YouSaid · the spoken record

Elon Musk

lines on the record
134
first
2024-08-02
most recent
2024-08-02
sittings or episodes
1
sources
podcast

Every line below is reproduced as it was said and linked to the record it came from. Nothing here is summarised or generated. Directory · Search · Corrections

  1. “You mean like different color targets, or are you being oh, multiple targets? Yeah, so BPS is log of number of targets times correct minus incorrect divided by time. And so you can think of different clicks as basically doubling the number of active targets. Got it. So you basically higher BPS, the more options there are, the more difficult to task. And there's also Zen mode you've played in before, which is like.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  2. “Obviously, the start of this journey. Still, hopefully, we get back to the places where you're doing multiple clicks and using that to control much more fluidly everything and much more naturally the applications that you're trying to interface with.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  3. “Yeah, it's a great question. So, maybe just to describe first how the actual update worked is basically an update to your implant. So we just did an over-the-air software update to his implants, and we could update your Tesla or your iPhone. And that firmware change enabled us to record sort of averages of populations of neurons nearby individual electrodes. So we have sort of less resolution about which individual neuron is doing what, but we have a broader picture of what's going on nearby an electrode overall. That feedback loop, I mean, basically, as no one described it was immediate when we flipped that switch. I think the first day we did that, you hit three or four BPS right out of the box. And that was a white ball moment for, okay, this is the right path to go down. And from there, there was a lot of feedback around how to make this useful for independent use. So what we care about ultimately is that you can use it independently to do whatever you want. And to get to that point, it required us to re-engineer the UX, as you talked about the dwell cursor, to make it something that you can use independently without us needing to be involved all the time. And yeah, this is.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  4. “The way in which we're measuring the behavior of individual neurons So we're switching from sort of individual spike detection to something called spike band power, which if you watch the previous segments with either me or DJ, you probably have some cons.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  5. “No one's like the first time Nolan played this game. He asked how good are we at this game? I think you told me right then you're going to try to beat me. I'm going to get there someday. Yeah, I think I fully believe you. I think I can.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  6. “There was like one Tuesday we were messing around, and I think I forget what swear word you used, but there's a swear word that came out of your mouth when you figured out you could just do the direct cursor control”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  7. “Up in those two different situations? Yeah, not necessarily. I think all these signals can still be represented in motor cortex, but the difference, I think, has to do with the naturalness of imaging something versus attempting the fatigue of that over time.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  8. “This point, by the way, is driven home. In a very painful way when you try to communicate with someone who cannot speak, because a lot of the time the last thing to go is they have the ability to somehow wiggle a lip or move something that allows them to say yes or no. And in that situation, it's very obvious that what matters is, are you asking them the right question to be able to say yes or no to?”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  9. “It's much more likely we find the meaning of human existence. And so, in the short term, as a heuristic, in the sort of search policy space, we should try to increase the diversity of people asking such questions or generally of consciousness and conscious beings asking such questions So, again, I think I'll take the I don't know card here. Let's say I do think there are meaningful things we can do that improve the likelihood of answering that question.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  10. “There's a TV show I really like called The West Wing. And in the West Wing, there's a character who's the president of the United States who's having a discussion about the Bible with one of their colleagues. And the colleague says something about, you know, the Bible says X, Y, and Z and the president says, yeah, but it also says A, B, C. And the person says, well, do you believe the Bible to be literally true? And the president says, Yes, but I also think that neither of us are smart enough to understand it. I think to like the analogy here for the meaning of life is that largely we don't know the right question to ask. And so I think I'm very aligned with. Or the hitchhi To the Galaxy version of this question, which is basically if we can ask the right questions.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  11. “Yeah, and I mean, great artist copy. And they also, you know, knowing which rules to break is the important part. And fundamentally, it must be about the listener of the piece. Like which rule is the right one to break? It's about the user or the audience member perceiving that as interesting.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  12. “They kind of play with exactly when the surprise happens and the expectations of the user. And that's even true through history as musicians evolve music. They take some known structure that people are familiar with and they just tweak it a little bit. Like they tweak it and add a surprising element. This is especially true in classical music heritage. But that's what I'm wondering. Is it all just entropy?”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  13. “I don't know either. I listen to a lot of classical music and also read a lot of poetry. Yeah, I do wonder if there is some element of the next token surprise factor going on there. Yeah, maybe. Because, I mean, like a lot of the tricks in both poetry and music are like basically you have some repeated structure and then you do like a twist. Like, it's like, okay, clause 123 is one thing. And then clause four is like, okay, now we're on to the next theme.