Neuralink Eyes 2026 Human Vision Trials as Patient Count Tops 20

Neuralink植入患者突破20人，拟2026年开展人类视力试验 脑机接口的终极上限是彻底绕过传统语言输出，实现人类原始意图向人工智能直接、无压缩的多模态传输。 Neuralink已将其脑机接口技术扩展至20多名人类患者，并通过其首款产品Telepathy成功帮助瘫痪人士控制数字界面。该公司高度依托垂直整合和激进的“全绿灯时间表”（all green light schedule），以应对复杂的实体制造与机器人自动化挑战。展望未来，Neuralink正为其视力恢复设备Blindsight筹备人体临床试验，同时致力于迈向这样一个未来：让AI充当高带宽的认知“外大脑皮层”（exocortex），能够借助神经基础模型解码人类的原始意图。 Neuralink 首款产品 Telepathy 已在超过 20 名患者体内植入，成功帮助高位截瘫和严重失语患者摆脱身体限制，仅凭意念即可精准操控电脑与辅助机器人设备。 (0:21) 公司的底层愿景旨在化解人脑与快速进化的 AI 之间的输入输出（I/O）带宽瓶颈，并通过自研手术机器人和工厂进行深度垂直整合，力求将侵入式脑机手术转变为如 LASIK 视力矫正般高度自动化、安全且可普及的过程。 (3:31) 正在研发的视力恢复产品 Blindsight 目前正推进临床前测试，有望在年底前启动人体临床试验；该技术通过直接电刺激视觉皮层绕过受损眼球或视神经，产生光幻视来帮助失明者重建视觉。 (9:24) DJ Seo 预计 AI 在长期将成为叠加在人类大脑皮层之上的“外皮层”（Exocortex），而脑机接口的终极突破是绕过键鼠和语言等传统媒介，实现机器与大脑之间原始概念意图的直接、无损且多模态互传。 (13:10)

Neuralink co-founder and president DJ Seo argues that brain-computer interfaces are essential to overcoming the severe communication bandwidth bottleneck between human cognition and artificial intelligence. While the company is scaling clinical devices to restore motor and visual function, Seo envisions BCIs eventually serving as an integrated cognitive layer capable of direct concept transfer. The Human-AI Input/Output Bottleneck (3:10) Co-founded in 2016 by DJ Seo, Elon Musk, and a small team, Neuralink was established on the thesis that human-to-machine input/output bandwidth is the primary constraint in interacting with advancing AI. Seo, who studied semiconductor principles and low-power miniaturized electronics at Caltech and UC Berkeley, views the brain as the only existing form of general intelligence. Rather than indefinitely translating thoughts into legacy outputs like keyboards, mice, or spoken language, Seo forecasts that AI will eventually function as an "exocortex"—a cognitive layer integrated directly with the neocortex. Neuralink is already training transformer architectures directly on neural data to build neural foundation models. According to Seo, fine-tuning large models on brain activity is beginning to uncover counterintuitive patterns, moving the technology closer to computing directly on raw, uncompressed human intent. Clinical Applications: Telepathy and Blindsight (0:01) Neuralink's near-term development focuses on medical restoration for patients with severe physical disabilities. Its first commercial product, Telepathy, has been implanted in more than 20 human patients, allowing individuals with quadriplegia or ALS to operate computer interfaces using thought alone. Alongside Telepathy, Neuralink's Convoy team is developing brain-controlled assistive robotic devices to help patients regain physical independence in daily life. The company's second product, Blindsight, is aimed at restoring vision for patients with damaged eyes or optic nerves. The device uses an external camera to capture visual inputs and directly stimulates the visual cortex at the back of the brain to generate phosphenes (perceived points of light). Seo reports that Blindsight is currently in preclinical testing, with the company aiming to transition to human clinical trials by the end of the year. Vertical Integration and First-Principles Scaling (6:31) To transition brain implants from experimental prototypes into widespread clinical tools, Neuralink is vertically integrating its manufacturing and surgical pipelines. The company manufactures its hardware in-house and builds custom surgical robots, pursuing a long-term goal of making implantation as automated, safe, and scalable as LASIK eye surgery. Seo credits Elon Musk's operational philosophy for driving this focus on scale from the company's inception. Operationally, Neuralink relies on Musk’s "all green light schedule," a first-principles framework that models delivery speed by assuming zero administrative, logistical, or human-caused delays. Seo explains that this exercise reveals that 80% to 90% of perceived bottlenecks are negotiable, accelerating physical hardware iterations. However, he acknowledges that this high-pressure approach brings severe organizational stress and rapid priority shifts. To support this work, Neuralink focuses its recruiting on software, firmware, and robotics engineers capable of handling complex data, rather than requiring formal backgrounds in neuroscience. Medical Risk-Benefit Limits on Elective Augmentation (22:41) Neuralink employs a "beachhead and expand" product strategy designed to create generalized interfaces across different cortical regions. Once its initial, most difficult