Learning about Neuralink w/ James Douma (ChatGPT x Neuralink)

you you saw all the intelligence we're done the whole human race can go on vacation now all right there are certain things about neurolink that are tough the material science behind your link is to like getting a level of biocompatibility and getting the insertion procedures mature to the point where we have no qualms about putting this in perfectly healthy human beings like my experience with how those kinds of texts get developed is they take time because we are very cautious in how we develop that stuff if we had mature nanotechnology brain computer interfaces trivial like it becomes a very simple problem I have the privilege of speaking with James doubler a self-described deep learning dork James's experience and Technical understanding are not easily found and I think you'll find his words to be intriguing and insightful this is one of several conversations that James and I plan to have the first conversation was recorded earlier on our channel so check that video out too if you haven't seen it hope you enjoyed this one let's talk about uh chat GPT and and pairing chat GPT with neuralink uh and like what does the future look like with a version 10 chat gbt with a version 10 in our link [Music] well so I think there's a stage of of brain computer interface development where the thing you want to plug into is The Language Center right so Chad GPT it's based on a large language model it processes it processes see like these are sequence models they don't have to be just sequence models but they're mainly sequence model sequence models basically they take a sequence in and they generate a sequence out and the mapping can be arbitrarily complicated um so you know if you think about it you know if you have a sufficiently sophisticated input to Output sequence converter if it's sufficiently sophisticated you can ask it any question and you can get the answer what is the cure for chat for uh cancer how do we travel faster than light you know like everything can be expressed as a sequence both input and output so uh language models they just process sequences but we've gotten them to this level of sophistication that they're starting to ban to knock on that door of being able to answer your questions whether the question posed is a unique riddle or whether the post question posed is just a factual inquiry right and so that's that's where we are right now with chai gbt that gets a lot better um in in a sense if you pick up a phone and you talk to a human being that's a sequence input output thing so you know imagine imagine you could call God on the phone right you can ask him anything and there's nothing he can't answer right so that that's like a sufficiently advanced version of chat GPT um and the limitation becomes the fact that you have to just talk right it if you want to learn um geometry and but you can only talk to your professor on the phone there are certain limitations there are certain things that you want to describe by drawing something on a Blackboard and pointing at a picture right they're just there's just a lot of stuff that's expressed that way there are lots of other things so right now at the level of you know human input output we got five senses and we mainly use two of them sound and and vision to learn almost everything those are those are the Really Big E's so chat GPT sort of works with abstracted sound it works with language um it can there's nothing to prevent you from building a version of chat GPT that takes audio in and produces audio out um it you know I'm sure we'll see that really soon and you could hack together a voice interface and people have already hacked voice interfaces onto it but actually building a version of chat GPT that takes the audio because uh like it's not well appreciated but you can't actually write human language down the thing when you when when I say something to you and you write it down there's a huge amount of stuff to get like that that huge like you don't write that down right it doesn't get there the the Nuance you know the posing the the pacing the body languages all of the stuff that doesn't get in there and when when humans speak to one another all of that carries a lot of meaning that's why a lot of people don't like remote work because they're the kind of people where the body language really matters and they want the full Fidelity of it's I I really like video over audio calls for the same reason I'm on the Spectrum and it's really hard for me to detect a lot of things in speech and if I can see your face I get a lot more out of the interaction than I get if if I don't have that uh so you know Vision adds a lot but then there are all of these additional layers of things that we have in our head for talking to ourselves and explaining things to ourselves there's for instance there's all of these nuanced levels of emotion you know we have how many words for emotion and in the English language I mean hundreds and hundreds of them melancholy uh you know so melancholy's a feeling you learn how to map melancholy the word Melancholy onto what you understand to be the experience of melancholy by reading other people's descriptions of experiences of melancholy and so Melancholy it's not it's not a single thing like you know melancholy's like this 50-dimensional thing that that could you know it could on the one hand it could you know you you have the Nostalgia aspect you know there's this Nostalgia aspect of it then there's there's a sadness aspect now there's all of these you know things that go on but we just have the one word if you could break out all those words like if we could communicate in that space of all of the component sub-components of emotion that go into that well now you've got this other layer of communication that you can do so language is easy in the sense that we know that language is already an abstract and represent representational form and we already have a lot of pretty powerful descriptors for it and we already have a lot of powerful tools even AIS that can work with language either in voice form or in text form so to me it's a natural like by the time we get to where we have neural link like I don't think we'll ever go through the keyboard and mouse phase with neuralink forward normal healthy people that's where we'll start in the early days for you know quadriplegics who need an output because it's the first really useful completely General thing you can do the bandwidth is just really low but there are advantages to the low bandwidth that make it easier to implement it but we're very likely to get to a level of maturity of the technology before we start talking about putting it in healthy people where you can just go straight to the language interface so that you've got a 20 000 or 200 000 word vocabulary and you can speak at you know 100 words a minute or 200 or 500 Words a minute so you just get this fantastically higher bandwidth like right out of the gate for healthy people and and because language is so well established already as a communication medium like it's just it's it's the Practical place to start but it's not the end of the game at all language like it's it human language is amazing it's just a really really powerful tool but it's still talking through a straw compared to what's going on inside your brain there's so much going on there that can't that you cannot share with other people because the pipe is too small and neural link has the potential to break open that pipe and if you want to break open that pipe you have to move away from languages and encoding because language was developed for this Ultra ultra low bandwidth interface that we have between each other and if you really want to get past it you have to step outside language as as a medium that you do and so and I expect like we'll go to that quick that we'll we'll go to representational systems that are that are higher dimensionality than language and higher bandwidth than language relatively quickly long but long before we do the full brain interface type thing I see and so in between like cat gbt or that really long-term interface there's kind of the in between which is the audio chat GPT and then there's also the video I mean an interesting question is whether we'll get to you know thought chat G because you know we like human human beings experience we experience time sequentially right and so your brain iterates through your your brain is a state machine that has a lot of sub elements they have their own States but those State you can describe the brain as as a system that has a lot of hidden internal state that evolves in coupled and independent ways yeah like you could describe the entire brain as a single state machine that's a very complicated internal State you have a latent variable which is your thoughts which is embedded in you like your internal experience is a latent variable that's embedded in your experience of reality and you could you could give the you could give GPT access to that so it could take that thought that latent reality variable stream process it and then hand it back to you in that format right that's probably the highest level of uh sophistication of an interface that you can get to with a human brain a biological human brain so you know that's the end game right that's where you get eventually and you know somewhere way down probably around the time that you that we can first start doing you know neuralink and healthy human beings you start with the language interface because it's you know it's the it's the super appealing low-hanging fruit apple right there on the branch for you you know you because languages like it's it's going to be a great uh place to start with this but but there will be things past it like way way past it if I understand what you're saying about a state machine correctly then uh you could manipulate like your feelings of time too right yeah let me explain that terminology it occurs to me maybe a lot of people don't so there are machines that have an intern that where in order to get them to do what you want they have an internal state which is represented which represents they have an internal state which changes over time so like your microwave oven your microwave oven it has a controller that's a state machine and the states are it's just sitting on the counter doing nothing it has us there's definitely a state where like I'm currently microwaving your food and aspects of that will be like this is the power and this is whether or not the light is on and this is whether the turntable is turning right these are aspects of the state at any instant in time you could describe the most important features of it in one or more variables that are its state that's what it's doing right now when we when you write software to control a microwave oven the it you frequently write it in the form of a state machine where um you know given you know you have a bunch of if if if then else you know whatever uh things there basically if I'm in this state and this thing happens my next state is this one so you know like you could be in the state of microwaving and the state machine could say if I'm in the state of microwaving and someone opens the door the next state is not microwaving and that's how it transitions from microwaving to not microwaving right so it turns out that most complicated things that we build in the real world is engineers and programmers to some extent you know you can represent the mistake machines where they have an internal State your computer is a state machine like all most of the software in your in your your computer desk could be thought of as a state machine certainly the control aspects are you know car controls and stuff they're State machines the car has an internal you know state which is I'm going this fast my steering wheel is at this angle the brakes are not pressed these seats are occupied the heater is on you've got all these variables that describe the state but the the totality of all those variables are the state and the control system of the car you know has a bunch of mechanisms that examine one or more of those state things and then given that you're in a certain State they interpret stimulus as requiring them to transition to some other state and that other state might have the airbags go off or it might have you know the door light turn off because you close the door or something like that like these are all state transitions so at the in the extreme case any arbitrarily sophisticated machine can be described as a state machine and a human being is one of these things that could be described as a state machine so what's the state the state is um when you write a piece of software you have this data structure inside it which is the state where you know the the the the the the the the parts of the software that are deciding what to do when a stimulus comes in they look at the they they get a stimulus they look at the state then they look at this table that says you know if I'm in this state and I get this stimulus then I go to this day and when you do your state transition then you change your inputs and outputs all the things you have control of to match the new state right it's a really powerful way of thinking about