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  14. “It's actually in the process of the user going through that generation that they understand what you mean. Like that's the beautiful part. It's also like when you look at a beautiful painting, it's not the pixels or the painting that are beautiful. It's the thought process that occurs when you see that, the experience of that, that actually is a thing that matters.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  15. “I think it's because the information bottleneck of language is. Pretty steep. And yet you're able to reconstruct on the other person's, in the other person's brain more effectively without being literal. If you can express a sentiment such that in their brain, they can reconstruct the actual true underlying meaning and beauty of the thing that you're trying to get across. The generator function their brain is more powerful than what language can express. The mechanism of poetry is really just to feed or seed that generator function.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  16. “And then there's a whole sort of like range of interesting neuroscience and brain questions, which is when you stick more stuff in the brain in more places, you get to learn much more quickly about what those brain regions represent. And so I'm excited about that fundamental neuroscience learning, which is also important for figuring out how to most efficiently insert electrodes in the future. So yeah, I think all those dimensions, I'm really, really excited about. And that doesn't even get close to touching the sort of software stack that we work on every single day and what we're working on right now.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  17. “You don't have that proprietary feedback loop. So, how can you make that intuitive for a user to control a high dimensional control surface without feeling the thing physically? I think that's going to be a super interesting problem. I'm also quite excited to understand, do these scaling laws continue? Like as you scale channel count, how much further out do you go before that saturation point is truly hit? And it's not obvious today. I think we only know what's in the sort of interpolation space. We only know what's between 0 and 1024, but we don't know what's beyond that.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  18. “I'm quite excited about basically everything we're doing. I think it's going to be awesome. The most prominent one, I would say, is scaling channel count. So right now we have a thousand channel device. The next version we'll have between three and six thousand channels. And I would expect that curve to continue in the future. And it's unclear what set of problems will just disappear completely at that scale. And what set of problems will remain and require further focus. And so I'm excited about the clarity of gradient that that gives us in terms of the user experiences we choose to focus our time and resources on. And also in terms of the Yeah, even things as simple as non-stationarity. Like does that problem just completely go away at that scale? Or do we need to come up with new Creative UX's still even at that point? And also when you get to that time point, when we start expanding out dramatically the set of functions that you can output from one brain, how to deal with all the nuances of both the user experience of not being able to feel the different keys under your fingertips, but still needing to be able to modulate all of them in synchrony to achieve the thing you want.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  19. “I think there's some I'm excited about, like something I'm excited about from the technology side and some I'm excited about for understanding how this technology is going to be best situated for entering the world. So I'll work backwards. On the technology entering the world side of things, I'm really excited to understand how this device works for folks that cannot speak at all, that have no ability to sort of bootstrap themselves into useful control by voice command, for example, and are extremely limited in their current capabilities. I think that will be an incredibly useful signal for us to understand. I mean, really what is an existential threat for all startups, which is product market fit. Does this device have the capacity and potential to transform people's lives in the current state? And if not, what are the gaps? And if there are gaps, how do we solve the most efficiently? So that's what I'm very excited about for the next year or so of clinical trial operations. The technology side,”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  20. “Yes, it is fundamentally an engineering challenge. This is important to emphasize, and it's also important to emphasize that it may not need fundamentally new techniques, which means that people who work on, let's say, unsupervised speech classification using CTC loss, for example, with internal Siri, they could potentially have very applicable skills to this.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  21. “Robustness against this kind of, you can think of it like overfitting, but really it's just that the model has not seen this kind of variability before. So you need to find some way to help the model with that.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  22. “Yeah, so it depends what version of this you're talking about. So, if you're talking about, let's say, the simplest example of just 2D velocity, then I think, yeah, data quality is the main thing. If you're talking about how to build sort of multifunction output that lets you do all the inputs the computer that you and I can do, then it's actually much more sophisticated nuanced modeling challenge because now you need to think about not just when the user is left clicking, but when you're building the left click model, you also need to be thinking about how to make sure it doesn't fire when they're trying to right click or when they're trying to move the mouse. So one example of an interesting bug from like sort of week one of a PCI with Noland was when he moved the mouse, the click signal sort of dropped off a cliff and when he stopped, the click signal went up. So again, there's a contamination between the two inputs. Another good example was at one point he was trying to do sort of a left click and drag. And the minute he started moving, the left click signal dropped off a cliff. So again, because there's some contamination between the two signals, you need to come up with some way to either in the data set or in the model build.