T。

它。

萨姆。

萨。

萨。

萨姆。

它。

萨。

欢迎来到Neuralink展示会。我们有大量令人惊叹的新进展要与大家分享，我认为这些进展令人无比振奋，同时也会向大家介绍我们计划在这里开展的未来工作。

现在，这本来应该是一档技术播客，大致上我会先做一个总体概述，然后我们会请Neuralink团队的多位成员上台，对各个领域进行深入的技术介绍。所以，是的，那我就开始做总体概述。现在。我要说的一些事情，是你们已经。嗯，如果你一直在关注Neuralink，就已经听过了，但外面有很多人根本不知道Neuralink是做什么的。

所以，对于一些你可能已经知道、但其他人并不知道的事情，我会稍微重复一下。Neuralink的总体目标，是最终创建一种全脑接口，也就是一种通用的输入输出设备；从长期来看，它确实可以与大脑的方方面面连接，而从短期来看，它可以与大脑的任何特定区域连接，并解决大量会给人造成严重功能障碍的问题。

所以，你知道，我们的长期目标就像，我是说，我会稍微谈一下长期目标。听起来会有点深奥，但它是。它其实算是我的首要动机，也就是，你知道，有点像，我们该怎么应对AI？比如，我们该怎么应对通用人工智能？

如果我们拥有比任何人类都聪明得多的数字超级智能，我们该如何减轻这种风险？

从物种层面来说，我们该如何减轻这种风险？那么，即便在AI非常、非常仁慈的良性情景下，我们又该如何甚至一起去。

一起搭上这趟车？

我们该如何参与？

而结论是，我认为，在搭上这趟车以及让AI与我们保持一致方面，那个、最大的限制就是带宽，也就是你能以多快的速度与计算机交互。所以，从某种意义上说，我们都已经是赛博格了，因为你。你的手机和计算机都是你自身的延伸。而且，如果你，我相信你会发现，比如你把手机落下时，最后会拍打自己的口袋。这就像患上了肢体缺失综合征一样，比如你知道手机在什么地方。

把手机落下有点像缺失了一条肢体。此时，你已经非常习惯与它交互，非常习惯成为一个事实上的赛博格，但是。那么手机或笔记本电脑的限制是什么？限制就是你接收和发送信息的速率，尤其是你发送信息的速度。所以，如果你在与手机交互，速度就会受到你移动拇指的速度，或你对着手机说话的速度限制。

这是一种极低的数据速率。

可能大约是每秒10比特，乐观估计是每秒100比特，但计算机可以按每秒千兆比特、太比特的速率通信。所以，我认为这是我们需要解决的根本限制，以减轻人工智能的长期风险，也让我们能搭上这趟车。而且。

是的，所以，但是就像我说的，这是一个深奥的解释，我认为它会吸引小众受众，其中一些人可能就在这里。但是。

而且这是一个非常困难的问题。所以，即使我们没能解决那个问题，我认为目前我们也有信心，在这个过程中成功解决许多脑损伤问题、脊柱损伤问题。所以。

是的，所以，不管怎样，其实贾斯汀·罗兰德就在观众席中。

嗨，贾斯汀。这里稍微引用了一下《瑞克和莫蒂》，那一集精彩的《瑞克和莫蒂》讲的是增强你家狗的智力，以及最糟糕能发生什么？

所以不管怎样，我推荐《瑞克和莫蒂》。

所以为了。所以你想要能够读取来自大脑的信号，你想要能够写入这些

信号，

你最终希望能够对整个大脑这样做，然后也将其扩展到与神经系统的其他部分通信。如果有一个，如果你的脊髓或颈部有点断裂。

那么现在，这段视频现在已经是18个月前的了。这是Pager，它正在玩“猴子意念乒乓球”。所以在这段视频中，Pager植入了一个Neuralink植入物。

有趣的是，你甚至看不到神经植入物。我们已经将神经植入物微型化到与被移除的那部分颅骨厚度相当的程度。所以本质上，这有点像用一块，你知道，智能手表来替代一块颅骨，就像戴着一块Apple Watch或Fitbit一样，实在找不到更好的类比。所以

你可以看到。

你真的看不出来。他看起来很漂亮，很正常。我认为这非常重要。如果你装有一个 Neuralink 设备，比如我现在就可能植入了一个 Neuralink 设备，而你不会，你甚至不会知道。我是说，假设，也许在这些演示中的某一个。

事实上，

这些演示中的某一个。我会的。

对。

所以，对。总之，这是。首先，这有点不可思议。嘿，猴子会玩《Pong》。比如，如果你给它们一个操纵杆，它们真的会玩《Pong》。所以 Pedro 最初学会了用操纵杆玩《Pong》。所以我当时就想，这挺新奇的。就像，我不知道猴子会玩《Pong》，但它们确实会。然后，所以我们先训练 Pedro 用操纵杆玩《Pong》。接着我们拿走操纵杆，并使用 Neuralink。

而现在这是。他在玩 Telepath。它本质上是一款心灵感应电子游戏。

所以从那以后，我们一直在走一段从原型到产品的非常艰难的历程。

我经常说，原型很容易，生产很难。真的，我会说，从原型发展为一种安全、可靠、能在各种各样的情况下运行、价格可负担并且能够规模化制造的设备，要困难 100 到 1,000 倍，难度大得惊人。

我是说，有句老话说，成功是 1% 的灵感和 99% 的汗水，但我认为可能是 99%、99.9% 的汗水。

对于想法容易、执行困难，我能举出的最佳例子就是登月。登月这个想法很容易，登上月球非常难。

我们一直在努力为首位人类受试者做好准备。显然，在把设备放入人体之前，我们希望极其谨慎，并确定它会良好运作。但我想，我们已经向 FDA 提交了大部分文件，而且我们认为，大概在约 6 个月后，我们应该能够为首位人类植入 Neuralink。

但正如我所说，在甚至还没有进入人体、甚至还没有进入动物体内之前，我们会尽一切可能测试这些设备。所以我们会进行台架测试。我们会进行加速的加速寿命测试。我们有一个假脑模拟器，它具有那种质感，就像是在模拟大脑，但它有点像橡胶。因此，在我们甚至会考虑把设备放入动物体内之前，我们都会通过严格的台架测试，尽一切可能做好测试。

所以，我们不会轻率地把设备放入动物体内。我们极其谨慎。而且每当进行植入时，无论是在绵羊、猪还是猴子体内，我们始终希望该设备是用于验证，而不是用于探索。这样，我们已经通过台架测试做了能做的一切。只有到那时，我们才会考虑把设备放入动物体内。对，实际上就在今天晚些时候，确切地说是几小时后，我们会向你们展示一个在大脑替代物中进行植入的演示。

如果现场观众中有人想当志愿者，我们的机器人就在那边。

所以，Lithien，自 Pager 演示以来，我们已经扩展到与一群 6 只猴子合作。我们实际上升级了 Pager。它们执行各种任务，而我们尽一切可能确保一切稳定且可复现，并确保设备长时间使用而不会发生性能衰退。所以，你们在那里看到的东西看起来像《黑客帝国》，但那实际上，那是神经信号的真实输出。所以那不是模拟，也不只是屏幕保护程序之类的东西。

那些是真正的神经元在放电。这就是其中一种读数显示的样子。

这里可以看到 Sake，它是其他正在键盘上打字的猴子之一。

现在，这是心灵感应式打字。所以说清楚，这是。他实际上并没有使用键盘。他正用意念将光标移动到高亮显示的按键上。严格来说，我们实际上还不能拼写，所以我不想过度吹嘘这个东西，因为那是。那是下一个版本。

所以，……

但这里真正酷的是，Sake 这只猴子仅凭意念就在移动鼠标光标，让光标移到高亮显示的按键上，然后拼出我们。我们想要的内容，任何我们想拼出的内容。但是。然后。

所以，这，这可以供某个比如四肢瘫痪或四肢麻痹的人使用。甚至在我们让脊髓方面的东西发挥作用之前，也可以用它来控制鼠标光标、控制手机。而且我们相信，某个基本上没有其他与外界交互方式的人，将能够比双手功能正常的人更好地控制自己的手机。

我提到了可升级性。可升级性非常重要，因为我们的首款量产设备会很像 iPhone 1。而且我很确定，如果已经有 iPhone 14，你不会希望脑袋里一直装着 iPhone 1。

所以它将会

能够

展示完全可逆性和可升级性。所以，你可以移除一个设备，用最新版本替换它，或者如果它出于任何原因停止工作，就将其替换。这是 Neuralink 设备的一项基本要求。我还应该说，Saki 和 Pager 都升级到了我们最新、最先进的植入物。所以实际上已经超过1年半了。现在 Pager 已经接受了首次植入，然后又接受了升级版植入，因此这是一个非常好的迹象，表明它可以长期使用，且没有观察到任何不良影响。

我认为，同样重要的是要展示 Sake 实际上喜欢做这个演示，并不是被绑在椅子上之类的。所以就是。是啊，猴子实际上很享受做这些演示，而且它们会得到香蕉奶昔，这有点像一个有趣的游戏。所以我想我要表达的是，我们非常重视动物福利，而且我很确定，比如，我们的猴子相当开心，你知道，所以正如你所看到的，在水果方面做决定很快。

所以，对于我们来说，我们最先打算在人类身上开展的2项应用是恢复视觉。而且我认为这一点很值得注意，因为即使某个人从来没有过视觉，比如他们天生失明，我们相信仍然可以恢复其视觉。因为大脑皮层的视觉部分，视觉部分仍然存在。所以，是的，即使他们以前从未见过，我们也相信他们可以，他们可以看见。

然后，另一项应用是在运动皮层，我们最初会让一个无法，几乎无法控制自己肌肉的人，你知道，有点像某种斯蒂芬·霍金那样的情况，能够比双手功能正常的人更快地操作手机。但接下来，显然，比这更好的是桥接这种连接。所以，从运动皮层获取信号，假设某个人颈部折断了，那么就把这些信号桥接到位于脊髓中的 Neuralink 设备。

所以我们相信，实现全身功能不存在任何物理限制。所以，我的意思是，尽管听起来可能像奇迹一样，但我们相信，让脊髓断裂的人恢复全身功能是可能的。

所以，是啊。

所以，是的。

好的。然后我想再次强调，这次更新的主要目的是招聘。

很多时候，人们认为自己没法真正在 Neuralink 工作，因为他们对生物学或大脑如何运作一无所知。而我们真正想在这里强调的是，你并不需要了解这些，因为当你把让 Neuralink 运作起来所需的技能拆解开来时，实际上其中许多技能与让智能手表或现代手机运作所需的技能相同。所以大致包括，你知道，软件、电池、无线电、感应充电，以及，你知道，还有我们特有的事项，比如动物护理、临床和监管事务，显然，机器学习这个说法被使用了很多。

但我们显然需要解读来自大脑的信号，而大脑是一个生物神经网络。解读生物神经网络的最佳工具就是数字神经网络。

所以这是。如果我有一条信息想要传达，那就是，如果你拥有制造手表、手机、计算机等先进设备的专业知识，那么你的能力将对解决这些重要问题大有帮助。

这比其他任何事情都更是我想传达的信息。所以，瞧，是的，那么就这样，我想请 DJ。

所以。

DJ 是 Neuralink 创始团队的一员，为公司作出了巨大贡献，接下来要发言的许多其他人也是如此。但我只想感谢 DJ 为 Neuralink 作出的巨大贡献，以及。