device secures regulatory approval, the company intends to use Premarket Approval (PMA) supplements and 510(k) pathways to streamline regulatory clearance for subsequent products. Seo notes that the primary bottleneck along this path is biological complexity rather than regulatory or commercial factors. Addressing consumer interest in elective cognitive augmentation, Seo emphasizes that Neuralink's immediate focus remains strictly on restoring lost biological function. Because brain surgery carries inherent risks, the current benefit-risk calculus justifies intervention only for patients with severe medical needs. For healthy individuals seeking optional enhancement, the benefit-risk balance remains uncertain, though Seo notes that theoretical off-label usage could emerge among private-paying individuals once the technology achieves standard commercial approval. Neuralink's immediate trajectory remains tied to clinical safety trials, regulatory pathways, and the biological challenges of long-term neural stimulation. While the company envisions a future where BCIs unlock elective human augmentation and high-bandwidth symbiosis with AI, the transition from therapeutic restoration to widespread consumer adoption remains dependent on proving automated surgical safety and navigating rigorous medical risk-benefit standards.

Neuralink 总裁 DJ Seo 阐述了公司通过高带宽脑机接口（BCI）克服人机交互瓶颈的核心路线，强调以全垂直整合与极致的第一性原理工程推动技术从医疗康复迈向通用化应用。 人机交互瓶颈与神经基础模型 (3:10) DJ Seo 指出，Neuralink 自 2016 年与 Elon Musk 等人联合创立以来，核心驱动逻辑始终在于解决人类与先进人工智能之间的输入/输出（I/O）带宽瓶颈。传统系统目前仍依赖键盘、鼠标和自然语言等旧有载体来传递神经意图，而脑机接口的终极目标是实现概念层面上直接、无损且多模态的高保真传输，最终让 AI 成为附着于人类新皮层之上的“外皮层”（exocortex）。 垂直整合与全自动化手术愿景 (6:31) 针对外界对研发进度的低估，DJ Seo 说明 Neuralink 迄今已在超过 20 位人体患者中完成植入。公司的关键战略是在初期便围绕规模化展开全产业链垂直整合，不仅自主研发植入芯片和解码算法，还自行搭建高复杂度工厂与手术机器人系统。这种模式大幅压缩了软硬件研发与迭代周期，公司的长远目标是将植入手术简化为类似近视激光手术（LASIK）的高度自动化流程，以具备面向数百万人乃至更大规模人群部署的可行性。 从运动控制到视觉恢复的产品矩阵 (9:24) Neuralink 当前首款产品 Telepathy 专注于帮助重度运动障碍及渐冻症（ALS）患者通过意图直接操纵计算机光标和语音交流，Convoy 团队则负责开发由脑控驱动的辅助机器人以重建肢体自主性。针对眼球或视神经受损的盲人，下一代产品 Blindsight 采用外部摄像头捕捉图像并直接电刺激视觉皮层产生光幻视（phosphenes），目前该设备正处于临床前测试阶段，有望推进至人体试验。公司采取“单点突破后再扩展”的管线策略，后续计划借助 PMA 补充申请和 510(k) 审批通道加快衍生产品的上市进程。 工程文化与“全绿灯”排期法 (16:17) 在研发管理上，DJ Seo 分享了源于 Elon Musk 的“全绿灯排期”（all green light schedule）思维工具：工程团队在剥离所有官僚延迟、运输拥堵等非物理限制后评估项目极限交付速度，往往会发现 80% 至 90% 的瓶颈纯属人为设定。这种第一性原理方法带来了极高压力的企业文化与快速变动的优先级，但也支撑了物理硬件的快速迭代。目前团队急需软件与固件工程师来构建神经基础模型，通过清洗海量神经信号训练底层 Transformer，而并不强制要求应聘者具备神经科学背景。 功能恢复与健康人群增强的边界 (22:41) 面对技术是否会转向人脑增强的疑问，DJ Seo 明确当前阶段的核心红线在于修复受损生理功能，整条研发路线完全受严苛的“风险收益比”约束。四肢瘫痪患者面临的医疗收益极为显著，因而具备承担手术风险的合理性，而健康人的自愿增强尚未形成明确的收益平衡。但他同时预测，一旦产品正式通过监管审批进入商业化阶段，理论上可能出现少数健康个体自费并通过神经外科医生进行“标示外使用”（off-label use）的情况。 Neuralink 将其拓展阻力归结为不可妥协的生物学限制而非市场或审批门槛；关于利用向大脑注入新感知来定量研究意识难题的设想，仍停留在前沿理论构想阶段，尚待长期验证。

[0:06] Kenneth: I'm talking to you with my mind. [3:10] Shawn Maguire: one of the most inspiring projects in the world [3:10] DJ Seo: the most interesting compute that we all carry [6:31] DJ Seo: Elon magic kind of sprinkles in [13:10] DJ Seo: direct, uncompressed, high-fidelity, and multimodal transfer of concepts [16:17] DJ Seo: all green light schedule [17:49] DJ Seo: this is not for everyone [17:49] DJ Seo: insane stress sometimes [17:49] DJ Seo: the best thing ever to just see it unfold [17:49] Shawn Maguire: you get reusable rockets and electric vehicles [18:49] DJ Seo: there's a saying at Neuralink that you don't have to be a brain surgeon to work at Neuralink [18:49] DJ Seo: human true intent [21:01] DJ Seo: beachhead and then expand [22:41] Dan: I know kung fu [22:41] DJ Seo: restoration of lost function [22:41] DJ Seo: there's a wonderful thing called off-label use once everything is approved

[0:01] Okay, guys, this one needs really no [0:03] introduction. We've got our brilliant [0:05] partner Shawn Maguire, president of [0:07] Neuralink DJ. Come on up, guys. Thank [0:09] you for coming. Welcome, Neuralink. [0:12] >> [applause] [0:14] >> We're actually going to stay on the side [0:15] cuz we're going to play a 2-minute video [0:17] just showing how badass DJ and his team [0:19] are. [0:21] We're starting off with reducing human [0:23] suffering. Our first product is called [0:25] Telepathy and that enables someone [0:27] [music] who lost the ability to command [0:29] their body to be able to communicate [0:31] with a computer. [0:32] I'm thinking [music] [0:34] and a cursor is moving on a screen. It [0:36] blew my mind. [0:37] It's not a matter of if I can do [0:39] something, but like [music] how. [0:41] It's been nothing short of miraculous. [0:43] Before the