control systems and it gets used almost universally for uh okay so that metaphor is also used in machine learning so um chat GPT is a machine it takes a sequence input it has this internal State variable which it generates from using a neural network this you know the neural network says here's my input sequence here's my internal State we call the latent variable it's a hidden it's called latent because it's hidden you can't see the state you can't interrogate the the state directly from outside only the software itself can interrogate the state and change the state right so it's this in variable that's embedded inside the thing but that variable encapsulates the entirety of the reality of the software right like everything is embedded in that so chat GPT like many many AI programs it takes the outside world converts that into an internal State variable and then the that state variable will be reflected in a particular set of outputs and when you get another thing like when you get the next word in the sentence right it takes that word it looks at the current internal state which will include the history of all the words it's seen up until now like how do I interpret this word it's the context for its interpretation and then it's new output is is uh generated you know is stimulated based on the new word added to the existing context and the state is updated to you know the state plus this word I just got right so that's the latent variable so if you apply that concept to a human being what's going on inside your head is your state right I see okay thanks for yeah going going through that um I have some other questions here that are not so related to what we were talking about but I think they're useful in understanding like the architectural decisions that neurolink has made and they're like broader goals for planning and scaling the device um so initially in 2016 uh Elon did an interview with Kara swisher and at the recode conference and he was initially saying that one way of doing a brain interface is going through the vasculature and you don't necessarily have to do a surgery um there are plenty of other companies that are working on wearable devices to get some sort of brain data and then there are other companies that are pursuing that vasculature Network implant inserted device but then neuralink is obviously doing this pretty invasive implant can you talk about the pros and cons of all these [Music] um well so Stan uh the inserting something into the vasculature the way I there's a company right now I think they're called stentrode I think you might have talked about them once before yeah they have a device called a yeah yeah that so they have like 16 probes on the end of a stent so stents are things that it's a medical Technology's been around for a while you use a lot you slide this uh sort of tube made out of wires into a blood vessel stents originally developed to help with people who had clogged arteries in their heart so you can slide a stent into a clogged artery you can twist it and the wire cage expands to hold the artery open so you solve clogs that way there are kinds of stents that can be inserted into the vasculature in the brain you can insert a stand almost anywhere if the blood vessel is significantly you just you know it's it's on the end of a wire you insert it into a blood vessel and then you can just push the wire until you get the blood vessel where you want they have a little uh you know set of controls on them so that you can cause the tip to bend left or right or something you use an x-ray machine or an ultrasound to to visualize the tip and as you work your way through the vasculature you steer it to get it to whatever part of the body you want of course it stays inside the blood vessel but blood vessel walls are pretty thin and and you know if you're going into the brain you know compared to a non-invasive system that has to try to uh you know extract signals from neurons through the skull uh you know being separated just by the membrane around a blood vessel obviously it's getting you a lot closer to the tissue it doesn't give you quite as much flexibility as as direct stuff does because you don't have big blood vessels everywhere in your brain you only have so many you know the the vasculature tree you know the branches are thinner and thinner the stent can only go down so small a branch existing stents they can't go into really really small blood vessels so you do end up being kind of restricted in terms of like where you could potentially put this put this down put that Stan and so you're not getting exposure to a really large amount of tissue you are getting pretty good exposure now the big upside to it is that the surgery associated with the stent compared is a relatively simple surgery I mean you go in it's a local anesthetic they Slide the thing into you you're out in an hour or whatever the uh you know I mean in the case of of uh using a brain computer interface the end of the stand you know the there's a Electronics package that goes on the end of the wire on that stand and that's also planted in inside the bodies in place right um but in a lot of ways it's a much less invasive uh surgery that you know in potentially entails significantly less it's not risk-free because stents can they can damage your blood vessels they can cause blood clots like that's the single biggest thing associated with with uh Stan let your body has a response to the stat you get clots blood clots can be bad especially in your brain um so so they're not entirely risk-free but it's considered relatively low risk compared to cracking your skull open and implanting something [Music] um so it's kind of an intermediate uh you know it's not non-invasive it's I I guess you'd call it less invasive and less invasive means it's going to be kind of an intermediate between the strengths uh of and you know the strength of a non-invasive thing is you can take it off you know you're not going to get an infection from it uh you know there's no immune response to the to it you know there's all kinds of things a downside of non-invasive stuff is you know your brain is is encased in a protective structure which is very thick and very durable and not at all amenable to uh transmitting signals to the outside world so uh so you you don't get a lot of signal resolution in either time or space when you are trying to talk to the brain through the skull uh and that's a huge limitation it sets it definitely sets a pretty low upper bound on the level of functionality that you can get out like you're never I mean you're not going to get I mean I guess it depends on what signals you're pulling through like fmri is non-invasive an fmri can see a lot of details uh inside a brain like you can get down to I think they're they're down to like two millimeter resolution or something like that on fmri now so you can see pretty small things and potentially you could get a lot of data out of that for a very highly trained individual who learned the biofeedback techniques necessary to you know Express themselves through an fmri machine so like in in some sense there could be some potential there eventually uh but people have been looking at that for a really long time and the amount of bandwidth that we are likely to get in the foreseeable future through non-invasive stuff is is really quite low um this you know the using a stent going through the vasculature with stent technology that we have today is [Music] um potentially uh significant improvement over non-invasive stuff but it's like you can't compare it to neural link just like the the difference in bandwidth is just enormous like what you what they will be able to do with neurolink like right out of the gate is just going to like it's just it's going to blow away past what you're theoretically ever going to be able to get out of out of a system which is built using mod current technology now uh I think Elon mentioned this concept of a neural lace neural lace which has been around for a while it was I most of the people who talk about it right now they they there's a version that [Music] um the same author that wrote the books that were that Nila Elon names all of his uh Landing his uh drone ships out of who is it I forget anyway so I haven't read it but is it Isaac Asimov no it's not asimovic yeah I've got all his books you'd think I'd remember his name this is what happens when you get old right um it uh anyway the idea is um at the extreme end it's a it's a nanotechnological device right so you you have something which is inserted into a blood vessel like if if if if I have arbitrarily good Material Science so that I can build a web that uses like single molecules for wires that it can grow itself it's got sophisticated processing all throughout its structure uh and you can get down to like single molecule scale for the web itself you can build something that can go through the capillary Network in the brain so like you know capillaries are the smallest blood vessels like you go through the the vasculature tree and at the far end you get to these minimal size tubes and they're the minimum size because they're just barely big enough for one blood vessel to squeeze through like you can't get any smaller and still pass blood so that's a certain scale that's a capillary and that's the you know those are the finest things well uh the capillary perfusion the number of capillaries you have per cubic millimeter of tissue in your brain is incredibly high so if you can build an interface that can fully exploit the capillary system of the brain you can get like neural link without you can get something that has the kind of resolution neural and calves without having to go do anything invasive at all like so there there's this you know concept for something that could be made if you had high enough materials technology that you could inject a neural lace into like the Carotid artery it could deploy into the full capillary vasculature system of the brain and it would give you access to all of the brain the deepest levels and it would not require any surgery Beyond you know injecting it into into the vein so you know there's we're really really far away from being able to do something like that like it's crazy it's a crazy high technology but um you know it's kind of the Holy Grail if you can do it like if you have the material science to be able to pull off something like that short of that you know the second best thing is um is basically building doing you know what neuralink does except you know ideally a much more refined version if you can get down to the single molecule wire type stuff then you know even doing something like neurolink uh is like what you can achieve um goes way up like you could you could connect to every neuron in the whole brain right if you had that technology and you're willing to go invasive okay um formerly neuralink was doing multiple holes with a pill behind the ear or a pillow shaped device been there I'm going to show you a picture of this let's see share my screen um so this is what it used to be and can you talk about the improvements that have been made and the advantages of changing it to how it is now so well so the advantage of the old I mean you can see why the attractions of the the existing system are the incisions that you're making through the brain you don't have to to remove as large a chunk chunk of the skull but you do have to go uh through the skin um and so like any place in the body where you have something that has to go transdermal and stay there there's a potential for infection there's a potential for injury at that site so if you're transdermal the way that this thing is um you know you do have those problems going like through the skin worrying about infection that kind of stuff they also have this if you have something external to the body the potential for mechanical damage to it you know it being bumped that you're having an accident you have a fall you put on sunglasses you know that like all of those things potentially can damage the device it's less likely for something which is entirely internal to the body but in this case you get uh you get to make much less invasive incisions in the cranium itself and you can distribute the electrodes more widely like one of the one of the disadvantages of neuralink in its current Incarnation is and you've got something roughly the size of a quarter you're going to and that's kind of the largest area of the neocortex that you can span when you so you couldn't like plug into the visual cortex and then all will also plug into The Language Center because they're too far apart you need two neural links in that situation same thing with like your primary motor cortex is too far from your Language Center so single narrow cortex wouldn't be able to do both of those things with this you could potentially do that I think the the downsides of this are the signal Integrity is going to be a lot lower lower I mean I don't know if the plugs they have in the bone here include Electronics but I think they do not so basically you have a passive transmission line going from the you know the Active Electronics which is behind the ear and the length of that transmission line is going to be a problem in terms of the number