好。

好的，很好。谢谢。

谢谢，埃隆。

我13岁从韩国搬来时，需要学习一种新语言来交流，我想知道，是否有更好、更有效的方式把我的想法传达给外部世界。

而在《黑客帝国》中看到尼奥学习功夫时，我记得自己当时想，哇，我想努力让那成为可能。

而如今，我相信这是一个可以解决的工程挑战，因为关于你的意图、思想和经历的一切都在你的大脑中，以动作电位放电统计数据的形式编码。如果你能把电极放在正确的位置，并具备适当的感测和刺激能力，那么这以及埃隆谈到的许多其他应用是可能的，而且我们可以帮助很多人。

我无比兴奋能在 Neuralink 投身于这项雄心勃勃却又重要的使命，让那样的未来成为现实。

我也无比荣幸能与一些杰出的同事、科学家以及横跨许多工程学科的工程师合作，共同研究生物学与技术的这一交汇领域。

今天你们会听到其中几位发言，了解我们所面临的各种技术挑战，以及我们过去1年取得的进展。

而且我认为你们会发现，对于其中大多数挑战，正如埃隆所说，你不需要事先了解大脑如何运作，而我们所做的很多事情，都是把工程学第一性原理应用于生物学。

那么，如何创建一个连接大脑的高带宽通用接口？

从第1天起，我们就专注于一套安全、可扩展且能够访问大脑所有区域的基础技术。这3个维度，即安全性、可扩展性和脑区访问能力，真正构成了我们在 Neuralink 设计产品的基础。

安全性，是因为我们想让设备以及安装过程都尽可能安全，从而推动这项技术的普及；可扩展性，是因为随着我们的设备变得更安全、更实用。会有更多人想要它。而且随着规模扩大，我们也希望让它更实惠；还有脑区访问能力，以便我们能够扩展我们技术的功能。

所以，我们的设备沿着这些维度迈出的第1步，就是我们所说的 N1 植入物。

它大约有1枚25美分硬币那么大，拥有超过1000个能够进行记录和刺激的通道。

它是在我们称为“线”的柔性薄膜阵列上微加工而成的。

它可完全植入并且是无线的，所以没有电线。手术后，植入物位于皮肤下方，看不见。

它还有一块可以无线充电的电池，你可以在家中使用它。

同样，为了将我们的设备安全地植入大脑，我们制造了一台手术机器人，称之为 R1 机器人。

它能够操纵这些宽度仅相当于几个红细胞的微小细线，将它们可靠地插入一个正在活动的大脑中，同时避开血管系统。

它非常擅长可靠地完成这项操作。事实上，由于我们从未展示过机器人实际进行从头到尾的完整插入过程，我们将现场演示机器人在我们的大脑替代模型上进行手术。那么，谁想看看插入过程？

它就在这里。那就是我们的 R1 机器人，还有我们的阿尔法患者，它正舒适地躺在患者床上。

这就是我们所说的靶向视图。所以你现在看到的是我们脑部替代模型的图像。粉色代表我们希望将电极插入的皮层表面。黑色代表我们希望避开的血管系统。你看到的这些带有数字的井号标记，代表我们打算放置每根导线的位置。

那么，我们要看看一些植入操作吗？

所以很快再看另一个视图。左边是插入区域的视图。而在右边，机器人将要做的是把阵列，也就是导线，从硅衬底上一根一根剥离，然后，然后插入我们在靶向视图中预先确定的目标位置。

所以。

好了，这就是第一次插入。

所以我们还会再看几次插入。

在我们的第一款产品中，插入大约64根导线的整个过程，这台机器人需要大约15分钟。所以，是的，第二根已经插进去了，我们还要插第三根。

好了。然后它将在后台继续进行，我们会在演示后面的部分再回来看它。

正如埃隆提到的，作为其中的一部分，我们一直在非常努力地推动从原型到产品制造的转变。我们所做的一件事，是将设备制造迁至奥斯汀的一处专用设施，以扩大制造规模。需要强调的一点，而且在这段视频中也很明显，就是让从事设计的工程师也在实体制造线上参与制造和调试，这对我们来说非常常见。

而这对于缩短我们的迭代周期时间一直极其、极其关键。

我们也扩大了手术规模，所以现在我们有一个专用的、我们自己的手术室，实际上在奥斯汀有一个双手术室。而这只是我们最终建造自己的Neuralink诊所之前的一块垫脚石。

所以借助这款产品N1和R1，我们最初的目标是帮助因脊髓完全损伤而瘫痪的人恢复数字自由，使他们能够像受伤前一样好地使用自己的设备，甚至比受伤前更好。

正如埃隆提到的，过去1年里，这一直是公司的核心重点，我们一直与FDA密切合作以获得批准，并希望在未来6个月内在美国启动我们的首次非人类临床试验。

所以希望这能让你们对我们的产品有一个很好的总体了解。

在接下来的1个小时里，我们将深入探讨这些主题的技术细节，向你们介绍我们面临的技术挑战，分享我们的一些进展，并预览接下来将会有什么。

那么接下来交给我团队的NIR，他将向你们介绍神经解码。

谢谢你，DJ

各位。

我叫尼拉万·陈，是脑机接口应用部门的负责人。我们的目标是让瘫痪人士能够像我一样操控计算机。甚至做得更好，我们希望提供快速而准确的控制，具备计算机的全部功能，并且能在任何时间、任何、任何地点使用。所以我非常兴奋地向你们展示，我们如何将N1设备与我们的软件和算法结合使用来实现这一目标。

去年，我们与各位分享了一段猴子佩德罗用大脑控制计算机光标的视频。那么我们是怎么做到的？简单提醒一下。首先，我们记录它运动皮层的神经活动。使用N1设备。

在它使用操纵杆玩游戏时，我们可以从超过数千个通道进行记录。然后，我们可以训练一个神经网络，根据它神经活动的模式预测光标速度。有了这个解码器，它只需想象就能控制光标，甚至不需要移动操纵杆。它可以使用这个解码器玩各种游戏。还包括一个网格任务，它要把白点移向黄色目标。

每当它成功完成一个，就会获得一滴它最喜欢的冰沙。它每天都会选择玩这个游戏。

这里可以看到他在2021年初的表现，大约就是我们发布上一次演示的时候。它相当准确，但比我们希望的速度稍慢一些。而光标控制是与大多数计算机应用程序交互的基础。所以从那以后，我们一直在努力提高光标的速度和准确性。正如你们所见，它快了很多、很多，几乎快了1倍。

不过，它仍然比我能做到的稍微慢一点。所以我们正在研究有创意的方法来改善这一点。

现在，只有速度还不够。你还需要一整套功能。几十年来，大多数软件都是为鼠标和键盘控制而构建的。至少就目前而言，为脑控重新创造整个生态系统是没有意义的。所以我们正在研究和设计供大脑使用的鼠标和键盘界面。我们的做法是训练 Pedro 和他的伙伴们完成各种计算机任务，然后设计预测其行为的算法。

这里可以看到猴子训练不同阶段中的几个任务示例。例如，左键和右键单击、单击并拖动、光标打字、精灵打字、手写，甚至还有手势。

现在，与计算机交互是双向的，反馈非常重要。我喜欢在点击按钮时能够切实感觉到按钮被按下。当潜在的 N1 用户尝试点击时，他们将无法感觉到这一点。我们解决这个问题的一种方式，是提供实时视觉反馈，通过改变光标的颜色来表示神经点击的力度。仅仅是在实体键盘上打字，就比在 iPad 键盘上打字快得多，也容易得多。

这会让脑控的使用变得更快、更容易，打字是最重要的功能之一。所以你们已经看过这条消息了，我想向你们展示这条消息是如何生成的幕后过程。这里你们可以再次看到萨凯使用虚拟键盘。输入这条消息。这个虚拟键盘和我手机上使用的键盘类似。凭借我们迄今实现的速度和准确性，在虚拟键盘上打字已经既快又容易了。

然而，我在电脑上打字时从不用虚拟键盘，因为它会遮住我的屏幕，而且也比我用10根手指能达到的速度慢得多。我们可以做得更好。例如，斯坦福大学的一个团队让一个人想象手写字母。然后，他们根据他的大脑活动解码出了这些字母。采用这种方法，他们得以提高打字速率。我们和猴子一起启动了这个项目，但它们当然不知道怎么写字。

所以为了模拟书写，我们训练了 Ranger，它是我们最喜欢的猴子之一，让它在 iPad 上描摹数字。这里可以看到它描摹数字 5 和数字 2。然后，我们用 N1 设备记录了它的神经活动。但现在，我们不再记录光标速度，而是实时解码它正在屏幕上描摹的数字。

我们从这个项目中得出了2个主要结论。第1，猴子非常了不起，能够学会非常、非常复杂的任务。第2，尽管它可以提高打字速率，但对于我们想要分类的每个数字和字符，它都需要数百个示例和样本。这无法扩展，我们解决这个问题的方式是间接处理。我们不会直接解码数字，而是先解码屏幕上的手部轨迹。

然后，当我们解码出手部轨迹后，就可以使用任何现成的手写分类器来、来预测数字和字符。例如，在 MLIST 数据集上训练的经典类。

为什么这如此重要，它很重要是因为现在我们有可能只用1个神经解码器，就解码任何语言中的任何字符。对于手部轨迹而言，这意味着你可以用英语、希伯来语、普通话，甚至猴语书写。而我们可以理解你想要一根香蕉？

所以，要改进功能和速度，我们前方还有许多挑战。我想把话筒交给 Bliss，请他谈谈第3部分，也就是我们如何让脑机接口随时随地都能工作。

大家好，我叫 Bliss，是 Neuralink 的一名软件工程师。当我使用电脑时，我的鼠标和键盘会按我的意图工作，至少在大约 99.9999% 的时间里是这样。我的目标是让瘫痪用户能够像我一样可靠地控制自己的电脑。下面就是我们希望这种体验呈现出的感觉。在这段视频中，你们可以看到 Saki 走到他的 MacBook 前，并选择进行他的打字任务。

整个解码系统开箱即用，感觉完全即插即用。实现这种高可靠性的第1步，是进行广泛的离线测试。使用 N1 Link 的典型流程是通过蓝牙连接，传出大脑的神经活动，然后利用这些神经活动训练解码器并进行实时推断。我们已经针对这一确切流程构建了一个模拟。但不是使用一只植入了设备的猴子。

我们使用一个模拟大脑，将合成的神经活动注入一个放置在服务器机架中的植入物。从该植入物的角度来看，它身处一个真实的大脑中。每次提交代码时都会运行这一模拟，以验证从硬件一直到神经解码器，我们的整个技术栈都能达到最先进的性能。然而，尽管这种模拟非常适合对软件和硬件进行集成测试，但它还不够精细，无法保证在现实世界中的高可靠性。

在现实世界中，我们试图解码的底层信号实际上每天都在变化。在这张图中，你可以看到在 psaki 的植入物一个代表性通道上检测到的平均放电率。每根柱子代表1天。你可以看到，每天的平均放电率都与前1天不同。这给我们带来了一个非常有意思的问题：如何让我们的解码器每天都保持稳健。实际情况可能是，如果你用某1天的数据训练神经解码器，然后尝试在第2天使用它，平均放电率可能会发生足够大的偏移，从而导致模型输出产生偏差。