implant, I was locked in, [0:45] non-verbal, quadriplegic. Now I control [0:48] my computer just by thinking and the [0:50] rewards have been immense for me. ALS [0:52] hit you like a pound of bricks. [0:54] I am not able to do things that other [0:57] dads can, but now he thinks it's so [0:59] [music] cool that I can do things that [1:01] other dads cannot. [1:03] It's great to control a computer. That's [1:04] a huge step. [1:06] The convoy team [music] is the team at [1:07] Neuralink that's working on assistive [1:09] robotic devices controlled by the brain [1:11] implant, physical independence in the [1:13] real world. I was moving the different [1:15] directions, the different axes [music] [1:17] with I don't know, it's pretty [1:18] incredible just to think about. Really [1:20] good feeling seeing what those things [1:22] capable of. It was incredible [music] [1:25] to be able to just gesture with an arm [1:29] again. [1:30] Gaining functionality [music] [1:32] that I thought was gone forever was so [1:36] incredibly life-changing. Like most ALS [1:39] patients, Kenneth was losing his ability [1:41] to speak. [1:42] The real value of the brain-computer [1:44] interface is that it would give Kenneth [1:46] [music] his voice back. [1:49] Hello. There we go. [1:52] I'm talking to you with my mind. [1:55] Everyone gets his own color. Will be [1:58] blue. You're going to do that? [2:00] Yes, please. [2:02] I'm done with my turn. [2:04] Our next product is Blindsight, which [2:05] will enable those who have total loss of [2:08] vision including they've lost their eyes [2:09] or the optic nerve to be able to see [2:11] again. The future of this technology [2:14] feels almost unlimited. We are going to [2:16] continue advancing the quality of the [2:19] sensors, the number [music] of the [2:20] sensors. Surgery will advance. Robotics [2:24] is going to advance. And we are finding [2:26] ways to apply it across all regions of [2:28] the brain as we find new ways to refine [2:31] and enhance its capabilities. It really [2:33] feels like the sky's the limit for what [2:36] we can do to help. [2:38] How does it feel, Audrey? [2:42] Life-changing. [music] [2:44] Life-changing. [2:47] I could not ask for more. [music] [2:50] All with my mind. [2:52] >> Mhm. All with my mind. [3:04] >> [applause] [3:09] [applause] [3:10] >> We were cracking a lot of jokes before [3:12] that video, but I mean, honestly, that [3:14] brought tears to my eyes. It's This is [3:16] one of the most inspiring projects in [3:18] the world. It's incredibly difficult [3:20] what they're doing and I mean, they're [3:22] like they're truly [3:24] saving people. They're they're changing [3:26] their lives. They're opening up [3:27] possibilities that had been lost to [3:29] them. It really is remarkable, DJ. So, [3:31] DJ co-founded Neuralink with Elon Musk [3:34] and a few other people back in 2016. [3:37] It's a different era in so many ways. [3:40] Um a lot has happened in AI since then. [3:43] Can you just tell us more, DJ? [3:45] What made you drop everything in 2016 to [3:48] build BCIs? And, you know, how were you [3:51] guys thinking about AI back then? [3:54] Uh first of all, thanks for having me. [3:56] Um it's a pleasure to be here. [3:58] Um I I would say AI has really been [4:01] central to the origin story of [4:04] Neuralink. [4:05] Um you know, really the key insight back [4:07] then was [4:09] um sort of the [4:11] the IO bottleneck between the human [4:13] output and AI capabilities. [4:15] And, you know, even now talking about [4:18] sort of bridging that gap seems a little [4:20] crazy and wild. Um although I guess like [4:23] with every passing week, it seems [4:26] more real and real. [4:28] Um but certainly back in [4:30] 2016, 10 years ago, it sounded insane. [4:32] Um [4:34] and, you know, as far as I guess my [4:37] personal involvement in in this um in [4:40] this journey, [4:42] I guess even as a kid, I was super [4:44] fascinated by the brain. I mean, how can [4:45] you not? It's the most uh interesting [4:48] compute that we all carry. It does [4:50] incredible things. It's the only form of [4:52] general intelligence that we know to [4:53] date. [4:54] And um [4:56] you know, as an engineer, I always [4:57] wanted to just understand, you know, [4:59] what makes us us and [5:01] uh really try to explore the inner [5:03] cosmos, uh so to speak. [5:05] And [5:07] you know, academically, I got introduced [5:09] to brain-computer interface or [5:10] brain-machine interface [5:12] um you know, in the late 2000s. Uh this [5:14] is a couple years after [5:16] um first human uh got implanted with the [5:19] Utah array, which is this rigid silicon [5:22] shank with sort of the through the [5:25] uh skin port. [5:26] >> [snorts] [5:26] >> And um [5:28] you know, just doing something [5:29] incredible. Uh someone who was [5:31] quadriplegic and be able to uh regain [5:32] some of that autonomy. And that was that [5:35] was very inspirational and you know, at [5:37] the time I was um studying at Caltech. [5:40] Go Beavers. [5:41] >> Go Beavers, yeah. He was more hardcore [5:42] though. The undergrads are way smarter [5:44] than the grad students. [5:46] Um [5:47] >> [laughter] [5:47] >> I I'd beg to differ, but [5:49] >> [gasps] [5:49] >> um [5:51] and um you know, I I was focused around, [5:53] you know, building miniaturized [5:54] low-power electronics. [5:56] And uh went to Berkeley to um pursue PhD [6:00] in in that program. Uh really with the [6:02] eye towards can you [6:04] sort of build a lot of the principles [6:06] from semiconductor to miniaturize these [6:07] systems. Um you know, take it out of the [6:10] lab setting, which was kind of [6:12] where where things were into the real [6:14] world. And um [6:16] you know, I think when I met Elon and [6:19] the team, uh you know, towards the end [6:20] of my tail end of my PhD program, I [6:22] mean, just the sheer ambitiousness and [6:25] the scale of this project was something [6:27] that I could not say no to. [6:31] So, [6:32] to me when I talk [6:34] something that's actually surprising me, [6:36] a lot of people when I ask them about [6:37] Neuralink, a lot of people have never [6:38] even heard of the company. [6:40] Even really smart people often times [6:42] they're not aware of how far along you [6:44] guys are. Just I mean, we just saw [6:46] video, you have 20-plus [6:48] patients where you've changed their [6:49] lives. What are some of the things you [6:51] think smart people most misunderstand or [6:54] underappreciate about the company? [6:57] Yeah, I think often times uh for those [6:59] that have heard about Neuralink, they do [7:01] kind of see these um patient stories on [7:03] X and also it's kind of natural to think [7:06] about the implanted device that's [7:08] recording neural intent and then some [7:10] sort of algorithm uh neural decoding to [7:13] convert those neural intent to something [7:14] useful for the patients. Um so, I think [7:16] that part is fairly fairly um well [7:19] known. [7:20] I think the part that's underappreciated [7:22] and I think this is really where Elon [7:24] magic kind of sprinkles in is [7:27] um you know, at even at the [7:29] the founding and day one of the company, [7:31] we we had scale in mind. We had not just [7:35] the device, but all of the [7:37] infrastructure around it to be able to [7:39] do the surgery, do the deployment, [7:41] build, you know, really, really hard [7:43] factories for building these devices. [7:46] And, you know, I think that sheer scale [7:49] obviously requires a lot of capital and [7:50] a lot of boldness. Um but I think [7:53] um you know, that formula seems to work [7:55] really, really well. Uh it it, you know, [7:57] vertical integration is something that [7:59] um you know, is really the lifeblood of [8:02] uh Neuralink and Elon companies and what [8:04] really enables us to have that fast [8:06] iteration loop from uh design, develop, [8:09] deploy, and then um just being able to [8:11] stack literally best talent across the [8:14] entire tech stack. Um and I think the [8:17] sheer scale of even the robotic surgery [8:19] that we had in mind of making it as a [8:22] LASIK surgery that can be deployed to [8:23] millions and billions of people [8:24] eventually, um I think it's something [8:26] that people fully do not appreciate and [8:29] uh there's a lot of effort that goes [8:30] into um building that machine. [8:33] Just in case anyone didn't understand [8:34] it, these guys build surgical robots and [8:37] like [8:38] there just even just on the robot side [8:40] itself, there's unbelievable [8:42] breakthroughs that make this a very [8:43] scalable approach and technology, but [8:46] this this kind of my experience as well [8:48] is just people underappreciate [8:50] when you build for scale from the [8:52] beginning, it seems like you're going [8:53] slow cuz you're building stuff under the [8:55] waterline. And then once the iceberg [8:57] pops over the waterline, it's