of channels that you can that you can transfer and the signal Fidelity that you are going to be able to get [Music] um so basically the version that they're doing now is more invasive and somewhat more risky but it has dramatically higher Channel capacity that's probably the trade-off um and then can you discuss the the pros and cons of of Bluetooth versus like maybe some sort of other alternative so Bluetooth is pretty good low power Bluetooth is what they're using right now um it like if you were going to pick something that you were going to develop especially for neuralink you'd probably make some changes you could like for example Bluetooth uh is designed to transmit at higher power for much longer distances than neurolink really needs you don't need to go that far with neurolink um Bluetooth probably doesn't have the bandwidth that you want you probably would like you probably would like lower transmission power higher data rates and maybe go for a protocol that takes less encoding so there's less heat generated as a byproduct because that like one of the problems they struggle with is uh heat generation approximate to brain tissue which is not good you could also pick a wavelength that was maybe better for transmitting through uh the body it's a you know the I guess 2.4 gigahertz that they use for neural link isn't bad there are better frequencies that you could pick overall I would say Bluetooth isn't it's not bad um certainly like if you're going off the shelf it's got a ton of of uh of great things going for it right now but if you were going to develop something from scratch you would probably pick something a little bit different and I I expect they probably will go to something different you know the thing about Bluetooth right now is it goes straight to your phone because the phones all have it like that's a huge piece of low hanging fruit if they go to anything different then they'll need a peripheral because or you know we'll need to wait for Apple to come out with the neural link peripheral for their for their embedded in the phones um let's move to knurling clinics uh so neuralink is is aspiring to convert from just like their current prototype to a mess uh delivered product um so what the clinics what would those entail and is it fair to compare that to like the Tesla showroom model versus being part of the whole dealership Network man that seems pretty far away doesn't it um well you know long term they want to be able to do neural length away we do Lasik today Lasik is you know it's mostly an automated surgery you go in you know you you go through your your pre-op you know you've you've had an exam before you go in um and a robot performs a surgery and it takes I don't know seconds or something like that and they're really accurate and they're really safe so I imagine a lot that long term that's what the you know an early link clinic in I don't know 2050 or whatever maybe maybe that's what it looks like is uh it's a it's a total outpatient procedure that takes 30 minutes it's all done by a robot and the road the setup all gets done ahead of time I think the the neural link the clinics that they're setting up right now I mean they're setting up operating theaters that I am expecting are in the short run they're going to be optimized for working on animal subjects and you know it looked like they had they have a setup for spinal implants and a separate one for cranial implants where in certain respects like the cranial implant is actually the much simpler procedure I would imagine but they probably do a lot more of those right now and they probably will continue to do a lot more of those and they'll do the spinal implants when you move to humans I would I would bet the early stuff that early things like I wouldn't be surprised at all if they end up happening in hospitals like with full-on you know cranial operating theaters and whatnot and then you know as they get more experience with it they're able to Tran to transfer it to a less sophisticated more outpatienty kind of facility but I would expect that to take a while okay um timeline so how how do you how long do you expect it will be before we're able to download our minds onto a Tesla Optimus bot or even or even like a thousand optimize yeah it's uh in um man the download thing there's there are a couple of different ways to get there uh I mean you know if you were trying to do it short term you need the full you need a you need an interface at least to the large majority of the neocortex and that's a pretty invasive that's like you know an early link the size of the whole top of your skull basically yeah yeah it's the whole so that's that's a lot uh you can uh you can do an approximation to that where you look at a lot less of the brain and then you use uh use algorithmic systems to predict what the other part must be but the Fidelity of that copy is going to go down if you do that and uh and probably the most significantly like the experience of the user would be that they were being copied not that they were transferring right the the so the other way of doing it is you get the full interface and you get plugged into an external AI and uh and over some period of time your experience of reality becomes split between your biological body and the external system and in that case the experience of the user would be that they would gradually transfer into that external vessel um so in certain respects that can happen sooner because you don't need as much of a brain interface but you need a lot more external hardware I think the external hardware is probably going to mature faster than neurolink is over the next few decades so I would guess like there's a pretty good chance that that ends up being like to the extent that people end up occupying non-biological bodies I would guess that's somewhat higher probability pathway right now and how long will it take I I like it it's decades away at least like it's it it doesn't happen super soon but in an interesting way uh you know neural link is probably the rate limiter on that because I think you know I the there are certain things about neurolink that are tough the material science behind neural link is tough like getting a level of biocompatibility and getting the insertion procedures mature to the point where we have no qualms about putting this in perfectly healthy human beings like my experience with how those kinds of texts get developed is they take time because because we are very cautious in how we develop that stuff so uh so it tends to proceed it like it could go a lot faster but um because it's a luxury device at least for now right like you know people we we're gonna have to see how Society evolves right and we'll have you know like a thing that could happen is if if if if if there becomes this perception that that that AI really isn't existential risk you know maybe it gets higher priority right and we decide that we're willing to take a certain amount of risk or if the economic benefits of getting it are sufficient and and that that there and there are people who you know really want those economic benefits then you know maybe it becomes one of those things where you know you go to Guatemala and you go to a black clinic and you get it done there like uh in uh you know William Gibson's cyberpunk novels from the 1980s right uh like you know that both of those probably happen faster than the FDA approved approach where we have to wait for three generations of doctors to die off before um you know you can you some super invasive thing uh gets approved for General use by the public over the counter right so I mean like my big my biggest concern about how fast it can happen is not how fast the tech can mature it's more about our regulatory apparatus and you know if you asked me to like go get an early link today I'd like no you know it's like the the risk reward trade-off is just not there today for that kind of stuff and we have we got a long way to go before perfectly healthy people who are in their right mind are going to sign up for having a procedure that invasive done for you know for the benefits that you're likely to be able to get in the short run you know I like I don't I you know it's you know if I can talk to my phone I can text you know I uh I have voice like we're gonna we're going to have voice rate communication with sophisticated AI really soon even without neuralink you know because uh AIS with arbitrarily good you know verbal skills they're in the near future now and we get those by talking to them um what neurolink offers you is the ability to go way beyond that you know and it's going to be a little while before um those interfaces where the physical interface is mature enough that you have a lot of people signing up for it and because that's a precursor to developing the software backend that gets you the really good uh you know access to higher level cognitive functions so both of those take some time uh when you were discussing having like part of your your own brain or your own feelings uh staying in your own biological body and then some being in the Tesla bot what if you had like five Tesla Bots and if you yeah if you have the external hardware to support it then it then it works one of the most amazing things to me like I remember back in the I think it was was it the late 80s or early 90s the first time that um they did an experiment where they um God I wish I could remember this guy's name too because he really deserves a lot of credit but essentially hooking up a primate motor cortex to a robot arm and saying you know if you could if if you know if a primate could learn to control the the motor cortex and there's this point in the experiment where the primate has learned to control the third arm right which initially in the experiment they you're mapping the the you know the primate is learning to control the external robot arm using the same set of motor neurons it uses for one of its actual biological arms but as the interface becomes more sophisticated the primate can use all three arms simultaneously right like initially the interface tries to mimic one of his biological arms but as as the primate learns more and gets more control eventually he's got three arms so think about that you know we tend to think of the human brain as and this was certainly true at that time like that that experiment hadn't been demonstrated before that there was reason to believe before we saw those that level of work that the capacity of the human brain was so tightly mapped onto the physical characteristics of our bodies that there was no reasonable mapping to a different body configuration like a third arm is a different body configuration right it's um like there were reasons to believe that we just weren't going to have that level of plasticity but what we've learned since then is the neocortex is not tightly wedded the neocortex itself the the white matter is very tightly wedded to the configuration of the body because it's a cabling that connects your your neocortex to your body right but the neocortex itself is not super tightly wedded to that so like you know you could have more eyes you could have more arms right you could have a second body right that the neocortex is it it's not you we don't have hard limits there now we do have a hard limit in there's an upper bound of the amount of processing that you can do with your biological brain right so just is you've only got so many neurons and it takes a certain amount of processing to do all of the say motor functions that that go on but you know if you take the motor function itself and you export that to an external system so that the external Optimus that you're controlling like there's an there's a brain in the Optimus that's doing the motor cortex parts of it and and the high level representation that's getting exchanged is much more abstracted um that's still like if if that external abstraction has to come back to your biological brain and get processed in your biological brain along with all the representations you're getting from your physical body there's going to be a certain amount of interference and you still have some limitations right at that point but you do in a sense get to occupy two bodies at the same time with maybe two sets of Senses two sets of motor skills and that kind of stuff right we we don't at this point have a good reason to believe that that's outside the set of capabilities now what that would feel like I got that I have a really hard time getting my head around like four eyes you know being able to see out the back your head like what would that be like that would really mess with your sense of space and you know um like there's a there's a level at which that's possible now you take the next step and you start taking some of those higher level functions and you export those too right so so not only is the motor functions but some of the high level function processing functions are also exported to the to the uh to the Optimus well now your OCTA you're you know you you are you and you are the Optimus robot or the 10 Optimus robots and you've just got like this sort of weird telepathic link where you're all sharing each other's Sensations and you're operating as a hive mind to decide what you're