在右侧，你可以看到这种偏差使光标很难移动到右上角。你可以看到，它在这里艰难地试图到达右上方，然后却轻松得多地向下移动到了左下角。

我们正在尝试许多方法来缓解这个问题。一些例子包括使用包含许多天数据的大型数据集来构建模型，以尝试寻找在不同日期都保持稳定的神经活动模式。我们正在尝试的另一种方法，是持续采样植入物上的神经活动统计数据，并在将数据输入模型之前，使用最新的估算结果对数据进行预处理。这确实是团队正在积极研究的领域；如果我们想让瘫痪人士能够像我一样操控计算机，这就是一个必须解决的关键问题。

我们面临的另一个大问题，是尽量缩短大脑中的一个尖峰影响屏幕上光标移动所需的时间。如果这个控制回路中存在延迟或抖动，光标就会变得难以控制，导致你在右侧看到的这类过冲。

我们在这个方向上取得的一项重大改进叫作相位锁定。相位锁定会将我们从植入物发出的每个数据包的边沿，与蓝牙无线电即将唤醒的确切时刻对齐。这最大限度地缩短了将大脑中的一个尖峰纳入我们神经网络预测所需的时间。在这里，你可以看到采用相位锁定后的延迟分布。不仅均值大幅降低，方差也降低了。

这使用户更容易预测其光标的行为。

在过去1年里，我们大幅提升了系统的稳定性和可靠性，并且已经能够在许多次会话和许多个月中展现出持续一致的高性能。然而，在这个系统真正达到即插即用之前，我们还有很长的路要走。所以，如果解决推出这项技术所需的难题令你感到兴奋，你应该考虑申请加入团队。

现在我要把时间交给阿维纳什，让他来来谈谈我们的定制低功耗 ASIC 如何检测大脑中的尖峰。

大家好，我是阿维纳什，是 ASIC 团队的一名工程师。我们设计了定制神经传感器，其中同时包含模拟和数字电路，可在1024个独立通道上进行记录和刺激。我们在性能、功耗和面积这3个主要方面都面临挑战。我们不仅必须把全部1024个通道装进一个仅有25美分硬币大小的植入物中，还必须测量振幅小于20微伏的尖峰活动。

今天，我想重点谈谈我提到的最后一个挑战。功耗。

功耗对我们很重要，因为我们希望未来的用户能全天使用植入物，不会因充电而受到任何中断。早在2018年，我们会将每个通道的每个样本都从设备发送出去处理，这消耗了大量电力。2020年，我们把尖峰检测搬到了芯片上。你可能知道，神经元通过放电传输信息，因此，只监测这些尖峰，并仅将这些尖峰事件从植入物发送出去，就是一种非常高效的压缩形式。

在过去2年里，我们继续在 ASIC 内部进行优化，将系统总功耗降至仅32毫瓦，并使电池续航时间翻倍。

让我们看看使我们的电池供电植入物成为可能的片上尖峰检测算法。首先，我们应用一个500Hz至5kHz的带通滤波器，去除频带之外的噪声。

接下来，我们使用对本底噪声的估算，为每个通道生成一个自适应阈值。最后，我们的尖峰检测器模块会识别一个尖峰的3个关键点。识别3个点，使我们不仅能够检测尖峰是否存在，还能检测尖峰的形状。这对于区分邻近单个通道的多个神经元可能极其重要。今天，我想重点介绍我们在最新芯片中所做的众多优化之一，这项优化具体将系统功耗降低了15%。

请注意，神经元产生尖峰的频率相对较低，这意味着我们的尖峰检测器会花大量时间寻找尖峰的第1个点，而只花很少时间寻找尖峰的另外2个点，因为它们只会在阈值被正确跨过后出现。我们可以利用输入波形的这一特征，将芯片内的内存访问减少30%。让我们看看它是如何运作的。我们的尖峰检测器被实现为由所有通道共享的单个功能单元，并配有 SRAM 来缓冲每个通道的状态。

当一个样本传入时，会从 SRAM 读取其通道状态，运行增量式尖峰尖峰检测步骤，然后将更新后的状态写回 SRAM。由于整个植入物每秒会进行2000万次这样的操作，每一次访问都会很快累积起来。在我们的最新芯片中，我们将状态分为2部分：热状态和冷状态。热状态在每个周期都会被访问，而冷状态只有在跨过阈值后才会被访问，从而减小平均访问宽度并节省功耗。

我们还在研发一款以刺激为重点的下一代芯片，它拥有4096个通道，但仍保持在我们当前芯片的占用面积内。除了增加通道数量之外，我们还在提高驱动电压，以便每个通道都能实现更好的激活效果。为了支持这一更高的通道数量，以及你们很快就会听到的广泛未来应用，我们正在芯片上加入一个 ARM 内核。

最后，由于这些芯片与我们当前的芯片尺寸相同，我们仍然可以把其中4块组合到单个植入物中，总计达到16,000个通道。尺寸仍然只有25美分硬币大小。

如你所见，我们一直在非常努力地改善植入物内部的功耗。但我们也一直在非常努力地改善植入物的充电体验，马特将对此进行介绍。不过首先，机器人刚刚完成了全部64根线的植入，所以让我们看看。

这是植入部位的视图，与 DJ 之前向你们展示的类似。但这里没有瞄准标线；如果仔细观察，你们可以看到，全部 64 根细线，每根携带 16 个电极，都已插入大脑替代体，同时避开了血管。而这一切都只是在过去 20 分钟内完成的。

现在让我们把时间交给 Matt

继续进行技术深度解析。

大家好，我是 Matt，脑机接口电气工程负责人。我们的全植入式 N1 设备依靠电池持续运行。当电池电量不足时，通过无线电力传输进行充电。

然而，与许多可以直接提供物理连接器、充电的消费电子设备不同，为全植入式设备充电会带来若干独特挑战。首先，系统必须能在较大的充电空间内工作，而不依赖磁铁实现完美对准。系统必须能够抵御干扰并迅速完成充电，以免造成过重负担。然而，最重要的是与脑组织接触时的安全性。

植入物外表面的温升不得超过 2 摄氏度。

为了实现这些目标，我们的充电系统经历了数次工程迭代。首先，如果你们看过我们在 2020 年 8 月进行的猪演示，Gertrude 植入的是一个使用我们第一代充电器充电的 N1 版本。该设备采用小型圆盘状封装实现，后来被拆分为远程线圈和电池底座。这款充电器使用起来很有挑战性。不过，我们通过它的实施学到了很多。

我们目前的量产充电器用于为当前这一代植入物充电，它采用铝制电池底座实现，其中还包含驱动电路。

一个尺寸为我们原始设备 4 倍的远程线圈。也可断开连接，这个远程线圈提高了开关频率，从而改善了线圈耦合。

这款充电器如今已投入使用，包括在我们的工程和动物测试设施内的多个应用中。我想在这里向大家展示其中一个应用。使用一种我们称为简易充电器的设备，并将线圈嵌入栖息环境中。通过增加一个新的外部控制回路，再加上一台香蕉奶昔泵，这群猴子已经接受训练，能够自行充电。

那么，让我们看看 Pager 如何给他的植入物充电。

在右侧，我们正在从 Pager 的 N1 实时传输诊断数据。当他爬上去并坐在线圈下方时，你可以看到充电器自动检测到他的存在，并从搜索切换到充电、充电。我们可以看到以 0 到 1 为刻度的稳压功率输出，以及输入这块电池的电流。

我之前提到过，我们改进了线圈耦合。然而，高品质因数线圈在相对较远的距离上表现出良好的充电性能。但当它们更靠近植入物时，你会看到一种峰值分裂效应，最佳、最高效率的功率传输被推向更高的频率，超出了为符合受监管的辐射发射要求而规定的 ISM 频段。在我们的下一代充电器中。

我们通过引入右侧所示的动态调谐来解决这个问题。这使我们能够实时调整发射和接收线圈的谐振频率，从而在性能下降前及时改变它们的特性。

电气工程团队目前正在开发第三代特许充电器。显著改进包括双向近场通信。这使我们能够降低控制延迟并改善热调节。

改善热调节会缩短充电时间。现在，朱利安将向我们介绍如何测试 N1。

非常感谢你，马特。我叫朱利安，负责脑机接口团队的嵌入式软件组。我们刚开始制造植入物时，只有一条小型生产线；要从植入物中收集数据，你需要拿着笔记本电脑手动走过去，连接设备并收集所需的数据。但我们的目标是制造极其安全、极其可靠的植入物。因此，为了做到这一点，我们扩大了生产线以及测试、吞吐量和数据收集能力。

首先，我们在生产线上增加了一整套验收测试。这些测试会检验每个组件及最终组装件的功能。下线的植入物随后会接受台架测试、加速寿命测试和动物模型测试。之后，我们全天候收集这些植入物的数据。这些数据由一系列云端工作进程处理，并以汇总形式显示。最后，所有这些信息都会反馈到我们的设计流程中，使我们的工程师能够随时回答有关任何植入物的任何问题。

现在，我将带大家了解这套基础设施的不同部分，先从固件测试开始。植入物包含一个运行固件的小型微处理器，用于管理其大量操作。在发布固件更新之前，我们希望通过单元测试和硬件在环测试对其进行严格测试，后者也称为 HIL 测试。要进行 HIL 测试，你需要为电池、电源轨和微处理器接入检测仪器，然后我们用蓝牙客户端连接每台设备，再让设备经历各种场景，以测试功耗、实时性能、安全系统、故障恢复机制以及许多不同的事项。

在这些系统的最初实现中，我们使用现成组件，以便迅速开始自动化测试。然而，这些系统是以一种相对手工作坊式的方式构建的，非常难以维护。这意味着测试很快成为开发的瓶颈。为了缓解这一问题，硬件和软件团队开发了一套新系统，将所有必需组件集成到一块基板上。

然后，我们可以将充电器和植入物硬件分别放在插接到这块基板的独立模块上，其中包括一块带有相对线圈的板卡，以便测试充电性能。这种架构让我们能够快速迭代不同的硬件原型，因为只需将它们放入系统，就能复用所有测试基础设施。此外，我们还可以将当前和下一代神经 ASIC 部署到 FPGA 上，再将它们也插接到这块板上。

这使我们能够测试一个完整的额外层面。这就是我们如何生成右侧这幅颇具开创性的图像。你所看到的是由我们的一些模拟神经传感器发出的尖峰活动，经由整个系统通过蓝牙进行流式传输，然后显示在手机上。这使我们能够在一个系统中测试从芯片到云端的一切。

这套系统的成本是原来的 1/5，体积是原来的 1/5，而且非常容易制造。这让每位开发人员都能在自己的桌上配备一台个人设备，也让我们能够将整套测试套件分片到大量安装在机架中的此类设备上。所有这些都大幅加快了我们的开发速度。

接下来看看我们如何监测植入物的电子器件、电池和外壳。植入物会定期采集其所有生命体征，并将其写入闪存。随后，当它下次连接到我们的某个记录站时，就会将这些数据传输出去。例如，通过查看湿度，我们可以了解植入物外壳的完整性。通过查看电池电压和功率测量值，我们可以评估电池健康状况。

所有这一切都自动完成，无需任何干预，让我们能够 24,7 掌握每一台设备的质量。此外，我们还可以利用这套基础设施按需请求高保真信息，以调查不同的异常情况。例如，在这个特定场景中，我们正试图追查在不同通道上观察到的一些虚假尖峰的来源。

所以，我们直接从那些通道请求了原始波形样本。

采集高质量神经信号需要完好无损的低阻抗电极。因此，这也是我们利用神经传感器上的专用电路密切监测的一项内容。那么，我们是怎么做的？首先使用板载 DAC 在单个通道上播放测试音。然后我们使用 ADC 同时进行记录，同时记录该通道以及物理上相邻通道的响应信号。

