Pod Insights The Cognitive Revolution Merging Man and Machine? The Neurotech Frontier with Dean W. Ball About this episode In this podcast, we dive into the fascinating world of brain-computer interfaces with Dean W. Ball. We learn about the latest technologies, from non-invasive EEG and ultrasound stimulation to invasive Neuralink implants. We also discover the state-of-art tools, technical challenges, and the big questions around thought dimensionality, ethical considerations, and the societal impact of brain reading and writing.
SPONSORS:
The Brave search API can be used to assemble a data set to train your AI models and help with retrieval augmentation at the time of inference. All while remaining affordable with developer first pricing, integrating the Brave search API into your workflow translates to more ethical data sourcing and more human representative data sets. Try the Brave search API for free for up to 2000 queries per month at https://bit.ly/BraveTCR
Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off https://www.omneky.com/
Plumb is a no-code AI app builder designed for product teams who care about quality and speed. What is taking you weeks to hand-code today can be done confidently in hours. Check out https://bit.ly/PlumbTCR for early access.
Head to Squad to access global engineering without the headache and at a fraction of the cost: head to https://choosesquad.com/ and mention “Turpentine” to skip the waitlist.
CHAPTERS:
(00:00:00) Introduction
(00:04:00) Background
(00:06:50) Journey to Neurotech Interest
(00:07:56) Concept of Human-Machine Merge
(00:12:01) Impact of Language Models
(00:15:53) Sponsors: Brave / Omneky
(00:17:20) Adoption of AI Tools like ChatGPT
(00:22:33) Neural Technologies as Sensor Fusion
(00:26:03) Brain Signals and EEG Devices
(00:29:09) Electrodes and Electrical Field Measurement
(00:37:04) Sponsors: Plumb / Squad
(00:39:01) Neurofeedback and Brain Training
(00:48:44) The Rise of Neural Networks
(00:57:29) Shared Latent Space for Brain Signals
(01:09:18) Trajectory of Brain-Computer Interface
(01:14:44) Near-Term Possibilities
(01:26:25) Limitations of Brain-Computer Interfaces
(01:38:31) Stimulating Brain Activity
(01:46:08) Policy Considerations
(01:52:02) Solved World and Artificial Constraints
(02:04:27) Acclimating to Merge Technology
(02:08:33) Advice for Understanding Neural Interfaces
Transcript 21,977 words · 3 speakers
Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Speaker 00 Computers already are neural interfaces. You don't control them directly with your brain, but they obviously interface with your brain and you are using your brain via touch control to manipulate them. Throughout the history of technology, one of the things we see is that information technologies don't just change the way information propagates, they change the way we think. The impulse that people have a lot of the time is to think, oh no, if there's an incentive to use this thing, then everybody will have to use it. What about the people that don't want to use it? I go there too. I have sympathy for those people. I think about the Amish. And I do think that there's probably going to be forms of digital Amish in the future that we
0:44 Speaker 00 need to be thinking about. At the same time, the people who want to enhance their cognition also should have the liberty to do that. And we should want there to be more cognition in the world, especially human cognition.
0:58 Nathan Labenz? Hello and welcome to The Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas and together, we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my co-host, Erik Torenberg. Hello and welcome back to The Cognitive Revolution. Today, I'm very excited to share my conversation with Dean W. Ball, research fellow at the Mercatus Center and author of The Hyperdimensional Substack. Dean is a kindred spirit who shares my obsession with frontier technologies and their near-term possibilities, and he's done a super deep dive into the current state of brain-computer
1:41 Nathan Labenz? interfaces. In this wide-ranging discussion, Dean explains the major approaches being pursued today, from the non-invasive techniques like EEG and ultrasound stimulation, to the invasive implants like those being developed by Neuralink. He gives us a sense for the current state of the art across important technical dimensions, such as the number of electrodes used in a given device, the spatial and temporal resolution that's achievable with different technologies, and the major technical challenges that still need to be overcome. Along the way, we touch on big-picture questions about the dimensionality of thought, the challenges of interpreting signal that's transmitted through the human skull, and the ethical and societal implications of being able to increasingly read from and even write to the human brain.
2:26 Nathan Labenz? Could we one day induce complex thoughts or download knowledge and skills matrix style? Will brain monitoring become a routine part of legal proceedings or job interviews? How might the widespread adoption of cognitive enhancement change human culture and even our biological evolution? While he maintains an appropriate degree of epistemic humility throughout, Dean's well-informed intuitions paint a picture of consumer devices that will soon provide powerful brain reading and crude but potentially useful brain writing capabilities. More advanced applications like high bandwidth two-way brain-computer communication will likely require invasive implants, but the pieces there too are coming together faster than many realize.
3:10 Nathan Labenz? Whether you're an AI expert looking to understand a related technology trend, a policymaker grappling with the implications, or just a curious observer of the human journey, this episode will give you a solid grounding in the emerging neurotech landscape. Special thanks to Dean for doing all the work to make this episode possible, and definitely check out his newsletter at hyperdimensional.substack.com. As always, we welcome your comments, questions, and suggestions for future episodes. You can contact us via our website, cognitiverevolution.ai, and if you do, you might also enjoy the chance to chat with my new AI clone, which has been trained on the full history of the cognitive revolution and is powered by Delphi AI.
3:52 Nathan Labenz? For now, I hope you enjoy my conversation about brain-computer interfaces and the potential for a man-machine merge with Dean W. Ball. Dean W. Ball, research fellow at the Mercatus Center and author of the Hyperdimensional Substack, welcome to the Cognitive Revolution.
4:07 Speaker 00 Thank you, Nathan. Great to be here.
4:09 Nathan Labenz? I'm really excited for this conversation. You and I are kindred spirits in that we both have had our day jobs and then we've had our obsessions with different areas of technology. We connected online after a couple pieces of writing that you put out were excellent. And as I know more about you, I was fascinated to see just how much time and energy you have put into learning all about the brain-computer interface space. And so today, we are going to do a survey on all things neurotech, and I'm really excited to learn from you in this domain and to share that with our audience.
4:46 Speaker 00 Awesome. Yeah, me too.
4:47 Nathan Labenz? So you want to start by just maybe introducing yourself? You've got a varied set of experiences and background interests. And then we can dive into your kind of high level thesis and dig into the technologies that will take us there one by one.
5:00 Speaker 00 Yeah, as the title of my Substack suggests, I have a varied background, a lot of different friends going into it. I've spent most of my career in the Think Tank world. So about a decade working in state and local economic policy for the most part, housing, infrastructure, things like that. But I've always maintained a very strong interest in technology. I started open source forum software. If there's anyone who remembers PHP BB from way back in the day, I was doing commits to that and writing technical documentation for that at a young age. And the coding bug hit me at that point. And I never went into it professionally, but it's always stuck with me. And just a general interest in technology and that morphed as the deep learning revolution
5:45 Speaker 00 unfolded in the last decade into following the ML literature pretty closely. And once ChatGPT came on the scene, I realized that this was a perfect moment for me because I have a background in policy and there's a lot of interest in the societal and policy implications of this technology. And at the same time, I've got a pretty good ground truth, technical sense of what's actually going on, which is often lacking in the policy space. So I try to bring that to everything that I do.
6:17 Nathan Labenz? I love that. I often say that we really need to figure out what is before what ought to be done about it. And there's all too much jumping to what ought to be done about it that is not based in a good ground truth, understanding of what technology actually exists today, what can be done with it, where that's heading in the near term and so on. So I appreciate that perspective and I'm excited to get into the details of.
6:44 Nathan Labenz? What you think is going to happen in the big picture of the possibility of a merge between man and machine?
6:50 Speaker 00 Yeah, so I'll explain a little more about how I came to this particular issue. I am a longtime fan of J. Z. R. Licklider, the DARPA senior researcher who wrote an article back in 1960 called "Man-Machine Symbiosis." It's not quite the physical merge that we're going to be talking that anticipates the development of something like AGI, as a lot of people did in the 60s, being like way easier. The thesis of that paper is AGI is going to be here in the next five to 10 years. And who knows what happens after that? Probably it'll become super intelligent and replace us, but in the meantime, we'll have a lot of fun. That's basically the thesis of this senior government official at that time.
7:32 Speaker 00 It shows you how much things have really changed in terms of just approaches to technology.
7:36 Nathan Labenz? Although I feel like I've heard similar quotes from some notable technology leaders more recently as well.
7:42 Speaker 00 From the technologists, for sure, but coming directly from the DOD is interesting. And it shows you how greenfield everything felt with computing at that time. And what's exciting about right now is I do feel like we're at a similar moment where a lot of things are possible. But I've always thought about this issue of symbiosis and thought about it in a fairly broad way. But in the last few years, is the concept of artificial superintelligence has become maybe a little closer than any of us thought, possibly eminent, if you believe some people on Twitter. The concept of the merge just keeps coming up. Prominent technologists, as you say, Sam Altman has a famous blog post called "The Merge",
8:26 Speaker 00 where he essentially describes it as an inevitability. Elon Musk is not just called it an inevitability, but is in fact working. On technical solutions. So he is probably taking this more seriously than anyone else. And just in general, in my conversations with people building in the AI space at advanced levels, there is a very common sense that this is something that will happen. And if artificial superintelligence is even remotely eminent, then the merge must be not too far off either, if that's going to happen. And so my fundamental question is, what does the technology really look like? And that's what I sought to answer in some articles that I've written.
9:10 Speaker 00 And to do that, I first looked at the actual scientific technologies that exist right now. And it's like, where is the literature? But also thought about it in a broader sense, because in a lot of ways, computers already are neural interfaces. You don't control them directly with your brain, but they obviously interface with your brain and you are using your brain via touch control to manipulate them. And throughout the history of technology, one of the things we see is that information technologies don't just change the way information propagates, they change the way we think. So that was my starting point of almost seeing the merge as something that started a long
9:52 Speaker 00 time ago, a path that we are still on and that will get more vivid and more visceral over time, but not a fundamentally new.
10:02 Nathan Labenz? Yeah, that's interesting. Certainly, I feel like I've been through a couple of waves already of technology changing how I think. It feels so quaint now, but there was the idea of we don't have to remember anything anymore because we can just go search for it on the internet. And certainly, that is true, or even GPS, right? Just we don't have to make spatial maps of our general environs because so many of us just default to using the GPS. For me right now, I wonder if you have experienced something similar with the current wave of technology. I do find myself getting into some pretty deep habits, at least with how I relate to language models. Coding tasks are a pretty good example of this.
10:44 Nathan Labenz? When I reflect on how I used to code versus how I approach coding tasks now.
10:52 Nathan Labenz? Probably making a mistake in the past because I've never been an outstanding coder, but I feel like I often used to go into the editor first and maybe start writing some comments and trying to break things down or starting to scaffold out the classes. But very often I was jumping too quickly into code. And one of the things I've learned from using the language models is I'm really rewarded for taking the time first to think, what do I actually want here? Can I define that in conceptual terms first before I start to diffuse, as it were, into the finer grained structure of the code? And in general, I think that's starting to happen to me. We recently did an episode with Short Wave, the email client as well, that has a really good email AI assistant.
11:32 Nathan Labenz? I see an evolution there too, where I used to remember keywords to navigate my vastly overflowing and unmaintained inbox. And now I don't have to be so keyword oriented anymore. I can ask like questions and that's definitely very helpful for me. But also I'm starting to feel already a little bit like my keyword search ability is beginning to atrophy. So certainly I recognize and experientially feel the idea that the ways of technology have changed how I thought. I don't know if you have any similar stories and this probably pales compared to what might be ahead of us still.
12:08 Speaker 00 Yeah, no, absolutely. And it is important to think about that kind of thing. For me, in a similar vein, LLMs just fundamentally changed the way I ask questions and made me much better at asking questions. Oftentimes, I remember it's about a year ago since my first experience is the GPT. And 3.5 was not quite as vivid in this way. But when GPT-4 came out, I would ask it a question and I would find, this isn't quite what I wanted. This isn't quite what I was looking for, but I know this model must be capable of giving me what I look for, so the problem must be with me. And so I kind of refining my question and getting down to what is it that you're really asking?