just it [8:59] just becomes massive so quickly and [9:01] I think people are going to be surprised [9:02] by once this thing really ramps, how how [9:06] quickly it goes. Yeah, and and I mean, [9:07] the reality also is that these things [9:09] take time. Like the world of atoms are [9:11] very unforgiving and um you know, I [9:14] think doing these sort of investment, [9:16] building out the team, having that [9:17] muscle, having that, you know, hard [9:19] lessons that you can learn from building [9:21] physical things is [9:22] uh I I think invaluable. [9:24] So, you guys now have over 20 human [9:26] patients. [9:27] Uh is there a particular patient story [9:29] you can tell us about that inspires you [9:31] the most or or just is the most [9:32] memorable for you? [9:36] Yeah, it's it's really hard to pick one. [9:39] Um you know, I think all these patients [9:41] that and participants that [9:42] [clears throat] we have worked with, uh [9:43] they they have incredible stories, [9:45] obviously, and um you know, there are [9:48] very special moments that I do have the [9:50] privilege of witnessing, especially when [9:52] they [9:53] I guess first encounter that digital [9:56] barrier [9:57] um drop with with the Neuralink and [9:59] that's always a very special moment for [10:02] uh the participants as as well as the [10:03] family. [10:05] Um [10:05] the thing that I will say though is that [10:08] you know, oftentimes I think what [10:10] sometimes become under appreciated is [10:13] the role of the caregivers and and their [10:16] family member usually their family [10:17] members. [10:18] Um you know, it's people like [10:21] Nolan's mother, Mia, uh Brad's wife, [10:24] Tiffany, and Ken's [10:26] wife uh Cheryl where [10:28] um I I think it's a really powerful [10:31] human story of love, sacrifice, and [10:33] resilience that you know, I think [10:35] is is really insp- inspiring to me. And [10:38] um you know, if I were to maybe um [10:41] take a slight philosophical tangent [10:44] um [10:47] you know, one of my core beliefs and [10:48] life philos- philosophy is that you [10:50] know, a lot of our [10:52] you know, fulfillment [10:53] comes from helping others. And [10:56] the reason I believe that's true is you [10:58] know, similar to how your present self [11:00] is very different than your past and [11:02] future selves. [11:03] Um sometimes maybe not unlike how [11:07] others are different. [11:09] And you know, if your goal and [11:11] fulfillment comes from striving [11:13] to do best for yourself, I you know, I [11:15] do think that you can apply the same [11:17] principle to say if you strive to help [11:19] others, there's a lot of fulfillment [11:20] that comes in. [11:21] And because I believe that this is [11:23] something that [11:25] um you know, I I personally and as well [11:27] as many others at Neuralink find um [11:29] extreme fulfillment being able to help [11:31] those that really cannot help [11:33] themselves. And I I think it's a it's a [11:35] extremely inspiring and motivating [11:38] things to kind of stand at the the team [11:41] human side and being able to uh ignite [11:43] that sort of fire of hope which which is [11:45] something that we hear from our [11:47] future or current participants and [11:49] future patients patients every day. [11:52] So, in the video we watched in the [11:54] beginning, it talked about this next [11:57] product, Blind Sight. What can you tell [11:59] you know, what can you tell us about [12:00] Blind Sight? You know, what's coming? [12:04] Yeah, I guess first of all, Blind Sight [12:05] is um [12:07] completely different paradigm shift than [12:10] our motor prosthesis. Um so, this will [12:12] be for patients that are blind uh or [12:16] have total total loss of vision from [12:18] diseases diseases of their eyes or [12:21] um optic nerves [12:23] where you would have an external camera [12:25] that captures the scene and then be able [12:28] to [12:29] uh directly write in the back of your [12:31] head in the visual cortex which you [12:33] know, through electrical stimulation [12:35] that changes polarization environment of [12:37] the the neurons can create these thing [12:39] called phosphenes. [12:41] And more electrodes you have, more [12:43] pixels you have essentially and be able [12:45] to uh regain vision. And [12:48] this is something that we're we're very [12:50] excited about and [12:52] you know, where where it is right now is [12:53] that we have currently