这样，我们不仅可以测量每个通道的阻抗，还可以将不同的物理现象映射到不同的特征签名。例如，开路通道会、会在该通道上表现为非常大的响应，而较短的通道会在相邻通道上表现为较大的响应。通过观察返回信号的纯度，我们还可以验证神经传感器自身的模拟前端是否正常工作。

在我们最初进行这些阻抗扫描的实现中，完成全部1,000个通道需要4小时。但通过使测试瘫痪、降低、采样、过滤，然后通过将大量计算移至固件端，减少我们必须从设备传出的信息量，我们现在只需20秒就能扫描全部1,000个通道。这意味着我们可以每天对每个植入物运行阻抗测试。

然后，我们的内部仪表板可以回放这段阻抗历史记录，使我们能够对生物与电子器件之间的接口获得非常充分的定量认识。

现在你们已经了解我们如何测试和监控植入物，我要把话交给乔希，他会告诉你们，我们如何通过加速植入物直至失效来更快地获得反馈。

大家好，我叫约书亚·赫斯，是脑接口团队的一名工程师。我们负责植入系统设计，以及许多制造和测试工具。朱利安刚才向你们稍微介绍了一些我们测试植入物电子器件、硬件和软件的方法。但就整个系统与其在组织中的使用寿命而言，又该如何呢？我们解决这个问题的方法之一，是开发内部加速寿命测试系统。

这个这个系统使我们能够加快并大规模捕捉长期植入物失效模式，从而迅速提高我们的迭代速度。更好的是，该系统还显著减少了需要动物模型的测试数量，这既适用于植入物原型，当然也适用于寿命测试。那么，这个系统如何运作？从非常基础的层面来看，归结为3件事。首先，我们要模拟组织内部的化学环境。

接下来，我们要加速这些化学相互作用，以及与我们的植入材料之间的扩散。最后，我们要对植入物的内部电子器件进行高强度循环。借助这些措施，主要是前2项，我们依据阿伦尼乌斯关系实现了保守的4x加速系数。换句话说，我们的植入物在加速系统中度过的每1天，相当于在体内度过至少4天。

一直以来，我们面临的最大挑战之一，就是对抗水分侵入植入物。因此，我们持续监测内部湿度，以观察异常上升。这里以白色显示的是植入物和我们一些动物体内超过1年期间的部分内部湿度数据。正如你们所见，我们的内部湿度传感非常灵敏，甚至能够检测到仅因水分通过植入材料扩散而产生的非常微小且缓慢的湿度上升。

现在，蓝色部分显示的是相同的内部湿度数据，但这些数据来自我们加速系统中的设备。

现在，如果我们根据加速系数调整这些数据，你们不仅能开始看到这些数据的一致性，还能看到这些数据延伸到了多远的未来。

现在，红色部分显示的是一个在我们的加速系统中发生故障的设备。在植入物电子器件发生故障之前的许多个月里，这个设备显示出湿度异常上升。

那么，我们如何构建这个系统呢？

嗯，我们是在新冠疫情封锁于2020年初开始后不久着手构建第一个系统原型的。因此，我们不得不发挥一点创意。如你们所见，我们的第一个系统原型有些简陋，而且是在我们其中一人的公寓里运行的，从地毯就能看出来。虽然简陋，但这个系统让我们能够通过最快的途径开始测试设备、调整工作流体的化学组成，并检查我们的约束条件。

我们还立即开始查找早期植入物原型中所观察到故障的根本原因，将这些信息反馈到下一版原型设计中，并且真的在短短几个月里不断冲洗和重复这一过程。这个系统完全是定制构建的，并经过高度迭代，通过2个系统版本和无数次小幅迭代，最终形成了我们目前运行的第3代系统。该系统能够进行高密度测试，并具备容器内自动充电和自动数据收集功能。

该系统还配有植入物托架组件，可以容纳大脑替代材料，使植入物能够由手术机器人安装并插入，就像你们几分钟前看到的那样。

我们还将该系统集成为高密度机架式形态，并配备了集中式流体管理系统，既能确保各容器之间的化学一致性，也能减少运行维护。该系统已运行了过去1年半，也遇到了不少挑战。由于系统本身承受着与其中植入物相同的加速摧残，要设计、构建并维护如此规模的系统，同时使其即便面对自身造成的影响也能保持稳健，一直极具挑战性。

那么，接下来是什么？

嗯，我们已经开始研发第4代系统，并从头对其进行了全面重新设计，采用每个容器容纳单个植入物且支持热插拔的设计，部分灵感来自高密度计算服务器。借助这一新系统，我们将在密度、稳健性和规模方面达到全新的水平。我们还计划让许多这样的系统投入运行，以求捕捉即使发生频率最低的边缘案例故障模式。

有了它，我们将让数千个植入物接受测试，以实现这些目标。

我们已经开始构建这个系统，但仍有大量工作要做。前方还有许多令人兴奋的挑战，例如引入机械应力、大脑替代材料的微运动，以及用EFIN复制电极丝周围的组织生长，从而进行更完整、更具代表性的加速测试。那么，现在你们已经听过我们在生产用于手术的植入物设计之前，对其进行严格测试的一些方式。

现在，克里斯蒂娜将带大家详细了解我们的手术流程。

谢谢，乔什。大家好。我是克里斯蒂娜，手术工程团队负责人。

要植入N1设备，基本上要经过这些步骤。确定目标位置并切开切口，钻出颅骨开口，移除称为硬脑膜的坚韧外层脑膜，然后插入细而柔韧的电极丝，将植入物放进我们打出的孔中，然后就完成了。你的皮下就有了一个植入物。看，妈妈，没有电线。开玩笑的。我的意思是，说真的，没有电线。但实际上我并没有植入一个。

手术机器人负责手术中的电极丝插入部分。这是因为手动操作会非常困难。想象一下，从头上取下一根头发，试着把它插进覆盖着保鲜膜的果冻里，同时还要达到精确的深度和位置，并且要在合理的时间内这样做64次。而且不。如果我们要求神经外科医生在手术中这样做，他们可能不会很喜欢。

所以，我们有了你们刚才看到的那个跳着小舞的机器人。我其实有点想叫它“小舞者”，但它叫R1，这个名字也很棒。手术的其余部分由神经外科医生完成。为了让这项手术易于获得且价格可负担，我们需要重新审视这一点。我来告诉你们为什么。

我上学时，父亲失去了行走能力，也无法使用双臂，甚至无法说话。他被诊断患有ALS。我们会在互联网上查看，你可能会看到这里或那里有某个人拥有某种很酷的定制机器人辅助设备。但这让人极度沮丧。可供他选择的方案是多么有限？而且有数十万人患有轻瘫，这甚至还不包括患有其他疾病、也可能从我们的设备中获益的人。

与此同时，神经外科医生并没有那么多，每100万人可能只有大约10名。培养神经外科医生需要大约10年或更久，而且他们通常已经非常忙碌了。你们可以想象，他们的时间非常昂贵。因此，为了让我们尽可能发挥最大的作用，并使这项手术价格可负担且易于获得。手术。我们需要弄清楚如何让1名神经外科医生同时监督多台手术。

这听起来可能有点疯狂，但在Lasik让激光眼科手术变得寻常之前，它可能听起来也很疯狂。Lasik已经存在了大约30年，而且还会继续下去。一开始，激光机器人只完成它必须完成的、最基础的核心部分，其余部分由外科医生完成。而且，而且随着一次次迭代，外科医生需要做的越来越少，激光机器人则完成其中的大部分。这是一项极具吸引力的手术，只需短短几分钟，而且往往能带来改变人生的结果。

自从我在2017年加入以来，我们也进行了少量迭代，以优化机器人的电极丝插入。其中一项我们必须面对的挑战与光机封装有关。所以，如你们在这里看到的，大约有3条主要光路，对我们实现可靠的电极丝插入非常有价值。其中一条用于对针插入电极丝的过程进行可见光成像。另一条则是名为octical相干断层扫描的激光干涉测量系统。

它能在大脑实时移动时向我们提供其精确位置。此外，我们还必须提供照明和光照，以便看清可见光摄像头中正在发生什么。要完成这一切，而针又位于颅骨开口的底部，特别是当它靠近颅骨壁时，要把所有部件都装进去并且还能看清，可能相当困难。因此，团队的解决办法是利用光子魔法，或者说偏振，随你怎么称呼，把所有这3条光路都放进同一个光学堆栈中。

这使我们能够实时避开血管。正如我提到的，大脑在移动，而我们一开始设定靶点的位置，可能并不是针头下降到那里时你想要插入的位置。因此，机器人实际上可以检测血管，然后判断下针位置是否会落在血管上，以及插入是否安全。这样一来，我们就能避免在主要血管上下针。接下来就是我们今天带到这里的机器人。

要实现那种减少神经外科医生参与、并让手术费用可负担且容易获得的操作，我们仍有很多工作要做。手术中对神经外科医生技能要求最高的主要的两个环节是颅骨切除术和硬脑膜切除术。亚历克斯和萨姆将进一步介绍我们认为可以如何取消硬脑膜切除这一步。这样就只剩下颅骨切除术了。

在神经外科手术中，如果颅骨切除范围足够小，就可以使用一种叫作穿孔器的标准工具，它能迅速完成这种形状的作业。但对于范围更大的颅骨切除术，外科医生必须依靠自身技能来应对不同患者之间颅骨厚度和颅骨硬度的差异；即使是同一位患者的同一次颅骨切除术，颅骨厚度也可能有所不同，例如。

此外，如果我们能制造出一种能够进行极高精度颅骨切除的设备，就能为未来将植入物固定到颅骨上的方式拓展设计空间。那么。那么我来展示几个我们的原型。屏幕上这种超声切割器和振荡切割器的优点是不会切割软组织。你可以切开骨头而不会切到大脑；但不过，正如你们在这里看到的，我们的超声切割器原型为了达到我们想要的切割速度，产生了相当多的热量。

那么再来看这里的振荡锯，我们设计了一种刀片，以尽量缩短切割时间，同时还传导声音，也会发热。正如你们所见，它可以切穿骨头之类的硬物，却不会切穿皮肤之类的软物。它很简单，而且有效。然而，如果你想切到任意深度或切出任意形状，振荡锯就是切不了。

我还担心没人能听懂。你们很聪明。那么，对于钻出任意形状，有一种经过时间检验的解决方案，那就是 CNC 钻机。我们在人身上这样做所面临的挑战是，需要确保它每一次都能可靠地切割，而且不会切得太深。我们用来确保不会切穿大脑的几种反馈方式包括力反馈以及阻抗。如果我能请一位志愿者。

不，开玩笑的。也许下次吧。不过，是的。这让大家得以了解我们正在开展的一些工作，以实现一种便捷且负担得起的手术。现在，亚历克斯将为大家介绍一下我们的下一代开发项目。

谢谢，克里斯蒂娜。我是亚历克斯。我是机器人团队的一名机械工程师。既然我们已经介绍了当前设备的技术和手术流程，接下来想介绍一些下一代开发项目。我和接下来的几位演讲者想谈谈其中一个项目，也就是实现设备的可升级性。

大家已经听到了我们在过去1年取得的进展。我们提高了植入物的稳健性、电池和充电性能，以及蓝牙的易用性。实际上，每一个新版本的设备都会显著改善。它的功能会更强，使用寿命也会更长。我们需要确保这些新技术仍能为早期采用者所用。这意味着，我们需要一种解决方案，让设备升级或更换与最初安装一样容易；正如许多医疗器械公司所发现的那样，这是一个很有挑战性的问题。