12:51 Speaker 00 And so often in fields of intellectual inquiry. That is actually the more important thing, but finding the right question is 90% of the intellectual journey. And so something that accelerates that, that makes us better at that, accelerates all kind of intellectual endeavor. And I am excited to see things like Claude 3 Opus, which it seems that the qualitative difference in that model versus GPT-4 is really at the frontier of all kinds of different fields, that the people that are doing middle of the bell curve activities, and not even coding necessarily, but just like various kinds of questions, um, science, history, whatever it may be.
13:34 Speaker 00 That those questions, the middle of the bell curve stuff, the difference between GPT-4 and Claude 3 Opus is not necessarily all that much, but that when you're asking things that are more at the frontier, Claude 3 really shines vis a vis GPT-4. I can definitely tell you, having used LLMs a lot to help me accelerate this process of, of researching these articles on neuroscience, that's definitely been the case with Claude. It's just so much better at answering those questions than GPT-4, where GPT-4 would get answers that I had enough context to know, like, eh, this is kind of BS here, this is just filling in the blank, whereas Claude 3 is more capable.
14:16 Speaker 00 I think as the LLMs get more capable, that will continue to be very relevant. And then on another level, obviously, anyone paying attention to this space is asking themselves philosophical questions about what is the nature of cognition? What is the nature of intelligence? What distinguishes my intelligence from the kind of intelligence that I'm seeing here exhibited on this screen? And it's actually a journey that we should all go through anyway, regardless of whether LLMs exist. Because asking, what special value do you bring to the table? And what unique perspective do you have?
14:57 Speaker 00 Is a question that everybody should be seriously asking themselves all the time, and not enough people do. And I think LLMs are a kind of forcing function for a lot of people, for labor market concerns, just as much as anything else, to just really probe at that. So in a way, for me, my experience is, as I look back on the last year using advanced LLMs, it's made me more reflective and, and just more inquisitive. And that is itself, that has all kinds of. Compounding benefits. And obviously, the other thing is that once so much about the history of technology is not how it impacts one person, or one building, or one place,
15:37 Speaker 00 it is instead about the nonlinear compounding effects when everybody has access to that same capability. And I think we're in the very early days of that. And that's part of why I find this to be such an enormously exciting time to be alive.
15:53 Speaker 02 Hey, we'll continue our interview in a moment after a word from our sponsors.
15:58 Nathan Labenz? Yeah, it's crazy. You hit on something there that I often think about as well, which is just that the future dynamics of the world at large as AI starts to infuse into everything are dramatically under theorized. When I go around talking to everybody in AI today, largely their implicit world model seems to be the world is the world. I'm applying AI here. Therefore, we're going to have the world plus my AI intervention. And that'll be sweet because we'll get this, you know, productivity gain, whatever. And that is not wrong, but it sort of misses the fact that everybody else is doing this at the same time. And therefore, there's going to be just a lot of, in my view, unpredictability to how all these new arrangements
16:43 Nathan Labenz? interact with each other at a higher level.
16:45 Speaker 00 I think about that all the time. And I think that we're in for just an order of magnitude complexification of the world beyond what we've already seen. If you think about the world has just become denser in the last 30 years. You think about what was New York City like 400 years ago, the island of Manhattan versus what it is today. It's the same geographic landmass, the day landfill parts of it, but basically the same thing. And the difference is that it's just dramatically denser and more complex. And I think that we are in for perhaps not geographical densification, but another substantial round of conceptual and cognitive densification. And one of the things I worry about, and another thing that attracted me to
17:30 Speaker 00 this issue area, is the question of how we're all going to keep up. With that, I worry now that people like you and me are already in a bubble. We're extrapolating multiple points out on this LLM, on the dynamics of AI agents being all over the world. Other people haven't even contemplated the idea of AI agents and have probably asked ChatGPT 3.5 one basic question 18 months ago and haven't thought about it really that much since. And so I worry about people being shocked and left behind. And I think that maybe neural technology is our way, at least I hope, for us all to just keep better pace with the complexity that I suspect is coming.
18:14 Nathan Labenz? Yeah, there was a recent study just this week that came out that said that 30% of young people are using ChatGPT for work or whatever. And it is a good reminder that that's a lot because it's only been a year and change since the first version and a year and just a few days since GPT-4, but it's also not that many yet. And the tools are still certainly relatively basic compared to where it seems quite clear that we're headed. In terms of going along for the ride, which is an Elon Musk framing on this. And this may be a little shocking for people that aren't aware of how much progress has already been made. Just one way I think about it is like, the way we use computers has, first of all, it's a bidirectional relationship, right? We put information into them, they give us information back.
19:00 Nathan Labenz? The way in which we have entered information into the computers historically has been this sort of finite action space of like you can type on the keyboard and you can click the mouse. And you can do a lot with that. But at any given time, you have a finite and often quite prescribed set of actions that you can take and outside of that, just nothing happens. And that's also interesting. That's one of the reasons that they cite for why the web agents historically didn't take off. They tried to do reinforcement learning on the web. And the reward signal was just too sparse. Agents couldn't get anywhere and they couldn't chain enough success together to get any reward. And so the whole project went nowhere for that reason. Of course, the signals that we get back have been getting richer and richer over
19:44 Nathan Labenz? time, like higher resolution, more realistic graphics, et cetera, et cetera. Now with this language moment, we have the ability to speak in natural language to the computers. And this really opens up a sort of richness of communication for what they can understand in terms of our mental states and desires. Ilya from OpenAI, once and possibly still from OpenAI, put that when he said, the thing that's most incredible about this is I speak to the computer and I feel I am understood. And that's a really good baseline reality check of what makes this different than before. But now recalling your Substack title, Hyperdimensional, we're moving around the verge of potentially moving
20:25 Nathan Labenz? from this language mode of communication, which has just opened up for us to talk to computers in our own natural language. That is obviously a richer space, much more wide open space than you can click on these particular buttons within an interface. But it is still, of course, the compressed form, token by token, that we're reducing our internal thoughts to this output stream that we can encode in language. And one big way to think about the set of technologies that we're going to talk about today is that it goes up yet another level in terms of the richness of the dimensionality of the space in which we can communicate with computers.
21:09 Speaker 00 Yeah, absolutely. So let's get into it. I think at a fundamental level, all of this, the neural technologies, it's really a story of very advanced sensor fusion and signal processing, which is the story of so much technological innovation over the last 100 years. We don't think about it in that way. But from Bell Labs on, it's been a story of signal processing. In this case, signals are the biomarkers of thought. Thought is fundamentally mediated by electricity and magnetism. And so you can measure electrical and magnetic fields generated by brain activity.
21:45 Speaker 00 That's the the primary way uh that this is done, or at least it's the most straightforward way. Let's put it that way. There are there are other interesting ways that we'll get into. The most common that you see for neuro monitoring, monitoring brain, say, not so much modifying them, but understanding what a person is thinking is called EEG electroencephalography. And it has a closely related cousin called MEG magnetoencephalography, which is measuring the magnetic fields. So kind of same idea. The difference is that EEG can be done with tiny electrodes placed on the head. Magnetoencephalography requires pretty substantial equipment, but I don't think is going to make it out of the lab anytime soon. The thing that a lot of people don't know, even in the tech space,
22:30 Speaker 00 is that it's not just Neuralink in this field shipping products. For more than a decade, there have been neural technologies using EEG, especially because it's relatively cheap to do all kinds of practical things. And it's obviously also been used in a lot of lab settings. The difference between really a lab EEG and a consumer EEG is just the number of channels that you get. So consumer EEG might be stereo, just two electrodes up to maybe at the maximum thirty two. That would be a lot. Thirty two would be like a whole helmet that you're wearing. Something like eight, sixteen is more common in the consumer setting. Lab can go all the way up to two hundred and fifty six.
23:15 Speaker 00 And this is just taking an enormous amount of data. That's something that is also not appreciated is just that the brain generates a torrent of data. A lot of EEG, you can change the sampling rate, but the sample rate is usually around a thousand hertz, so a thousand samples per second. And eight channel EEG in a consumer headset is going to be generating eight thousand data points from your brain per second. So you quickly get into millions if you're using it for for any substantial period of time. Fundamentally, the problem with all of these technologies really is the human skull. The skull was designed to protect your brain and it does a good job of that. And it is not particularly conductive. So it tends to attenuate signal coming out of the brain and going in. There's other things too.
24:00 Speaker 00 There's cerebrospinal fluid. There's other mediating layers between the actual surface of the cortex and the brain and the scalp. But the skull is the most important one. And so modeling that, modeling the skull, which also varies between people. In thickness, density is very difficult. It also varies not just between people, but cross your skull. So your skull is not homogeneous in terms of its density or thickness. And so it has to be modeled in real time for an ideal reading. And that is where I think we're at the frontier of these technologies. But I'll talk a little about what has been done so far, particularly with EEG.
24:41 Speaker 00 EEG devices can do things like screen for cognitive, for signs of cognitive decline. They can read, for example, Parkinson's. There are biomarkers of mental illness that can be picked up in EEG devices. Generally, you don't see that on the consumer side of things. And the reason for that is that relates to FDA regulation. We can get into that later. But for consumer devices, you'll often see devices to help you meditate better. The brainwaves, they fall into a few broad families. And it is actually kind of intuitive. Your brain, the electrical fields that your brain is generating oscillate at higher frequency. When you are thinking more intensely, that is an intuitive relationship.
25:25 Speaker 00 So when you're in deep sleep, the frequency is very low. And when you are focusing or in a panic, the frequency is very high. And so those kind of crude measures of is this person sleeping? Is this person in a deep meditative state? Is this person attentive and focused? Or are they scared? These are the kinds of things that can be read right now in consumer devices, still fairly crudely, but it can be done.
25:56 Nathan Labenz? So a couple of just very practical rookie level questions about exactly what we're measuring and exactly what we're doing with it here. The electrodes that you put on your head to measure these electrical signals. You said the the frequency with which they can take a measurement is about a thousandth of a second, a thousand hertz is a thousand cycles a second. So it's one one thousandth of a second measurement. The rate at which a neuron can fire is also roughly a thousand times a second. Is that right?
26:28 Speaker 00 It can be quicker than that. Neural activations can be quicker. So I should actually just step back. First of all, we're talking here about non invasive technology. There's a whole school of invasive stuff which we can get into later. But on the non invasive side, yeah, that's right. And EEG is also useful because of that temporal resolution, as it's called in the literature, is excellent for EEG. As a comparison, the kind of gold standard for spatially imaging the brain is FMRI. And that, at its best, can take one image every two seconds. So in the space that FMRI, if you're using FMRI, you're in a lab because FMRI requires superconducting materials.
27:11 Speaker 00 You're not wearing that on your head anytime soon. If you're wearing, for example, let's just say a 64 channel EEG, then in the space that an MRI can take one image of your brain, the EEG has generated more than a hundred samples from the brain. So the temporal resolution is great, not quite at the neural activation level from a temporal perspective and far from it from a spatial perspective. So an electrode at its most precise can see millions of neurons. Still down to a spatial resolution of centimeters or millimeters in some cases.
27:47 Nathan Labenz? I want to understand a little bit better what the electrodes are actually measuring, like what their output is, and then how that gets translated into this sort of spatial model of what's going on in the brain. Is it accurate to say that an electrode creates one number every thousandth of a second that would represent sort of the strength of the electrical field?
28:09 Speaker 00 A point in time. Yeah, exactly.
28:11 Nathan Labenz? And it seems like there's Fourier transform kind of math that must be happening here, where this is similar to like how cellphone towers work from what like limited understanding I have of this branch of physics. But basically, we have all of these neurons firing off with these sort of spikes, right? And there's a very short duration, hot, relatively high voltage signal that is getting sent very locally. Stop me if I'm wrong about this at any point because I'm really not very expert in this at all, because that's now happening in all these billions of locations concurrently throughout this whole region of the brain. That then sends out this mess of signal to the outside world, much like
28:51 Nathan Labenz? a cellphone tower is sending out a mess of signal that is communicating with all of the phones in the area at once. And then in a similar way, all these electrodes are receiving this sort of messy signal where it's like, okay, I'm here, and this is what I feel right now. And that is the aggregate of all of the signals that have been created from these individual neuron firings. And then there's like a real computational challenge after that to say, okay, we've got 64 different signals because we've got 64 electrodes on the head. What does that translate into? So can you tell us a little bit more about how that is happening? Like, how does that sort of 64 numbers get translated into a
29:35 Nathan Labenz? spatial understanding of what's going on inside?