going to do collectively this is getting really weird no I mean we're talking about what's possible right we're talking about what's possible here this is I mean will people do this like it's going to depend on the past it's a path dependency like which technology is mature at which times which ones do people decide to adopt what gets prohibited what gets encouraged what's cost effective what's not uh you know all of these things are going to influence the shape that these Technologies take when we finally get them yeah I see I think knurling could do something so similar to what Tesla's doing with their artificial driver uh like like Tess is basically taking all of their Fleet data from the biological drivers and all the like they're basically using everybody's driver brain to recreate their own Tesla artificial driver I like I suspect that neuralink would be able to do something really similar if they have neuralink implants and millions of people and they have like everybody's brain feeding this artificial brain and then all of a sudden like you sell this artificial brain to everybody or you license it as a software package does that seem feasible to you I have to like put some rails on this description to okay sort of make it match up with my understanding of how this stuff works and what the technology the way it works today um so uh I mean there are a lot of ways that that Tesla's uh autopilot team uses the fleet to uh help them develop um FSD technology that can drive the cars lots of different ways the one that most people think about is the way the cards are used to gather training data um and that can be used to train neural networks that are used for perception contextual awareness you know um planning they they're starting to use neural networks for planning to that so that that's not the only way there are lots of ways you can use the fleet right you can use it to test run like in Shadow mode or not in Shadow mode you know people people who use FSD they drive the technology they're testing it and they provide feedback in real time as they're using it because every time they intervene or they correct it or whatnot they're adding a signal they're identifying a circumstance where the behavior is inadequate and it needs to change and you know the Big Data backroom machine uh learning processing that they do on all of these examples because they've got you know hundreds of thousands of people using the system so they get a lot of detailed statistical data about where the system performs well where it doesn't and they can close that Loop of you know training the system to get better so that you have fewer interventions okay so that that's kind of the course thing you know that that is going on so if you have a lot of people with neural links can you extract can you do an analogous thing to the thing that I just described so the thing that I just described is not the only way that they use the fleet to create the software it's just the most kind of notable one that most people talk about most generally which is like Gathering training data and then using the fleet to evaluate the current state of the system with kind of detailed feedback could you do that with neuralink sure if you if you have if you have a function you want to add to a neural link or you have some external system that you can get a bunch of people to interact with uh you can gather data from their interactions and you can use it to refine the behavior the system chat GPT is doing this right now is one of the training um you know the way chat GPT gets trained Chad gbt is a GPT Transformer neural network that's a language model which is actually really similar to gpt3 which was a previous generation and the biggest difference the thing that makes it that added all of the amazing capabilities to recently is that they change the way that they train it well in the last stage of training they added these additional stages of training but one of them is reinforcement learning from Human feedback uh and when they want to do the reinforcement learning they uh they train the system in a sandbox to produce useful answers and what constitutes they need lots of examples of useful interactions with human beings in order to build this reference system that they can use for doing this training where do they get that data they're getting it from you if you're using chat GPT right now you know and it's in the disclaimer when you sign up to use it they capture that data and they use it to train it and make it better so a lot of you know Google this is how Google chat this is how Google searches get better these days right part of what happens when you type something to Google is stuff that gets processed through neural network and they look at how people interact with Google and they gather that kind of stuff so on that level yeah sure you could do that with neural link by you know essentially having functions that you're trying to train and using user interactions with neuralink or probably more likely people interacting through neuralink to some other system and that other system being the thing that you're trying to make better well you know being able to monitor the stuff that's going back through between people using their neural links and that system is going to be a super valuable vein of data to tap and making that external system work better for people because you're able to understand how easy how easily or how difficult it was for somebody to understand something or use something yeah there's there are lots of ways actually there's not like a single thing it's just you know if we picked One external external hypothetical system then we could talk about you know hypothetically like what you get out of it the fact that there's a lot of data and that you're observing this complex interaction um that needs to be refined uh in a non-obvious kind of non-linear way like having access to that data is really useful and neuralink would definitely provide a lot of that data so there are lots and lots of ways that you could use it like it's hard to narrow it down to just one it's kind of what I'm saying yeah see well yeah I mean I think that's to me like the most exciting thing about the longest term future of norlink is like Tesla with Optimus may be able to use a lot of how they view like yeah literally view the outside world and then work off of that but then neural link would be much more detailed I think because you're just recreating what a human would do and what they're thinking anyway yeah depending on what you have access this is going to be a really interesting world when we get to that there are a lot of really foundational Concepts that our society is based on that um they're not fundamentally true they just they're true in the world that we exist in right now and when things [Music] um you know when the technology starts challenging those like free will you know it's like in a world with neuralink that whole question of whether people have free will it gets a lot more complicated than it is right now right we we just like uh we build our world with rules assuming people have free will we know we have a long history that if you build society that way things work better right now so whether people actually have free will or not is a completely separate issue like you've got to build the world so that they do but what happens when you start to get these things where they clearly don't right we we already have a struggle a lot with mental disease addiction there we have all of these phenomena that basic where we have to start asking in those particular situations do these people have free will do we need to treat them as if they have free will like are they do we need the same incentive function systems for them that we do for a full-fledged normal sane human people operating normally in the world because in many cases you have to change rules for those people one way or another in order to have a functional relationship between them and society and uh in a neuralink world like you know you get a neural link you can flip on the happy switch and go sit in the corner you're happy right you don't have to do anything does that person have free will well you know they did when they turned on the happy switch yeah yeah yeah yeah exactly like they chose to they chose to get in the first place most likely yeah you can say this about alcoholics and drug addicts and right too but then you know you you and you you do it does it it is going to create these situations right and the thing about a thing like neuralink is it's a much more complicated nuanced thing because you have a much finer control like most of the issues with addiction that we have is people have one or some basic set of needs that are getting met by their addiction you know and we sort of know what the substitutes for those are and the mechanisms behind that kind of stuff but with neural link you know you you kind of have there's an unlimited number of things that could specifically in groups be satisfied or not like the thing I was saying about how like wouldn't it be great you know I don't know if it would be great it might be terrible if you know if there were a lot more people like Elon right who like you're not satisfied with success well neurolink we can make them all that way all right yeah yeah do they have free will well did they decide they want to do it you know it if you live in a world where you feel like everybody should be you know a a Pious Christian you know that's a thing you can neural link you can make them all Pius Christians right did they have free will if they chose it you know you know it there are some super dystopian sort of angles to this when we when we because a lot of the boundaries of what constitutes a human being where you're going to start Crossing those lines right everything having to do with the definition this is why this is why abortion is such a difficult Topic in the society that we live in right because at some level you can see it as a definition of like at what point is a fetus a human being and do we have to treat it like a human being and people argue they they disagree about where that boundary is and you and the disagreement over that boundary leads to us to this really fundamental Schism between the two sides in that argument and and that's you know that's just are they all are like you know do people have a right to commit suicide you know if you've got terminal cancer patient like this is another thing we struggle with like what rights do people have over themselves we people don't have unlimited rights over themselves and we don't have unlimited rights over other people and so you know Society doesn't have unlimited rights with respect to us and as long as we're all natural biological human beings that's a much more constrained discussion with neuralink the you know the these the number of different dimensions upon which we can disagree is Jeff going to explode sure so that's okay that's another answer to Warren's question by the way you know how you know how could it destroy us is you know any sufficiently large change to to the to uh um the collected common sense about what constitutes the relationship of the individual to society and their rights with themselves and society's rights with respect to us like these Technologies both Ai and and neurolink in some like really visceral ways it challenges a lot of those uh a lot of truths that have been foundational to human existence forever yeah Tom S is the exploding rate of change in AI developments a positive or negative for neuralink for for neural links ability to achieve its objectives it's definitely going to be a positive um the a neural Link in particular is interesting because uh it's a they're they aspire to have a lot of electrodes placed in tissue and uh and making the most of interpreting what you get back from those electrodes is a really complicated um information processing problem and it's really well suited to what neural networks in particular but you know artificial intelligence techniques in general are bringing to the table so improvements in that technology are going to directly reflect in uh you know in the post-processing algorithm's ability to make the most sense possible of uh what they what the implant can get from the electrodes that are in there for whatever number of electrodes there are if you only have one or two electrodes there are lots of you know kind of traditional signal processing techniques you can throw at trying to understand if one or a small number of neighboring neurons is is firing or not because that's kind of the level of information that you get but say for instance that you or planting a set of electrodes in like the speech centers of the brain and what you want to get out is you know you want to turn you want to generate words you want to generate language and and so that's going to be a a very complicated function of what the electrodes are reading um and it's going to be very fluid it's going to change over time right so it's which adds this extra dimension of complexity to it and AI techniques are going to really help a lot with being able to make that work so in that sense it's it's a it's a big Advantage now there's this thing that we talked about a