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  23. “For every single channel simultaneously, you can actually get better validation metrics, meaning you're fitting the data better and it's generalizing better on offline data if you use this convolutional architecture. You're reducing parameters. It's sort of a standard procedure when you're dealing with time-series data. Now it turns out that when using that model online, the controllability was worse, was far worse, even though the offline metrics were better. And there can be many ways to interpret that, but what that taught me at least was that, hey, it's at least the case right now that if you were to just throw a bunch of compute at this problem and you were trying to sort of hyperparameter optimize or, you know, let some GPT model hard code or come up with or invent many different solutions, if you were just optimizing for loss, it would not be sufficient, which means that there's still some inherent modeling gap here. There's still some artistry left to be uncovered here of how to get your model to scale with more compute. And that may be fundamentally labeling problem, but there may be other components to this as well.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  24. “Not all of them are equally controllable by the end user. And so it might be as simple as saying, oh, you could just add auxiliary loss terms that like help you capture the thing that actually matters. But this is a very complex nuanced process. So how you turn the labels into the model is more of a nuanced process than just like a standard supervised learning problem. One very fascinating anecdote here, we've tried many different sort of neural network architectures that translate brain data to velocity outputs, for example. And one example that stuck in my brain from a couple years ago now is at one point we were using just fully connected networks to decode the brain activity. We tried AB test where we were measuring the relative performance in online control sessions of sort of 1D convolution over the input signal. So if you imagine per channel, you have a sliding window that's producing some convolved feature for each of those input sequences.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  25. “Yeah, good question. There's a couple different ways to answer this. So maybe I'll zoom out briefly first, and then I'll go down one of the rabbit holes. So the zoomed out view is that building the decoder is really the process of building the data set, plus compiling it into the weights. And each of those steps is important. The direction, I think, of further improvement is primarily going to be in the data set side of how do you construct the optimal labels for the model. But there's an entirely separate challenge of then how do you compile the best model. And so I'll go briefly down the second one, down the second rabbit hole. One of the main challenges with designing the optimal model for BCI is that offline metrics don't necessarily correspond to online metrics. It's fundamentally a control problem. The user is trying to control something on the screen. And the exact sort of user experience of how you output the intention impacts their ability to control. So for example, if you just look at validation loss as predicted by your model, there can be multiple ways to achieve the same validation.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  26. “This is like a profoundly game changing thing for folks in that situation. And this is even before we start talking about folks that may not be able to communicate at all or ask for help when they want to. This can be potentially the only link that they have to the outside world. And yeah, that one doesn't, I think, need explanation of why that's so impactful.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  27. “Maybe one comment on that too. For folks that aren't familiar with assistive technology, I think there's a common belief that why can't you just use an eye tracker or something like this for helping somebody move a mouse on the screen? And it's really a fair question and one that I actually was not confident before Noland that this was going to be a profoundly transformative technology for people like him. And I'm very confident now that it will be. But the reasons are subtle. It really has to do with ergonomically how it fits into their life. Even if you can just offer the same level of control as what they would have with an eye tracker or with a mouse stick, but you don't need to have that thing in your face. You don't need to be positioned a certain way. You don't need your caretaker to be around to set it up for you. You can activate it when you want, how you want, wherever you want. That level of independence is so game changing for people. It means that they can text a friend at night privately without their mom needing to be in the loop. It means that they can like open up and browse the internet at 2 a.m. when nobody's around to set their iPad up for them.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  28. “I actually don't know the full reason why, but I can imagine several explanations. One such explanation could be that the context effect difference between some open loop task and some closed-loop task is much more significant with no one than it is with a monkey. Maybe in this open loop task, he's watching the Lex Freeman podcast while he's doing the task or he's whistling and listening to music and talking with his friend and asking his mom, what's for dinner while he's doing this task. And so the exact sort of difference in context between those two states may be much larger and thus lead to a bigger generalization gap between the features that you're normalizing at sort of open loop time and what you're trying to use at closed loop time.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  29. “Yeah, this is a great question. So with monkeys, we have found various ways to do this. One example I'd do this is you ask them to do some behavioral tasks like play the game with a joystick. You measure what's going on in the brain. You compute some mean of what's going on across all the input features and you subtract that in the input when you're doing your BCI session. Works super well. Whatever reason that doesn't work super well with Nolan.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  30. “And so, if what you're trying to measure is how much rice is in the pot, you're going to get a different measurement different days because you're measuring with different pots. So that based on rate shifting is really the thing that, at least from a first order description of the problem is what's causing this downstream bias. There can be other effects, nonlinear effects on top of that, but at least at a very first order description of the problem, that's what we observe day to day, is that the baseline firing rate of any particular neuron or observed on a particular channel is changing”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  31. “And what it looks like when it's modulated. And what we have observed, and what has also been observed in academic work, is that that baseline rate, sort of the, if you tar the scale, if you imagine that analogy for measuring flour or something when you're baking, that baseline state of how much the pot weighs is actually different day to day.