29:39 Speaker 00 Yeah, this is a great question. First of all, yeah, your intuition here and what you sketched out is basically correct. And yeah, a lot of the basic analysis that's happening here is your standard Fourier transform.
29:50 Nathan Labenz? Just to define that, that basically takes a composite signal and breaks it down into the intensity of the signal at different frequencies over time. So you have some wobbly signal that you're measuring, and this says, okay, how could you re-represent that signal as the sum of intensities at certain frequencies throughout the frequency spectrum?
30:16 Speaker 00 Yeah, the way I think about it is almost like if your brain is playing a chord, the Fourier transform separates it into individual notes. And so you can see in that way. And that's essentially what's going on with EEG signal processing. I would just add that there's competing electrical signals, right? Your eyes, when you blink, are generating electrical fields. All of the muscles in your face are doing the same thing. The device itself that you're wearing is generating electrical fields. So there's all of this competing electrical activity around the head, which is being sampled. The electrodes are gonna sample some of that. That's gotta be processed out. And then there is what I was alluding to earlier,
31:00 Speaker 00 that the electrical field itself from the brain is going to be attenuated in different ways dynamically by the skull and other sort of cranial media. Difficulty of processing that, that's where there have been a lot of traditional techniques, traditional statistical analysis techniques that have been used, and older MEL techniques, convolutional neural nets, recurrent neural nets, all that kind of stuff that have shown some promise. Basically, everything that I'll talk about, actually, is really using those kind of older forms of statistical analysis to break this apart and figure out what's relevant and figure out what it means. There hasn't been a lot of transformer-based work and other types of things.
31:45 Speaker 00 And that's why I think there's just a lot of low-hanging fruit. And it's actually still an open question to me. The transformer has been around for a while now. There are some papers that show EEG transformers and other types of neural signal processing with transformers. But why aren't there much more? Is it the conservatism of academia? My first intuition was maybe it's a data problem. But these things generate a lot of data. So I really don't know. I'd love to know. If any of your listeners have insights, I'd love to hear.
32:14 Nathan Labenz? Yeah, I have a couple of intuitions, which may or may not be right. And people can correct me on this as well. But we just did an episode on the first 90 days of Mamba literature. And one of the things that is really interesting about this new mechanism, the selective state space mechanism, is that it does have different strengths and weaknesses compared to the attention mechanism, both in terms of how much memory it consumes, where The attention mechanism is quadratic in the length of the input. And that might be, by the way, one of the reasons, like, just as you talk about, like, a torrent of data and a thousand samples per second. If that were to be naively translated to a thousand tokens per second,
32:59 Nathan Labenz? then very quickly you're getting to a level of tokens that we have only very recently reached with frontier-grade transformers. It was only with GPT-4 a year ago that the public first got to see a quality 8,000-token transformer. And before that, it was like, just a couple of months where we had just seen the 4,000. And before that, as of like 18 months ago, 2,000 tokens was what you could really get from like the OpenAI API. So just the sheer volume of data may not lend itself super well to the transformer. But also, another interesting thing is that when they break down these microtasks and look at what the transformer can do and can't do.
33:40 Nathan Labenz? One of the things it really struggles on is the hyper-noisy environment. There was an interesting result in this one Mamba versus transformer comparison paper. It's more about the selective states-based mechanism and the attention mechanism. Those are really the two things that are more dueling it out than the higher-level architectures. And they're not even dueling it out because they actually work best together, spoiler. But in the super noisy environment where what actually matters is quite rare in what you're signaling, then the transformer sometimes has a hard time converging. And the intuition I've developed for that is because it's changing all the weights at the same time across like the entire range of the input,
34:25 Nathan Labenz? it may be that the gradient is often dominated by noise and has a hard time converging on the signal. Whereas, when, I don't want to make everything about the selective states-based model, though I do have an obsession about this, as folks know. It is updating per token. And so it seems like it has a more natural mechanism when the actual signal hits to say, oh, and this is where I start to violate my no anthropomorphizing policy. But it has an ability to recognize when the signal hits and update in a more focused way on that one thing that really was supposed to matter. Whereas a transformer is updating everything all across. It's considering everything at once. And so it seems like the signal can get lost in all that noise.
35:09 Nathan Labenz? The recurrent nature of the selective state space mechanism allows you to kind of zero in and do the gradient on the signal when you have the signal. And then, of course, there's still a lot of noise, but that maybe can get separated from the signal because of this bit by bit level processing and updating. I'm not 100% confident in that theory, but it is consistent with all the evidence that I know of so far. So we'll see how that evolves through time.
35:36 Speaker 00 Then we might be in for an exciting couple of years if that's true. Yeah, that's a great point. Very well.
35:42 Nathan Labenz? Hey, we'll continue our interview in a moment after a word from our sponsors.
35:47 Speaker 00 Maybe that. I've talked a lot about the downsides and what's hard. Once you get into this literature, what ends up happening is you just admire the very, very challenging problems that there are and just how complicated the brain is. But there is a lot that you can do. And what's shocking about EEG is that it works as well as it does non-invasively, especially when you consider that the only electrical signals really that can reach the brain or that can reach the electrodes on the scalp from the brain are very much on the surface, very much cortex, maybe one to three millimeters below at the most. Everything else just gets totally attenuated.
36:30 Speaker 00 So any deep brain structure, there's just nothing that EEG can really read. So everything we're talking about is stuff that's coming from just the surface of the brain, and that can do things like predict seizures with nearly 100% accuracy up to one hour before. That was demonstrated, in fact, about eight years ago in a lab setting. And there are devices that do this for people with epilepsy. So the telltale signs during which the patient
36:58 Speaker 00 Feels absolutely nothing. There's no external sign for the patient that a seizure is coming, but there are brain patterns that can be interpreted by consumer EEG hardware that can predict with almost 100% accuracy whether a seizure is going to happen. Same thing with Parkinson's. Early onset Parkinson's has a pretty distinct neurosignature that can be read. That's something like 90 plus percent accuracy. So already the potential of a medical device, I mean, I'm wearing AirPods right now. In fact, Apple has a patent that they've explored of putting electrodes on AirPods to do this exact thing. Just simply something, a pair of AirPods that can
37:40 Speaker 00 monitor for important medical conditions like that and maybe help me relax. That's already something that could be quite useful and seems like it will happen within the next few years. Maybe not exactly in the AirPods form factor. And when I said, you know, make you relax, that refers to something called neurofeedback. So rather than a direct modulation by hardware of your neural state, neurofeedback is basically giving you the data and allowing you to change your patterns of thought in response to the data. So there are neurofeedback devices that have EEG sensors on a headband that you might wear and it pairs with a phone app. And the phone app will
38:25 Speaker 00 show you your neural activity in real time and maybe give you a game or some other kind of cognitive stimulator that you play with. And it will tell you how focused you are. And you can get focused more. The idea is that you can actually train circuits in the brain to become more focused. There are devices that do this for sleep without you having to do anything. A lot of devices, for example, there's one called the Vea, I believe, that resembles like a sleep mask that you wear around your head. And I've never tried this, by the way. I don't think the Vea is actually shipping yet, and I'm certainly not endorsing it, just to be clear. But you can wear it like a sleep mask and it will play music. And that music is dynamically tuned in response to EEG signal and meant to bring you into a more relaxed state.
39:11 Speaker 00 And ultimately, I don't think there's a lot of scientific literature backing this up, but there is general scientific literature to support the idea that neurofeedback is a thing. Your brain will develop circuits to get into a more relaxed state more easily on its own with enough practice of that. So that's the kind of thing that we're talking about in consumer devices. There's also motor control, which is a whole different interesting field that maybe you want to go into. But first, if you have any questions on the first set of things.
39:39 Nathan Labenz? Yeah, this sort of same paradigm seems to happen all over the place. There's like echoes of this going with the increasingly rich signal. Then we also have the sort of increasingly meaningful states that we're able to identify, right? Within the transformer, this is pretty well studied now where in these sort of late middle layers, the inputs have been worked up to these rich concepts and you can even identify the direction in activation space that corresponds to justice or fairness or love or these kind of things. You're like, wow, how did this AI learn to represent that when all it's doing is next token prediction? And here there's like a somewhat analogous thing where as the ability to read this signal gets better and better,
40:24 Nathan Labenz? we're able to see like, okay, it takes two electrodes on the brain to detect if you're sleeping or awake. With four, you can get to, are you stressed or are you relaxed? With eight, you can get to general emotional state like fear, happiness, disgust, anger. And then with 16 and 32, more advanced things are starting to happen. So I am actually really interested in the motor thing because that seems like an interesting lens into just like, how much resolution do we already have in the ability to decode this stuff? And then I also wonder if you have thoughts about the limits of that with this technology. Of course, but we're going to discuss some of our technologies too.
41:04 Speaker 00 Yeah, so motor control, it really does turn out that when you think about moving something, it is tantamount to sending the electrical impulses to actually move that thing. And that can be translated for people that are unable to make that motion themselves. So there's demonstrations, maybe it's just my algorithm, but I see this on X all the time, of a person wearing a helmet, controlling a robotic arm, for example. And oftentimes, the motions are jerky, or slow, or you'll see sped up footage. And those are the signal processing issues that we have now. If you see someone wearing a whole helmet, that's going to be a lot of EEG signal that they're reading. But you can interpret some basic motor control from the cortex, from cortical activity.
41:50 Speaker 00 Now, a lot of motor control is coming from deeper brain structures. The limits on that, at least to me, as someone who's not a neuroscientist, unclear exactly. But I actually think it's important to say at this point that some of the highest quality data sets for this kind of application are data sets that combine EEG readings with FMRI or PET scans. Now, PET scans and FMRI, again, not the kind of thing that are going to make it into consumer use cases. Maybe ever, certainly not anytime soon. In lab settings, you can record that. And then you have an interesting relationship because the FMRI in particular
42:32 Speaker 00 is a 3D voxel, as it's called in the literature, representation of brain activity. FMRI uses blood flow to infer a blood oxygenation, to be precise, to infer brain activity. And then you can connect that with EEG signal. So all of a sudden, that's an interesting AI application because you've got a much richer, but more sparse signal coming from the FMRI, and then a super noisy signal coming out of EEG. And there's an interesting connection. What are subtle differences that we might not notice, subtle patterns that we might not notice between different states of mind that we can easily see on the FMRI that we can't see on EEG, but perhaps it can be pulled out of the data?
43:17 Speaker 00 I think that's A very rich area of research. And we'll talk later about some companies that are using models exactly of that kind to do interesting things. But that could really advance, especially on the motor control. But non-invasive EEG-based brain-computer interfaces already exist. And that's to understand. Are they super useful? No, not really. There are four enthusiasts of this kind of thing, and they're not cheap. But you can today buy a 32-channel EEG headset and pair it with software that you can use to manipulate objects on the screen of your brain. It requires some calibration. Can't just do that out of the box, but it can do it.
44:03 Speaker 00 It's like they have a block that you can move left and right and things like that. Nothing like the cursor control that Neuralink has displayed, has shown. But that is possible in digital environments. And then certainly also it's possible to move robotic devices. And there are prosthetics that work in this way already.
44:21 Nathan Labenz? One question I've been pondering here is, what is the kind of breakdown? We talked a little bit about traditional signal processing and Fourier transforms and whatever. And then I'm also reminded of Elon Musk saying on the Neuralink show and tell day a year and change ago now, that turns out the best thing to decode a neural net is another neural net. And so I'm wondering what parts of the interpretation process? You've got like 1,000 numbers coming off each electrode per second. And then ultimately that needs to be translated into something, right? A classification of your state or like a direction in motion space. You're going to try to move your robot arm or whatever you're trying to do.
45:05 Nathan Labenz? And between there, there's, you could imagine like decoding with all traditional methods. You can imagine an entirely neural net thing that just like takes in these raw numbers and translates that to an output with basically no principled approach other than just throw a model and a bunch of data at it and let it figure it out. Do you have a sense for what the right balance is there? And whether today's balance is likely to hold? Or are we headed for another bitter lesson where just all of this gets thrown into neural nets ultimately?
45:35 Speaker 00 It does feel that way. We see problem after problem falling to the unprincipled application of neural nets. Obviously, it's more complicated than just throwing data into architecture. But basically, we've seen that over and over again. And every time it happened, the experts and the scientists who were specific to that field told us that will never work. It's impossible. And then it works. So I've learned to be humble in this regard, and I wouldn't want to hazard too much of a guess. My intuition right now is that the field frankly is probably relying on a little bit more primitive methods of doing the signal processing than it could.