couple of times which is uh you know elon's aspiration that neurolink becomes a hedge against um you know Humanity losing control to you know intelligent AIS running various aspects of our society because it's likely to be the case that at some point we're just not going to be able to keep up with them at least not as individuals and so having neuralink as uh as a way to expand human capabilities to try to keep up with the machines as they get better at this kind of stuff well in that sense it's kind of working against you know that particular goal the faster AI goes the sooner the deadline for neural link to make a big difference you know occurs and and you know we can hope that neurolink comes together quickly and and shows results for say you know people who don't critically need it it it's it's very it's probably already efficacious for some population of patients that you know that require that have some very difficult trade-offs in terms of their you know what their injuries or their diseases are uh doing to their to their lifestyle but if you want to get an early to a point where it's a clear benefit for a healthy human being that's a lot that's there's a lot of Road that's got to be covered between here and there and you probably have to get past that before neural link starts to become the Hedge against the um you know the potential for too much too powerful AIS and in our world so like in that sense I guess I overall I think it's a big win I I think you probably couldn't do neurolink without AI that we have today like if you if you rolled the clock back even 10 years from today um probably the best AI techniques at 10 years ago wouldn't have been good enough to do a really good job even if you had you know if you have the if you could implant neural link you know and you had the physical interface working and it was compatible and that kind of stuff even 10 years ago the difference between then and now in terms of how good AI is would make a really dramatic difference in how useful that implant was going to be today and by the time they're able to uh you know cross that threshold of something that's functional for a healthy human being we're probably not going to be talking about a thousand electrodes it's they're probably going to be talking about you know 60 000 electrodes or something like that implanted at that point maybe a million electrodes by that point and it's definitely going to be the case that if you want to make the best use of 60 000 or a million electrodes so really uh milk the dynamic you know bandwidth that's available out of that it's going to take some pretty sophisticated AI to do a good job with that probably stuff we don't have today but probably stuff we will we will have a lot more in 10 years and it's going to help a lot with that problem can you explain in a little more detail what the advancements in those uh neural networks like training uh could could look like well I guess you don't even have to know what the reason would be right like there there could be some phenomenon occurring in the brain and just because you have an advanced uh neural network ability ability to train the neural network then you could learn a lot of stuff about the brain that you wouldn't be able to today is that true well if you're talking about from a research standpoint of trying to understand the brain and what it does yeah that that's certainly true I'm thinking more you know when we're talking about neuralink I'm thinking about an implantable device where the objective of the device is to provide a high bandwidth output and maybe a high bandwidth input Channel also to uh to a human being who's engaged volitionally in in conversation right like you want to say something and you want to understand something which is being said to you or that you're receiving and neural Link's ability to facilitate that in ways that like your five senses and your ability to speak or your ability to type can't so there's I mean there's there's kind of two aspects to this one of them is the bandwidth right the direct bandwidth that you can get from a neural implant in principle is much much larger than the bandwidth that we that can that you can get in through your eyes or your senses and much higher much much higher than the bandwidth that you can express with your voice speaking or with your hands typing especially the typing thing is that's in a specialty severe constraint right now it one so that's one part of it the other thing is like if you close the loop it gives the there's this other level of stuff which can happen where you've got signals going into the brain you've got the the brain's ability to respond directly to those signals and provide feedback directly back through the interface so the uh the brain itself is very plastic it learns and it adapts really quickly and existing brain computer interfaces they all take advantage of this to a greater or less lesser extent I somebody who receives a brain implant you have to learn to use it there's a training process that you go through and the the biological manifestation of that training process is that your brain tissue is reconfiguring to enable itself to do what you want to do as to the best of its ability with that interface whether it's understand what's coming in or clearly Express what it is that you want to send out if you can close that Loop tightly right at the level of the brain tissue uh it's very likely that that learning process is going to go much more quickly it's going to be much more nuanced and adaptable and respond quickly so uh AI techniques are allowing us to make the machine you know there's there's a brain computer interface as a brain on one side it's got a computer on the other side and AI techniques the brain is already plastic and adaptive learning machine with a lot of bandwidth and an ability so to get these two to really talk to one another they have to learn a protocol you know when you plug this every every human brain is unique like fingerprints are unique but even more unique than that when you're born there you your your brain has a rough plan for like this part of your brain and which mostly comes from like your your genetics determine uh the the course wiring of your brain like you know your eye the nerves from your eyes they go to a particular part of your brain the nerves from your ears they go to another particular part of your brain your hands you know all the muscles in your body they go to a particular part of your brain but when you're born like very little of that this is for human beings in particular for some for some animals this is different but when you're born very little of that neocortex is pre-programmed to do very many of those things at all which is why you know infants have no control of their arms or legs no very little volitional control over their arms and legs they develop it really really quickly but it takes a lifetime to develop really nuanced control of all these things and that's your brain that's the gray matter in your brain you know the part that's really adaptable adapting to the fact that the white matter the wiring is is all in there you know like you know the contractor came in they built the building they ran all the cables and that kind of stuff and now you've got some people in there programming the computers and the computers are figuring out which wire connects to what and how they're going to control you know all the functions in this building or or what or whatever I mean the the big wires they are kind of uh constrained like you you kind of get a set of those but one of the amazing things about about mammals and human beings in particular because our neocortex is so big and so many of our functions are out to the neocortex but almost all of that is incredibly plastic and reprogrammable so and luckily for you know the people who want to make brain computer interfaces that neocortex tissue it's totally Exposed on the surface of the brain just underneath the skull so you we can get access directly to all that wiring so when you build a I mean it's very likely it's you can't if you had to do a thing where you for a particular user like say you want to build um interface where a person can think sentences thoughts you know that kind of stuff and the external system interprets it but you're not going to rely on the brain's ability to adapt at all right in that case you have to figure out which neuron is for the word the and which neuron is for the word you know and so on and unfortunately the brain doesn't organize words in that simple kind of sense even you know it's not like you have a neuron per word or whatever the deal is much more complicated than that so if you can't rely on the brain to adapt at all you need just like so many electrodes and so much processing so much time to like figure out what the existing hard wiring is because it's different for every uni individual I mean we know you know to a first approximation that a particular part of your brain is where most of the language stuff goes on and various you know the the reason that that's where the language happens is because that's where the most important nerves in your body that are associated with language that's where they terminate in the cortex so we so we can reliably say well your Language Center is in this part of your brain but we can't reliably say for any into for any single individual exactly which neuron is doing what thing and so if we're sort of shotgunning a bunch of electrodes into the brain which is kind of what neural link does they don't figure out ahead of time with any specificity what you know any particular point in the brain is trying to do they pick a section and they know that it roughly correlates with some function that they're interested in whether it's motor or speech or Vision they put a bunch of electrons in there and now the system has to learn how to talk to the brain and the brain has to learn how to talk to the system how well that learning works like to what degree you are able to express nuance and express it quickly to quickly Express a sequence of thoughts and with with specificity and not you know and accuracy so there's not noise that's a process that the brain half of the interface has to go through and it's a process that the machine half of the brain half of the interface also has to go through both sides have to learn to talk to one another there's a handshake and they're going to have to figure out what that protocol is so you know the brain has its learning algorithm and we're not going to be able to mess with that very much luckily for people who build BCI interfaces it's already an amazing algorithm it's incredibly flexible um and especially in the case of like human beings have primates uh this is probably turned dolphins and whales and stuff too I don't know as much about them but you know you know that tissue it not only is it really flexible and really dense and really high capacity capacity but humans also have an amazing amount of volitional control and we know all these tricks like visualize this imagine you're smelling that imagine moving your arm say in your head this word I mean we've got all of we've got this very powerful imagination uh that uh that we can use to learn to activate all these different parts of our brain and in in complicated ways primates can do this too and that's why monkeys can learn monkey Pond right it's you know pigs can do it less mice can do it less and so on pretty much anything with a with a cortex we know approximately how that goes but animals with simpler brains and and less of the rain dedicated to the cortex they're in their volition is less capable of expressing itself with complete plasticity so humans are an extreme of that which is great for planting these systems in the brains so now there's the other half um what does the machine side need to do well we we want a similarly plastic similarly dynamic system on the machine side to make the most of that interface as quickly as possible if if you get a neuralink implant and then you have to spend 20 years you know learning you know to use it before you can like move a mouse cursor around a screen the ROI on that isn't as good as if you can implant this and people can get the mouse cursor control in hours or in a day or two and if you want to be able to get I mean the mouse cursor is kind of low hanging fruit that's why they start with it with primates because you know we you we know with some amount of specificity what part of the brain controls like your right arm or whatever and so you know if we look at some electrodes there in a primate's brain or whatever um and you do something that trigger you know you engage in some activity that involves moving that right arm around well you can you can look at how the right arm moves around and you can correlate that with the signals that you see coming out of the brain you can figure out what the relationship is that's what they did with monkey Pawn right I mean the monkey starts out playing Pawn with a joystick and they have a set of electrodes implanted in the part of the brain that's for the arm that's controlling the joystick and they figure out what the correlation is between the signals in that part of the brain and how the joystick moves