萊克斯獨白 #2：Neuralink 的未來

Lex Fridman Podcast – Lex Solo #2: The Future of Neuralink Machine transcription by Whisper large-v3-turbo (MLX, local). May contain recognition errors. This is a solo episode of the podcast, one of many, I think. I'll add these to the RSS feed so you can listen to them as well. Hopefully it's interesting to some folks. These will probably have a Lex solo number before the title of the episode, as opposed to just a number for the regular interviews. The aim is for these episodes to be focused on a particular topic, at times challenging, at times personal, at times exciting to me on a technical or philosophical level like the episode today. This episode is on the long-term future possibilities of brain-computer interfaces in general, and Neuralink specifically, based on the recent update on progress from the Neuralink team. I have a basic outline in front of me with some ideas, but most of it is just off the top of my head, so I hope you're okay listening to that kind of thing, about my general thoughts about Neuralink and the recent update of progress. I was fortunate enough to attend the demo in person as a random visitor, really, and chat with another random, but much, much smarter visitor, Jim Keller, with whom I did a podcast a while ago, and we agreed to do another with him soon. He's one of the most interesting and brilliant people I know, so it was great to catch up. But outside of that, I was just a spectator like everybody else watching online. I have no insider information, have no interest in insider information. I'm just a fan, long-time fan of the human brain, and anyone who's working hard to understand its inner workings. The general sense I got is that there's a lot of exciting engineering and scientific challenges that the big and varied team there is tackling. I think it's a really exciting place to be. Just lots of ideas swimming in the air and lots of brilliant people. It's always exciting to me to sort of be in the presence of great engineering teams, so it's exciting to see that. But what I found especially exciting for my romantic and scientific soul is the long-term vision, the dreams, the possibilities that were mentioned by the team in a spontaneous final question that was asked where every member of the team up there answered their own version of what they're excited about to see in the next decade, two decades, long-term, future possibilities of this technology. So this video is about that. My thoughts about the possible ways that Neuralink might change the world and the human condition. I'll try to stick to some categories, some structure, and try to discuss off the top of my head my thoughts of the possibilities that fall within those categories. I should note here that a lot of the things I'll discuss are long-term visions of the future. To make all of these visions a reality is exceptionally difficult, there's a journey of many breakthroughs required. But I think we are now in the realm where a lot of these things are scientific and engineering challenges that can be solved by great teams, by bold innovation from many companies, not just Neuralink, hopefully many others, hopefully many competitors, that push the boundaries of what is possible. But this video is about the visions of the possible futures. And I think great efforts of humankind start with a vision. Let me give you a quick outline of categories within which I see some exciting possibilities. So first is alleviation of human suffering. Second is understanding of consciousness, intelligence. Third is augmentation of the body and mind, and generally augmenting reality. Fourth is gaming, and beyond that, virtual worlds, virtual reality. Fifth is all the engineering challenges around merging biological systems and computational systems, basically tech. Sixth is telepathy, much richer forms of different communication. Seventh is saving and replaying of memories, but also saving and replaying of mental state or mind states, period. And finally, eighth is merging with artificial intelligence, all the exciting possibilities around that that I'd like to discuss. So since I'm Russian, let's start by discussing human suffering. I think first and foremost, as was mentioned by the team, is the possibility that Neuralink might help alleviate human suffering. The nervous system, the brain, at the very basic level is the source of pain. That's both physical pain and psychological pain. So you can talk about anxiety, depression, trauma of all different kinds. The ability to measure signals from the brain, and perhaps more importantly, the ability to send signals and in a closed loop interact continuously with the brain, sending signals in both directions. It seems like it provides a very rich toolkit with which to start to deeply understand the human brain generally, but in the nearest term, to focus that exploration on the understanding of neurological diseases so that we may first, of course, understand, and second, to treat them. A huge number of mysteries yet to be uncovered at the very basic level of how do we treat some of these diseases. And that falls into the category of human suffering. You know, we often think about suffering as arising from the environment within which the individual lives. So by placing the focus on the environment, it allows us to kind of be hopeful because we can make the environment better. The source of trauma, the source of anxiety, the source of depression, all of the things that come up in political discourse, these are all things that we can do something about. So that's what we