Neuralink's DJ Seo: Inside the Race to Connect Brains and AI

[00:00] Okay, guys, this one needs really no introduction. We've got our brilliant partner, Sean McGuire, president of... [00:07] Neuralink DJ. Come on up guys. Thank you for calling. Welcome Neuralink. [00:14] We're actually going to stay on the side because we're going to play a two minute video just showing a badass. [00:18] DJ and his team are. [00:21] We're starting off with reducing human suffering. Our first product is called telepathy and that enables someone who lost the ability to command their body to be able to communicate with a computer. I'm thinking and a cursor is moving on a screen. It blew my mind. It's not a matter of if I can do something, but like how. [00:51] been immense for me. Aeros hits you like a pound of bricks. I am not able to do things that other dads can, but now he thinks it's so cool that I can do things that other dads cannot. [01:03] It's great to control a computer, that's a huge step. The Convoy team is the team at Neuralink that's working on assistive robotic devices controlled by the brain implant physical independence in the real world. I was moving the different directions and the different axes with [01:17] I don't know, it's pretty incredible just to think about. It's a really good feeling seeing what that thing's capable of. It was incredible to be able to just gesture with an arm again. Gaining functionality that I thought was gone forever was so incredibly life-changing. Like most ALS patients, Kenneth was losing his ability to speak.

[01:42] The real value of the brain-computer interface is that it would give Kenneth his voice back. [01:49] There we go. [01:51] I'm talking to you with my mind. Everyone gets his own colors. We'll be blue. You're going to do that? Yes, please. I'm done with my turn. Our next product is Blindsight, which will enable those who have total loss of vision, including if they've lost their eyes or the optic nerve, to be able to see again. The future of this technology feels almost unlimited. [02:15] We are going to continue advancing the quality of the sensors, the number of the sensors. Surgery will advance. Robotics is going to advance. And we are finding ways to apply it across all regions of the brain as we find new ways to refine and enhance its capabilities. It really feels like the sky's the limit for what we can do to help. How's it feel, Audrey? How's it feel, Audrey? [02:39] life-changing life-changing i could not ask for more all with my mind all with my mind [03:04] Thank you. [03:10] We were cracking a lot of jokes before that video, but I mean, honestly, that...

[03:14] brought tears to my eyes. This is one of the most inspiring projects in the world. It's incredibly difficult what they're doing. [03:22] I mean, they're truly... [03:24] Saving people. They're changing their lives. They're opening up possibilities that have been lost to them. It really is remarkable, DJ. So DJ co-founded Neuralink with Elon Musk. [03:34] and a few other people back in 2016. It's a different era in so many ways. A lot has happened in AI since then. Can you just tell us more, TJ, what made you drop everything in 2016 to build... [03:49] BCIs and how were you guys thinking about AI back then? [03:54] First of all, thanks for having me. It's a pleasure to be here. I would say AI has really been central to the origin story of Neuralink. You know, really the key insight back then was sort of the... [04:11] the I/O bottleneck between the human output and AI capabilities. [04:16] Even now, talking about bridging that gap seems a little crazy and wild. Although, I guess with every passing week, it seems more real and real. But certainly back in... [04:29] 2016, 10 years ago, it sounded insane. [04:32] Um... [04:33] Thank you. [04:34] And, you know, as far as... [04:36] I guess. [04:37] My personal involvement in this journey is, [04:41] I guess even as a kid, I was super fascinated by the brain. I mean, how can you not? It's the most interesting compute that we all carry. It does incredible things. It's the only form of general intelligence that we know to date.