46:19 Speaker 00 And that sort of moving in the direction of a fully neural net is at least worth experimenting with. And I think part of the reason that the scientific community doesn't do that is because you want comparability with studies that have come before. There is this like innate conservatism in this field of science that a fresh crop of startups focusing on this will be entirely uninterested in. And I think we're at the stage where we need to get this research out of scientific labs and more into at least corporate R and D labs. And obviously, something that's important to understand here is that every
47:02 Speaker 00 one of the big tech companies employs neuroscientists who work on things that are at least adjacent to this. Apple's got a lot of patents in the field. Apple's exploratory design group is probably looking at this. That's their kind of skunkworks operation. Meta's been more public about it. Meta's public reason. And in fact, Mark Zuckerberg, it's worth noting, in his review of the Apple Vision Pro, which we should talk about, by the way, because that's like an interesting kind of half step towards all this. But Mark Zuckerberg, in his review of the Apple Vision Pro, said, Basically, something along the lines of the eye tracking interface they've built is okay until we get the neural interface hooked up. Basically viewing it as an inevitability.
47:46 Speaker 00 There are probably, right now, inside of both startups and large corporate R and D labs, fresher approaches to this being tried than at least what I have seen in the academic literature, though in fairness, academic literature doesn't always go into a ton of detail about exactly how the signal processing is done at a statistical analysis level. But yeah, that's my general sense.
48:08 Nathan Labenz? Availability of data also may be a real issue for some of these things. It seems like the way that this has progressed in neuroscience has backfilling a narrative here, but I feel like it is, the brain is the ultimate black box, right? Even more messy and black boxy and just hard to get into, obviously for all sorts of medical and ethical reasons as compared to a digital neural network, which you can be into with clarity and chop parts off and see how it works with different permutations. So there's been just a huge initial challenge of figuring out in rough terms what is going on and how does it work? And so people are doing these very small sample sizes of looking at people in FMRI's
48:52 Nathan Labenz? and trying to deduce what brain region does what and like what frequency of wave seem to correspond to what. And this super low data regime figuring out in general terms what's going on then seems to naturally lend itself to, okay, now let's try to apply principled approaches to identify when that's happening versus I could see this shifting very quickly and that probably would happen in private companies first, just because if nothing else, they're going to see the opportunity and really invest in scaling up the data. Having the data to say, what were you trying to do? When we recorded that brain state, what were you trying to do with your robot arm or with the cursor on the screen? Like that data just has never existed.
49:36 Nathan Labenz? So you can't really take the better lesson approach until you have the scale of recorded data. To make it go. And it seems like we've only recently realized that this is going to work for everything and we need that scale of data and nobody has really collected it globally. We just did this episode with Paul Scotty, who is the author of the Mind I2 paper. They are beginning to work on a multimodal brain state interpretation model, which would take in different kinds of signals and try to output different kinds of predictions. And it sounds like that's really only getting underway, like there's, you know, data is super fragmented. It's all kinds of different places and forms.
50:18 Nathan Labenz? And that also seems like a big part of why this hasn't happened yet. Would you agree?
50:23 Speaker 00 Yeah, I do. And I think we're at a very primitive state when it comes to the data. I don't think we have a good sort of population level modeling of the variation in neurosignatures, even things like skull thickness. Right now, it's like the literature is, well, everyone's skull is different. And that's probably not true. It's probably not literally true that it's like a snowflake, right? Like my skull is, it might be unique at a very granular level of detail, but you can probably model population dynamics for this sort of thing. And that would make your imaging challenges substantially easier. Same thing goes for the neural activity. It's definitely quite likely that high dimensional thought
51:08 Speaker 00 in particular is probably a pretty unique signature. My associations that I have with the concept of flawed three opus, or Mozart or something, are very different from yours. And it's not entirely clear that we're going to be able to get to the point of like, oh, Dean and Nathan are both thinking about Mozart. That seems hard to be able to reach, but there's a lot lower level thought that is still interesting, that it seems like you could model, that's also more, that's higher level, somewhere in between Mozart and You know, I'm scared of this tiger. I'm trying to run away from it. We can read that pretty easily. We know what that looks like. That has a pretty common signature that's easy to pick up on.
51:52 Nathan Labenz? The MindEye 2 paper actually may shed a little light on this because, first of all, the data set is only from eight people that they work on. So people, it should be right before this on the feed. Basically, what they're doing is looking at your brain state as measured, in that case, by fMRI, and reconstructing what you were looking at. They had an earlier version of this where they created a custom model for each of the eight people that are represented in this one open source data set. Each person had to spend like 30 to 40 hours in an MRI machine over presumably a bunch of different sessions, and they're being shown an image every few seconds.
52:33 Nathan Labenz? And then they just had to sit there and click a button if they recognized that they had seen that image before, just to kind of make sure that people are, you know, engaged in the task on an ongoing basis. So the jump between MindEye 1 and MindEye 2 was that instead of creating a single model trained on all of the 30 to 40 hours of data per individual, they were able to train a single model based on seven of the eight individuals' data. By the way, each of the brains is quite different in shape. They report the number of voxels per individual, voxel being roughly speaking, a centimeter or a two millimeter cube.
53:16 Nathan Labenz? The lowest number of voxels from one person is like 12. How they choose the number of voxels is by like the anatomy of the brain. So they're looking at basically the visual cortex, segmenting that off, and then just splitting it into voxels. And the resolution of the voxel is constant. So how many voxels you get depends on the total space of the portion of the brain that they identify as your, the relevant visual cortex for this purpose. So 12,000 and some is the low end, and the high end is over 17,000. So you have a 12,000, a couple in the 13,000 range, a couple in the 14,000 range, a couple in the 15,000 range, and then one over 17, close to 18,000 is the highest. So you see like a
53:58 Nathan Labenz? Roughly 40% difference from the lowest number of voxels to the highest. And that requires then a little bit of a bespoke adapter portion of the model for each person, because they just have literally different numbers of input. Number via vector that is measured is a different length for each of these individuals. So the adapter to then get to the shared latent space has to be a bit different for each individual. But once they create that shared latent space, then they are able to do an additional person with just one hour of data. So the key finding there is that they go from a like a previous technology, same data set, interestingly, same raw information.
54:42 Nathan Labenz? These are like the same sessions that were recorded, but they're able to go from a version where it works at 30 to 40 hours, which is obviously prohibitive for most usage, down to now they can get it to work with one hour of data because they're tapping into this like shared latent space that they've created from other individuals. So I think that is pretty interesting. I asked him a question about can we say anything about how similarly people perceive the same thing? And he said, unfortunately, in the data set that they were working with, there's very little overlap between the different images that people saw. So on the one hand, you would say like, in some sense, that kind of maybe suggests that there is like a high level of generality because they're able to get this shared latent
55:27 Nathan Labenz? space to work even when people mostly didn't see the same images. On the other hand, it kind of creates a limitation in terms of can we say what the inner product of my Claude three conception and your Claude three conception is like, unfortunately, in this study, people just didn't see the same thing enough for them to be able to do that sort of analysis. But I did find it quite interesting, truly profound, quite interesting is an understatement. It's a profound observation that you can get with only seven individuals whose brains vary in size by up to 40% to a shared latent space that is general enough that you can just come plug another individual onto it with basically one hour's worth of calibration
56:11 Nathan Labenz? data from an fMRI.
56:15 Speaker 00 Yeah, no, that's wild. It really is. And I think we're just scratching the surface there. Like I said, I don't know that we're going to get to my Mozart and your Mozart or whatever. But I also think that there's a really wide space. There's a lot of surface area that you can reach. So, yeah, and anything that makes it easier to collect data is particularly appealing because having to have people go through these things for thirty, forty hours is an unreasonable collection. So I suspect that we will see an acceleration of all this pretty soon. It's going to happen pretty soon. And I don't know that EEG will be the technology specifically.
56:57 Speaker 00 There's one other approach called Esmer's Functional Near Infrared Spectroscopy. If you almost think of it as like FMRI, but that you can wear around in a consumer device. It's not going to give you the 3D depth that FMRI is, but it's doing the same thing where it's blood oxygenation to measure brain activity. It is actually pretty simple. It's pretty comparable to EEG in terms of what it can do and in terms of its costs and things like that. So it could be that, but I think that there are devices that I can imagine existing that are not the full brain-computer interface that we're all dreaming of, or like what Neuralink is doing. But it can get you to some pretty interesting directions.
57:40 Speaker 00 And by the way, just as a side note, what Neuralink is doing is really kind of just an invasive version of everything we're talking about here. So instead of putting the EEGs or the electrodes on your scalp, they're putting a lot of them, very high density, implanted in the brain. With some silicon onboard, they can do some of this ML sensor processing right there for latency purposes. So what you saw with the guy playing chess on a computer with an implanted device is very impressive. And it's amazing to see some awareness of these issues coming up. Also, not shocking.
58:14 Nathan Labenz? You want to describe that just in case people haven't actually seen it?
58:17 Speaker 00 Yeah, there I go thinking that everyone's as engaged in this as I am. So yeah, this is the first patient of Neuralink.
58:26 Speaker 00 Is a young quadriplegic, a very sad case. He had a diving accident, something like that. Doesn't have any use of his arms and he had the Neuralink implanted and was able to manipulate standard Windows computer. If you've ever used Windows, it's exactly the same interface. It's not some speck of software. He's just using Windows to move the cursor around to play chess. The first night that he got all this hooked up, he stayed up all night playing Civilization VI, which I can relate to. And yeah, so just this amazing ability to more or less fully use a computer. And with devices like this implanted, typing has been demonstrated, not in a way that a consumer would ever want to use,
59:08 Speaker 00 but typing has been demonstrated with non-invasive technology. You can do that with non-invasive EEG. But being able to just fully manipulate a computer and live at least a digital life, surely with the brain, was what he was able to do. And again, amazing, but not something that would come as a galloping shock to anyone who's been paying attention to this. We've seen stuff like that before. And a lot of what Neuralink actually has done in kind of a classic Elon Musk fashion. One of the most important innovations that they have done is the automated surgery device, the robotic surgeon that can do this more or less without human intervention.
59:47 Speaker 00 There's Elon Musk thinking about not just how do I bring innovative technology, but how do I change the cost structure? He's thinking along those lines. So he's obviously thinking about wide-scale deployment or something like that.
1:00:00 Nathan Labenz? Okay, so can we summarize, or maybe there's a couple other high-level data points that could bring all this into focus to zoom out and give this the sort of survey view of where we are in the taxonomy. So it's two directions in which the information can flow. We have focused on reading of the brain states, not yet the... Ability to change the brain states, accepting as much as very sort of a feedback thing of if you measure and then you show that to the user, then they can get into a certain rhythm and react to the measurement that they're seeing. But that neurofeedback is not direct manipulation of the brain state. And then within this reading half of the equation, there's a lot of different technologies.
1:00:40 Nathan Labenz? We've got stuff that's outside the skull is subject to just really noisy signal, a lot of challenges with that. But with the EEG, you do get high frequency of signal. Then you've got the FMRI, which is a million dollar machine or whatever, and certainly not something you can wear around, gives you a much better spatial resolution. You can really see what's going on at a finer grain level, but less fast. There's some idea that these two modalities might be merged and there could be some really interesting generalization. Based on that, you can maybe unpack that a little bit more. And then in terms of what we can do with it, it's like with just a couple electrodes on the head, you can do basic stuff like are you asleep or are you awake?
1:01:25 Nathan Labenz? With a lot of electrodes on the head, you can do reasonably advanced stuff, although it's still kind of clumsy and slow, like you can type with your brain. Neuralink then is going inside the skull and that gives them a much cleaner signal and gives them the ability to have higher bandwidth. That's kind of the whole value prop that Elon Musk has talked about over time. Are there other memorable kind of striking demos or products that you think people should understand that are coming out of all this work that they could go check out or watch a little demo videos of?
1:01:57 Speaker 00 Yeah, I would say robotic control is definitely always worth looking at and you can find, if you just Google EEG robotic control, they're not really devices, they're not marketed devices to do this, but you can find a lot of laboratory settings where it's been done. Beyond that, there are all kinds of interesting headbands, things like that that you can wear sometimes, like headphones too. Over your headphones is very common with a band over them and they integrate the EEG into the band that goes over your head. There's a lot of products like that out there and they're not fantastically expensive, they're not super cheap. Usually somewhere between $500 and $1,000 would be roughly my estimate.