focus on when we try to alleviate human suffering. But from another perspective, the real source of suffering and pain is the human mind, which creates the experience, the lived experience from the perception of the external environment and the perception of the internal environment. Now, there's a lot of discussion of meditation, exercise, a lot of social programs and education, all kinds of things that aim to help the mind. But in addition to that, the exciting possibility with a brain-computer interface is that we might be able to accelerate our understanding and treatment and control of the internal environment of the mind. Now, of course, it's also important to say that there's injustice in the world, there's evil in the world. Neither Neuralink or any other piece of technology will be able to get rid of hatred in the world. But the hope is that at the individual level, you'll be able to aid in the alleviation to some degree of all the sources of neurological suffering. So second category of future possibilities in Neuralink have to do around our understanding of how the brain and the mind works and all of the things that derive from that. So basically, we'll supercharge research going on in neuroscience today. So first is understanding how the brain works at the functional level. So all the different modules from memory to perception to cognition and all the sub-modules of that. And as we untangle those pieces, it's possible that it will inspire or instruct us on the engineering side of how to build smarter and smarter artificial intelligence systems. So inspire totally new algorithms for learning systems, for reasoning systems, for knowledge base, knowledge acquisition, and so on. And as you push that further, of course, to me, as an artificial intelligence researcher, the exciting possibility is that we may be able to understand human intelligence where, not location-wise, but functionally, where intelligence arises in the brain or good answers to the question of what is intelligence. And the next step is, beyond just engineering AI systems, that may help us understand how we enhance it. You have all these productivity hacks, all these kind of life hacks, understanding from where our ability to reason about this world comes from might help us to really have some nice brain hacks to improve our ability to reason in a purely natural way I'm referring to. This is before any kind of augmentation from a computational device. Now, the next level of understanding the human brain, as was mentioned by the team, as I bring up often, is the fascinating, hardest, most interesting problem, I would say, of the mind is the hard problem of consciousness. Beyond intelligence, where does consciousness arise from the brain? Again, not location-wise, but functionally. And again, to be able to, with more scientific rigor, answer the question, what is consciousness? Is it a property of matter? Is it a unique, emergent property of the human brain? Is it something totally different that we don't even understand? Like, our mind is some kind of key into an alternate dimension that only psychedelics and a device like Neuralink may be able to unlock. So, at the risk of sounding crazy, it's an exciting possibility to take consciousness from, I would say, a field of philosophy in the 20th century to a field of science and engineering in the 21st century. To me, that's deeply interlinked with intelligence because I think there's a beautiful dance there between consciousness and intelligence in the human mind that's not easy or even necessary to untangle, but I think understanding one will help us understand the other. And finally, perhaps, interconnected with consciousness and intelligence, it might help us take the question of, is there free will into the realm of science and engineering versus the realm of philosophy? To try to make a rigorous study of where does this experience of making a choice, making decisions, like we humans have a control of the way the future unfolds from where that arises. Is that a real part of the fabric of reality or is that something that the brain conjures up? What I see Neuralink as, as I talked with Elon the second time on the podcast, I see it as a way to sort of get beyond the factory walls and see how the inner workings of the factories operate. As a scientist, as an engineer, and a bit of a philosopher, that's truly exciting. Third future possibility of Neuralink is augmentations of all different kinds. So, regaining the ability to move for people who can't move parts of their body. I mentioned neurological conditions that affect the mind, but certainly there's neurological conditions that affect the body. I mean, giving people who can't walk the ability to walk again or to walk for the first time is such an exciting possibility. If you've seen videos of people who for the first time are able to see color or gain a function that they didn't have before through technology, the bliss in their eyes is magical. Now, the augmentation doesn't have to be just in regaining the physical function of the body. It could be augmentation to the mind. It could be, for example, regaining the ability to see by stimulating the visual cortex, connecting a camera to the visual cortex. And perhaps more than regaining regular visual function, it could lead to superhuman level vision, whether that's expanding the spectrum, like ability to see infrared, or it's doing some basic augmented reality kind of things where some of the detections are done for you about moving objects, about the categories of objects, and all that kind