the next gen [12:55] version that we want to go to human with [12:58] uh in preclinical testing. So, hopefully [13:01] you know, I might eat my word, but [13:02] hopefully by the [13:03] >> do not need to make any predictions that [13:05] you're uncomfortable making. [13:07] end of this year. [13:10] So, everyone in this room is thinking [13:12] about AI. It's a stacked audience. [13:15] How do you think things play out at the [13:17] interface of BCIs and and AI? [13:21] >> [snorts] [13:22] >> With the simplest thing being just, you [13:23] know, thinking straight to Grok prompts [13:26] >> [laughter] [13:27] >> coming, you know, I'm sure no one has [13:28] this. [13:30] Yeah, I'd be careful what you wish for, [13:32] I guess. Um [13:34] Yeah, I I I [13:35] I do think [13:37] in some time horizon, [13:39] um AI basically becomes a exocortex [13:43] similar uh to our neocortex which is [13:46] basically a layer above our limbic [13:48] system. [13:49] And um really that's what kind of drives [13:52] a lot of the cognitive abilities that we [13:54] we have. [13:56] And you know, it's ultimately all about [13:58] the bandwidth. [13:59] Um it's really all about that interface [14:01] and [14:02] >> [snorts] [14:03] >> you know, I think [14:05] if you really think about the ultimate [14:07] ceiling of this technology, I I think [14:09] it's really [14:11] enabling [14:13] direct [14:15] uncomp- compressed [14:17] high fidelity [14:18] and multimodal transfer of concepts. The [14:21] I learned kung fu in the Matrix. That [14:24] that's that's one, but even maybe beyond [14:26] that, right? Is you know, your your um [14:29] you know, where this this is obviously [14:31] still far out in the future. I you know, [14:32] our our current system still is [14:34] converting neural intents into legacy [14:38] systems like keyboard and [14:40] mouse control and language. [14:42] Um but [14:44] you know, I think the real breakthrough [14:45] is going to happen when when you can [14:47] bypass all of that and be able to uh [14:50] you know, sort of compute on the raw [14:52] intent itself and [14:55] I I [14:56] I think it's actually not super [14:58] distant in the future. In in many ways, [15:00] I think we're seeing this with modern AI [15:02] systems where [15:04] um you know, these transformer [15:06] architectures are doing incredible [15:07] things. There's nothing fundamentally [15:09] that prevents them from [15:11] um say training on [15:13] and learning about the actual latent [15:16] manifolds of the neural [15:18] systems um and [15:21] you know, I think [15:23] the key component here really is scale. [15:26] Um I I think again, we're seeing this [15:27] with AI revolution where [15:30] things just seem impossible without [15:32] scale, but things just become inevitable [15:34] with scale. And I think similar things [15:36] are going to happen and [15:38] you know, we're even at our scale of 20 [15:41] or so [15:42] participants, we're starting to see some [15:45] interesting results in building you [15:48] know, what's called neural foundational [15:49] model where you have these [15:52] state-of-the-art LLM models or [15:54] transformer networks that are being [15:55] fine-tuned with neural data and there's [15:57] a lot of really interesting patterns and [15:59] things that it's learning that's very [16:01] counterintuitive. [16:02] Um so, I think [16:04] you know, being able to sort of think [16:05] about and again, I think this is where [16:07] scale is key um and that's where in the [16:10] world of atoms [16:11] it's it's very unforgiving and there's a [16:13] lot of uh work to be done on the on the [16:15] ground. [16:17] What is something you've learned from [16:18] working with Elon? And bonus points if [16:21] it's something under appreciated by the [16:22] public. [16:26] Yeah, I think [16:28] um [16:30] I certainly also had different [16:32] conception of what Elon time is. [16:35] Um the insanely aggressive schedule. [16:38] Um what I've learned is that it's not [16:41] really some [16:43] arbitrary management [16:46] tactics, but it's really used as a um [16:50] rigorous engineering and first principle [16:52] tools. [16:53] Um [16:54] you know, again, this is more applicable [16:56] in the world of atoms. [16:58] But um we have this concept called all [17:01] green light schedule [17:02] uh which is you know, kind of forcing us [17:05] to think about it if every single light [17:06] is green [17:08] you know, there's no administrative [17:10] burdens or delays or no shipping [17:13] delays or risk or you know, no legacy [17:15] bottlenecks, how fast can you build it? [17:18] Um [17:19] and when you think from that [17:22] point of view and advantage point and [17:23] literally strip the strip away any [17:25] man-made [17:27] um bottlenecks [17:29] and you ask a question of how fast can [17:31] you actually manufacture a chip? How [17:33] fast can you design a robot and test it? [17:36] Um it [17:38] it turns out like I don't know I don't [17:40] know what the right number is, but at [17:41] least 80-90% of it is things that you [17:45] think are [17:47] the way it is just cuz you don't really [17:49] think about it from that perspective. [17:50] So, [17:51] I think having that ability to um [17:54] you know, frame everything from what is [17:56] physically possible first principle [17:59] baseline for building something like [18:00] that. [18:02] Um I I think it's a really really useful [18:04] just framework for thinking. [18:06] Um I I I should warn people that this is [18:09] not for everyone and it it obviously [18:11] comes at a cost um in terms of [18:15] uh [18:16] you know, shifting priorities and uh [18:19] insane stress sometimes and um [18:23] you know, [18:24] cultural ramification of having this [18:26] sort of system, but [18:28] um [18:29] when it works [18:31] uh boy, it's um [18:35] it's something that is very empowering [18:36] and and I think um [18:39] it's it's also uh [18:41] yeah, just like the best thing ever to [18:44] just see it unfold. [18:46] When it works, you get reusable rockets [18:47] and electric vehicles. I'm ask one more [18:50] question and then open it up to the [18:51] audience. Uh and this is a a layup. You [18:54] know, part of how convinced DJ to come [18:56] cuz this guy is very hard was, you know, [18:59] that maybe we can help him with [19:00] recruiting. Um [19:02] like there's all these AI foundation [19:04] model representatives here where [19:05] everyone's competing against each other. [19:07] It's hard to do recruiting. This is [19:09] something different. They you know, [19:10] they're you know, an on an island that [19:14] scaling brain computer interfaces. If [19:16] any [19:17] The question for DJ is [19:19] are there any particular roles that [19:20] would be very helpful to fill? And for [19:22] any of you, if you have, you know, [19:23] brilliant neuroscience friends or you [19:25] know, people doing BCIs that [19:27] Stanford or Berkeley or anywhere [19:29] send them like send them to me and I'll [19:30] send them I'll screen them and then send [19:32] them to DJ, but any roles that are [19:34] particularly burning? [19:36] Yeah, um [19:38] there's a saying I'll I'll answer that [19:40] question, but I'll start by saying that [19:41] there's a saying at Neuralink that you [19:43] don't have to be a brain surgeon to work [19:44] at Neuralink. Uh we really don't require [19:46] any background in neuroscience. [19:49] Um you know, [19:51] most people and especially like some of [19:53] the best people at the company are just [19:55] hardcore engineers. [19:56] So [19:58] I'll state that you know, we're we're [20:00] just in general looking for just [20:01] hardcore engineers and you know, you you [20:03] can learn the neuroscience and I mean [20:05] we'll we'll learn as we go. [20:06] And there's again like a lot of our [20:09] problems that we're solving are real [20:11] hardcore [20:12] engineering challenges in the [20:13] manufacturing and the robotics and [20:15] things like that. [20:16] I would say [20:18] Um [20:19] in their software and firmware right now [20:21] is is a critical role for us to fill [20:24] and as I mentioned I think we're [20:26] starting to enter into an era where the [20:29] scale of the data that we're getting the [20:30] quality of the data that we're getting [20:32] is extremely interesting to train some [20:35] of the most state of the art models. And [20:37] there's also some inherent challenges [20:39] associated with how do you even label [20:41] this data set? Like how do you actually [20:42] know what the human true intent is [20:44] unless [20:45] how do you how do you design it in a way [20:47] such that [20:48] you know, you can you can sort of clean [20:50] up the data which is obviously a big [20:51] part of [20:52] AI. So [20:54] So any really smart person that doesn't [20:56] want to work at a frontier lab. [20:59] We have 4 minutes for questions. [21:01] Eunice is bringing you