1:02:37 Speaker 00 In terms of the brain-computer interface, though, the company that I know of that is furthest along in this regard is a company called Emotive, E-M-O-T-I-V, and they are the ones that I was referencing earlier. The little V I Y that you can buy them and use them at home and pair it with software that allows you to do some very low-level manipulation of objects. You mentioned, though, like, what might be the next sort of level of generalization that we reach as we get better data and we apply better model architectures? Obviously, I don't know. There's been some interesting work on this in the lab is better decoding of language. So the best way to put it is mind-reading. So we've seen that with image.
1:03:20 Speaker 00 There have been a lot of, you know, mind eyes as an example. Meta has had some research in this regard. There's a lot of people that are basically decoding images that you're imagining in your mind's eye into using AI into a predicted digital image, basically doing the same thing with text, word by word at first, eventually getting higher bandwidth and higher latency. That seems achievable and that seems very interesting, particularly when you think about it in the context of communicating with an LLM, prompting an LLM with thoughts. As we were saying at the beginning of this conversation, the challenge, but also sort of opportunity with a frontier LLM is how good is your question?
1:04:02 Speaker 00 And the better your question, the more precise and tailored your question, the better an answer you will get out of these systems. So some people just don't have the communication capability to ask exactly the question that they want to, but they can surely think of it. So can that start to be translated into questions that they could prompt in an LLM? I don't know, maybe. Certainly, it seems to me that basic motives could be. So I think about something like this, this device that came out at CES this year, the Rabbit R1, which uses that thing that they call a large action model.
1:04:44 Speaker 00 An LLM that translates what you want into intense, basic intense. I also think about something like the iPhone and Apple devices in general have these automation layer called shortcuts where you can make little modular pieces of software to do a simple thing. Could we start to translate thought into basic actions like that? Maybe not my thoughts on a painting by Picasso, very complex set of thoughts, but maybe I want to see what my schedule is for today. I want to see what the weather is for today. I want to turn the bedroom lights on, things like that. Can we translate that and then use existing AI infrastructure, sort of automation infrastructure currently being built to actually just perform that?
1:05:30 Speaker 00 And essentially, that might feel shockingly like telepathy. If you can think, I want to turn the lights downstairs off and they will turn off. That might feel shockingly like telepathy. So in a certain sense, the really stone cold reality of where we are, but everything feels like it's converging to this point where a real qualitative leap is possible. At least it seems to me like it could be in the relatively near future. But to your point, to get to the, the brain computer interface that we all really want, you need write access, not just read access.
1:06:04 Nathan Labenz? Yeah, let's sketch out a little bit more the, how does this tip over the next couple of years? It seems like the model is basically, you know, this is a bit hackneyed now in the AI space to put everything in terms of what GPT level are we at? But it seems like we're at GPT one on this, maybe between one and two, where it's like, GPT two was just barely starting to be useful. When fine tuned, you could do some interesting things with it, but you couldn't do much with it, right? It certainly wasn't doing much in the way of reasoning. It didn't have the few shot learning ability that emerged with GPT three.
1:06:46 Nathan Labenz? And so you could do classification type tasks or some like sentence completion perhaps, but very limited application.
1:06:56 Nathan Labenz? It seems like we've sort of managed through mostly basic science to get to the point where we've got a similar level of capability. And you've got links here in the show notes that we can include so people can go check out in visual form like what these devices look like. But there's eight or so different consumer devices that have seen enough there where they're like, somebody's going to want to buy this. Let's get started building a company on it. And where I see this really changing is akin to robotics too. The amount of data that is going to be generated as the products hit the wild, even with the early adopter set, seems like it just goes vertical compared to what has existed before.
1:07:38 Nathan Labenz? Keeping in mind that Mind Eye II is eight patients, a total of 200 and some hours in an MRI across eight individuals. Now you've got things that are going consumer. So you've got orders of magnitude more data flowing in and you have people that are actually attempting to use them. And so you're going to start to get feedback on what is working, what is not working. And so the regime, the data regime is just totally changing. So if there's ultimately a very simple story, it's like we've gotten through enough basic science to get to the point where there's just this kernel of utility, which is going to tip us into a much faster bootstrapping dynamic where we're going to soar through orders of magnitude of data available.
1:08:22 Nathan Labenz? That of course, we've already got architectures that can probably decode that signal. If there is enough data to learn from, the same machine learning techniques are working everywhere. So unless there's some very odd reason why they won't work here, the data is about to come online and that's going to be the big unlock that's going to allow for just tremendously more utility. And we might hit some limits around just how good of a signal can we ultimately get? But it seems like at a minimum, we are headed for this sort of stored procedure triggering that you mentioned for image decoding is obviously what you're seeing is already quite decodable. What you're thinking about in turning that into verbal form seems like very achievable.
1:09:05 Nathan Labenz? And then even control seems like it likely gets refined to the point where it's practically useful, if not like super smooth, just based on collecting a ton of data and applying known machine learning techniques to those signals. Is that basically your world model? Or how would you refine what I just said?
1:09:24 Speaker 00 I think that's you're pretty much spot on. The GPT two comparison is an interesting one. I recall back in my housing policy days, kind of YIMBY adjacent, build more housing is what YIMBY means. Build more housing and make it cheaper, to put it simply. I tried to use GPT two to analyze municipal zoning codes, segments of municipal zoning codes to give me just a thumbs up or down on how restrictive was it? How not restrictive was it in an automated way? And the answers that you could get out of it compared to now are just so simian was like good, bad kind of thing. And you look at where we are now. The difference, of course, is that language is a kind of ground truth that you can refine
1:10:09 Speaker 00 your understanding of over time. And the question really is, how consistent is the language of thought across people? I think that's the fundamental barrier, potential barrier, is how calibrated does this have to be? And ultimately, if I have to go get an FMRI scan to make any of this useful to me, then that kind of really changes it a lot. Maybe I'll do that. That's the craziest thing in the world, actually. How long does it take to go buy an Apple Vision Pro? About half an hour.
1:10:41 Nathan Labenz? Demo is 25 minutes.
1:10:43 Speaker 00 Yeah. Yes. How hard would it be to put an FMRI in the back of an Apple store or something? I don't know, but still it changes it for sure. The level of calibration to me is really the big open question here, both on the interpretation side and the imaging of the skull side. But I have to think that substantial progress can be made and that it is possible to at least translate impulses, desires, basic desires, and that that can be fed into various kinds of AI architectures that can take actions on your behalf. That just seems super possible in the near term. I would be shocked if that didn't happen in the next five to 10 years, maybe less, maybe
1:11:28 Speaker 00 a lot less. That basically is normal.
1:11:30 Nathan Labenz? Yeah, I see a parallel also between the fact that all of these devices exist and the current state of AI agents broadly, where it's like, everybody sees where the technology is going. In the case of the agents, it's like, yeah, GPT-4 wasn't quite trained. There's this weird juxtaposition where it's like, closing in on an expert level on things like medical diagnosis, but then it sometimes can't click the right button on a very simple user interface, or it gets stuck, it gets into like loops that it sometimes can't break out of, like, why is that disconnect there? Presumably, it's because there wasn't really the sort of task completion mid-length episode data available to train the first version of GPT-4 on.
1:12:14 Nathan Labenz? I strongly believe now that the big tech leaders are investing heavily in creating that sort of mid-length episode data, and that in all likelihood, I would bet reasonably confidently that the next big shift is going to be that those sorts of things are going to start to work, even if like the MMLU score doesn't go up that much in the immediate term. And so to close that thread, when that happens, then all these agent products start to work dramatically better all at once. And I guess back to that dynamics question we talked about earlier. On the brain-computer interface side, it does seem similar, where people are developing all these different form factors. You've got glasses, you've got helmets, you've got headbands, and they all kind of don't
1:12:56 Nathan Labenz? work that well because there wasn't enough of the right kind of data to train them on, but you also had to get these things out there to get that data. I don't know that OpenAI necessarily had to have all these agent products created. I think they probably could have created their own data. But in this field, you actually need readings off of a lot of brains. And so it seems like we're in this kind of similar space where the hardware is starting to ship. It doesn't quite work, but again, it's going to collect a lot of data. And then you can imagine a lot of these things turning on relatively quickly and being just a lot more advanced, perhaps without even necessarily needing major upgrades to the hardware. You already have a 32 channel thing. That may well be enough for a lot of use cases if you can decode the data effectively.
1:13:41 Nathan Labenz? So maybe it'll take more than 32, but it's, you know, what, what is the number of electrodes that the guy from the Neuralink patient had?
1:13:47 Speaker 00 That's a great question. I don't know exactly, but it's a lot. They're very dense. It might be like a thousand plus. I'm not actually a hundred percent sure, but it's quite a bit because of the sort of fibrous way that they're doing the implant is like very dense and fine.
1:14:03 Nathan Labenz? Google, both traditional search result and generative search result comes back with a thousand and 24 electrodes.
1:14:11 Speaker 00 Yeah, so it's a lot. Yeah, the problem though with that, obviously, is that it's a lot in a very specific place. And there are long-term issues with that. Invasive is very promising in terms of capabilities, but you got to have them all over the place. The EEGs, they do have the same problem that the non-invasive ones have when they're implanted in the skull. They can't read everything in the brain. They can read in very local areas. They can read down to the neuron level invasively, but they can only read locally. So same problem. And also, obviously, like the idea of connecting an EEG in my brain to the internet is absolutely terrifying, just from a cybersecurity perspective. So there's obstacles there too.
1:14:53 Speaker 00 Yeah, but one other thing that I do think about though is inasmuch as these various platforms, the problem with foundation model agents is that their test environment is the real world, the actual full, they don't have a baby version of the internet or computing environment that they can interact in. They have to use our computing environment, which is weird and has all kinds of history to it and affordances that are for us that they don't necessarily need, that probably ultimately serve to confuse them and complexify the environment for them. It is why I mentioned the shortcuts idea from Apple's platforms is what they have kind of done is modularize the functionality of not just their whole operating system, but developers can plug into this too.
1:15:38 Speaker 00 So third-party apps are modularized too. And Android, I believe, has something similar. It's not just Apple. That's the kind of thing that again, it could like take off pretty quickly. And it could actually be the local, so it could be fast. You could do this inference locally in theory, if you had a sufficiently powerful compute onboard, but it wouldn't be like, you don't need like an H100 to do this kind of thing. Yeah. And I mean, the other shocking thing is just how little compute has been applied to this problem. Yeah, there's low hanging fruit. I have no doubt about it. So yeah, that takes us to the right section of this, which from a technological perspective is actually pretty simple for me to explain in a narrative way. Because it's the same basic principle.
1:16:21 Speaker 00 We want to manipulate electrical and magnetic fields. The problem is that everything I described in terms of getting signal out of the brain in a clean way. Those problems are magnified substantially for getting signal into the brain because that's really what your skull is intended to do. It's not so much to keep things in, it's to keep things out. So there's direct current and alternating current stimulation transcranially. This is non-invasive tech. Transcranial magnetic stimulation would be the magnetic field equivalent of that. The problem is that there's been some promise, there's been some stuff that's been shown in the lab. But it's either very expensive hardware, it's not writing that good hardware curve like the electrodes are.
1:17:05 Speaker 00 By the way, which are getting denser, cheaper all the time. And the signal just diffuses. So it's just erratic, it's impossible to focus. And it's very varied between people. So I don't really see the magnetic or electric field manipulation as being all that promising in the long term for neuromodulation. Some people disagree with what's called transcranial magnetic stimulation. Some people disagree and think that is going to be the path if we can just get the cost down. But right now the cost is very, very, very high. So it's just not what I'm focusing on because I'm interested in things that seem like they could happen in the relatively near future. There is a technology, though.
1:17:51 Speaker 00 This really was my exclusive focus when it comes to neural, non-invasive neuromodulation. It's called transcranial focused ultrasound. It's a very old technology, it goes back a hundred years. But really, its rebirth is in the past ten or twenty years. What this is, is firing very high frequency sound waves at the brain, far outside the range of human hearing, far outside the range of the hearing of dogs and cats, though there are animals that can hear it, rats, bats, I think can hear it too. But that you can fire at a sufficient frequency and amplitude and pulse to target pretty precisely into the brain and deeper than anything else can.