of stuff. Many of the ideas here are the same as those explored by the work that people are doing in augmented reality devices, but it's very possible that the difference between a brain computer interface and glasses, for example, or heads-up displays is that BCIs might be able to create a much richer high-bandwidth experience with a fast, closed loop of perception, more so than the constraints that you have to operate under with glasses or HUDs. Fourth, a super exciting possibility for those of us who were once gamers, or still are gamers, is by creating an immersive gaming experience. So, BCIs might be able to, once again, read the brain and stimulate parts of the brain that enrich in some way the gaming experience. This could be very shallow kind of basic enrichment, just being able to measure levels of excitement, emotion, those kinds of things that can aid in the experience of the game. But also, again, as I said, with augmented reality, being able to stimulate the visual cortex in order to create an immersive visual experience. So, with a brain computer interface, beyond just gaming, you can start to think about creating virtual worlds, virtual reality. That's very useful for games, but just creating an immersive experience of all different kinds. Again, this is an open question, but there could be technical barriers in creating an immersive, rich, high bandwidth experience with a virtual reality headset versus a brain computer interface. It's an open question, of creating a fully immersive experience, what is easier to do in the long sort of arc of history. With the technology we have today, it seems clearly more doable in the short term to create virtual reality experiences with a headset as opposed to something that requires brain surgery. But that's not to say if we look at the long arc of technological progress, that the much easier solution won't come from the direct access to the brain through something like a brain computer interface. And again, I think bigger than gaming, a lot of people write to me about psychedelics, for example, which I've never done. But this would be an example of something where you can create visual experiences that are safe and controlled and can take you perhaps to some of those different multiple dimensions or wherever the heck you go when you take psychedelics in a more controlled way perhaps. And maybe even taking a step back into more kind of vanilla experiences of visualizations and meditation. So imagine the closed loop of being able to write and read from the brain in aiding the meditation experience. sort of emptying your brain from thoughts figuratively and literally. The fifth exciting future possibilities of Neuralink and brain computer interfaces is all the innovation engineering around the two-way communication between a human-made electrical computational system and a biological system. Just the fabric, the nature of the two design paradigms, not saying biological systems are designed, but they are designed through evolution. Whatever that resilient mess, mush of biology to the more structured, architectured electrical systems that are programmed explicitly and clearly, the communication between these two different worlds and bringing them closer and closer together is super exciting. First, at the very basic level, that could be all the innovation around robotic neurosurgery or even surgery in general. So, allowing robots to do what narrow AI systems do best, which is for basic tasks that have vision and control where everything is controlled in the environment, fully actuated system to be able to minimize the risk of injury, maximize the probability of success. so there's a lot of interesting innovations around just the robotic side of that. The next layer of that, when you look at some of the materials engineering and even the computational side of connecting the laces to the brain, so connecting the electrical device to the biological device, we may be able to understand how to engineer sort of physical computational systems that have some of the same nice properties of resilience that biological systems have. And in so doing, be able to work better with biological systems, but also just be able to be more resilient, more robust, more adaptable perhaps, or maybe come up with totally different ways that such systems can learn about their environment, just like our biological systems can at multiple levels. And another layer of that, when you look at what Neuralink is currently doing, they have 1,024 channels, the engineering around scaling that to, I don't want to put numbers out, but any number above that is super exciting. It's already 100x anything else that's out there, but you can imagine, especially long term, it being 10,000, 100,000, I mean, it could be millions, maybe billions. I mean, there's so much possibility of engineering breakthroughs about the number of channels that are possible that we can't yet imagine. And that's an engineering challenge of how to scale these number of connections, which are tricky to do because they have to live, exist successfully in cooperation with biological systems for months, years, you know, for long periods of time. That's really interesting. I feel like that's a forcing function for us to understand really how we can engineer systems that in the best possible ways are not only able to work with other biological systems but become more like those biological systems. So, sixth possible future of Neuralink and BCIs was mentioned a few times by the team under the flag of telepathic communication or telepathy, conceptual