tissue that affect that at the margin where you've got all these complicated things kind of interacting there's there's all these chemical aspects of the brain uh of the tissue in the brain and how it's going to interact with the thread and the needle of you know the water not just the water content but all of these other uh chemicals some of which are electrical electrochemical and they're going to interact with these metal things that you're putting in there like coming up with a brain proxy that actually mimics all that stuff well enough that you feel like well if it works in the brain proxy it'll work like that seems like a real significant challenge to me so I was surprised to see them but you know that's also they said they were committed to trying to not using animals for exploration and so this you know to me I that it sounds like they're very serious about that because developing a brain proxy is going to be really really hard like it's a big significant side project once you get it sorted out it's really valuable I mean you know not just because you don't have to use as many animals but you I mean you you can only use so many animals and using animals like even if you take the ethical aspects of it off the table it's really they're really hard to use it just takes you know uh it's just a lot of work to use animals to do a lot of this kind of stuff and by developing a brain proxy you can you can do stuff at scale and much more quickly that you don't have to so like it totally makes sense that they're doing this it's just it's really impressive that they're doing it now right they're they're not fooling around they're going to scale um so they had a near term so I had noticed like in the previous uh animal stuff that we saw and uh they weren't doing any stimulation like we didn't see any you know neural link injecting signals into uh into the brain the body previously and this time we saw we saw them actually they had a spinal implant in a pig and they were demonstrating a certain degree of control over the muscles in the in one of the rear legs that was actually pretty impressive like I had some pretty basic questions about how how you were going to be able to like get enough control over motor I think I mentioned this before but you know uh individual muscles in the body they have a large number of motor motor units so an individual motor unit is like a it's a it's a collection of muscle fibers that have a common innervation like there's a there's a single neuron that comes down and tells this group of fibers to contract and muscles in the body have anywhere between like you know you know a couple hundred to several hundred motor units in them you know your bigger muscles have a lot of motor units right so if you want to do a good job of Contracting muscles you you want to get control of a bunch of different motor units right and then the other thing about motor units is like when you fire a motor unit you get it like 100 millisecond contraction and then it has to rest a minute and then you need another one right so your brain actually rotates between the different motor units and a muscle I'm not actually sure if that occurs in the brain stem or if that occurs um in the uh in the there's a place where the upper motor neuron and the lower motor neuron come together in the back of the spine right uh so I don't know if the processing for that happens down there I think it doesn't I think it actually happens in the brain so so you know if you want fine smooth control of Contracting a muscle which you're going to need if you want to walk you you know you need access to a bunch of motor neurons and you need you know you need to develop an algorithm that can use them so that can sequence them appropriately to get the right duration of the right amount of force with the right onset and all that kind of stuff uh and if you're doing that with a bunch of different muscles you need a lot of interface too a lot of things and and and contraction seems kind of complicated uh and yet in the video it looks like they've got I mean I I suspect they're not getting activation of a lot of motor units when in the pig demonstration that they did right now and certainly you know it's in no way coordinated there's nothing like walking involved they're just demonstrating that they can activate individual muscles um but they are activating individual muscles and they're getting a sustained activation of the muscle which I thought was pretty impressive so they're doing that they're talking about uh going into the visual cortex right there was a talk there you know so they're thinking about it they have plans to do it they've done enough looking at the structure of the tissue that they need to interface and examining whether or not they can do that uh you know and they've done it in a couple of animals right like they've got some animal experiments that basically demonstrate that they uh that they're able to stimulate the visual cortex and that the animal will perceive and respond to it as if it were in you know a flash of light in its visual field so that I thought was impressive because we didn't see to the best of my recollection they weren't doing any input before and now they're doing some input that that's super important because like the potential for changes in behavior of that of the tissue both an immune response and also that the nerve tissue itself will re will change its structure and that that will actively interfere with your ability to inject a signal into it is much more likely for outputs than it is for inputs because input is basically passive right you're definitely not stimulating any electrochemical response in the tissue if you just you have a passive wire that's sitting there or it's very minimal because it change the presence of the metal itself will change the electrical field and and the area but as