so this is a really simple experiment in in a sense right you just you watch how the monkey's brain tries to move the arm you interpret those uh and when your accuracy is good enough like once you've trained the algorithm to the point where the brain where the algorithm can look at the brain tissue and predict accurately what the joystick is going to do you know then what you can do is you unplug the joystick and you can just use the brain to control the thing and that's what they did with monkey pawn and this is an approach that's been used a couple of times and that's one of the reasons why this motor cortex thing is a really popular way to do the experiment because you can train an animal to do something physically with its body you can monitor that you can learn to interpret that uh and you know that it makes her pretty simple but but imagine that what you want to do instead is is do language right so that you have the equivalent of chat GPT or some some language interface AI that you're talking to directly this is a much better interface than a keyboard and a mouse because it's much faster like almost everyone can speak much more quickly than they can type and even for people who can type really really fast you know like if you try to speak quickly especially if you try to speak in a really abbreviated fashion to someone you know well you know uh people who've been living together for years and know each other really well they can have complicated conversations with relatively few words because they under because you know they look at context and that kind of stuff and so the bandwidth that you get is going to go up a lot if you have more than you know a keyboard has 50 60 keys on it something like that and an action is choosing one of 60 keys but when we speak every word that comes out of your voice even if you don't consider like tone and context and that kind of stuff it's one of twenty thousand like your average adult frequently has a 20 000 word vocabulary that they use frequently and and most people my age you know they've got 50 000 words that they can draw so this is a keyboard with 50 000 buttons like that's very high bandwidth if you can you know if you can if you can hit several buttons a minute you know that's a crazy high bandwidth and that's that's kind of table Stakes for where we want to get with neuralink is to have that kind of bandwidth so that so that the bandwidth coming out is hundreds of kilobits or something a human brain is like it's like one of these really big Google data centers if you look at the processing capacity I mean there's really an enormous amount of processing capacity that every human being is carrying around inside their head but you may be too young to remember this but when I when I started the first time I was able to get on the internet I had a hundred baud modem right this is like this this is this is back when uh modems were even before modems became popular like this is really really slow 100 pot modem right so imagine that you've got a Google Data Center and the only link out of it is like a single 300 Baht modem link like that's the human humans right now that 300 bod is what we can talk but there's all this Computing capacity behind it right so you know we have human language uh has you know evolved to allow us to get the most value we can out of a relatively small number of words we're able to say we do this through all kinds of tricks that some of them are signal processing a lot of more context in dictionary like when we're having a conversation and and I'm not just saying words according to like some dictionary thing but I'm modeling you and I'm modeling your model of me and and I'm modeling uh you know you can go around these Loops right and we we do this complicated modeling of each other when we have a conversation because it allows us to say to say more with fewer words because our pipe is so small and we have so much we want to say right neurolink is in at a real fundamental level it's trying to break that thing so that instead of having a hundred Baht modem you know you have a 10 10 megabit cable link you know or 100 megabit cable link or something like that Vision input is a little bit easier uh Vision vision is close to a megabit so you know the thing is you can only you know human vision is designed to recognize certain kinds of things so you can only encode so much visually like a human watching a soccer game can see a whole lot of things going on you can watch a room full of kids playing and extract a lot of mean but if you're reading text it's kind of slow you know because your attention focuses from word to word to letter to letter for most people as they go so even though in principle you know the bandwidth coming into your eyes is really high as soon as you move away from naturalistic phenomena like objects moving around in the real world and we get to these abstract things like words on paper it slows down tremendously right and neural link can potentially break all of those gaps because it can go around your visual cortex's need to see kids and trees right it can your visual cortex isn't so isn't it's not great with really abstracted inputs and text is highly abstracted and so so reading is slow reading is super slow compared to a lot of other things that we can do that's why a picture's worth a thousand words all right because it just takes a lot of words to express what you can see in a glance if the thing you're looking at is some natural phenomenon that people know how to interpret easily right if you're looking at noise on a TV screen like a staticky channel that might be megabits of content but you're incapable of absorbing it because your visual cortex hasn't is not organized to grab you know noisy data off a screen and turn it into something but um so uh AI is going to be really important for making that interface work because both sides of that interface have to learn to talk to each other and that you have to do that from scratch every time for every single human who gets an implant so we want those we want those algorithms to be efficient we want them to be really high bandwidth we want to be quick to learn we want them to automatically adapt because your your brain is constantly remodeling itself uh every time you go to sleep you wake up in the next morning and your brain is different like the functions in your brain that do things they've all migrated slightly a little to the left a little light they've reallocated a whole bunch of stuff and the you know a brain computer interface is going to have to be adaptive enough that it can keep up with your brains uh constant imperative to remodel itself it's interesting that you had said uh you want to make it like it's a small pipe but you wanted to make it much larger but felon was talking about how that communication is like a straw it's going through a straw right now and you want it it's nothing it's shockingly I mean it literally it's like you know you imagine one of these Warehouse size Google data centers worth of processing capacity because that's what that's what brains have if you're talking about CPUs certainly that's what a human brain has and yet one single dial-up line you know 300 bod coming out of it like that's that's all the communication it's got right it's amazing that we can make good use of of it and yeah of course we we end up in this space because we have you know we start from you know Evolution started out with no language the fact that that we adapted language that we have it at all and that it's so useful is a testament to the utility of even a little bit of communication when you have almost none mm-hmm but it you know human ability to communicate is severely severely constrained by our i o limitations and that Evolution will just accelerate with uh with all the advancements in Ai and then with neuralink2 like we'll be able to communicate even faster and and more efficiently than the rate of change was before right and that's the that's the idea there's I mean there is kind of an organization to the brain of an adult human being which it comes to after a lifetime of experience uh and you know in the case of me and you we got to this point in our lives without having a neural link so you know if we're not going to relearn all of that stuff from scratch the uh you know we are to some extent you know adults like you and me we're going to be constrained by the patterns our brain already learned and the way it already organizes itself not having neuralink so people who grow up with neural link they're very likely to have a very different set of constraints and they're very likely to be able to have much higher bandwidth than an adult implanted with this even even if you were 10 years old when you got it as opposed to 20 years old when you got it that would be and 20 would be way better than being 30 right in terms of the Fidelity that you'll be able to get to but um so they're definitely organizational constraints because the human brain is really plastic but it's not infinitely plastic it can't it has to reorganize itself according to the stimuli you know at every when it reorganizes it's not like you can go to bed and you go into a coma for a month while your brain reorganizes itself you got to be able to wake up in the middle of the night and run from a lion if you have to right so every single point of change your brain has to you know your all of your your mental functions they have to remain functional as they adapt it's not like when we remodel a house and you know the owners can move out and then you tear everything down a month later maybe or two months later or you know if you're a normal contractor six months later you actually get your house back and it's working again right the brain has to stay functional all the time so that means that um you know if you're an adult and your brain is is built a certain way then it can it the the foundational functions of the brain they can only change so fast so if we're doing a version of neural link that's really trying to get access to those foundational things and you would like to remodel those foundational things in order to make it you know what the brain is doing better suited to what neuralink is uh you know to the purpose of using neuralink effectively that's going to take time it takes much less time for a 10 year old than it than it's going to take for a 40 year old so when somebody has sleep paralysis or they're Dreaming or they're lucid dreaming uh is is it known like why yeah is it known why why they're paralyzed or why they're not able to like have a lot of their functions like in general yeah it's I um so the so sleep paralysis is it's a side effect of the Persistence of your of your uh your motor cortex disconnect so you when when you sleep at night you know you dream and it in your like one of the things that happens in your dreams is your brain tries to move your arms and legs around like it will it will attempt to act out the things that you're dreaming about and uh as a per you know presumably protective function your brain disconnects your motor cortex from your body when you go to sleep so you're not thrashing around in bed this is just a normal function everybody's got it so there are various kinds of dysfunctions of this switch that can occur they're rare but they do occur and one of them is that the switch cannot turn on when you go to sleep or it can be slow to turn on so those people they move around a lot in their sleep some of them get up and walk around in their sleep right it's you know there's there are manifestations of sleepwalking that are literally like this thing isn't disconnecting the way that it should um the other one is what we call sleep paralysis which is you wake up and the switch is still on your your brain is still disconnected from your body but you have become conscious so then you're you you go to move your body and it doesn't move because the switch hasn't disconnected yet and people who experience this they experience it as being paralyzed and I I guess there's there's this uh um I don't know if it's a hallucination there there's this kind of way of experiencing this which is that it feels like someone is sit or you have a great weight on you holding you down and you are unable to move so of course people experiencing this they're waking up in the middle of night they're kind of groggy they're just emerging from a dream or something like that and their body's not moving right and suddenly they have this kind of panicky moment and the in your intuition immediately goes to I'm being held down right like this is you can go back hundreds of years and you can find people writing in their diaries about how like they woke up in the middle of night and a demon was sitting on them or whatnot and a minute later the demon vanishes right and they they get their movement back after that after the switch like reconnects the way that it's supposed to when you get back so those are dysfunctions of sleep paralysis that's what you're talking about right yeah I'm I'm like again in shock at uh your breath of knowledge just like you happen to know this um let's see uh so Michael S on Twitter can nanotechnology be used as a part of neuralink specifically as electrical conductors Computing