and consensual telepathy. So, I think in general to enrich the bandwidth in quantity and quality of the communication between two human beings. So, you can imagine being able to communicate not just through this kind of 1D realm of words but to communicate visual concepts first of all but also kind of mind maps of like multi-dimensional concept maps that are in our mind when we're trying to reason through things to be able to communicate those in some way. It doesn't even have to be kind of perfect replication but any kind of improvement, increase in the bandwidth of the communication between humans on the visual or on the conceptual side is super interesting. I think somebody on the team mentioned kind of art to be able to communicate creative, artistic the kind of things in your mind and share them with others without having to learn the skill of converting that art into something in a physical world that can be observed by others. You can sort of directly without learning the skill be able to communicate all the crazy beautiful things that are in your mind. I think for my world of like programming for example, it'd be exciting to think that two human beings at any level could sort of collaborate together. It gives a whole other meaning to pair coding where two people can collaborate together as they work on a project of any kind whether it's in the programming world or any kind of design world, architecture, any kind of illustrations, all that kind of stuff. Collaborations between humans for intellectual labor, for design, for engineering work or any kind of collaboration in the intellectual space. And finally, I think it'd be pretty good for podcasting. So for those of us who don't like the sound of our voice and funny enough don't like to be in front of the camera, instead of having to convert my thoughts awkwardly in a monotone voice into a microphone, I can somehow communicate them in a much richer way, which I think at least for an introvert, I think the kind of things going on in my mind seem to be much more eloquent and interesting than the kind of things that come out of my mouth when I perform the conversion. So from like a car mechanic or maybe I should say like brain mechanic perspective, my converter is not working very well between the brain thoughts and visualizations and concept space to mouth speaking speaking different English concepts. So I look forward to this podcast being consumed and generated telepathically. So seven possible future application of Neuralink will be the ability to save and replay memories or save and replay mind states. it's a way to do what Daniel Kahneman, for example, talks about as many of us kind of live life through memories of previous events. So kind of the memorable, special things that happen to us are experienced more deeply and more frequently through our memories than directly when we actually experience them for the first time. And the exciting possibility of Neuralink is basically improving the resolution of that memory replay that we generally do anyway. As people should check out Daniel Kahneman's work, he describes it quite eloquently. And it's true. Many of us live in our memories. It's also from a certain perspective, nice to be able to modify, delete, or alter some of those memories. memories. So for example, on a darker side, it could be traumatic events that, you know, from a psychological perspective could be haunting. You can remove or at least alleviate the impact of those memories onto your cognition. Or maybe pull stuff from the subconscious. You can think of it as a Freud's favorite kind of toolkit to play around and explore with our own mind to discover our demons. demons. So as opposed to the David Goggins approach that I've taken recently of doing insane amounts of exercise to discover and have a conversation with my demons. Could do it in a more controlled and safe environment of brain computer interfaces. As a quick side note, there's interesting echoes of the memory replay that you've seen our reinforcement learning systems. So it's kind of interesting to think that instead of just us being able to replay our memories, it could be our own little machine learning systems that can learn something from our previous memories by replaying them over and over to try to give us maybe a strategy of how to avoid those memories in the past. So it's basically converting our prior experiences into data. And once it's converted into data, that could be used for all kinds of applications. So you can think of like a personal machine learning system that can replay your memories and try to figure out, try to be a personal executive assistant to you to advise you what to learn from those experiences. With a lot of these applications that I've already discussed, privacy and security is of paramount importance. I mean, like with actually a lot of our technology, but this is very much at the forefront of what Neuralink is working on currently and always will be. And I think a lot of companies in general in the tech space will sink or swim based on how much they respect privacy and security. I think in the early days of our development with social networks and so on, you could get away easier by being careless with people's data. I think my long-term, perhaps optimistic, but I think it's a realistic view of the future that people will demand much more control over their data, demand much more transparency around privacy and security, transparency and clarity. So of course, that's underlying all the different futures that I'm discussing. And finally, to move a