far as we know neurons they're not super sensitive to local electrical Fields right yeah it takes a pretty big pulse of current to actually trigger a neuron but the flip side like that's something you can expect a significant amount of pushback from the tissue of the brain when you start stimulating it and like I don't know a lot about about how that stuff responds but they've taken the step of doing it and they're getting that learning under their belt so that is very cool I thought um sorry you have to go no I don't need to go that was a voice activated system responding to me um before we go to the next topic um is it possible that they were just so like the way I understand correct me if I'm wrong um if I'm holding this cup and I just hold it for an hour then during that whole time like there are different muscle groups that are being activated and turned on and off and there's many of them even if you've pulled it for one second right because individual activations are only about a tenth of a second so if you hold it for one second there's at least 10 different now that doesn't necessarily mean it's 10 different muscles once a muscle fires after after it rests for 100 milliseconds you can come back to it so like at a minimum to be able to smoothly apply a continuous amount of force you probably need three different sets of motor units that are giving you approximately the you know the same amount of force as you rotate between them right so it's not physically possible for like any stimulation to occur at just one specific muscle unit and just continue that one for like three seconds on some scale it is like if you just activate a motor unit and you hold it what will happen is the motor unit will contract exhaust contract exhaust right there will be it'll be jiggling at about 50 hertz or something like that in fact um you can you can observe this yourself like if you do weight lifting or whatnot go and go in the gym load up as much weight as you can do or you can actually another really easy way to do it is like clench your jaw put feel your jaw muscle and then clench it tight and when you get it really tight it'll start oscillating and it's that thing where they're exhausting and releasing and then recontracting right so the thing is if you need to not have that vibration or not have that vibration be big enough to enter like if you want fine control that vibration will mess with you right so if you only have a single motor unit what's going to happen is the muscle well first of all it's going to get you're you're not going to get a good power profile it's going to get really tired really fast right but it's also going to be vibrating right and so like if you want smooth stuff what you want to do is you want to smoothly transition between a set of fibers and you want to give them time to recover before you ask them to contract again to get good control to get fine control you right so like you know if you want like they're not going to be able to activate one motor unit in the thigh muscle and be able to walk with that you know they're going to need a little better control okay so okay now so an interesting question is to what extent do they need to build into their algorithms the stuff that your brain is doing when your brain does this with your muscle like to your brain your bicep is 200 different muscles that all kind of have the same action not quite pretty similar because different parts of your bicep can contract separately right and they do generate slightly different forces and your brain learns all this stuff it's just you know your bicep is is uh like it's like 380 co-located muscles with a common insertion and origin points all right it just looks like a single muscle and when you cut when you cut it open but to your brain it's 400 muscles okay okay uh yeah keep going with the topics that you've I think it looks like you're reading off of the list yeah so I just I wrote this stuff down to help me remind so that they're working on a spinal implant okay right that's a really interesting thing because the existing neural link like it's a really good fit for the skull like it it replaces a piece of the cranium right it's designed to sit in that gap between your skin and the dura and you know the probes are aligned you know they come out of the bottom of it it's designed so it's got a bottom that faces the brain and the top and the top is where the inductive interface and all that kind of stuff go on I mean uh okay so now you want to go to the spine well the spine is really different structurally first of all it's a really dinette your brain is more or less a static bone I mean it's got joints and they do move around a little bit but the amount of movement is super small it's you know it's more or less a single structure um but your spine isn't like that at all and not only is it super flexible it's load bearing and it's critical to normal function right so it moves it's also got a really complicated shape your brain you know it's like your cranium it's a more or less Universal thickness if you look at any small subset of it right so you know if you get the thickness about right in the Earl you know the spine's not like that at all it's got a really complicated shape to it the the spine itself then you go into the spinal column and that's also like as you work your way down the spinal column like every single vertebrae it's a different shape right and the you know the composition is different as you work your way down there's less and less white matter more and more gray matter uh and then um you know the the gray matter it's like uh neural link like maybe they can plug it into white matter but white matter is that it's the myelinated axonal sheaths coming out of neurons right yeah you could and you could trigger that I suppose but mostly neurolink it's really you know it's really clearly designed