[04:54] And [04:56] As an engineer, I always wanted to just understand what makes us and really try to explore the inner cosmos, so to speak. [05:05] And [05:06] You know, academically, I got introduced to brain-computer interface or brain-machine interface in the late 2000s. This was a couple of years after... [05:16] first human got implanted with the Utah Ray, which is this rigid silicon shank with sort of the through the, [05:25] skin port. And [05:28] Just doing something incredible, someone who was quadriplegic and be able to regain some of that autonomy. And that was very inspirational. And at the time, I was studying at Caltech. Go Beavers. Go Beavers, yeah. He was more hardcore, though. The undergrads are way smarter than the grad students. [05:45] I'd like to differ, but... [05:51] And, you know, I was focused around, you know, building miniaturized low-power electronics and went to Berkeley to pursue a PhD in that program, really with the eye towards, can you – [06:04] sort of [06:05] build a lot of the principles from semiconductor to miniaturize these systems, you know, take it out of the lab setting, which was kind of, [06:12] where things were into the real world. [06:16] I think – [06:18] When I met Elon and the team towards the tail end of my PhD program, I mean, just the sheer [06:24] ambitiousness and the scale of this project was something that I could not say no to.

[06:29] Thank you. [06:30] Thank you. [06:31] So, [06:32] To me, when I talk, something that's actually surprising me, a lot of people, when I ask them about Neuralink, a lot of people have never even heard of the company. Even really smart people oftentimes are not aware of how far along you guys are. I mean, we just saw a video. You have 20 plus people. [06:48] patients where you've changed their lives. What are some of the things that you think smart people most misunderstand or underappreciate about the company? [06:55] Thank you. [06:56] Yeah, I think oftentimes for those that have heard about Neuralink, they do kind of see these patient stories on X. And also it's kind of natural to think about the implanted device that's recording neural intent. And there's some sort of algorithm neural decoding to convert those neural intent into something useful for the patients. So I think that part is fairly, fairly well known. [07:20] I think the part that's underappreciated, and I think this is really where Elon magic kind of sprinkles in, is, you know, even at the – [07:29] the founding and day one of the company, we had scale [07:33] in mind, we had not just the device, but all of the infrastructure around it to be able to do the surgery, do the deployment, build... [07:42] you know, really, really hard factories for building these devices. And, you know, I think [07:48] That sheer scale obviously requires a lot of capital and a lot of boldness, but I think that formula seems to work really, really well. [07:57] Vertical integration is something that –

[08:00] is really the lifeblood of Neuralink and Elon companies, and what really enables us to have that fast iteration loop [08:07] from design, develop, deploy, and then just being able to stack [08:12] literally best talent across the entire tech stack. And I think the sheer scale of even the robotic surgery that we had in mind of making it as a Lasik surgery that can be deployed to millions and billions of people eventually, I think it's something that people fully do not appreciate and there's a lot of effort that goes into building that machine. [08:32] Just in case anyone didn't understand, these guys build surgical robots, and even just on the robot side itself, there's unbelievable breakthroughs that make this a very scalable thing. [08:44] approach in technology, but this is kind of my experience as well as just people underappreciate [08:50] When you build for scale from the beginning, it seems like you're going slow because you're building stuff under the waterline. And then once the iceberg pops over the waterline, it just becomes massive so quickly. And I think people are going to be surprised by... [09:03] Once this thing really ramps, how quickly it goes. Yeah, and I mean, the reality also is that these things take time. Like the world of atoms are very unforgiving. And, you know, I think doing these sort of investment, building out the team, having that muscle, having that, you know, hard lessons that you can learn from building physical things is, I think, invaluable. [09:24] So you guys now have over 20 human patients. Is there a particular patient story you can tell us about that inspires you the most or just is the most memorable for you?