身体的愈合反应并不会让这件事变得容易。所以这个问题还没有解决。但在实现这一目标方面，我们已经取得了显著进展，今天想向大家介绍一下。

现在，我们必须先介绍一些背景，说明是什么让设备升级如此困难。我们先从解剖结构讲起。皮肤下面是颅骨。再往下是硬脑膜，这是一层将骨骼与大脑分隔开的坚韧薄膜。在硬脑膜与大脑之间，则是软脑膜蛛网膜复合体，这是一个充满液体、用于悬托大脑的结构。为了安装设备，外科医生会移除一块圆盘状的颅骨和硬脑膜，以暴露大脑表面。

随后，设备会取代被移除的材料。挑战就出在这个界面上。几个月后，所有空余空间都会被组织填满，将设备和丝线包裹起来。

由于丝线尺寸很小，移除设备本来会轻而易举，它们会直接从大脑中滑出来。正是表面上方形成的组织层让移除变得困难。我们在内部制造了工具，用于研究并表征这种反应，例如组织学和微型 CT。在这些图像中，可以看到大脑表面上方已经形成的那层组织，它包裹着丝线，并与周围组织粘连。

我们探索了许多不同的途径，希望在设计上应对这一愈合过程，并找到一种让设备升级无缝进行的解决方案。

我们最成功的做法，是降低手术的侵入性。我们不再直接暴露大脑表面，而是将硬脑膜保留在原位，维持身体天然的保护屏障。这可以防止大脑表面出现包裹现象。实际上，这对简化手术并提高其安全性来说是一个巨大的胜利。不过，正如克里斯蒂娜所暗示的，这并非没有代价。硬脑膜是一层非常坚韧且不透明的薄膜。

正如大家在这些扫描电子显微镜图像中所见。它由致密的的胶原纤维网络构成。这给我们插入电极带来了一系列技术挑战。

其中一个挑战是透过硬脑膜成像。正如左侧所示，我们目前的定制光学系统在对暴露的大脑表面进行成像时，具备相当惊人的能力。然而，正如右侧所示，一旦硬脑膜处于原位，就看不到大脑表面密集的血管系统了。泥土挡住了视线。衰减实在太严重。为了解决这个问题，我们正在开发一种新的光学系统，使用医疗标准的荧光染料对组织下方的血管进行成像。

这里可以看到染料灌注通过血管，使血管凸显出来。要证明这一系统的准确性和可重复性，仍有大量工程工作要做，但完成之后，我们就能瞄准并避开硬脑膜下方的血管。

我们还在探索将激光成像系统应用于更深层的组织结构。在左下方，可以看到硬脑膜下方各组织层的一个切面。这幅图像由我们的光学相干断层扫描系统采集的多个体积数据汇编而成。可以在上方看到这些体积数据的拼接图。未来，将这些新系统与术前成像（例如 MRI）的配准相结合后，就能在不直接暴露大脑表面的情况下实现精确定位。

现在，成像并不是坚韧的硬脑膜解剖结构所带来的唯一挑战。接下来，我想把话筒交给萨姆，请他谈谈将电极穿过这层薄膜插入时面临的一些挑战。

谢谢，亚历克斯。大家好，我是萨姆，负责针具制造和设计团队。正如亚历克斯所说，硬脑膜之所以能够很好地保护大脑的那些特性，也让我们很难将丝线插入其中。在人体中，硬脑膜的厚度可能超过1毫米，听起来似乎不算多，但与我们的40微米针相比，其实已经很厚了。例如，如果把针放大到铅笔大小，那么硬脑膜的厚度就会相应放大到超过4英寸。

看看需要放大多少倍才能看见它。当针的特征进入画面时，在同一个画面里已经可以看到单个红细胞了。

这，这只是。

等等。

这是我们最新设计的一张真实扫描电子显微镜图像。在左边，可以看到丝线的末端。中间是针，右边实际上是我的一根头发。所以，是的，它极其微小。除了非常小之外，它的设计还面临许多其他挑战。其中一个挑战是，我们必须使用针以及容纳针的保护套管来抓住丝线，并在将它从这层保护性硅背衬上剥离时将其固定住。

然后，我们还必须继续夹持它，把它移到表面，并在插入过程中将它从套管中释放。另一个挑战是，坚韧的硬脑膜下方的大脑非常柔软。因此，如果针不够锋利，它只会不断压出表面凹陷，却无法刺穿。如果这段自由长度变得过长，实际上就会像这样直接让针屈曲。另一个挑战是，我们不仅要让针穿过去，还必须让丝线也穿过去。

所以，我们确实必须专注于优化针和丝线组合在一起后的轮廓。这些只是设计这类东西所面临的一部分挑战。到目前为止，我们发现，解决这个问题的关键在于提高我们的迭代速度。但我们先来看看这些东西最初是如何制造的。我们从一段直径40微米的钨丝开始，其中加入少量镭制成合金，以提高延展性。

我们在内部设计了这台飞秒激光铣削机，用来切割针和套管的结构特征，它能以亚微米精度完成这项工作。今年，我们花了大量时间，把这东西从一个科学项目变成了一套工业系统。

就在几个月前，熟练操作员制作一根针需要22分钟，而且即使是熟练操作员，也只能达到约58%的良率。如今，同样的流程只需要6分钟，任何人只需接受几分钟培训，就能达到91%的良率。只需点击一下，铣削机就会切割并测量针和套管，然后把测量数据上传到我们的lim系统，这样机器人就能使用其所用每根针的精确尺寸。

不过，这些全都是针对我们当前的设计，而且我们已经用了几年时间来优化它的制造流程。到目前为止，当前设计一直很好地满足了我们的需求，但它对丝线的保护还不够充分，无法穿过那簇东西。所以，就像我说的，我们必须想出一种新方案，而且需要能够快速迭代设计。

不出所料，《机械手册》中没有关于这种东西的页面。因此，我们深入研究了飞秒激光烧蚀的科学原理，并摸索出一套工作流程，使我们能够像使用CNC铣床一样使用我们的激光铣削机。这让我们能够进行数次迭代。这让我们可以在不到1小时内完成新设计的迭代，当我们真正进入状态时，每天可以迭代好几次。因此，右侧仍能看到的最新设计实际上可以穿过9层Durock，在台面上总厚度为3毫米。

这远远超过了我们在人类身上可能预期的程度，并且留有很大的余量。

不过，针并不是这道难题的唯一部分。可以想象，这里的所有这些设计都要配合不同的丝线使用。所以，我们也需要一种对丝线进行迭代的方法。我们通过在内部进行微加工来做到这一点。

今年夏天，我们用了大约9周时间彻底重建了洁净室，这项工作带来的变化包括大幅降低颗粒物数量，从而使良率和产能能够大幅提升。再加上微加工团队取得的所有其他重大改进，我们只需几天时间就能迭代新设计。

不过，拼图的最后一块是测试。我们可以提出任意多种新设计，但除非有办法在适当的条件下实际测试它们，否则我们就不知道该调整什么。更糟的是，我们会花时间针对错误的东西进行优化。以这种故障模式为例。几个月前，我们已经能够相当可靠地穿过硬脑膜进行植入。但当我们拿这些替代模型进行微型 CT 成像时，我们意识到，我们对细丝末端的夹持力其实太强了，正把它们从棒状表面下方稍微拉出来一点。

等我们解决这个问题时，我们意识到，这个问题对周围材料或组织的特性非常敏感。我们可以制作一个从不发生这种情况的替代模型，也可以制作另一个每次都会发生这种情况的替代模型。这凸显了为什么我们必须花时间，让台架测试尽可能准确地匹配组织。现在我要把话筒交给莱斯利，她将介绍我们一直是如何做这件事的。

谢谢。

大家好，我是莱斯利，负责微制造研发。我们感兴趣的部分工作，是了解我们的植入物和细丝在完全植入体内后所处的生物环境。然而，直接从生物学中学习本质上很慢。因此，为了快速推进，我们正在开发模拟生物环境的合成材料。这让我们能够尽可能多地在台架上获取认识，并开始逐步摆脱动物测试这一行业标准。

不过，开发准确的替代模型很有挑战性。植入环境由许多解剖层次构成，每一层都有独特的特性。随着时间推移和植入部位愈合，新组织会形成并填满任何可用空间。此外，与心血管活动和头部运动相关的运动也增加了复杂性。

因此，为了开始应对其中一些挑战，我们正在利用来自生物学的反馈来设计材料。这可能涉及组织的力学表征，或分析细丝与组织界面处的相互作用。其中许多表征工作甚至是在手术过程中完成的，方法是使用定制硬件和软件改造我们的手术机器人，让它兼作一种灵敏的表征工具。

然后，我们使用收集到的数据，并将其反馈到材料优化中，使这些材料在力学、化学以及如此处所示的结构方面，都表现得与生物组织一模一样。

从这里展示的、放在一个盘子上并由琼脂和一张封口膜构成的简陋初代大脑替代模型开始，我们已经取得了长足进步。虽然它很简单，却让我们得以通过无数次台架车间测试来完善机器人植入。

如今，我们的替代模型稍微复杂了一些，我们升级为基于复合水凝胶的大脑替代模型，能更好地模拟真实人脑的模量。我们还加入了硬脑膜替代模型，并开发出一种可注射的软组织替代模型，迄今为止，它使我们能够进行台架模拟 X 植物测试。我们对未来替代模型有一份非常长的愿望清单，其中一些项目包括集成软组织、大脑、骨骼、皮肤，甚至整个身体的手术替代模型。

一种能够模拟运动、血管活动和电生理活动的大脑替代模型，以及一种用于测试生物相容性和电刺激的生物替代模型。

有大量工作正在进行，让我们更接近未来的替代模型，其中包括如此处所示的实验室培养大脑类器官研究。所有这些都将让我们更接近这样一个未来：我们能在台架上了解更多、迭代得更快，并减少对动物模型的依赖，甚至有一天完全取代它们。说到这里，我要把话筒交给丹，他将介绍一项非常令人兴奋的下一代应用。

谢谢。

谢谢你，莱斯利。我叫丹，在从事视觉神经科学研究的职业生涯之后，我来到 Neuralink 工作。

研究。

我受到启发加入这家公司，是因为我在我们的设备中看到了让因眼部损伤或疾病而失明的人恢复视力的潜力。我们的设备有许多特别的特性，使其特别适合这项应用。首先，除了能够记录每一个通道之外，我们还可以通过每一个通道注入电流，刺激大脑中的神经活动。这一点很重要，因为它让我们能够绕过眼睛，直接在大脑中生成视觉图像。

其次，对于视觉假体而言，我们的设备可以拥有数量极其庞大的电极。这一点很重要，因为电极越多，能在大脑中生成的图像密度就越高。第三，得益于我们的机器人，我们可以将这些电极深深植入大脑。对视觉假体而言，这一点很重要，因为人类视觉皮层深埋在大脑内侧面一个名为距状沟的褶皱中。

在这张图像中，我在一幅 MRI 图像上用红色标出了距状沟。它包含一幅视觉世界、视野的映射图。其表面积大约等于每侧一张信用卡。如果将它展开并铺平，你会看到图像是倒置的，是上下颠倒的；但更有意思的是，它发生了扭曲，使视野的中央部分，也就是注视点，被大幅放大。例如，如果你看这张林肯的图像，如果你直视他的右眼，那么注视点左侧的一切都会被传送到你的右侧视觉皮层，右侧的一切则会进入你的左侧视觉皮层。