to plug into gray matter and enter you know essentially influence or read what's going on from neurons and gray matter well the gray matter on the spine spinal cord is on the inside and it's got a pretty complicated distribution of stuff and and this is really important you know there's it's kind of it has this dorsal part which is the part that's closer to the outer part of your back well that's where all your senses are the motor control stuff it all comes from the inside of your spine facing in toward your you know so so if you want to get a probe into you know the dorsal sorry the ventral horn the vent the inside part is the ventral the outside part is the dorsal the if you want to get to the ventral horn which is what you need to get to you have to like you know you have to deal with the fact that you're in this highly flexible Dynamic load-bearing you know like you can't just cut off a piece of vertebrae and replace it with a piece of plastic right that like at least I don't think that's going to work at all it's also really complicated like the vertebrae the back of the vertebrae they have all these spines that come off and there's this very complicated uh set of ligaments and muscles that that make it flexible and protect it right and you have to respect the Integrity of this complicated structure when you do that if you just cut a ligament off of the Horn of one of the vertebrae like the animals is not going to be functional if you do that right so so uh so you know when I thought about them doing the thing in the in going into the spine I thought that's because you could you can like to a first approximation you can implant the neural the neural link relatively close like if you have two or three inches of wire that can go to your probe you know you can find a space between the disc or you can drill a really small hole maybe in a vertebrae and then feed through that but then but now you've got to feed all the way through that feed all the way through all the white matter you know there's a distance you got to go and then you got to go through the dorsal horn and into the ventral horn right and then get your probe placed there and it's all possible but it's not like just I mean the cranium is so much easier right it's just on on that you know it it's impressive that they seem to be doing it right they they showed examples of them you know placing probes into spinal tissue and obviously they're doing it the pig demo shows that they even get functional uh functional stuff out of it so I thought that was that's a really impressive result to me and it's a really big difference from the last show until uh so you said that the white matter is my myelinated myelinated axonal sheet yeah yeah so the axon that comes out of a neuron like in in Gray matter when the axon the output from a neuron that you know the neurons when one neuron talks to another what happens is the axon of one neuron forms part of a synapse which is a junction between two things and then the other side of the snaps is a den the dendritic half of the snap so it's the input to a different neuron an electrical you know this electrical domino effect thing propagates down the axon and then to the axonal side of the synapse that synapse it responds by releasing a chemical across the snap a neurotransmitter right that's where the word comes from it transmits neural information from one cell to another and then that triggers you know the the same sort of cascade function on the other side of the synapse okay so if you want to send a signal very far uh you kind of have this transmission line problem right it's a a an axon like if you're just going a short distance you can just have an have an axon but say you want it to go really far like there's a you have a lot of neurons in your brain that are like a meter long like it's a single the axon on a single neuron is like feet long and uh so if you want good signal Integrity the axon needs to be insulated from the surroundings and so and this is what there's a there's a sheath that nerve cells can wrap I think they're called Schwann cells or whatnot they wrap myelin around the axon so if you want to send a signal a long way well if you have bundles of nerve fibers going from one place to another these it this is what we call white matter you know you got a bunch of axons of a bunch of nerves and they're all wrapped in their own little sheaths they're all traveling together and it looks like a you know when you if you look at a brain it's got gray matter which is like lots of neuron bodies and and with that with with relatively short axons and then you've got white matter which is mostly just the axons of nerves traveling a long way taking a signal safe from the brain to a muscle in your jaw or the the you know the an upper motor neuron in fact I think the longest one in your brain there's a sensory neuron that runs from the tip of your big toe all the way to the base of your all the way to the base of your brain right so it's like you know 1.5 meters long or something like that hmm okay uh well so and then the white matter in the brain is uh is is it's closer to the inside yeah and the gray matter surrounds that white matter yeah put in the spinal cord it's reversed yeah so you know the the the your neocortex which is a huge chunk of your brain and it's the outer part that we were talking about it's this sheet the sheet well the way that the signals travel from one column to another in the brain if they have to go very far as they come off the inside and they go because you know you walk you walk it the short path is through the brain right so if you look at the if this this neocortex watered up inside your brain it has all these nerve fibers on the underside of it that cross connect to other regions of the brain like if if a signal is going a short distance it can just go laterally across a couple of