little bit beyond the ability to save and replay memories is to save mental states. And that's essentially a path towards digital immortality. So you can think of being able to save the contents or the critical contents of your mind into digital form and then being able to transfer to other systems, to robots, planets. Or as in, for example, my discussion with Sarah Seeger, who searches for habitable planets outside our solar system, exoplanets, we discussed the idea that one way, perhaps, to reach far away, livable planets that might have extraterrestrial intelligent life on them is by sending digital humans there. So being able to save essential or entire contents of the human mind and to be able to reload it once you arrive into any kind of, whether it's a biological or a robotic system. So that's the kind of stuff that Ray Kurzweil thinks about. It's also a kind of stuff that I think a lot of people are excited about is the ability to store and digitally transfer the contents, the important, the beautiful contents of the human mind. Finally, the eighth future possibility of Neuralink and also one of the original motivations behind the company is the methodology by which the human mind, the human brain, the human society can merge with artificial intelligence systems once they're able to achieve human level and superhuman level intelligence. Since the origins of the field of AI, most, if not all, of the progress that's been made has been in what might be called narrow artificial intelligence. But as a lot of people have discussed, now there's a lot of debates around this, there's a lot of thoughts, but it seems very possible that humans, limited though we are, will one day be able to engineer systems that are far more intelligent than us human beings in some dimension that fundamentally changes the fabric of human society. So we already have AI systems that are much better at a lot of things in humans in a narrow way. But there might be a set of dimensions where an intelligence system is able to generalize better than humans in a set of tasks that can lead to existential risks to human beings, where artificial intelligence systems essentially become a kind of direct or indirect competitor. Whether that's a paperclip paperclip manufacturing AI systems that destroys all humans, just to make its manufacture paperclips a little bit more efficient. Or if it's a much more complex distributed system, kind of like our social networks of today, but much smarter, with some kind of combination of GPT-3 or GPT-20 systems that kind of creep up on us, like the boiling water creeps up on the lobster and overwhelms the resources or the capacities of human civilization in a way that's fundamentally traumatic or destructive or poses an existential risk. Even if that point is far away in time, and that's difficult to predict, I think it's very difficult to rationally say that we will never reach that point. So once you allow that as a possibility, you start to think from an engineering perspective, how can we minimize the existential risk associated with that? And then creating ways to merge with the AI. So we kind of ride the wave of AI and they ride the wave of the functionalities of the human brain is an interesting possibility. I think it's a beautiful vision of a future that's mostly filled with mystery. So we don't know how AGI systems will evolve, but it's an interesting idea that as AI systems become smarter and smarter, it is one way to ensure our survival is to expand the capacity of the human mind to communicate with AI and with the AI to communicate with the human mind. At the basic level, to me, that's super exciting because AI systems can learn from the brain. The brain can learn from AI systems. And I'm, as a person who is a big fan of deep thinking, of sitting for multiple hours and focusing on a single idea and just thinking with a sheet of paper and thinking about an idea, I find myself needing to look up things a lot. And that's actually a huge distraction and it's a huge drain on my mental resources and the kind of distraction. The timing of the thinking is disrupted by having to look up different kinds of information, to look up different kinds of papers, to look up even basic information on Wikipedia. So the ability to kind of close the loop, to increase the bandwidth of thinking of the lookup of the information that's available online is super exciting to me. Now that's not even AGI. That's just like basic recommender systems, basic search engines, basic even like GBT 3++ type of communication back and forth. I think it's really exciting to empower the brain as it's doing the usual kind of deep thinking that it's capable of. And then, of course, but then of course you could take that farther as the AI systems get smarter and smarter and smarter. If we completely open the gates of the communication in two ways, then it increases the likelihood of AGI not leaving us behind. And I think that's a scary and exciting future. And that's probably where we humans do our best work. I hope these thoughts were interesting, useful to some of you. In these difficult times of economic pain, of political division, I personally, and I hope others do too, draw a lot of inspiration from companies, from people, from scientists that are boldly pushing forward the limits of human knowledge, the limits of human capability, and just engineering and building, doing their best to engineer and build a better future for our world. So I hope you find it inspiring as well. And as always, I love you all and hope to see you next time. Thank you.