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  32. “Yeah, so maybe let's talk briefly what the actual underlying signal looks like. So, again, I spoke very briefly at the beginning about how when you imagine moving to the right or imagine moving to the left, neurons might fire more or less. And the frequency content of that signal, at least in the motor cortex, is very correlated with the output intention, the behavioral task of the user is doing. You can imagine, actually, this is not obvious that rate coding, which is the name of that phenomenon, is like the only way the brain can represent information. You can imagine many different ways in which the brain could encode intention. And there's actually evidence like in bats, for example, that there's temporal codes, so timing codes of like exactly when particular neurons fire is the mechanism of information representation. But at least in the motor cortex, there's substantial evidence that it's rate coding, or at least one first order effect is that it's rate coding. So then if the brain is representing information by changing the frequency of a neuron firing, what really matters is sort of the delta between sort of the baseline state of the neuron.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  33. “And that last point is very related to channel count discussion. So as you scale out number of channels, the relative importance of any particular feature of your model input to the output control of the user diminishes, which means that if the sort of neural non-stationarity effect is per channel, or if the noise is independent such that more channels means on average less output effect, then your reliability of your system will improve. So one sort of core thesis, at least I have is that scaling channel count should improve the reliability system without any work on the decoder itself.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  34. “One other nuance there that is worth mentioning. So, again, our goal is really to enable a user worth process to control the computer as fast as I can. So that's BPS with all the same functionality I have, which is what we just talked about, but then also as reliably as I can.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  35. “Each channel corresponds with a specific represented intention in the brain. So for example, if you have a channel 254, it might correspond with moving to the right, channel 256 might mean move to the left. If you want to expand the number of functions you want to control, you really want to have a broader set of channels that covers a broader set of imagined movements. You can think of it like kind of like Mr. Potato Man, actually. Like if you had a bunch of different imagined movements you could do, how would you map those imagined movements to input to a computer? You could imagine handwriting to output characters on the screen. You could imagine just typing with your fingers and have that output text on the screen. You can imagine different finger modulations for different clicks. You can imagine wiggling your big nose for opening some menu or wiggling your big toe to have like command tab occur or something like this. So it's really the amount of different actions you can take in the world depends on how many channels you have in the information content that they carry.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  36. “Yes, short answer is yes. It's a bit nuanced how that curve or how that manifests in the numbers. So what you'll see is that if you sort of plot a curve of number of channels that you're using for decode versus either the offline metric of how good you are at decoding or the online metric of sort of in practice how good is the user using this device, you see roughly a log curve. So as you move further out in number of channels, you get a corresponding sort of logarithmic improvement in control quality and offline validation metrics. Important nuance here is that.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  37. “Either completely changing the thing you're decoding or just extending the number of things that you're decoding. So this is serving the direction of functionality. You can imagine giving more clicks. For example, left to click or right click, a middle click, different actions like click and drag, for example. And that can improve the effective bitrate of your communication processes. If you're trying to allow the user to express themselves through any given communication channel, you can measure that with bits per second. But what actually matters at the end of the day is how effective are they at navigating their computer? And so from the perspective of the downstream tasks that you care about, functionality and extending functionality is something we're very interested in because not only can it improve the sort of number of BPS, but it can also improve the downstream sort of independence that the user has and the skill and efficiency with which they can operate their computer.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  38. “The next bottleneck after that was actually just sort of software stability and reliability. If you have widely varying sort of inference latency, In your system, or your app just lags out every once in a while, it decreases your ability to maintain and get in a state of flow. And it basically just disrupts your control experience. And so there's a variety of different software bugs and improvements we've made that basically increase the performance of the system, made it much more reliable, much more stable, and led to a state where we could reliably collect data to build better models with. So that was a bottleneck for a while is just sort of like the software stack itself. If I were to guess right now, there's sort of two major directions you could think about for improving BPS further. The first major direction is labeling. So labeling is, again, this fundamental challenge of given a window of time where the user is expressing some behavioral intent, what are they really trying to do at the granularity of every millisecond? And that, again, is a task design problem. It's a UX problem. It's a machine learning problem. It's a software problem. Sort of touches all those different domains. The second thing you can think about to improve BPS further is”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  39. “We've seen historically is that different parts of the stack become bottlenecks at different time points. So when I first joined Erlink like three years ago or so, one of the major problems was just the latency of the Bluetooth connection. It was just like the radio on the device wasn't super good. It was an earlier revision of the implant. And it just like no matter how good your decoder was, if your thing is updating every 30 milliseconds or 50 milliseconds, it's just going to be choppy. And no matter how good you are, that's going to be frustrating and lead to challenges. So at that point, it was very clear that the main challenge is just get the data off the device in a very reliable way such that you can enable the next challenge to be tackled. And then at some point it was actually the modeling challenge of how do you just build a good mapping, like the supervised learning problem of you have a bunch of data and you have a label you're trying to predict, just what is the right neural decoder architecture and hyperparameters to optimize that. That was a problem for a bit. And once you solve that, it became a different part of the thing.