[09:34] Thank you. [09:36] Yeah, it's really hard to pick one. You know, I think all these patients and participants that we have worked with, they have incredible stories, obviously. [09:48] very special moments that I do have the privilege of witnessing, especially when they... [09:53] I guess, first... [09:54] encounter that digital barrier, [09:57] drop with the Neuralink and that's always a very special moment for the participants as well as the family. [10:05] Um, [10:05] The thing that I will say though is that oftentimes I think what [10:10] sometimes become underappreciated is, [10:13] the role of the caregivers and their family, usually their family members. You know, it's people like [10:21] Nolan's mother, Mia, Brad's wife, Tiffany, and Ken's... [10:25] where [10:28] I think it's a really powerful [10:31] human story of love, sacrifice, and resilience that, you know, I think [10:35] is really inspiring. [10:37] to me. And if I were to maybe-- [10:41] take a slight philosophical tangent. Um, [10:47] One of my core beliefs and life philosophy is that a lot of our [10:51] you know, fulfillment, [10:53] comes from helping others and [10:56] The reason I believe that's true is [10:58] you know, similar to how your present self is very different than your past and future selves, sometimes maybe not unlike how

[11:06] Others are different. [11:09] And if your goal and fulfillment comes from striving, [11:13] to do best for yourself, I do think that you can apply the same principle to say, if you strive to help others, there's a lot of fulfillment that comes in. And because I believe that this is something that... [11:25] You know, I personally, as well as many others at Neuralink, find extreme fulfillment being able to help [11:31] those that really... [11:33] cannot help themselves. And I think it's extremely inspiring and motivating things to kind of stand at the team human side and being able to ignite that sort of fire of hope, which is something that we hear from our [11:47] current participants and future patients every day. [11:51] Thank you. [11:52] So in the video we watched in the beginning, it talked about [11:56] This next product, Blindsight. [11:59] What can you tell us about Blindsight? What's coming? What's coming? [12:02] Thank you. [12:03] Yeah, I guess, first of all, blindsight is a completely different paradigm shift than our motor prosthesis. So this would be for patients that are... [12:15] blind or have total loss of vision from diseases of their eyes or optic nerves, where you would have an external camera that captures the scene, [12:27] and then be able to directly write in the back of your head, in the visual cortex, which, you know, through electrical stimulation that changes polarization environment of the neurons can create these,

[12:39] thing called phosphenes. And more electrodes you have, more pixels you have, essentially, and be able to regain vision. And this is something that we're, we're, [12:50] very excited about and [12:52] Where it is right now is that we have currently the next gen [12:55] version that we want to go to human with in preclinical testing. So hopefully [13:01] I might eat my word, but hopefully by... And you do not need to make any predictions that you're uncomfortable making. [13:07] end of this year. [13:10] Thank you. [13:10] So everyone in this room is thinking about AI. It's a stacked audience. How do you think things play out at the interface? So it's. [13:18] BCIs and AI. [13:20] Thank you. [13:21] With the simplest thing being just, you know, thinking straight to Grok prompts coming, you know, I'm sure Nolan has this. [13:30] Yeah, be careful what you wish for, I guess. [13:34] Yeah. [13:35] I do think [13:37] In some time horizon, [13:39] AI basically becomes an exocortex. [13:43] similar to our neocortex, which is basically a layer above our limbic system. [13:49] And really, that's what kind of drives a lot of the cognitive abilities that we have. And, you know, it's ultimately all about the bandwidth. [13:59] It's really all about that interface. [14:02] Thank you. [14:03] You know, I think... [14:05] If you really think about the ultimate ceiling of this technology, I think it's really enabling

[14:13] direct. [14:15] uncompressed [14:17] high fidelity, [14:18] and multimodal transfer of concepts. The I learned Kung Fu in the matrix. That's one, but even maybe beyond that, right? [14:29] This is obviously still far out in the future. Our current system still is converting neural intents into legacy systems like keyboard and mouse control and language. [14:43] But... [14:44] I think the real breakthrough is going to happen when you can bypass all of that and be able to – [14:50] you know, sort of [14:51] compute on the raw intent itself and [14:54] I think it's actually not super... [14:58] distant in the future. In many ways, I think we're seeing this with modern AI systems where [15:03] these transformer architectures are doing incredible things. There's nothing fundamentally that prevents them from [15:11] say training on [15:13] and learning about the actual latent manifolds of the neural, [15:18] systems and [15:21] You know, I think [15:23] The key component here really is scale. I think, again, we're seeing this with AI revolution where [15:30] Things just seem impossible without scale, but things just become inevitable with scale. And I think similar things are going to happen. [15:38] you were [15:39] even at our scale of [15:40] 20 or so participants, we're starting to see some interesting results in building, you know, what's called neural foundational model where you have these