processors or as a biochemical interface with the human physiology uh you know nanotechnology is super well suited to assisting Us in making effective brain computer interfaces um in in a sense the the single biggest challenge of uh brain computer interfaces is that the things we want to interface to are so small and we are so big and all the things we make are so big by comparison so nanotechnology I mean the word gets used a lot of different ways um you know once upon a time when it was first coined it meant you know uh molecule size complex machines and if you uh you know certainly if we had those those would be incredibly useful in fact like if you if if we had mature nanotechnology brain computer interface is Trivial like it becomes a very simple problem to uh you know to overcome like all most of the problems of you know that that we that we really struggle with are in large part because the smallest devices that we can make flexibly and that we can control well aren't all that small and the components we want to try to talk to in the brain are really really small we were talking about how um you know the the needle size that they inject that neural link is currently using is on the order of 40 microns and 40 microns is the width of an entire neurocortical column with like 100 neurons in it right so you're I don't know it's like imagine that I'm trying to build a you know I'm going into a Data Center and I you know I want to shove a sensor down into the data center but the sensor is the size of a whole rack of computers right so I'm going to ram it down through the ceiling and you know the racks are just going to have to move out of the way right it's you know the things we would like to be able to talk to are really really small and it and it's you know we have a hard time so another way of interpreting uh nanotech which is a the way the term gets used more today is talking about making things with a degree of precision and control that that we are we are manipulating the structures down below the scale of about 100 nanometers or so that's that's just generally what people consider nanoscale right once you get below 100 nanometers um so you know everything that we can do at less than 100 nanometers like every significant manufacturing technology that we have that will let us get down below that scale is a potential contributor to bring computer interfaces uh so in that sense we have nanotechnology today there are many we have many technologies that can get down to that scale like we can make transistors really small um an interesting thing to consider is the possibility that that a a somewhat more mature version of neuralink today right neuralink today you know all of the electronics are in the little puck thing that gets implanted you know at the in the cranium and then the you know the probes that go down they're just wires I mean they're complicated super super skinny wires but it's just a wire it doesn't have any Active Components in it it's an insulator it exposes the conductor at some point near the tip and that that's all there is but imagine that you could embed an amplifier in that so that it was an active component not just a passive component so that the wire that goes back up to the electronics is a is um is a controlled impedance transmission line and not just you know not just I mean you know every wire is a transmission line but it you know if you can do a controlled impedance transmission line you can you can get more signal more reliably at lower sensitivities you can transmit it farther with less degradation of signal fidelity and to do that really well you want to be able to embed an amplifier right in the tip like right where the wire is well physics says you can do that I mean we we make transistors in integrated circuits now we can build amplifiers they're smaller than 40 microns much smaller than 40 microns so in principle you could make an amplifier using existing integrated circuit technology and stick it on the end of one of these things and it would be as small as the tip of one of these probes that we're sticking in the brain right now and you could have all kinds of signal processing right down there and then you could you could do the pre-processing right at the end and go now even if we're not talking about complicated machines like that electronic machines you know on the scale of what the one micron scale of the end of the tip of one of these things you could imagine for instance for instance one of the things we have right now is if you look at the you know 100 nanometer scale of one of these probe tips it you know it's like a telephone pole being shoved into you know a complicated you know shove a telephone pole into a bush right that's a probe going into your brain right now where all of the leaves are you know neurons and or synapses and that kind of stuff like the the probe itself is big and crude compared to the structure that you're trying to get access to but the just you know the fact that cells can organize complicated structures at much much finer scales then we can build the probes too right now it says physics allows that to happen it's a it's a matter of us working out the manufacturing technique so nanotubes are something we know how to make now they're super useful uh nanotubes they're there so nanotubes uh there are lots of different ways to make them basically anytime you make a wire or a tube you take a sheet and you fold it uh on a molecular scale and you get it down to the order of you know 10 or 15 nanometers or something like that or even smaller you can get really small um we call those nanotubes and uh if you make them out of if you use carbon as a substrate carbon is a semiconductor just like Silicon is a semiconductor and uh if you fabricate a you know a nanotube you can you can you can take a sheet of graphite you can take it you can do this with silicon too it's just easier with carbon um you can roll you can make a single layer nanotube or multi-layer nanotubes if you make multi-layer nanotubes then the the you know where where you take the sheet of graphite and you roll it in the tube and then if you if you look at graphite it's it has a hexagonal pattern you get like six carbon well it's a sheet and the carbons embedded in the sheet they have three links to other carbons in their layer and the fourth link every carbon carbons all have four bonds carbon you put crystalline carbon has four bonds and so when you make graphite the crystalline carbon every crystal and carbon has three bonds that connect to three Neighbors and if you do that across a grid you get a hexagon you get this hexagonal pattern these repeating hexagons and every other every other at carbon atom its fourth Bond goes up or down so you know if you go along the molecule you get you know Carbon spare Bond up carbon spare Bond down carbon sphere Bond up so you take two layers of graphite and you sandwich them against one another so that the UPS from one connect to the Downs on the other and the downs from but well that's what graphite is so you get graphite you look at the structure graphite that that's how it's built so the thing is so graphite looks kind of like chicken wire right you just if you roll out chicken wire it's used for fencing on farms or whatnot you know it's this hexagonal pattern of wires you you roll it well you can imagine that if you roll up this hexagon you can roll it up a bunch of different ways like you know if the hexacons go across this way you can just roll it and match one against another and that makes a really strong uh pattern that has a certain alignment there's a certain amount of stress between the layers because each subsequent layer is a little bit larger so they don't match perfectly you get stress inside the structure and you can you can slightly tweak the way that the that these layers roll against one another and Vary the crystal structure that forms between the adjacent layers as you as you tweak these things and you can generate almost any kind of band Gap that you want band Gap is an electronic property that we use that defines the electrical characteristics so you can make carbon nanotubes into semiconductors with whatever band Gap that you want and they can become an active electronic components like all of this is within the realm of possibility it's been demonstrated there are companies now working on trying to make integrated circuits where the transistors are you know nanotube layers that are deposited down on substrates that are made with more conventional semiconductor manufacturing techniques so now can you do you can also do experiments in the lab where you make you get single carbon nanotubes you build them at some kind of scale and you tweak their electronic properties and it doesn't have to just be graphite there are lots of other materials that you can use also graphites popular it's very flexible we know how to make it it's a semiconductor it's really strong you know it's got a lot of really great properties for doing this kind of stuff but one of the things you get is you get a really small wire out of this because the graphite is conductive you can tweak the conduction also by doping it or by changing the crystal structure uh carbon nanotubes are potentially orders of magnitude smaller than the smallest wires that we can build today so the you know this this like I said it's like you know the Probe on a neural link today is like a telephone pole jabbed into a bush right and the uh the carbon nanotubes are much closer to the scale of wires coming off of that telephone pole so like if you wanted to connect with more Fidelity resolution proximity to all to all of the cells and synapses and axons and dendrites that are in this tissue that you're trying to interface to well in principle a carbon nanotube can go right down to a particular synapse and Sample it directly it could go right down to a particular branch of a dendrite or an axon and you could sample that directly so you know when we get the control and we will eventually get the control to be able to make these things work you know the way that we want them to to be able to build structures at that scale yeah that's going to be amazing for brain computer interfaces because now you'll be talking about I mean you might still at some scale be shoving a telephone pole into a bush but once you get the telephone pole down there you know you will have a a very large number of very smart little wires come out of that telephone pole that will be able to interface to the surrounding nerve tissue with a lot of fidelity and precision that will give you much better um signal characteristics for both your input and your output on that so yeah Nano I mean nanotechnology is already to some extent playing a role I mean we talked about how they're trying to use amorphous silicon carbide uh as an insulator uh for these electrodes right now I don't know if they're doing that right now if they're just looking at it but you know if you you're probably talking about a layer of silicon carbide which is on the order of nanometers thick in order to make that work because of the characteristics of silicon carbide you can't put a very thick layer on and still use it as a insulator for a flexible electrode so in certain respects there are already going to be aspects of what they're doing that are nanotech and certainly the the silica the ICS these days are getting really really close to to nanotech it um if you're you know when we talk about the I guess the the TSM the the process they're using right now is uh you know 20 so-called 28 nanometer uh integrated circuits um which means that the smallest transistors the smallest to a first approximation this is a very the terminology is it used to be that when you said uh you know 100 nanometer or 90 nanometer or 300 nanometer like back when we were at that scale it was really true that when you looked at a transistor the smallest feature was like 300 nanometers or whatever as we've gotten smaller and the trend the shape of transistors in ICS has gotten more complicated that simple metric we used to use doesn't doesn't work as well and it is now the case that when you talk about a five nanometer silicon process you're no longer talking about a process where the smallest transistor component is five nanometers it's roughly approximate to that but that's that's no longer true but you know we are talking about single digit nanometer scale it's minimum feature sizes for the electronics that are going into this and of course neurolink it wouldn't be possible without the nano technology of semiconductors that we're using in it right now because you you have to get a certain level of processing capability you have to be able to do it with a certain level of efficiency you have to be able to do it in a certain package size and you know the nanotechnology that we enjoy with semiconductor processes today is totally critical to to making what we do with nanotech what we do with neuralink possible now and it's going to become even more important going forward there it's you know we will have one nanometer or subnanometer scale semiconductors by the time neurolink is really a mature technology and it'll get used for this kind of stuff and I wouldn't be surprised if uh if a