1:18:34 Speaker 00 You can't get to deep brain regions, you can't write to the hippocampus, which is memory, from, with this, but you can go a few millimeters, as much as a centimeter, into the brain and you can target it very precisely. So it's a few millimeters to a centimeter of depth and then millimeter level resolution spatially. So it's a very promising technology. Also, it's cheap and also, much like the electrodes, ultrasonic transducers are on a good hardware curve. They are getting cheaper and denser year by year. The weird thing about TFUS, as it's called, it's not very well understood how it works.
1:19:17 Speaker 00 And it's just weird in principle that firing high frequency sound waves into the brain makes you, that changes your thought in any way. That's just weird, but it does. There's some who theorize that it changes the electrical receptivity of brain tissue in certain ways that just make electrical signal flow more readily. There's a theory that it just stimulates neurotransmitter release in the targeted regions of the brain. But no one's really found any safety problems with it at reasonable dosages. At high dosages, it has a thermal effect, as we see with other ultrasound. At a very high dosage, it can cause brain tissue to heat in a way that you probably don't want.
1:20:00 Speaker 00 The other thing from a safety perspective, of course, is that all of these things are done one time in a lab. Nobody has modeled what the long-term effects are of using this, for example, every day.
1:20:14 Speaker 00 It's very promising. The FDA says it's safe below a certain level. And it's been shown to do some shocking things in my mind. So we've talked a little bit about improving bandwidth. Maybe the single most interesting finding to me of this whole chain of research that I've done is this finding that with TFUS at the somatosensory cortex, fired at the specific regions of basically your tactile sensation, the part of your brain that measures tactile sensation, you can perform tactile discrimination tests. So basically, you cover the user's eyes and rub a pin on their hand and then rub two pins really close together.
1:20:56 Speaker 00 And if they can distinguish between the one pin and the two pins. And TFUS was shown to improve that when it was being fired at the brain. And also had some offline effect, too. An offline versus an online effect. An online effect is do we observe an improvement in the desired cognitive activity when the treatment is actually being applied? Offline is do we observe it afterwards? And 40 minutes afterwards, the discrimination had attenuated a little bit, but was still there. So it had a long-term effect on what it would seem to me is the brain's channel bandwidth to the hand. So that is very interesting. There's a lot of other things that it's done, too.
1:21:39 Speaker 00 Mood, focus, visual acuity and visual discrimination tests in the same way that I described with the tactile discrimination. You can actually induce complete changes to the visual field or at least noticeable changes to the visual field by inducing these, what's called phosphines. If you imagine when you shut your eyes and you lightly press with your fingers on your eyelids, those kind of blobby shapes you see, those are phosphines. You can induce that in a person's visual field using TFUS. Auditory discrimination also goes up. Another finding that interests me, too, is the people can discriminate between vibrations
1:22:24 Speaker 00 at different frequencies that are very close together. They can do that discrimination, not substantially, but marginally better. With TFUS and all of these studies, for the most part, are TFUS applied at fairly low dosage levels, well below where that thermal effect is seen that I mentioned. So there might be an interesting range of capabilities that you could obtain if you were to go beyond that dosage level. And I've talked to both of the leading scientists in the world who have had the top TFUS papers in the last 10 years, and they've both said, we are very confident that
1:23:03 Speaker 00 you can increase dose safely at this point. That's TFUS. And can it write thoughts to the brain? Like, absolutely not. And I don't even want to suggest that I think that's remotely close. I think that's really hard. That's my intuition. But this idea of being able to improve not just mood, but things like channel bandwidth in the senses, that's really interesting to me. So yeah, any questions?
1:23:32 Nathan Labenz? So you had said that the skull is a huge barrier in the most obvious way to getting these signals into the brain. It seems like that's also a problem on the way out, but we sort of compensate for it by just like having a bunch of electrodes all around and collecting the mess and then trying to clean up the mess. But on the way in, where the precision really matters, that becomes a big barrier. Is there an evolutionary story for why the brain would be good at this? Or do you think this is just like an accident of biological history that the brain blocks these electrical signals?
1:24:07 Speaker 00 I think it probably has to do with the fact that it's probably useful for all kinds of other reasons for bone, just in general, to be low conductivity. And the skull is made of bone and you just benefited there. I have no idea, though. That's an interesting question. I'd love to ask an evolutionary biologist that question. I will say that with TFUS, the exact same imaging problems are present. Weirdly enough, in certain cases, the skull can actually have a beneficial effect on getting ultrasound signal into the brain. It's the kind of signal you're trying to create and the specific region of the skull is just right. It can actually have a lensing effect, but in general, you still have to deal with signal processing and sort of attenuating,
1:24:51 Speaker 00 modifying your signal based on a dynamic model of the stuff you saw to do that.
1:24:56 Nathan Labenz? It is crazy that this ultrasound is basically just vibration, right? Sound is vibration. This is just high frequency vibration. And I don't know if you know the frequency off the top of your head that this operates at, like middle A or middle C. The tuning note in the orchestra is like 440, right? So this would be like, you go up several octaves, we can definitely still hear up into the range of a few thousand hertz, it seems. And then I guess this would be like north of 10,000, maybe?
1:25:27 Speaker 00 I think it's five to 10 usually in that range, five to 5,000 hertz.
1:25:33 Nathan Labenz? Yeah, it's weird, right? At a neuron level, this basically amounts to shaking the neuron, right? It's like this thing can sort of fire off its electrical signal a thousand-ish times a second, or sometimes even faster, as you said earlier. Now we bring in a literal physical shaking that comes in at an even higher frequency than that. And that seems to kind of wobble things loose and just, this is stimulus, right? Is there a suppressive effect that can be achieved this way? Or is this purely a stimulus technology?
1:26:07 Speaker 00 In most of the literature I've seen is for excitatory signals, but no, you can do inhibitory effects as well. So yeah, you can block certain things too.
1:26:16 Nathan Labenz? And do you have a sense of what is the difference? If I'm imagining shrinking myself down to neuron scale and sitting in this region of the brain, it seems that the vibrations coming through, like, how would I know whether they're supposed to make me do stuff or not do stuff in my particular local area?
1:26:35 Speaker 00 Yeah, that's a very good question. Love to ask one of the scientists who helped me with this research, I'd love to ask them this question. That's a great practical question for them. My general sense is that most of the inhibitory stuff is not that much inhibitory stuff that actually gets done. Let me actually quickly pull up my notes. I'll get you an answer to that question. On what specifically has been inhibited at least? Cause that might answer the question. Yeah, so the inhibitory TFUS, most of it has been used for pain attenuation. So it's looking at the parts of the brain that are sensing pain and are targeting that. And that seems to work for whatever reason.
1:27:19 Speaker 00 I don't think that is well understood at all. I don't know for sure, but that also has the side effect of reducing motor time, reaction time for various kinds of motor tasks and can reduce some of the things we were talking about.
1:27:33 Nathan Labenz? Yeah, it's interesting. There's also, even if it was purely stimulating activity in a particular region, it's an important fact to keep in mind that the brain is self-regulating in all sorts of ways, right? So we see this at the cellular level too. There is a gene that gets expressed to suppress the expression of another gene. And in at the brain level, there are regions that activate to suppress other regions. And so you can imagine, even if you could only turn things on, in many ways, that might allow you to turn things off if you can figure out the indirect pathway to get there. Doesn't sound like that's what's going on here, but.
1:28:08 Speaker 00 It might well be what they're doing. That that that could very well be the case. It is definitely inhibitory in terms of activity, though. And I don't have a good model for why that works. The excitatory is at least intuitive. If anything, it's a little like, well, if you shake the TV set, it generally fixes the signal. There's a primitive part of us that can understand how that works. But yeah, the inhibitory is is a bit more of a puzzle to me as to why that works.
1:28:36 Nathan Labenz? So this question of writing thoughts, I'm not sure we ever would want this exactly anyway. But putting that aside, I can start to imagine how you might close the loop here. Maybe there are some barriers that I'm not immediately seeing. I'm sure there are plenty of challenges that would have to be overcome. And I'm not seeing any fundamental barriers. The resolution on this technology was down to a couple of millimeters as well, right?
1:29:03 Speaker 00 That's right.
1:29:04 Nathan Labenz? And so that's notably like the same scale as the voxels that come out of the fMRI. And so if I connect this to the MindEye paper again, and I'm like, okay, there are 12,000 to 17,000 voxels per person back in the visual cortex. And I got a mini lesson on this in that episode around how much semantic information is encoded in the visual cortex. Is that just like raw sensory workup until kind of lines and edges? And then the front of the brain does the sort of that's a tiger type stuff. And Paul, the author of this paper, he said, no, still in the back of the brain as part of the visual cortex, there is a lot of semantic information.
1:29:44 Speaker 00 Yeah.
1:29:44 Nathan Labenz? That understands like what it is that you know, and actually recognizes conceptually what it is you are looking at. So from these 12,000 to 17,000 voxels to ish millimeters cubed. They're able to extract both a general sense of the image that you're looking at. Like, what are the colors and what are the regions of the image? Which part is dark? Which part is light? Whatever. But then also. Straight from that reading, they can predict a caption for the image from the brain state. And that's because there is enough semantic information there that not only do you have this kind of blurry, purely visual information, but you also have this conceptual
1:30:26 Nathan Labenz? information. So anyway, point there is that you could read that at a two millimeter cube level. And now you're saying that you can also focus this signal down to that same kind of scale. It seems like the possibility, at least conceptually, exists to create a feedback loop where you might say, okay, I want to send some signal. How do I know that I sent that signal? I also need to then read the signal back. So you can focus the signal down to a couple millimeters region, but maybe you can't do that 10,000 wide. Maybe you'd have to do that 10,000 voxels at a time to induce meaningful higher order concepts in the brain. But it does seem like you at least now start to have some ability to make
1:31:11 Nathan Labenz? a perturbation and then also read it on the other end. And if you have a sense for what it is you're targeting, I want you to be thinking about Claude three. There seems to be at least some ability to start to be like, okay, I'm going to send a signal and then decode what states arise from that. And was the person thinking about Claude three? Yes or no? Then I update my network that is deciding what signal to send in based on the signal that is later read out. And this seems like a leap, but given what we have seen work, it doesn't seem too crazy to think that you could start to close this loop with, was I able to induce what I was trying to induce? Maybe the reward is too
1:31:52 Nathan Labenz? sparse. Maybe you need to have like broader regions, subject to receiving a signal at the same time to really get anywhere. If you can only target one, two millimeter cube, maybe that just gets lost in the overall broader state of the brain. But I'm starting to at least imagine how you can close a loop and begin to reliably induce certain things because you can read those things and that gives you some.
1:32:20 Nathan Labenz? Ability to correct or to gradually learn how to do the inducement in the first place. I'm not sure what other big barriers there would be to doing something like this. But it also seems like there's plenty of possibility that things could just be quite odd. And that maybe you don't actually need that. If you really could figure out how the brain works, like all these sort of thoughts arise somewhere. And they presumably arise locally first. There's some sort of trigger that ultimately propagates through broader regions of the brain. But it starts with some sort of input or some sort of signal that, you know, becomes dominant in the moment. And ultimately rises to the level of consciousness. It seems like it's not crazy to think that maybe you could find the key levers through this process of attempting to.
1:33:04 Nathan Labenz? Now, would we do this on humans? Maybe not, but you could definitely do it on monkeys. And I bet we're similar enough to monkeys. We could even create like a shared latent space with some monkeys at some point to get to the point where you could say, yeah, I'm just going to run this loop a ton. I'm going to try to induce these things. I'll measure how close did I come? I'll update on that. And even if I don't have the ability to stimulate that much of the brain at a single time, potentially I can find with a key that unlocks the lock to induce the states that I want to induce. Possibly crazy, but what, if anything, would say that can't happen?
1:33:37 Speaker 00 The main thing is that we don't really understand at all. As far as I know, we don't have a good understanding of how information is encoded in the brain. And it can be read back when you're doing things like the mind eye, the FMRI, and especially when you're pairing it, like when you're looking at the visual cortex, you can kind of like extract out of it. But actually, the right process, like how do you do that? Where does the right happen? Probably more than one place, right? Like the thing that distinguishes brains from even the most sophisticated neural networks is just how much crosstalk there is and how much weird resonance and just all sorts of feedback. It's constantly being conveyed.