a couple of columns but if it needs to go very far it exits the bottom of the column enters a nerve fiber and then that signal will hop across to some other you know it'll it'll go across to some other column in the brain and when a signal needs to leave your brain to go to some part of your body right it has to travel through one of these long you know isolated myelinated axons and that will you know in the case of your spine there's a signal that will go it goes from your brain it goes down your spine until it gets to the vertebrae where the nerve exits the spine and then there's a junction there so you got the upper neuron the upper neuron is the one that goes from the brain to that point in the lower neuron is the one that goes from that point and then exits the the uh exits the spine and might go you know if it's like a motor unit it'll go all the way to the muscle fiber and that'll have a the muscle fibers have little input things switches on their surface and the nerve connects to that switch and so the signal travels down and when it hits the end then the muscle the motor unit contracts and then okay and so then they talked about like going the reverse so that you have some sort of feedback and then so let's say like I want to stimulate my ankle moving this like like this uh and then and then I do that and it hits a fuzzy uh surface then I get that sensory input and it sends it back through another lower motor neuron from from that so motor neurons are the ones that go out to your muscles so the sensory neurons that come back from like you know the touch sensors in your skin or something like that we don't they're not called motor neurons but you know it's a similar thing it runs in the opposite direction okay yeah so so it's it's also getting a lower lower sensory neuron yeah yeah that's literally what it's called uh maybe yeah okay it could be and then sends that back and then uh hits like an on off switch in the the spine area in that same spine area yeah so the dorsal the dorsal Horn of gray matter inside the spine usually there would be an interface there so yeah so the sensory neuron will come in it'll enter the dorsal you know there's a there's like a little tube in the back of the vertebrae that it enters through and then and you know it it goes into that the dorsal horn which is where the sensory neuron interfaces are and then that switches to another neuron that that um I think are they called Rising neurons or descending neurons I forget but it's a neuron that goes up you know and that scent brings a signal back up to your brain okay and that when that upper neuron hits your brain again that is hitting in the white matter or gray matter well the white matter is the is the wire and the gray matter is like the computation right substrate so like the the bodies of neurons that do the processing you know that make decisions that do the switching or whatever the deal is that's the gray matter and the white matter is cabling that runs from one part of the body to another so your spine you know if you look at the spine it's got a ton of gray it's got a ton of white matter because you know the a really big part of his job is just you got this giant bundle of signals and it's got to go down your body right so there's all of these axons for these neurons that are just going down the spine and then but so this is the that that the thing is like uh on your brain right the gray matter is on the outside and the white matter is on the inside yeah the white matter is on the inside because that's the right way to architect it you know it's because you know you don't want to have a wire running around the outside of the brain and it can just go through right it's shorter to go through the metal right when you get to the spine you know but you know it turns out the evolution decided to put the white matter on the outside and the gray matter on the inside um so it's a little bit harder to get to the gray matter in the spine because it's surrounded by thick layers of white matter which is sensitive you know you have to be really careful you can't just cut holes in white matter because you're cutting the axons that you yeah it's a signal you're severing the spinal cord right if you if you slice into the white matter but if the white so white matter being external to the great gray matter in the spine that that's actually more inside of our bodies right yeah well it's first of all it's inside the horny protection of the spine itself right because the spinal cord it runs down a channel in the stack of vertebrae vertebrae vertebral bones uh and uh so like you know you got a bone and then you got a bunch of white matter and then outside that it's not even just that problem I mean you have like you know ligaments and muscles and connective tissue and then skin you know outside that you have to go through all of that stuff but yeah it's uh you know the spine is just this Flex it's this powerful rugged flexible Dynamic structure which is really different from the brain from the skull right which is rigid mostly it's approximately rigid doesn't move much you don't really you know there aren't a lot of muscles on the top of your brain that are on the top of your skull that are powerfully moving your scalp around right but the muscles in your spine they have to hold your entire body weight um there I see okay so I guess I want to just make sure I verify my understanding so the the neurons are in the the gray matter up in the brain that's like when we say neuron bodies right I mean the the neuron itself there's a body which is the cell body and then it has dendrites that come in and that hasn't one axon typically that come out and that the axon like if a nerve if a signal has to travel a long way it's the axon that'll be long so the nerve body it'll be at the input near the dendrites and then you'll have that this axon that's