他的眼睛虽然在图像中非常小，但在大脑中被放大到几乎占据视觉皮层表面积的四分之一。

在过去半个世纪里，视觉神经科学家对大脑中的视觉处理形成了深刻的理解。推动这项研究的主要方式，是记录皮层中单个细胞的活动，通常是猕猴的皮层细胞。其中一项奠基性发现是，视觉皮层中的每个细胞只代表视野中很小的一部分。你的知觉由微小感受野组成的镶嵌图案构成，每个感受野都属于你视觉皮层中的一个细胞。

因此，如果你记录猴子体内其中一个细胞的活动，比如这个位置的细胞，就能找到屏幕上一个非常小的区域，在那里施加光刺激会引起该神经元的调制。视觉皮层中的另一个位置会对应屏幕上的其他位置。在这个例子中，是下方视野。这些区域被称为感受野。

我们已将设备植入2只名叫 Code 和 Dash 的恒河猴的视觉皮层。

这意味着，当它们四处活动时，我们可以记录由其正常居家环境在视觉皮层中引发的活动。但众所周知，猴子喜欢香蕉奶昔。这意味着我们可以轻松教会它们注视屏幕上的点，并给予它们奖励。我们可以非常精确地奖励它们，因为我们可以使用红外摄像头追踪它们眼睛的位置。这使我们能够做到的一件事，是绘制我们用单个设备能够记录到的每个神经元的感受野。

现在，我们这样做的方法是：让动物稳定注视屏幕，同时给它播放随机棋盘格影片。然后，我们只取影片中在该细胞内引发反应的那些帧，并将它们全部平均。这是一种称为反向相关的技术。为此目的，它通常在视觉神经科学领域得到相当广泛的应用。这是使用这种技术绘制的一个感受野示例，中央的十字是注视点，你可以看到兴奋性感受野和抑制性感受野的红色和蓝色小区域。

这些区域赋予皮层细胞一些特征性属性。因此，我们可以同时记录所有电极对应的全部感受野。

如果我们把所有这些感受野汇集到一起，让它们重叠，并把它们放到一台、一台处于典型观看距离的电脑显示器上作为尺度参照，你就会开始了解到我们能覆盖多大范围的视野。使用这个初步设备，许多感受野都靠近中央凹，

靠近注视点。

这部分是由于我之前谈到的中央凹放大效应，但周边也散布着一些感受野。这些来自大脑更深处距状沟内的记录位点。

到目前为止，我只谈到了从皮层记录信息。但要制造视觉假体，我们需要进行刺激。因此，如果我们刺激感受野位于这个位置的细胞，就会在那个位置产生闪光感知，而这种闪光只有猴子能看见。

我们怎么知道猴子看见了它？我们怎么知道它看起来是什么样的？嗯，很遗憾，我们不能问它们看到了什么，但我们可以训练它们告诉我们关于那种磷化氢的一些信息。我们先训练猴子注视屏幕上的一个中心点，比如这个白点。我们先在屏幕上呈现真实的视觉刺激，并在猴子朝那些刺激做出眼球运动时给予奖励。

这里我们闪现一个白点，猴子会将眼睛转向它，绿色箭头表示的就是这一眼球运动。然后我们选择另一个随机位置，并在猴子将眼睛转向那里时给予奖励。一旦它熟练掌握这项任务，我们就可以开始把这些真实刺激与对电极的电刺激交错进行，并产生一个磷化氢。猴子看到闪光后，自然会朝它进行一次扫视。

这不仅能告诉我们闪光出现在视野中的什么位置，我们还可以改变注入该电极的电流，看看它进行这种扫视的频率，以及我们产生的刺激磷化氢有多显眼，或者可能有多大。

让我们看看执行这项任务的代码。我想先以1/4速度展示给你们：出现了一次视觉闪光，它将眼睛转向了那里。猴子只能看到这个屏幕上的白色内容。它看不到自己的眼球运动，当然也看不到我们何时进行刺激，但这里我们进行了刺激，它便朝同一个位置做出了同样的扫视。因为我们刺激的是同一个电极，当时屏幕上什么也没有出现，而且它没有任何其他线索来做出这一眼球运动。

让我实时展示给你们看。可以看到，猴子喜欢非常快速地工作。而当我们进行刺激时，它会实时做出那次扫视。

看起来它已经受够了。

所以，我向你们展示的是一种在视野中产生磷化氢的方法。这在视觉神经科学中并不是什么新事物，但如果你把那个磷化氢看作视觉图像中的一个像素，我们只需要扩大规模，产生更多得多的像素，并让它们覆盖整个视野。这是一幅示意图，展示了使用我们N1设备的视觉假体可能是什么样子。一台摄像头。比如，摄像头的输出将由一部iPhone处理，随后由它把数据流式传输到设备。

图像将被转换成对视觉皮层内电极进行刺激的模式。有了1,000个电极，我们也许能够产生一幅类似于你们在右侧看到的图像。但正如Avinash告诉你们的那样，我们的下一代设备将配备16,000个电极。如果在视觉皮层两侧各放置一台设备，就能提供32,000个光点，为盲人形成一幅图像。

我们的目标将是为一个在黑暗中生活了数十年的人点亮光明。

非常感谢。接下来交给Joey，他现在将介绍我们设备的另一项非常令人振奋的应用。

谢谢你，Dan。

那么，我叫Joey。我是一名神经工程师，也是Neuralink下一代团队的负责人。

对于脊髓损伤患者来说，大脑与身体之间的连接被切断了。大脑仍然正常运作，却无法与外界交流。你们已经听说了我们如何使用N1 Link作为一种沟通假体，帮助脊髓损伤患者控制电脑或手机。但它也可以用来让身体重新活动起来。让我展示给你们看。

首先，介绍一点神经解剖学知识。运动意图产生于运动皮层，并沿着长神经纤维向下传递，穿过脊髓。这些是脊髓中的上运动神经元。它们形成突触，也就是说，与另一个运动神经元，也就是下运动神经元建立连接；后者把这些运动意图传递给肌肉，肌肉随之收缩，进而产生运动。当然，随意运动还涉及许多其他回路。

你可以把脊髓想象成由很多很多对这样的2个连接构成。在脊髓损伤中，其中一个连接被切断，无法使肌肉收缩。

让我们再放大一点。这里左侧可以看到脊髓的一个横截面，其中示意性地画出了一根向下延伸的纤维。它穿行于白质束中。这是上运动神经元，随后它在这片蝴蝶形的灰质区域内形成突触，这一区域被称

为运动神经元池。

在运动神经元池中，下运动神经元经腹根向下延伸至肌肉，使肌肉收缩；随后，这些运动产生的感觉后果，例如你的手触碰物体的感觉，会经背根返回脊髓，并沿脊髓上行至大脑的感觉区域。

同样，在脊髓损伤中，这一连接被切断了。

如果我们能把电极置入脊髓，比如放在邻近下运动神经元的运动神经元池中，就可以刺激这些神经元，激活它们，进而使肌肉收缩并产生运动。但这很难做到。脊髓相当脆弱，而且会在骨性椎管内显著移动。这可能损坏电极，也可能损伤组织，或者两者都会发生。

但我们的电极又小又柔软，而且我们的机器人能够将它们深深插入组织，也许可以一直深入脊髓腹角。因此，我们正是这样做的。这里可以看到R1机器人的视图。这是一个定位视图。我们已经在脊髓横跨许多毫米的范围内放置了电极。R1机器人能够将这些电极深入插入腹角，进入与下运动神经元距离非常近的运动神经元池。

这一点很重要，因为它让这些电极能够与那些神经元建立局部连接，并激活非常精确的运动。

现在，为了追踪运动，使用你们可能在电影制作中见过的那种动作捕捉标记是非常常见的做法。可以用黏性较弱的黏合剂贴上这些标记，你们可以看到我正在把它们贴到手上。接下来的几张幻灯片中，我们将利用这些标记来放大观察运动。

好，这里有一头正在跑步机上行走的猪。你们以前可能在之前的一次Neuralink演示中见过类似的内容。但与以前不同的是，这头猪拥有不止1台Neuralink设备。大脑中有1台设备，脊髓中也有1台。我们可以从这台设备、这些设备实时流式传输神经数据，并用它们来做一些事情，比如，比如解码猪的关节运动。

这里左侧可以看到髋关节、膝关节和踝关节的时间序列。我们正在解码这些运动。这非常酷，但实际上并不是我们想做的。我们想朝另一个方向推进。我们希望刺激脊髓并使运动发生。

好的，那我们就来做吧。这是一头猪，一头快乐又健康的猪，正在做猪喜欢做的事，也就是四处拱土寻找食物和零食。正如你在地面上看到的，那里有一个蓝色屏幕。这是一个自愿参与区，猪会自行进入其中，表明它愿意接受刺激。当它处于该区域内时，我们就进行刺激。如果猪离开该区域，我们就会停止刺激。

和之前一样，你可以看到，我们能够追踪关节的位置，同时也能传输神经数据。好的，那我们来刺激一个电极。这是一根导线上的一个电极，当我们刺激时，凝块会引起腿部的屈曲运动。在左侧，你可以看到关节的运动，也可以看到以黄色显示的刺激模式时间序列。腿正在向上移动。这是另一个电极，当我们刺激它时，会引起伸展运动。

实际上，这个稍微更难看清，因为腿正在伸直，髋部也在移动。但如果你仔细看，就能看到这是怎么回事。腿正在移动。

我们可以在多种不同的导线上进行刺激，产生不同的动作，而且，而且实际上还能在空间和时间上对它们进行排序，从而形成模式。在左侧，你可以看到对不同电极进行不同刺激的时间序列。你可以看到关节的运动。在右侧，我们正在放大观察肌肉活动。这也让我们了解到这些动作的力量、力度和特异性。

所以，除了进行序列刺激外，我们还可以实现持续运动。这些是强有力的肌肉收缩，可能是站立或其他承重活动所需要的类型，而且对于在这个世界中进行互动确实至关重要。

好的，所以刺激脊髓只是整个故事的一部分。你还必须获得，比如说，用于在脊髓上进行刺激的指令信号。不幸的是，我们有办法做到这一点。我们有你们已经听说过的 M1 Link，它被植入运动皮层。

那会如何运作？

所以，我们将导线置于运动皮层并记录脉冲。这些脉冲会通过无线方式传输，实时发出，并被解码成刺激模式。随后，刺激会被传递到脊髓腹角，传递到我们想要激活的肌肉所对应的运动神经元池。然后，我们刺激激活这些下运动神经元，从而使肌肉收缩并产生运动。

当然，事实上，没有感觉的运动有点困难。试想一下，如果你的四肢麻木了，尝试移动它们会是什么样子。但我们也可以获取感觉信息。因此，你的运动所产生的感觉后果可以以脉冲的形式记录在脊髓背角中。例如，这里是一根羽毛触碰手部。这些脉冲随后可以被实时解码，转化为刺激模式，发送到大脑中的同一个 N1 设备，或者可能是感觉区域中的另一个设备。

刺激大脑的那个部分会产生触觉和本体感觉，从而形成闭环。

所以，把这2个环路结合起来，我们就能从大脑中解码运动意图，并用它来刺激脊髓，引发运动；然后在脊髓中记录这些动作产生的感觉后果，用来刺激大脑，从而产生感知。

现在，要实现这一完整愿景，我们还有很多工作要做，但我希望你们能看出，实现它所需的各个部分都已经具备了。如果你们也和我一样对这一前景感到兴奋，我希望你们考虑加入 Neuralink。