1:34:20 Speaker 00 And I don't think we have a good understanding of that at all. Would I be shocked? If you were able to induce a simple thought in someone's mind by firing ultrasound at V2, the visual cortex? Well, no, I would be pretty surprised if that worked. It wouldn't be a galloping shock to me because of the informational aspect of that. I'm not sure that high dimensional information can be, first of all, encoded in instructions to ultrasound, to ultrasonic transducers, and then successfully transmitted into the parts of the brain, probably several, where it would need to go in order to actually write information, particularly of any.
1:35:00 Speaker 00 I'm thinking of the moment in The Matrix where Trinity downloads the instructions for how to operate a helicopter. Thinking of that as being your North Star. That seems hard to encode in all the various ways they would need to be encoded. But I wouldn't rule it out. At the very least, though, like, brain activity is possible to stimulate, and it is possible to change the mood of the wearer, the level of focus that they have, things of that nature, all possible. We don't know how much and how high dimensional that can get. Obviously, like, a good mood is a pretty general concept that could be refined quite a bit.
1:35:43 Speaker 00 There's a few TFUS devices that are on the market. There's a company called Prophetic AI that is based in New York. They've had a lot of buzz on X and other social media with some of their various announcements. And there's a couple of things that are really interesting about what they're doing because it is actually kind of the convergence of a few of the things that we've talked about in this discussion. First of all, they have developed what they call an ultrasonic transformer, which is a transformer-based system trained on EEG and fMRI data, which takes as its input EEG readings from the headband that they're going to sell
1:36:28 Speaker 00 and outputs instructions to the ultrasonic transducers also on that same headband. So it's inferring your brain state, and then it's turning that into instructions for neuromodulation. And Their ambition is to do this with lucid dreaming. It's actually a pretty easy signature. From an EEG perspective, this is not hard. Like you can absolutely recognize that somebody is lucid dreaming while they're sleeping. Because the neurosymmetry of sleep is pretty consistent, pretty easy to recognize even for low channel EEG. And then there's a spike of gamma waves in the prefrontal cortex.
1:37:09 Speaker 00 It recognizes that spike and then it directs the transducers to just keep firing to maintain that brain state and keep you in that lucid state. They might have slight corrections to that description, but the basic idea is that it can keep you inside of a lucid dreaming state. They also have ambition to do quite a lot of other stuff. Conscious experiences of all kinds, much kind of higher dimensional conscious experiences. They won't start with focus and a positive mood, but they have ambition to go beyond that. A lot of other people that are operating in the space have ambition to go beyond that. And again, you're writing a positive part of your record. So just like you are with EEG.
1:37:51 Nathan Labenz? Yeah, the how high dimensional is thought is it's really, I'm thinking about the representation engineering work that recently came out of case and other, there's been a bunch of authors on that paper, but they identify these high order concepts through clever, but honestly, it's not even that crazy of a technique where they create a bunch of different contrasting pairs that show like the presence and absence or the positive and the negative of a concept of interest. And then create enough pairs of those and then run them all through and then look at these intermediate states. And then take the averages and are able to basically say, okay, this appears to be the general direction, trying to kind of let the noise cancel out.
1:38:31 Nathan Labenz? That represents the direction from unfairness to fairness or the direction from sadness to happiness. And then they can start to use that vector direction to detect those states in further downstream inferences, blind, right? Just looking at the activations and saying, is this a happy state or is it an unhappy state? And then they can also start to layer those in and shape behavior based on just adding it to the whatever the thing is doing at a given time. If you just add on the happiness direction, then you can see that you actually steer the downstream model behavior in a reasonably intuitive direction where it seems like, okay, now it is actually generating happier
1:39:13 Nathan Labenz? outputs. So that is way, way easier to do in a neural network than it is to do in the brain. And the dimensionality of that is pretty high. So it might just be an engineering challenge that is like, at least in a non-invasive way, just too crazy to get to. But it does seem like we have something pretty similar going on. And you're citing all these examples where it's like, you're doing that at basically a low dimensional, very crude sort of way. And as we get into these higher and higher level concepts, I also think about the anthropic sparse auto encoders with this, where they show that basically from these high dimensional
1:39:56 Nathan Labenz? states, you want to identify these human intuitive concepts. And the more space you give the sparse auto encoder, the more concepts it will find. They call this feature splitting. So if you don't give it that many, if you give it like a limited number of concepts that it can differentiate its activation space into, then you'll get high leveled kind of more general concepts out. And if you give it a lot more space, then the features will split and you'll start to see these very granular level representations. And it seems like we're just operating right now at a very, very general level. We have not yet got to the level of resolution where the features can split. And so we're making
1:40:37 Nathan Labenz? these very basic modifications to mood, but it seems like it's really a question of resolution more than anything else over time.
1:40:47 Speaker 00 Yeah, resolution and it might also just be data, the same data problems that we talked about earlier. And then obviously with TF US in particular, that's even more limited of a data set. It might be some combination of all those things. Yeah, but I I don't know how far it can go. In principle, any conscious state is inducible. But how far can that go with non-invasive? My guess would be somewhere between exactly what I'm experiencing right now at the high end and being happy at the low end. I would suspect you can do more than that, more than being happy, less than what I'm experiencing right now. But where exactly? I think we just have to find out. It is, though, possible.
1:41:30 Speaker 00 It takes some time to to even get your head around what that would be like. The closest thing is probably drug use. And I really don't want to say that this is like using a drug because I think that this might be mu- much more than that. And I don't even necessarily mean illicit drug use. I just mean something that's meant to enhance your mood. But even then, my guess would be that because of the sort of slower mechanism of action that most pharmaceutical mood drugs have, it's going to feel like it's coming from inside you more. Whereas something like TFUS, you will feel good or focused suddenly. I don't know exactly how long the feelings persist.
1:42:13 Speaker 00 I would be shocked if it were like a steep drop off, like a completely like cliff. But it is also probably fairly quick that it drops off. And certainly, I I would guess that the experience of using something like the prophetic headband to lucid dream, if you use that every day, it would be like taking LSD every single day and you would eventually go into a Sid Barrett doom loop would be my guess. But yeah, how far can you go? I have no idea. It's very early days. This is very much at the frontier. At the same time, immediately gradable into consumer hardware, not expensive. And FDA approved below a certain limit. I'm happy to explore this more.
1:42:54 Speaker 00 That might be a good segue into some of the policy. I am a policy person.
1:43:01 Nathan Labenz? Before we go into policy, let me just ask one more kind of technical intuition question. If I had to make a high level guess miss on everything we've talked about, it feels like I would put my money on we're going to have. Pretty good brain reading in consumer devices over the next few years. Cause the hardware is already there and the cost curve is good. And the data is about to explode. And the best thing to interpret one neuron net is another neuron net. On the other end of the right or the modification to brain state side. It seems like resolution and just the overall ability to accurately send the signal that you want to send
1:43:46 Nathan Labenz? is limited and seems like it's probably going to continue to be limited relative to the bandwidth that you would actually need to create really rich input signal to the brain from outside the skull anyway. And so we're probably headed for something that is like fairly crude and sort of valence level, but we're not going to be invoking or reducing thoughts of Claude three opus with precision with those sort of non invasive techniques. And so if those things are to be done, it seems like they would require invasive techniques. And at that level, then you can actually send the signal where you want at the obvious cost of having to
1:44:29 Nathan Labenz? implant thousands or potentially many thousands of electrodes into the brain. But what would the path be if there is to be a path to something like the matrix moment where you could randomly download skillsets or knowledge bases or whatever, or even just create the sort of mind control to get people to act in whatever way you want them to act. With the current technology landscape, that seems only plausibly feasible through invasive methods.
1:44:55 Speaker 00 Yeah, I totally agree. The only thing I'll say, just as a caveat, is that the spatial resolution of TF US and the spatial resolution of EEG are basically the same. There might be practical differences in spatial resolution that I'm not thinking of. The obvious one being that with EEG, you can't differentiate between all the neural signals between individual neurons, but you aren't picking up. The collective, the aggregate result of their electrical activity. Whereas with TF US, maybe it is the case that you need neuron level or close to neuron level targeting to create really rich experiences. But I don't know.
1:45:38 Speaker 00 At the same time, neural networks tend to have like a fair amount of redundancy. Would you say that's like a fair observation? That there's possible that you don't need to get down to the level of the individual neuron to do quite a bit. So yeah, that, but in general, I agree with your intuition. That's about where I am too.
1:45:54 Nathan Labenz? Cool. Then let's get to policy. And also I wonder, do you have any intuition for sort of as this technology maybe follows its natural course? How does life start to change? Have you started to game out dynamics at all? And policy will be a part of shaping those dynamics. I, I used to ask people all the time, if the Neuralink had reached a million patients and was generally considered to be safe, would you get one? And interestingly, for people that are like very much on the frontier of AI, I got very different answers there. But some of the interesting answers that I got were rooted in the idea that I'd have to, how else would I keep up? So I feel like there's some game theoretic aspect to this and policy can shake the sort
1:46:39 Nathan Labenz? of game theoretic environment perhaps, and maybe can do other things too, including encouraging or discouraging the development of different kinds of technologies. But I'm interested in both kind of how you think this in the absence of governmental intervention, how it evolves? And then what questions government can and should be asking and what maybe the most likely things are to happen in that respect.
1:47:01 Speaker 00 Yeah, one of my favorite anecdotes, I'll go back to the 1910s for one second. The interior design of the average American's house before, and you can just imagine like an old house, right? Victorian house, whatever. Beautiful houses in Detroit, dark red, dark green walls. You often think about those kind of like burgundy. Why, why were all the walls dark back then? That seems weird. The reason is that interior elimination was provided by kerosene lamps, which stained walls. So you had dark colored walls to hide the stain. White walls are a luxury enabled by electricity. My point only is that it is very hard to predict the outcomes of what happens when these things are at scale.
1:47:46 Speaker 00 Obviously, like the things I can predict are going to pale in comparison to what will actually happen. That being said, a couple of things do seem apparent to me. I always try to think about the legal system first. And this is an area that interacts with the legal system, I think, in really interesting ways. The concept of being under oath changes if we want it to. And that's the question. And a interesting way of thinking about the next 20 years of technology in general is will we want to impose artificial constraints on various aspects of social life and technological development? And I'm not saying that we should.
1:48:30 Speaker 00 I don't have a strong model of that yet, but it does occur to me that if we're headed towards the Nicholas Bostrom solved world, which I don't think we are, by the way, but if he's right, then one of the things you will need is artificial constraints. Because my strong intuition is that a solved world quickly devolved into hell. I really do believe that.
1:48:54 Nathan Labenz? Can you unpack what exactly you mean by solved world? And then you're probably going there anyway, but more kind of granular, specific questions, I take it that you're asking, you can imagine a lot of different regimes, but would we require you to wear one of these headsets to testify? So we can also look at your internal brain states in addition to your like spoken testimony. That's the sort of thing that you are getting at, right?
1:49:18 Speaker 00 Yeah, exactly. The solid world is a concept from Nick Bostrom. Essentially, post singularity techno utopia style thinking, where every conceivable good is available in enormous abundance. And humans have godlike powers to assemble atoms in whatever way that they find desirable. Or self-dame does, some form of conscious life does. Maybe not us, but I'm not a big subscriber to any of that sort of techno utopian thinking. But yeah, ultimately, it's a very different kind of society if we can actually know at a biological level whether or not you are lying, for example, right? That's just a profoundly different society.
1:50:03 Speaker 00 Do we want that? Is a society with no deception actually a desirable thing? Is an AI model with no deception actually a desirable thing? I'm not confident the answer to either of those questions is yes. In fact, I'm like pretty confident that the opposite is the case. I'm pretty confident some degree of deception is an important part of life. I think some degree of deception is probably an important part of judicial life. In court, we assume that there will be some degree of bending of the truth. And I, I think that thinking about court is a fascinating way to think about how society will digest this. It's not the only way, but I just think it's a very concrete way to think about it.
1:50:44 Speaker 00 What are the limits on something like a warrant or a subpoena? In a world where varying degrees of dimensionality into human thought is recorded and in theory, something that you can be examined for a court case. My instinct on that kind of a question is that it's probably beneficial to a certain point. But there is an extreme where if everyone really has Neuralink and really truly every thought you have can be recorded in a high resolution compression, you probably don't want that to be available to, certainly not for advertising purposes. That's another thing I think about. Within limits, it seems fine. At the extreme, it seems dangerous to have to figure that one out.