[15:52] state-of-the-art [15:53] LLM models or transformer networks that are being fine-tuned with neural data, and there's a lot of really interesting patterns and things that it's learning that's very counterintuitive. So I think... [16:03] Thank you. [16:04] Being able to sort of think about, and again, I think this is where scale is key, and that's where in the world of atoms, it's very unforgiving, and there's a lot of work to be done on the ground. [16:17] What is something you've learned? [16:18] From working with Elon. And bonus points, if it's something underappreciated by the public. [16:26] Yeah, I think... [16:30] I certainly also had... [16:32] different conception of what Elon time is, the insanely aggressive schedule. [16:38] Um, [16:39] What I've learned is that it's not really some... [16:42] arbitrary management, [16:46] tactics, but it's really used as a... [16:50] rigorous engineering and first principle tools. [16:53] Um... [16:54] Again, this is more applicable in the world of Adams. [16:58] But we have this concept called all green light schedule, which is... [17:04] kind of forcing us to think about if every single light is green, [17:08] You know, there's no administrative burdens or delays or no shipping delays or risk or, you know, no legacy bottlenecks. How fast can you build it? [17:19] and [17:20] When you think from that perspective, [17:22] point of view and vantage point and literally strip the strip away any man-made, um, bottlenecks,

[17:29] and you ask a question of how fast can you actually manufacture a chip, how fast can you design a robot and test it, [17:38] It turns out like [17:39] I don't know what the right number is, but at least 80%, 90% of it is things that you think are... [17:47] the way it is just because you don't really think about it from that perspective. So, [17:51] I think having that ability to [17:54] you know, frame everything from what is [17:56] Physically possible. [17:58] first principle baseline for building something like that. [18:02] I think it's a really, really useful just framework for thinking. [18:06] I should warn people that this is not for everyone, and it obviously comes at a cost in terms of – [18:15] uh, [18:16] you know, shifting priorities and insane stress sometimes and, um, [18:23] you know [18:24] cultural ramification of having this sort of system, but [18:29] When it works... [18:31] uh... [18:32] Boy, it's – [18:34] It's something that is very empowering, and I think it's also... [18:41] Yeah, just like... [18:42] the best thing ever to just see it unfold. [18:46] when it works to get reusable rockets and electric vehicles. I'm going to ask one more question, then open it up to the audience. And this is a layup. Part of how I convinced DJ to come, because this guy is very hard, was maybe we can help him with recruiting. There's all these AI foundation model representatives here, where everyone's competing against each other. It's hard to do recruiting.

[19:08] This is something different. They're on an island that... [19:13] scaling brain-computer interfaces. If any... [19:17] The question for DJ is, [19:18] Are there any particular roles that would be very helpful to fill? And for any of you, if you have brilliant neuroscience friends or... [19:25] You know, people doing BCIs that [19:27] Stanford or Berkeley or anywhere. [19:29] send them to me and I'll screen them and then send them to DJ. But any roles that are particularly burning, [19:36] Yeah, there's a saying, I'll answer that question, but I'll start by saying there's a saying at Neuralink that you don't have to be a brain surgeon to work at Neuralink. We really don't. [19:46] require any background in neuroscience. Most people, and especially some of the best people at the company, are just hardcore engineers. So I'll state that we're just in general looking for just hardcore engineers. And you can learn the neuroscience. And I mean, we'll learn as we go. And there's [20:08] Again, like a lot of our... [20:09] problems that we're solving are real hardcore [20:12] engineering challenges in the manufacturing and the robotics and things like that. I would say, [20:18] Um, [20:19] embedded software informer right now is a critical role for us to fill. [20:25] And [20:25] And as I mentioned, I think we're starting to enter into an era where the scale of the data that we're getting, the quality of the data that we're getting is extremely interesting to train some of the most state-of-the-art models. And there's also some inherent challenges associated with how do you even label this data set?