谢谢。

谢谢。

这显然非常了不起，并且具有明确的治疗潜力。对于

神经科学研究界来说，能够使用其中一些工具也会很棒。

你们是否有计划向神经科学家开放这些工具？

有，有，我们有。

这是个很好的问题。我认为，如果我们把手术机器人和设备提供给大学和医院的神经科学研究部门，可能会有很多东西可以被研究出来。所以我认为，在我们已经达到的时候。我们需要让这些机器投入生产，显然还要获得 FDA 的批准，但我认为把它们提供给研究型大学和医院会非常合理。

问题是，在我们所拥有的数据中。

在你们收集的数据集中，有没有一些是你们计划向科学界开源、开源的？

是的，我认为那会是。那应该没问题，我想。对，当然，完全可以。

因为我认为，对于从事 AI 研究的人来说，在这些数据的基础上继续开发并为大脑构建基础模型，可能会非常有意思？

对，这一点很好。对。事实上，我完全不介意直接把它发布在我们的网站上。想用就用。

期待。

很好。

谢谢。

感谢你们所做的非常精彩的演示。我有1个问题。正如我们都知道的，对于植入式电极，无论是用于刺激还是记录，在

我们植入电极后，瘢痕组织

会在电极周围生长，尤其是对于记录而言，长期植入后，我们获得的信号会变得越来越小。你们如何解决这个问题？

先介绍一下背景，我叫扎克，负责脑接口微制造团队。我认为我们无法具体地看出解决它，但有一点，我们拥有的1项优势是，我们的线既柔韧又尺寸小，可以借此尽量减少瘢痕组织和损伤。我们已经开始着手、并将继续推进的一些未来工作，是进一步缩小线的尺寸，只为尽量限制免疫反应，并切实限制瘢痕组织的生长。

其实，我想追问一下。

那么你认为这是否会

有助于实际在电极表面施加一些阻力，或者采用其他某种方式？

嗯，我想，也许这个问题其实就是，随着时间推移，我们观察到了什么样的信号衰减？

而且，你知道，基本上就是1年后它还能用吗，2年后还能用吗？能。

所以，对，对，这一点很好。具体就线的使用寿命而言，我们可以用来评估的真正黄金标准，是从动物参与者那里获得的数据。关于这一点，我不确定之前是否提到过，但我们目前最长的数据来自1名动物参与者，其有用的功能通道维持了600，持续了600天，在此期间，我们一直利用这些信号为 BCI 做一些有用的事情。

然后，对于我们最新版本的设备，我们有一批参与者拥有达到或接近1年的数据，而且由此实现的 BCI 功能也完全有用。

谢谢。

如果可以的话，我再补充一点。你提到可能使用药物来减少炎症。我们实际上正在积极研究的一件事，是采用某种生物编码，以减少炎性炎症，或者让它们变得光滑。所以，你知道，你提到过，也从演示中听到了，我们面临的挑战之一，是将这些线从植入后形成的新生膜组织中移除。

所以有一些这样的项目，我们确实在研究如何将从生物学和这些涂层中学到的一些东西融入我们的细丝中，希望既能减少炎症，也能让取出变得更容易。

另外还在继续缩小电极的尺寸。所以当电极变得非常小时，瘢痕组织的那种炎症反应就会变得微乎其微。所以它就像一个非常微小的电极，身体基本上会忽略它。

这真的令人印象深刻。

祝贺整个团队。那么，当然正如你们所知，当前电子电极的一个问题是，它们很坚硬，而且会四处移动。

所以你会遇到这些神经非平稳性

而且我想，我们很多人原本希望，借助这些非常细的细丝

它们或许会更多地随着

大脑一起移动，这样你就不会看到那种情况。

但从我们展示的持续数百天的数据来看，存在很大的变异性。那么你能谈谈它们会移动多少

吗？你是否知道，比如它为什么会移动？

你能阻止它移动吗？

这些信号在每小时之间和每天之间有多稳定？

你好，我是布利斯，我是脑机接口团队软件小组的负责人之一。在你刚才提到的那张图中，我们展示的是某一特定通道上每天记录到的平均放电率。正如你非常清楚的那样，要理解这一点相当复杂。如果你日复一日地记录完全相同的神经元。举例来说，也可能是你实际上每天捕捉到的是不同的神经元，而这就是放电率发生变化的原因。

我们认为，这至少不是这里这种情况的主要原因。原因是，如果你逐日观察那种脉冲形状，即使平均放电率发生了很大变化，你仍然会看到相对稳定的脉冲形状。显然，这并不是一个完全无懈可击的解释，但至少能让我们有一定信心，认为你实际捕捉到的并不是不同的神经元。不过，在稳健性、非平稳性这个问题中，至少有一部分仍然很有可能确实属于这种情况。

好的，很好，谢谢。是的，谢谢你的问题。

是的，澄清一下，这个电极的位置其实相当稳定，因为你有这些非常微小的，基本上是非常微小的导线，带着。

而且其中还有一些活动余量，比如你有，你有最初固定在颅骨上的设备，但接着你又有这根很长、非常细小的导线，其中有一段像线圈一样卷绕着。所以它确实倾向于基本保持在同一个位置。

我们也在 Neuralink 的 Twitter 上征集了问题，所以我们会穿插提出其中一些。Supe 想问，Neuralink 能在哪些大多数人没有意识到的方面帮助人们？

嗯，我的意思是，一旦进入那里，你知道，就有很多事情可以做。所以，你知道，你显然可以测量温度，因此可以非常早地检测到发烧；你无法测量压力。我认为你大概能检测到那个

在

中风非常早期、刚刚开始的阶段，因为你可以看到电信号开始变得有些失控。所以实际上，一旦进入那里，你知道，借助非常简单的传感器，可能可以进行大量常规健康监测。

嗨。

你们大家都做得非常好，把大量复杂的工程

和科学内容提炼出来，并讲解得极其清楚。

所以做得很好。我想问一点关于模拟的问题。我想是关于磷化物和诱发运动的。

你们认为这更像是局部刺激吗？是近细胞刺激吗？你们是在引导电流流动吗？你们激活了多少个细胞？使用了多大电流？我只是好奇它的尺度有多大，以及你们是否拥有很高的精度，或者很大的，你知道，你们也产生了非常显著的行为效应。

嗨，是的，我是Dan。而一根电极仍能刺激多少个细胞，取决于电极的阻抗、导电焊盘的尺寸、输送的电流量、频率，以及所有这些因素。因此，我们可以利用非常大的可变范围来定制磷化氢的形状，或者必然是形状，但也可能是磷化氢的强度。我们认为，用我们目前的电流电极，至少在代码中，根据粗略计算，在视觉系统中受到刺激的细胞大概是一个直径约50至100微米的球体。

这个球体越小，你就能让特定的磷化氢越小、越有针对性。基本上，……中的像素越小

你能生成的图像。

所以这方面有很大的定制空间。

实际上，也可以通过控制电极之间的场，也就是电场，得到高得多的、比如说有效像素数。所以它不一定，只是不是一对一的关系。实际上，你可以动态调整这个场，并模拟出远……有一个、有一个非常高的神经元与电极比率。所以试试，比如说，你能不能达到，你知道，或许 10 比 1，甚至可能 100 比 1。所以基本上是百万像素类型的。

你能正常看见吗？我想人们会想知道这一点。我认为那是可能的结果之一。

嗨，Lon。这太不可思议了。

你能谈谈植入物本身的使用寿命吗？

还有，植入物的材料会如何与脑组织发生反应

或者与骨密度或骨骼结构发生反应？

谢谢。

好的，很乐意谈谈这个。我是Jeremy，是脑机接口团队的工程师，我认为最好从数据说起。就像Zach提到的，我们有一个植入物，你知道，一只猴子使用它进行了617天的BCI操作，那是Pager在升级为激光设备之前使用的。我们当前版本的设备已经持续运行了将近1年。至于Josh刚才谈到的加速寿命测试仪，我们有植入物的数据：上一版本相当于8年的加速时间，当前版本相当于4年的加速时间，而且还在继续。

所以这算是从数据说起。这些设备仍在持续运行，从理论上说也还在继续。影响设备使用寿命的因素基本上有3个。第1个是密封，也就是设备的气密封装。第2个是电池和内部电子元件。第3个则是Zach稍微谈到过的线，以及这些通道能否正常记录来自大脑的信号。

我们认为，就瓶颈而言，密封的寿命会远远超过另外2项。所以只从理论上来说，我想Josh提过，它是一种热塑性聚合物材料。因此，随着时间推移，会有极少量的水分扩散穿过它。就这一特性而言，我们认为它轻松就能维持20多年。而且正如我所说，到目前为止，我们还没有看到当前版本设备的密封失效。

所以对于电池和内部电子元件，我们其实还没有真正触及这里的极限。这实际上取决于使用情况以及你想要多长的运行时间。我们目前正在获取数据，以便做出时间跨度更长的推算。但现在我们相信，在第3年这个时间点，我们可以达到80%的运行时间，对于4小时的运行时间来说，这大约是，你知道，3个半小时。但正如Avinash所说，我们很快就会把它翻倍，也会提高到4倍。

我们有这样做的计划。所以内部电子元件其实也不是瓶颈。因此，我们真正着手解决的是线本身，以及Zach提到的那些通道的使用寿命，Zach可以谈谈我们为延长这一寿命而正在进行的一些改进。

谢谢。是的。所以就像之前提到的，正如Jeremy所说，我们对线的测试不一定有一个终点。话虽如此，我们正专注于使用寿命，因为我们认为这是一个需要解决的重要问题。因此，在推进当前设备的同时，我们正在做的一件事，是积极推进使用非晶碳化硅对线进行绝缘，我们认为这会让使用寿命大幅超过5年，但当然仍有待测试。

与此同时，我们才刚刚开始研究原子层沉积，我们认为这甚至可以进一步大幅延长线的使用寿命，并能沉积非常薄的层，从而保持线的柔韧性及其在这方面的优势。因此，除此之外，我们当然还必须设计并验证非常可靠的台架测试，以真正模拟体内条件并观察通道退化情况。所以这就是我们在线线的使用寿命方面正在研究的内容。

然后我想你还问到了BioComp，关于Biocomp，我可以说，基本上我们目前使用的所有材料至少都是生物稳定的。我们也经常把材料送出去进行生物相容性测试。基本上，我们所做的是，在许多情况下使用文献中已有的材料，学术实验室已经开始研究这些材料，而我们算是顺势跟进，并以此作为起点。

谢谢你回答这个问题。

很好。

那么，我们还有一个来自 Twitter 的问题。

这是戴维提出的，他问团队：自上次演示以来，你们学到的最重要的经验是什么？

已经大约2年了。我相信完成了大量工程工作。所以，是的。有人想回答一下我们过去2年学到了什么吗？

所以，过去2年里我们学到的一件事是，是与从小规模着手时相比，人脑尺度下的大脑移动幅度有多大。当你制作大脑替代模型时，很多研究从啮齿动物开始，大脑并不会移动那么多；而换成人类，大脑可以移动数百微米甚至更多。我们的丝线和针如此之小，以至于放大来看时，那种移动看起来就像有1英里。

我想补充一点，我觉得是植入物所处的环境实际上有多么动态。所以我们谈过，比如当这个植入部位愈合时，瘢痕或新组织可能会生长并填满空间，而这会影响我们的丝线可能如何在那个空间里相互作用。所以这就是为什么我们如此强调设计精确替代模型的重要性。这样你就不必等上数月让植入部位愈合，而是有望在数小时内获得这些信息。