significant part of the technology isn't making really small ICS that go on the ends of the probe tips that go that get injected directly into the tissue so that you can get uh more IO with better signal to noise ratio than you can get by just sticking a telephone pole in a bush right this is James doma's prediction for neurlink technology we have at hand to make the to make it as good as we can and um neuralink is today and is going to continue to be for quite a long time limited by the limits of Material Science and nanotechnology is the bleeding edge of Material Science sure um okay so I have this thing that I want to share can you see my screen that's kind of small let's see well I I could stretch my window a little larger okay that's better go ahead okay um so Warren redlick posted this poll on Twitter um and then he said what's the greatest risk of artificial intelligence um and then you responded with this and I'm gonna say let's choose the the expectation value of impact so total area under the curve I guess of impact um what do you think is the greatest risk okay uh I think the greatest risk is the uh is it spoils us to the point What's it Elon has some expression like it's like uh you uh the the fear that Human Society ends not uh with a bang but with a whimper and he said something like a whimper in adult diapers or whatever that um I think the biggest risks the biggest kind of intermediate duration risk that we have is that we get all we want and we stopped caring right um and uh like as in terms of area under the curve right like human society if humans stop caring if we stop working if we stop trying to make the world better if we stop trying to help each other you know if if we if the if Society can you know if if people can individually get what they want out of life and they don't need other people they don't have to interact with other people you can live in your own bubble where you just have all the things you want and everything in the bubble caters to you like that is a kind of death of society and uh you know and it does die in adult diapers because the you know the incentive for having kids having kids is a super risky undertaking right it takes a lot of time you get really invested in it um you know if people are honest with themselves it it there's no it's there's no guarantee that it's going to turn out great for you it might turn out great it turns out great for a lot of people but for a lot there are a lot of people it doesn't turn out great for and you know if you can live in your virtual reality fantasy and you know you can have exactly the kid you want if you want to have kids right the computer will will fabricate the kid that you want in your fantasy world and you can have the experience of raising them in that world if you want to right you don't you don't have to take any risks you get everything you want right that's um area under the curve in terms of like actual risk likelihood of outcome that's the thing I think is probably the biggest risk it's not that we do something terrible it's that we get everything we want and stop striving because of that and that at some point not long after that you know that that's my favorite answer to the Fermi Paradox by the way um the you know the Fermi Paradox is Enrico ferme uh was the this the story behind it the reason it's called the ferment Paradox is because Enrico fermay first posed this question apparently he did it over lunch with some other guys that I forget where it was I think it might have happened at the Manhattan Project but anyway there's a you know a bunch of Nobel Prize winners sitting around the table and they're talking about how big the universe is and like what the probability is of life being out there and all that kind of stuff and for me says where are they right which is you know it's it sounds like a simple question but uh there's a really profound reality that underlies it which is the universe is really really astoundingly mind-bogglingly big I mean you think the Earth is Big the Earth is nothing compared to the Scale of the Universe it's really unbelievably big and it's been out there for 13 billion years like we're nowhere near the first on the scene so if there's any likelihood to intelligent life at all the Galaxy I'm assuming they don't die off quickly right that's another thing like if intelligent life arises only to vanish A Thousand Years Later well yeah there aren't going to be any out there because they never survived long enough to to uh to get to get out there there are lots of different answers that have been posed to the ferment Paradox but the thing that seems most likely to me is that uh you know you is it before you become a kardashev one civilization everybody gets what they want and they don't have to do anything for it and you know we're we're the product of Natural Evolution you know our instinct is to you know is to cover our needs and be grateful for that right once you're we were talking about the limbic system and how your neocortex is kind of like your limbic system is kind of like this dumb brute just wants things right and your neocortex is like this you know uh this amazing tool that your limbic system has and your limbic system uses it to get what it wants out of the world well if we get if we get the tertiary cortex and tertiary cortex is dramatically more capable than our neocortex is then our limbic system is just going to get what it wants and you know one of the things you can observe about people is once they've got what they want you know they just watch The Flintstones right you know they watch I mean so anyway that might be the fate of civilizations right so that's what I think the big risk is I think the big risk is that we get what we want right and that you know the striving that all the generations before us well in western civilization we think of uh in western civilization we think of like man struggling against nature to achieve you know a greater great purposes for himself his species his family his country and uh so these are the higher callings that we work towards whether it's you know being cancer or uh feeding everybody or ending War you know whatever these things these are all these these great things but what happens when you get all that right you don't have cancer and people don't die and there is no war there's no you know we we're you get to a post-scarcity civilization that means that nothing scares like any everything anybody wants they can have and Once Upon a Time you know the the counter argument to post scarcity civilizations was well people will still be scarce right and opposed scarcity civilization you might have all the material stuff you want you still got a finite number of people like every single person has one person completely dedicated to them which is them right but AI totally solves that problem you know if a if if if you know if there are machines dedicated to being your friends and they're as smart as people they can be every bit as funny interactive they can they can mimic any person that you want with perfect Fidelity because they're a superset of that person now all of a sudden there's nothing scarce right everything is abundant and there's nothing that you have to work for you can have anything you want at any point in your life and you don't have to do anything for it except want it right I I you know in a very real sense that's the end game for technology and an important question is now what when you get to that point yeah yeah I see I I had thought that when Elon brought up this um this claim about like people or civilization dying off with a whimper I thought it was primarily based off of like people are continuing to live longer and longer and they're having fewer and fewer kids but actually it's both what you're describing and also like this whole people living longer and lower birth rate yeah yeah because the whimper that's the extreme case right and people live infinitely long there's arbitrarily low birth rate you know uh a universe has been around for 13 billion years the Earth formed four and a half billion years ago or so um the first bacterial life is two one billion years ago something like that we got complex multicellular stuff on the order of a couple hundred million years ago with the Cambrian explosion that kind of stuff dinosaurs were 60 million years ago 60 million years ago like the human humans have been around as a species you know most people would say on the order of a million years before between before where you draw so like for a million years there have been people here for at least hundreds of thousands of years people have had language and societies as complicated as the societies and languages that we have today for a hundred thousand years we were still basically Aborigines right and then you know in 1400 we had the Renaissance and and we've been on you know like a train to the Future ever since then and here we are and it's just getting faster and faster and faster you know so there's all of this history behind us bringing us to this point and we're really really close to just like everything humans ever wanted and we're getting there really quickly and it it's worth asking you know like where does that leave you when you get there trains pulling into the station now right and you know we got to decide where we're going to go when we get off this train yeah it seems like uh there's definitely like a certain level of fulfillment that's needed after like the initial basic things are covered basic needs are covered it's like that Maslow's hierarchy of needs right where it's like the food and sleep and shelter and then after that the Fulfillment it's like a lot of people now they're just retired early and there's no more like a Siobhan posted this from an early shoes like yeah people need quests to pursue and yeah the side quests yeah right yeah yeah but it's like a lot of the Silicon Valley Tech nerds or anybody now nowadays with all these different Creator tools anybody could hit that Financial escape velocity super early and then it's like what do you do there needs to be something yeah that's one of the things that makes Elon such an anomaly right is that it is like you know the common path is what Jeff Bezos did right or what Bill Gates did you know you get your 100 billion dollars and then you disengage like why should you keep suffering why should you struggle against all these things haven't you done enough don't you and deserve to like enjoy your life now you've done plenty for people just relax you know life is short you struggled for so long you know and that's a super attractive uh you know people who get to that point and they have an obsession which is strong enough to keep them productive way past the point that they have everything they need I think they're a rare breed and you know one of the solutions potentially too uh to this dilemma that I just uh you know posited about us getting everything we want and it being our undoing is uh either culturally or you know with some sort of genetic intervention making people more making everybody like Elon right so you just don't you're never satisfied you always want more right it's I mean there's there's good sides and bad sides to that right like if you're if you're never satisfied right that can also lead to all kinds of pathologies and the destabilization of society like even a relatively small number of people who are unsatisfiable can be really destructive to a society if they have influence um but on some level if everybody or virtually everybody is easily enough satisfied that we soon get to a point where we can satisfy them that itself that's also a really big problem potentially but anyway I mean so we're going a long way with Warren's question like I could come up with all kinds of other potential downsides to AI the reason I asked his question is because I've literally heard a hundred different plausible you know uh problems that that AI can do societies are really complicated things they're really really complicated things and they're super Dynamic and they evolve all the time and every time you bring a new technology into society that messes with the uh the Dynamics you know the stability of aspects of that Society you you are you know the society is going to undergo changes that some people in that Society are going to see as very destructive and very negative thing other people might see them as really positive things um uh but so you know AI is going to be the the most powerful technology like a lot of people have said it's the last technology right it's a deep mind it's right there in their motto it's like solve solve intelligence and use it to solve everything else and there's a sense in which that's really true right you you solve intelligence we're done the whole human race can go on vacation now all right yeah this was my second conversation with James if you missed the first conversation check it out here and if you've already seen that one too there's another video right here that YouTube's recommending specifically for you hope to see you at the next episode

Channel: Neura Pod – Neuralink

Views: 27,458

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: xOmDm477NPQ

Length: 117min 32sec (7052 seconds)

Published: Fri Feb 17 2023