box. [21:04] >> [snorts] [21:05] >> Question on how you choose when to like [21:10] when to go deep on a product versus [21:12] expanding to others like [21:13] what you built with telepathy seems like [21:16] already transformational product and yet [21:18] you're already kind of on to the next [21:19] thing. Like how do you decide what when [21:22] to expand into another area and what are [21:25] the biggest blockers right now to taking [21:28] even what you've already built and [21:29] making that more widely available? [21:32] Yeah, that's a great question. You know, [21:34] our our whole strategy is beast head and [21:36] then expand [21:37] and then we try to make everything [21:39] generalizable enough so that you don't [21:41] just build [21:43] a very specialized system for motor [21:44] cortex and you know, visual cortex, [21:47] auditory cortex etc. [21:49] I mean for us there is a pretty huge [21:52] bottleneck that is difficult to overcome [21:55] which is biology [21:56] and [21:58] things that are somewhat easier to [21:59] overcome but still difficult is [22:02] regulatory and payment and that market [22:04] dynamics. [22:05] So a lot of our [22:08] strategy sort of hinges on what is the [22:10] optimal time at which we should put some [22:12] amount of resources so we can pipeline [22:14] the crap out of things so that once one [22:17] of them is approved which is going to be [22:18] the first and the hardest one to get [22:20] approval to build that clinical evidence [22:22] of safety, you can then point to it, go [22:25] through a much faster process that's PMA [22:27] supplements, 510k and things like that. [22:30] But in order to do that you you want to [22:32] build up some amount of preclinical and [22:35] clinical data. So that's kind of how we [22:36] think about the timing of launching [22:38] different products. [22:41] Dan [22:42] So the video was incredibly moving and I [22:46] guess a lot of what you talked about [22:48] maybe these [clears throat] are not [22:48] exactly right words but are like their [22:49] piece for patients that are have [22:51] diseases and and that's incredible and [22:54] then Shawn mentioned something about I [22:55] know kung kung fu. I'm wondering like is [22:58] there a near or I'm curious what the [23:00] near and maybe even longer term [23:02] view of like more things that are [23:04] optional and for augmentation as opposed [23:06] to like fixing problems. [23:09] Yeah, I think yeah, so [23:11] certainly our focus right now is [23:13] restoration of lost function and this is [23:15] obviously a very serious medical device [23:17] and serious surgery that you would have [23:19] to go through. The way the way I think [23:21] about all of this is that at the end of [23:23] the day it's all about benefit risk. [23:25] And that's how everything's framed as. [23:28] Obviously for people that are [23:29] quadriplegic their [23:31] risk appetite for potential benefit is [23:32] much much higher. [23:34] I do think that [23:36] it's still [23:37] not entirely clear what that transition [23:39] to [23:40] non-medical use case would be. Um [23:44] you know, at what point do you cross the [23:46] benefit risk for someone who's otherwise [23:47] healthy? [23:48] Um [23:49] Again, I think this is we have some [23:51] ideas as to what what that benefit side [23:54] would be and then obviously figuring out [23:55] a way to reduce the risk as well. [23:57] But one thing I will say is that there's [24:00] a wonderful thing called off-label use [24:03] once everything is approved where if you [24:06] can find a neurosurgeon, if you can find [24:09] a way to pay for it yourself, you could [24:10] get it for someone who's otherwise [24:13] healthy. So I do think that once we [24:16] get to sort of market there there may be [24:18] some handful of people that might walk [24:21] around with a neuralink. [24:24] We have 30 seconds to rapid fire [24:26] question then rapid fire answer and [24:27] we're done. [24:29] Do you think the hard problem of [24:30] consciousness is a hard problem and if [24:32] you do is there a pathway to solving it? [24:35] Yeah, hard problem of consciousness is [24:37] extremely hard problem. [24:39] I do think that [24:41] if you are able to inject new senses, [24:44] there may be ways to [24:45] quantitatively understand that. [24:47] That was very fast. DJ one of the most [24:50] inspiring companies in the world. Thank [24:52] you for being here. Really appreciate [24:53] it. [24:53] >> [applause]