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  40. “Yeah, I think the nature of this work is the first answer that's important to say is I don't know. This is the edge of the research. So, again, nobody's gotten to that number before. So what's next is going to be heuristic, a guess from my part.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  41. “I don't know what the limits. I mean, the limits you can calculate just in terms of screen refresh rate and cursor immediately jumping to the next target. But there's, I mean, I'm sure there's limits before that with just sort of reaction time and visual perception and things like this. I'd guess it's in that below 40, but above 20, somewhere in there, it's probably the right thing never to be thinking about. It also matters how difficult the task is. You can imagine some people might be able to do 10,000 targets on the screen and maybe they can do better that way. So there's some task optimizations you could do to try to boost your performance as well.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  42. “Yeah, please track my record. I mean, the reason I did this literally was just because I wanted the bar to be high for the team. Like I wanted the. The number that we aim for should not be like the median performance. It should be like it should be able to beat all of us at least. That should be the minimum bar.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  43. “And then it has to be really late at night. This is, again, a night owl thing, I think we share, but it has to be like midnight, 2 a.m. kind of time window. And I have a very specific physical position I'll sit in, which is I used to be, I was homeschooled growing up. And so I did most of my work on the floor, just like in my bedroom or whatever. And so I have a very specific situation. On the floor that I sit and play. And then you have to make sure there's not a lot of weight on your elbow when you're playing so you can move quickly. And then I turn the gain of the cursor so the speed of the cursor way, way up. So it's like small motions that actually move the cursor.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  44. “Yeah, it's true. So, what I do is I actually don't eat for a little bit beforehand, and then I'll actually eat like a ton of peanut butter right before I get like this is a real thing. This is a real thing, yeah.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  45. “I have like a whole ritual I go through when I play WebGrid. So it's essentially like a Daya Plan associated with this whole thing.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  46. “Which is about if you imagine that 35 by 35 grid, you're hitting about 100 trials per minute, so 100 correct selections in that one minute window. So you're averaging about between 500, 600 milliseconds per selection.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  47. “It's so funny. This is, I've heard this reaction so many times before, sort of first user was implanted, we had an internal perception that the first user would not find this fun. And so we thought really quite a bit actually about should we build other games that are more interesting for the user so we can get this kind of data and help facilitate research that's for long duration and stuff like this. It turns out that people love this game. I always loved it, but I didn't know that that was a shared perception”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  48. “We make there, which will hopefully translate then to even people who can't speak but don't feel comfortable doing so because they're in a public setting like their doctor's office.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  49. “Yeah, great question. So the underlying signal we're trying to decode is going to look very different in P2 than in P1. For example, channel number 345 is going to mean something different in user one than it will in user two, just because that electrode that corresponds with channel 345 is going to be next to a different neuron in user one versus user two. But the approach is the methods, the user experience of how do you get the right sort of behavioral pattern from the user to associate with that neural signal, we hope it will translate over multiple generations of users. And beyond that, it's very, very possible, in fact, quite likely that we've overfit to sort of Nolan's user experience desires and preferences. And so what I hope to see is that when we get second, third, fourth participant that we find sort of what the right wide minimas are that cover all the cases that make it more intuitive for everyone. And hopefully there's a cross-pollination of things where, oh, we didn't think about that with this user because they can speak. But with this user who just can fundamentally not speak at all, this user experience is not optimal. And that will actually, those improvements.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source

  50. “It'll make the EQ bigger because it knows Lux is the person I'm going to go see. And so that kind of predictiveness can make the experience much more smooth, even without improvements to the underlying decoder or feature detection part of the stack. So we do that with a feature called magnetic targets. We actually index the screen and we understand, okay, these are the places that are very small targets that might be difficult to hit. Here's the kind of cursor dynamics around that location that might be indicative of the user trying to select it. Let's make it easier. Let's blow up the size of it in a way that makes it easier for the user to sort of snap onto that target. So all these little details, they matter a lot in helping the user be independent in their day-to-day living.”

    2024-08-02 · Lex Fridman Podcast · #438 – Elon Musk: Neuralink and the Future of Humanity · IDENTIFIED FROM THE TRANSCRIPT · source