1:51:28 Speaker 00 And my general sense is that from a policy perspective, sure. I mean, if this exists, the Europeans will regulate it, right? As can we know that? And half of America at least will. What to regulate it, whether or not they'll be able to is a separate question. Actually, it's probably worth talking about just for a moment, what the current regulatory state of all this stuff is. So these are not considered to be high-risk medical devices by the FDA by and large. They fall into this category two, this mid-level of risk. But there's an exception to that, which allows you to evade most of the FDA's regulatory processes. If you are marketing a general wellness device.
1:52:11 Speaker 00 So I remember when the Apple Watch came out with its blood oxygen sensor a few years ago. It was during COVID. It was like in 2020 that watch came out. Apple was extremely careful and blood oxygen is an important measure for COVID. Apple was very careful to say this has nothing to do with measuring for COVID. This is purely telling you your blood oxygen level. Because if you connect the device to diagnosing or treating any specific medical condition, then all of a sudden you're in a whole different world from a regulatory perspective. Neural technologies have generally gone for this general wellness exemption from the FDA.
1:52:54 Speaker 00 The FDA has not actually been clear that that exemption applies to them. They've been asked to make that clear and they have refused to do so, which is a common thing that the FDA and other aspects of American bureaucracy tend to do. Anyone who knows cryptocurrency will also be familiar with this, that like a regulatory tactic is actually uncertainty. That is a tactic, that's a form of regulation that regulators use, is creating uncertainty and creating gray areas. So that's where most of these neural technologies exist right now, is sitting in a gray area. They're being marketed and the reason that the FDA does that, I suspect, is maybe they don't know. Maybe they don't know how they feel about it. That could be true. But also for sure, they probably want to preserve the optionality to cut back on this stuff if they want to
1:53:36 Speaker 00 and to remove all the general wellness devices from the market in a heartbeat. But at the moment, if You could make a noninvasive brain-computer interface that can induce all kinds of conscious experiences. And as long as you're not trying to treat or diagnose a specific medical condition, and as long as you're staying below certain thresholds of TFMS and other things, like there's certain safety guidelines. As long as you're mechanically underneath those things, then you can just sell this on the general market. That's the way, that's how it currently applies. And the final part of your question was about long-term dynamics of adoption.
1:54:15 Speaker 00 And yeah, I think if, if these things are cognitive enhanced devices, then there will be evolutionary incentives. There are evolutionary incentives to do all sorts of things right now that most people do not do. Right? Like, we have enough data to know that I have an evolutionary incentive to go on a jog after this podcast. Will I do that? Probably not. Or that I should only eat whatever Andrew Huberman recommends. Like, all that stuff is true, and yet we don't do it, right? We don't do it all the time. We shouldn't be drinking alcohol, but a lot of people drink alcohol. So, I don't know. And I see this in the AI safety community in general.
1:54:53 Speaker 00 A lot of people think that legible Darwinian evolutionary impulses are going to drive technology adoption in society. And I think it's more complicated than that. I don't think it is a straight Darwinian evolutionary algorithm that's being applied here. So, that's my read on that. And I also think that there's a flip side. I think the impulse that people have a lot of the time is to think, "Oh no, if there's an incentive to use this thing, then everybody will have to use it. And what about the people that don't want to use it?" I, I go there too, right? I have sympathy for those people. I think about the Amish.
1:55:33 Speaker 00 And I do think that there's probably going to be forms of digital Amish in the future that we need to be thinking about. At the same time, The people who want to enhance their cognition also should have the liberty to do that. And we should want there to be more cognition in the world, especially human cognition. There's a part of me that says, my God, like, the idea that we would want there to be less human cognition in a world where GPT-5 is right around the corner? I am not sure from what world model that impulse derives, but it is a world model that I could poke holes at. Not that it's wrong, but I could certainly poke holes at it. So that's like, my mind is not at all made up on any of this, but that is how I model this right now.
1:56:16 Nathan Labenz? I know you're potentially just a couple doors down the hall from Robin Hanson, who's also a recent guest on the podcast. And he makes a extremely compelling case that we are in a, what he calls strange dream time, between what are almost sure to be much longer eras, both before and after us, in which evolutionary dynamics and just the practical constraints of like available resources are in fact the dominant drivers of how things unfold. And we're in this weird moment now where we've created way more capital per capita and birth rates are down. And it seems like the sort of strangeness probably can't last in the course of evolutionary time scales.
1:57:00 Nathan Labenz? Then he also, it's funny you mentioned the Amish too, because he also has a part of his near term world model, which I think is less compelling personally. But you can debate that with him over lunch, perhaps where he thinks because of the lower birth rates in general, but higher among the Amish, we may be headed for a period of technology stagnation and Amish domination. So that gets a little fine grained in terms of like how precise the crystal ball has to be to advance a theory like that for me. But there's definitely a couple of interesting lunch debate topics for you now at the Mercatus Center with him.
1:57:30 Speaker 00 Well, for sure. And I think part of his point is like, not to put words in his mouth at all, but we have a lot of artificial hyper parameters on the way society works, and those things are called laws and policies. And that's what I study for a living. And yeah, they've created all kinds of unintended weird effects. And that's not to say they're bad or good or really any of these, to say that they exist. And to say that things don't proceed according to mathematical models of the way that history, nature has unfolded in the past for those reasons. Because they didn't have occupational licensing a million years ago. And that does change the way that evolution works at a certain level, at least the evolution of society.
1:58:10 Speaker 00 And yeah, like just one other point about the kind of digital Amish thing. And just in general, my read on, there's some polling that gets put out about AI and things like that. I don't put a lot of stock in issue polling personally, and I don't think they're coming from organizations that are motivated to find the truth. I will just put it that way about what Americans think. I think the questions are along the lines of someone's going to make how 9,000? How do you feel about that? It's coming to kill you. It's gonna be here next year. What do you think? 98% of Americans are opposed to like, okay, that's not that useful of a question. But issue polling is also not that useful of a field unless you do it really carefully.
1:58:51 Speaker 00 But I do think that these things, adoption of technology is going to start to have a political valence to it. And whether that is coded as AI or whether that ends up getting coded in terms of these neural technologies or the Apple Vision Pro, is the Apple Vision Pro a political statement of a sort? Yeah, I see it as one already. I think a lot more people will in the future. So I don't know exactly how to model that, but I do think that we should expect this all to become more political. Not just about who gets to control the existing platforms, but about sort of whether you are interested in these technologies at all.
1:59:29 Nathan Labenz? Yeah, unfortunately, I would love to see the discourse around AI and technologies like this remain separated from day-to-day political discourse as long as possible, just because it seems like everything gets worse once it gets cast through the lens of certainly partisan politics. I'm not sure I agree with your view on polling in as much as I would say like, my read of a lot of those answers is that I certainly would agree that they're subject to framing effects majorly. So definitely buy that argument. My sense of just the people that I talk to in life and they're outside of the bubble, when I get outside of the bubble. Their kind of gut reactions to things is that it is negative by default.
2:00:11 Speaker 00 Yeah.
2:00:11 Nathan Labenz? And arguably, that has been shaped by culture and fiction and the Terminator and whatever. But I do think there is something quite real there that those answers are getting at. Though I also do think they're coming largely from a position of ignorance, certainly when you're just doing general public opinion polling. So I always advise people, I don't think, we don't have too many chat GPT novices listening to this show. But again, when I get outside the bubble and present like business leaders or whatever, I'm always just like, okay, first thing you gotta do. Spend some actual time with the technology, like you need to develop your own experiential understanding of what this stuff is. You can't just have it all be filtered through the media for you. Because the surface area is just so vast and the range of different use cases.
2:00:56 Nathan Labenz? And it's just so big compared to what you can get in an article or whatever. So you really do need to get hands on with it, feel it, mess around with it, give it your data, see how it reacts to stuff that is really personal to you. And if that's useful to you, and that also gives you a great sense of its strengths and weaknesses in doing that. Maybe in closing, I wonder if you could recommend some things that people might do to start to get themselves acclimated to this merge technology tree. Again, a lot of people follow this show will be very familiar with the language models. They'll have a sense for where AI is and some sense of where it's headed.
2:01:35 Nathan Labenz? I would guess that most people have never worn any of these devices, even starting with an Apple Vision Pro. I would still guess that some 5% of the audience has even done the demo at the Apple Store at this point. So what would you, you know, suggest? Is it like going and watching the demo videos on the company websites that we can, you know, put links to these companies in the show notes? Is it going and trying an Apple Vision Pro? Is it, what, what do people do to start to orient themselves to this and start to develop their own? Not just like through the media, not just by listening to you and me, but how do they get their own sense for how they should start to feel about this technology?
2:02:11 Speaker 00 Yeah. First of all, I will just say in response to the polling thing, I totally agree with you. People are pessimistic about this. That is a fact. I think that the techno optimists like myself, we have enough to fight. So I don't want to suggest that the polling is like manipulating the reality. I think it's exaggerating a tendency that already exists. That's how I would put it. Anyway, it's a very good question. I've never thought exactly about what other people should do. I guess my somewhat myopic reaction is like the path I took was useful enough for me. There's a good book about all this that came out recently called The Battle for Your Brain by Nita Farahani.
2:02:55 Speaker 00 She's done some great reporting almost about this kind of stuff, and I would recommend reading that. I actually would recommend trying on an Apple Vision Pro. And the reason for that is that you would be shocked how close your gaze is to a kind of neural interface. And it is something we didn't talk about at all, but it is probably the case that the ideal form factor for any of this kind of neural technology we're talking about is probably something that wraps around the head, much as the Meta Quest and the Vision Pro do. And so when you start to combine the idea of, well, we've already got eye tracking and we've got hand tracking and now we're adding in the neural technology.
2:03:40 Speaker 00 Maybe the neural signal doesn't need to be that good to do something really interesting, right? That's a the whole different angle to approach this is sensor fusion. So that's just one note. But yeah, I would try a Vision Pro.
2:03:51 Nathan Labenz? It's, by the way, a breathtaking experience. I was really floored by just the demo. And I haven't bought one yet, only because I'm not sure how much content there is and how much time I would really spend in it. But it was definitely an eye opener for me in the sense that it was like akin to a GPT-4 type experience where I have the Oculus 2. And when I first got the GPT-4 access, I had been using a lot of GPT-3. But the step up in terms of the quality and the like, oh my God, like this is just an arrestingly different experience. It is a similar, like you have to experience it to feel
2:04:36 Nathan Labenz? the difference. I could tell you how much better GPT-4 is versus GPT-3, but the best way is to get your hands on, I would say the same thing, even if you have done like relatively recent VR, but not done the Apple Vision Pro yet. I would say you do kind of owe it to your own worldview, if not necessarily to buy it, but at least to get that sense of what this thing can do, how high resolution it can be, how immersive it is, just how compelling the overall sensory experience of it is, because it is genuinely next level and it absolutely feels to me like a part of the future. Haven't quite figured out how to use it yet. The experience of it is, yeah, this is definitely going to be a thing.
2:05:12 Speaker 00 Yeah, no, I totally agree. And I think that it's not entirely a neural interface, but it's also not, not a neural interface. So I would say, like, reading that book, trying an Apple Vision Pro, and beyond that, unfortunately, my answer is boring, which is always try to proceed from ground truth of actual empirically demonstrated knowledge and not narrative, because the narrative surrounding this technology in particular is likely to be quite toxic and quite misleading in a variety of different ways. So model it for yourself. That is my best advice. Don't rely on someone else's model, including mine.
2:05:51 Nathan Labenz? In the future technology scheduling business, it is important to maintain always a high degree of epistemic humility. And I think that's a great note potentially to close on. Dean W. Ball, research fellow at the Mercatus Center and author of the Hyperdimensional Substack. Thank you for being part of the cognitive revolution.
2:06:11 Speaker 00 Thank you, Nathan.
2:06:12 Nathan Labenz? It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please, don't hesitate to reach out via email at tcr@terpentine.co or you can DM me on the social media platform of your choice.
Previous Automating Circuit Board Design Using Reinforcement Learning w Sergiy Nesterenko, Founder of Quilter on Autopilot Next Organic Holodeck with Prophetic AI co-founders Eric Wollberg and Wesley Berry Page updated October 26, 2026 · transcript and summary are machine-generated