really long and then the end of the axon will end in one or more synapses where it connects the dendrite of other neurons or in the case of a motor unit the the synapse actually it interfaces to uh motor to a bundle of muscle fibers right so the white matter is nerve is neurons it's just just the axon part of the neuron but then how is that connected or how is it in close proximity to a dendrite of another neuron if that dendrite is not in the white matter so that it's so the uh well so if you have say the say that you know you have a particular vertebrae you know uh t7 you know right here in the middle of your back and uh so you gotta so you you know processing happens in your brain in the in your motor cortex it decides it wants to contract this muscle so a neuron somewhere in the cortex whose job it is to contract this to make the final decision on Contracting this motor unit switches on at the you know in in the brain and that motor neuron it has a really long axon on it and that axon will come at it'll go through the white matter inside the brain it'll go down to the top of the spinal cord it'll keep going it runs down in this channel right so keep going down until it gets to t7 and then right at t7 it'll it'll emerge into this ventral horn where there will be a nerve body with nerd with dendrites right for the nerve it's going to activate so now it's going to trigger that nerve now that nerve is going to exit the spine run down I guess in this case t7 well this would be like a breathing muscle or something if we were at that nerve but anyway it goes out to the you know your diaphragm or whatever it's activating and then it will it it uh takes a signal along to a motor unit in that muscle and then that muscle contracts right so you've got I mean you're yes there's one tissue in your brain that makes a decision to do this then you have a single neuron whose job is to take it from the base of the brain down to the vertebrae and another neuron whose nerve body is in is in that ventral Horn of gray matter inside the spine approximately at t7 which we're talking about and it's axon will leave you know it'll exit the vertebrae through a passageway and it'll go all the way to the muscle typically let's see okay so I think I I think I'm getting why I was confused because the axons that that are in the brain and only in the brain so uh like that sounds that don't go very far aren't white matter because they're not myelinated axons that have to go very far if they have to take a signal a long way they need insulation on the wire and the insulation is what makes it white matter I see okay uh in the in the example that you provided with the diaphragm is there sensory information Sensory neurons that get sent back uh uh let's see generally yeah for skeletal muscles that's definitely true for the for the uh diaphragm the diaphragm is an unusual piece of muscle so it might be a little bit different but like regular skeletal muscles um so skeletal muscle fiber itself I believe does not have feedback but what what does have feedback is that there are sense cells that are in the ligaments so the muscle doesn't connect to Bone directly there's an intermediate tissue called the ligament that connects to the muscle on one side into the bone on the other side ligaments frequently have sensory cells in them that sense the tension that the ligament is under so that's appropriate deceptive feedback so there would be a sensor in that and that thing will send a signal up you know that that thing will report back to the spinal cord and ultimately back to the brain and that goes through the dorsal horn uh and then triggers a switch there to then send a different upper sensory neuron up through the white matter again till yeah so I mean the axon is the cable part of the neuron and the gray part is the body and the dendrites right I mean there are also bringing that just that just make decisions like if you you look in the neocortex right like a mini column has a hundred neurons in it and most of those neurons are just talking to other neurons in the mini column or to adjacent many columns right so they have dendrites and axons and their axons are super short so they don't need to be insulated right but at the bottom of the mini column or towards the bottom of the mini column there will be you know there may be one or more neurons that that need to send a signal to another mini column far away or maybe all the way down the body to uh to a motor unit somewhere and and so so that part that the axon of that nerve when it when it exits it'll get routed and the thing is because these signals tend to to travel in clusters right what you see is a whole lot of different neurons axons get in a bundle that becomes a nerve fiber and it runs you know to some part of the body but it'll be a whole bunch of neurons that are all going to approximately the same place and so they're you know they travel together um but that each one has a separate Axon and each of those axons is isolated and if you get a sufficient bundle of them together it becomes white matter right and it it's so it's just the signals that have to go a long way go white matter and the signals that are traveling tiny distances microns or you know a millimeter inside the brain or something like that they they you don't need white matter for that because the signal is not going so far that it needs to be uh protected and let's see is it true that myelin sheath can form around shorter axon lengths I'm well so the myelin sheath itself is not part of the neuron like there's another kind of cell that wraps itself around the uh around the axon to form the that insulating myelin sheath and I think those those kinds of cells they're just basically not present in the gray matter parts of your brain and they are present in the white matter parts okay okay uh yeah I keep uh if you have more