[20:42] Like, how do you actually know what the human true intent is unless... [20:45] So how do you design it in a way such that you can sort of clean up the data, which is obviously a big part of-- [20:52] AI, so [20:54] So any really smart person that doesn't want to work at a frontier lab, um, [20:58] We have four minutes for questions. [21:00] Younis is bringing you Vox. [21:04] question on how you choose when to like when to go deep on a product versus expanding to others like [21:13] What you built with telepathy seems like an already transformational product, and yet you're already kind of on to the next thing. [21:20] Like, how do you decide what, like, when to expand into another area? And what are the biggest blockers right now to, you know, taking even what you've already built and making that more widely available? Yeah. [21:32] Yeah, that's a great question. You know, our whole strategy is beachhead and then expand. And then we try to make everything generalizable enough so that you don't just build... [21:43] a very specialized system for motor cortex and, you know, visual cortex, auditory cortex, etc. I mean, for us, there is... [21:51] a pretty huge bottleneck that is difficult to overcome, which is biology. [21:57] And [21:58] things that are somewhat easier to overcome but still... [22:01] Difficult is regulatory and [22:03] payment and that market dynamics. So a lot of our [22:07] strategy sort of hinges on what is the optimal time at which we should put some amount of resources so we can pipeline the crap out of things so that once one of them is approved, which is going to be the first –

[22:19] and the hardest one to get approval to build that clinical evidence of safety, you can then point to it, go through a much faster process that's PMA supplements, 510K, and things like that. But in order to do that, you want to build up some amount of preclinical and clinical data. So that's kind of how we think about the timing of launching different products. [22:40] Thank you. [22:41] Dan. So the video was incredibly moving. And I guess a lot of what you talked about, maybe these are not exactly the right words, but are like therapies for patients that have... [22:51] diseases and that's incredible. And then Sean mentioned something about, I know Kung Fu. I'm wondering like [22:58] Is there a near or I'm curious what the near and maybe even longer term. [23:02] view of more things that are optional and for augmentation as opposed to [23:07] fixing problems. [23:09] Yeah, I think, yeah, so certainly our focus right now is restoration of lost function, and this is obviously a very serious medical device and serious surgery that you would have to go through. The way I think about... [23:22] all of this is that at the end of the day, it's all about benefit risk. And that's how everything's framed as. Obviously, for people that are quadriplegic, they're [23:30] risk appetite for potential benefit is much higher. [23:35] I do think that [23:36] It's still not entirely clear what that transition to [23:40] non-medical use case would be. [23:44] At what point do you cross that benefit risk for someone who's otherwise healthy? [23:48] Um...

[23:49] Again, I think we have some ideas as to what that benefit is. [23:53] side would be and then obviously figuring out a way to reduce the risk as well. But [23:58] One thing I will say is that [24:00] There's a wonderful thing called off-label use. [24:03] once everything is approved, where if you can find the neurosurgeon, if you can find... [24:09] a way to pay for it yourself, you could get it for someone who's otherwise unhealthy. So I do think that once we get to sort of market that there may be some handful of people that [24:20] might [24:21] walk around with the Neuralink. [24:23] Thanks. [24:24] We have 30 seconds, so rapid-fire question, then rapid-fire answer, and we're done. [24:29] Do you think the hard problem of consciousness is a hard problem? And if you do, is there a pathway to solving it? [24:35] Yeah, hard problem of conscience is an extremely hard problem. I do think that [24:40] If you are able to inject new sensors, [24:44] there may be ways to quantitatively understand that. [24:47] That was very fast. DJ, one of the most inspiring companies in the world. Thank you for being here.