more stuff please keep going I think no that was I think I got through the list of just like uh what okay okay so one thing uh the software the uh the software is still super simple super early days right like my the scent now you know they didn't show us the software and they didn't talk in great detail about how they were doing stuff but it sounds like they're mostly have pretty simple algorithmic techniques that they're using both to uh to extract signals from the brain and now they did show they did they clearly showed some very simple neural networks that they were using to process outputs to make decisions about like moving the the cursor around and that kind of stuff but this is like you know it's a 32 neuron these are neural network neurons not neurons in the brain but they'll have like a 32 neuron single hidden layer with like two output neurons or four output neurons removing uh the cursor around but the so you know they have this incredibly incredibly simple little neural network that's basically looking at the firing statistics and then making a decision you know on which direction the cursor should move and they they train up that network but the complicated part of that network is mostly how do you turn the the firing statistics into a signal that you feed into the neurons and I think that's probably a really simple when I say there's a lot of potential information and nuance and how you interpret those firing statistics and it seems like right now they're just using really really really simple methods to like get stuff basically working so so one of the things I did not see was I didn't see any signs that they were trying to do sophisticated interfaces and in fact there was one there was an audience question at the end uh and I forget uh the name of the woman who asked the question but she basically said you know you've got this you know dynamic tissue that you're interfacing to that you know it changes over all these various time frames hours days months weeks whatever and you know how does how do you uh how do you deal with that and can you take advantage of it and uh the you know the answer that basically came back was well you know we can reprogram our software right so uh yeah it's I mean it's I'm not doing good credit to the response on the thing but they're not one thing that they do not appear to be doing right now is Building Systems that dynamically change their own behavior in response to changes that they see and they're they they gather a bunch of Statistics they train up a very simple neural network to do the thing they test it to see if it's working that kind of stuff and then you know if it stops working then you gather more statistics and you retrain it right whereas like long term what you're you know that the external system is going to be constantly wanting to learn from and and you need to somehow close the loop in order to do that which is going to be easier with human beings so that so that you so that you have ways of understanding when your training is starting to drift and and then and then improving the training on the system and along that path you'll get to you know after you go down that path a little ways you'll get to a place where effectively the neural networks outside that are interfacing to the tissue and the networks of neurons in the biological tissue they'll start to act as a single unit learning from each other okay and adapting to each other and they're like they're not doing that right now I can see why they're not they don't need to do it for any of their short-term goals but like it's going to be such an important and Powerful capability like I would have been encouraged to see to hear them talking about you know what they were thinking along those lines and how they were thinking about approaching that problem and this this could well be the kind of thing where like doing this kind of stuff is going to be much easier when you're in a human subject because you you can give pretty abstract instructions to a human or you can ask them to tell you with Nuance like what they're experiencing when they have it so you just got much deeper and richer feedback to the experience of the user with a human being than you have and it might be that that is the right time to start like saying how can we get the nervous system and the external you know sort of simple artificial intelligence technology that we're using to plug into it how do we get those two to play well together so that they so that we we now have a complete learning and adaptive system that learns all the way through from one end to the other because nerve tissue can do it and the outside neural network that plugs into it it can do it too right so getting them to cooperate when they're doing it closing the feedback loop all the way out through the hardware so that the learning can propagate through that whole system like that that's going to open up the it's going to significantly improve the signal the noise ratio that you get through your link because the systems will learn to get better at it just like automatically and they'll automatically compensate for um you know as the behavior of the implant drifts you know the nervous tissue and the external system will automatically compensate for it

Channel: Neura Pod – Neuralink

Views: 45,230

Keywords: Neuralink, Elon Musk, Max Hodak, What is Neuralink?, Neura link, Neura Pod, Tesla, SpaceX, starlink, Nueralink, Nuralink, Brain computer interface, What does Neuralink do?, brain machine interface, Artificial Intelligence, Metaverse, Facebook, Neuralink News, Neuralink news 2021, Neuralink 2021, Neuralink Update, Neuralink Update 2021, Neuralink news and updates, neauralink update, Neuralink monkey, neuralink presentation 2021, neuralink pig, neuralink demos, Neuralink stock

Id: 7RXmljnqkfw

Length: 119min 58sec (7198 seconds)

Published: Wed Jan 11 2023