我是机器人团队的亚历克斯。我认为，我们在工程团队内部确实学到的一件事，是持续进行验证和测试的重要性。当我们构建比如精确到个位数微米的运动系统时，就需要我们甚至更加信赖的验证和测试系统，以证明它们能够可靠地工作。我认为，我们确实学到的一件事，就是同样重视这些验证和测试系统，并与我们的产品同步设计这些系统。

我认为，我们学到的另一件事，作为 BCI 或弯曲控制与算法的一部分，是同样地，构建一个原型并让它只在1只猴子、1只 Pager 身上奏效，可能算是一次很大的成功，但与让它每天都能在其他所有猴子身上奏效相比，也相对容易。所以，真正把它做成产品并不容易，但我们正在学习该怎么做。

我的意思是，我了解到大脑真的很柔软，比你想象的柔软得多。它不像，你知道，花椰菜、西兰花或类似的东西。它更像一个水气球，而且它在你的头骨里移动，移动幅度很大。所以，把它想成椰子里装着一个柔软的水气球，或许是个不错的方式。

你好。

考虑到蓝牙的带宽限制，你们是否考虑过采用其他无线通信技术？

嘿，是的，这个问题的第一部分可以由我来回答，然后第二部分我会让 Matt 来回答。这是个很好的问题，尤其是当你考虑如何增加并扩大我们想要记录的通道数量时。对于我们想开展的这类工作来说，这正日益成为一个瓶颈。我们正从几个方面考虑这个问题。一个方面就是直接改进底层无线电接口，稍后我会让 Matt 谈谈这一点。

我们考虑这个问题的另一个角度是，怎样才能更高效地传出植入物中的数据？我认为，最初的方案是压缩。也就是获取你的数据，查看它的特征，找出一种更高效的表示方式，然后只传出经过压缩的数据流。供参考，目前我们的蓝牙带宽约为每秒150千字节。我们从植入物传出的压缩数据流约为每秒50千字节。

所以到目前为止，我们在这方面做得相当不错。但当你开始考虑16,000通道的设备时，这还不足以让你完全实现目标。因此，在压缩方面能起到帮助的其他一些做法，是实际上只传出机器学习模型的输出，而不是运行模型实际所需的输入。所以，我们在后台尝试的一项技术叫作“头端解码”，本质上就是把目前运行在猴子玩游戏所用的 MacBook 上的机器学习模型，转移到植入物上实际运行。

而这就像是一个超级酷的工程问题。如果你想讨论如何让复杂的神经网络在相当于车库门开启器的设备上运行，就来找我聊聊。这很有趣。

是的。所以，解决这个问题的另一种方式，基本上就是完成计算密集型工作，只得到你实际在意并要用来控制某个东西的原始信号，然后把那个东西从植入物传出来。无线电方面，我把话交给 Matt。

是的。回答你的问题，我们确实在研究其他无线电技术。技术。其中一个特别的方向，是在几个不同频率上采用500 MHz频段和超宽带。所以，它的优势体现在你能够实现的比特率上。大约是6到8到10兆比特。

延迟方面也有相当显著的改善。此外，我们还在研究另一种无线技术，以及 W 波段。

你好，感谢各位所作的非常清晰且令人信服的演示。在最早的某场演讲中，有一点让我印象深刻，我想可能是 DJ 的演讲，那就是这样一种愿景：通过这些脑机接口获得新的复杂技能，比如施展功夫的能力。这反映出大脑从根本上说是一台学习机器。然而，后来介绍的许多技术解决方案，在框定时却被框定为试图校正大脑在较长时间尺度上发生变化的方式，比如数天过程中的漂移，或者组织可能随时间愈合的方式。

我很好奇，对于开发这种与一个从根本上具有可塑性、会在多种时间尺度上以复杂方式发生变化的系统相连接的技术，你们共同的愿景是什么。数天、数月、数年。

这是个很难的问题。我认为其中或许会以某种方式存在双向学习。有时，我们会的分数可能会修正我们的算法，而我们更希望获得某种更稳定的表现。但当然，如果随着时间推移，大脑中的人会学会如何更好地使用脑机接口。我们就需要更新模型。所以，为了学习甚至是新的任务，会以某种方式形成一种互动关系。

随着时间推移，这些可能会成为我们需要学习的东西，了解这个人如何学会与计算机互动，然后构建合适的交互体验和用户界面，并构建能帮助他控制我们想要控制之物的算法。

对于 Nir 刚才所说的，我只想补充一点。是的，在某些方面，大脑具有可塑性并能够学习其实是一项优势，而且这能帮助我们，因为我们实际需要做的工作更少，而处于环路中的人实际上会学会如何更好地使用我们的设备。但我们这种特定方法和设备的一个优势是，我们正试图打造一款通道数量极高的设备，这样我们就能，你知道，将电极均匀分布在一个功能区域上。

这样一来，随着时间推移，某些东西是否移动或偏移就没那么重要了。我们可以把这项工作转移给软件，因此也可以构建会随时间变化的算法。所以我认为，这两点实际上都是我们这种特定方法的优势。

我们还有一个来自 Twitter 的问题。Juan 想知道，如果有人刚刚高中毕业，将来想在 Neuralink 工作，你建议他选择什么职业道路？

实际上就是我们介绍过的任何一种技能。我们正在开发新的芯片。

有材料科学，有软件，显然还有动物护理。实际上，我们在 Neuralink 招聘职位中列出的所有内容都会

是一份很好的指南。

是的，我其实很喜欢这样说：当你翻阅任何大学的手册、浏览所有专业时，我认为你可以拉出、指向其中的每一个专业。而这家公司里都有某个人要么是该领域的专家，要么，你知道，主修过那个专业。所以，这确确实实是一项真正的跨学科事业。而且我认为，你知道，只要专注于你所热爱的任何事情，或者你擅长的任何事情，然后，你知道，尽可能深入地钻研下去。

那么，在你自己构建神经接口的过程中，肯定会有一个属于你的位置。

嘿，我们看到了猴子进行心灵感应。但你能否再多谈一点动物行为训练，比如它们的生活和日常流程？当然。我叫 Autumn。我是研究服务主管，其中包括我们的动物护理项目。作为一名动物福利科学家，这是我非常感兴趣的话题。因此，我们的训练项目主要配备行为分析师，他们帮助我们思考如何从训练中消除任何潜在的厌恶刺激或挫败感。

我们认为条件作用是主要方式，其中包括将正向强化作为主要训练方式。

让我想想，我还能和你们分享什么？

是的，是的。我的意思是，这可能不属于行为训练本身的一部分，但我们是在“三个R”的框架下考虑动物福利评估的，这指的是优化、替代和减少。因此，当我们考虑优化时，行为训练确实以这种方式发挥作用。而我们想要消除的事项中，尤其是在研究中，束缚是我们最优先希望消除的事情之一。所以你们今天看到了很多视频，动物会自己走到它们的工作站，因为我们非常努力地去、去消除任何束缚动物的要求。

还有别的吗？

嗯，就上一点再补充一下，你刚才说到，作为这里的一名工程师，这个地方真正令人振奋、也非常酷的一点是，我们确实能够开展许多技术创新，而这些创新可以直接转化为动物在参与这些任务时更大的自主性。所以，正如你们所看到的，猴子只需自愿走到一根树枝旁就能充电，它们也会自愿在自己的生活环境中用笔记本电脑玩游戏。

而完全植入式、完全无线的设备、感应式充电器，所有这些都促成了这种体验。所以，在这里工作非常酷的一点，就是我们确实能够在这类事情上进行创新。

能和一群工程师合作确实很有帮助，他们真的能为猴子做出很酷的东西，让猴子能够更轻松地进行行为训练。

所以我想，回答之前那个关于学什么才能加入Neuralink的问题，我想除了猴子工程，你还可以加上猴子生意。

你好。

我的问题是关于可升级性的，你们多次提到了这一点。所以在那个手术中，会有某种取出植入物的手术，然后你们会植入一套新的植入物。那么你们能否谈谈取出植入物的手术可能造成的损伤，如果有的话，也就是组织损伤？需要等多久？你们会植入相同的区域吗？而且在升级方面，你们为植入手术进行的脑部扫描大概是什么样的？

我不知道有多少个问题我

可以问，

所以我可以先回答其中一些问题。我在可升级性、那些取出植入物的流程以及如何把这些流程设计得更好等方面做了很多工作。

我们正在努力实现的目标是，正如我在演示中提到的那样，让升级植入物真的像最初安装它一样简单。我们今天没有、没有展示太多那种取出植入物的例子，但我们已经非常接近直接取出一个植入物，然后在完全相同的位置重新安装另一个植入物了。这绝对、绝对是目标。我们正将植入物安装在初级运动皮层，这是与此类设备进行交互的一个重要区域。

所以我们。目标是在同一个位置植入。也许如果你扩展到其他应用，那么你会有兴趣转移到别的地方。但我们肯定希望能够插入同一区域。

就损伤而言，这个。我认为我们最关心的是脑内损伤。而我们已经发现，并且也谈到过，大脑顶部那层组织带来的挑战。我们在解决这个问题的道路上已经取得了很大进展。

但由于细线尺寸很小，大脑内形成的那种瘢痕包膜非常轻微，因此实际上很容易将它们移除。所以，即使是第2次或第3次植入，我们也能看到有用的信号。我想我们的一些脑机接口人员或许可以谈谈这一点。确实有猴子参与者正在使用它们的第2个设备，而且确实在充分利用这些设备。

那么，1个，2个问题。其中1个是有人问了关于塑料的问题。你们有没有从行为角度在任何猴子身上观察到任何可塑性？

还是说现在下结论还太早？或者还没有任何观察结果。

从猴子的行为来看。我们看到，它们需要一段时间来学习如何，当然，要针对测试进行训练，但植入后也是如此。而它们提升起来相对很快，能达到很高水平的脑控表现。比如 Pedro，几天后他就能够，像是，才3天就已经能够非常快地学会使用这个设备。当然，由于先前的植入，他已经接受过这项任务的训练，但换上新的植入物后，经过3、4天，他就能控制到接近使用前一个植入物时的表现。

但你们有没有在进阶方面注意到什么，也就是说，大脑已经超过了你们正在运行的神经网络？

很难说。不，没有，算不上。

好的，我还有另一个问题，更偏向电气方面。你们谈到有10、24个通道正在记录。你们传输的是原始信号，还是只有你们所说的那3种尖峰事件。也就是低、中、高，还是你们传输的是原始的、完整的原始波形？

是的。

大家好，我是 Julian。我可以稍微谈谈这个，也许 Avinash 也想补充。不过，我们的芯片会读取原始信号，但我们通常传输出去的是尖峰，并且会在芯片上实时检测这些尖峰。我想，这极大地压缩了数据。对，继续。我们正在对此进行改进，但我们可以请求。我们可以请求原始样本。有时我们也会直接在芯片上处理特定的统计数据或其他数据，然后把计算出的数值发送出去。

所以，有很多种方式可以处理这些数据。

是的。至少就我们目前使用的 N1 系统而言，它依赖 BLE 无线电，因此存在带宽限制。所以你实际上无法从全部10、20、24个通道流式传输原始数据。不过，简单介绍一下我们的压缩算法，也就是尖峰检测算法的开发历程，我们确实曾有过一种类似有线的系统。我们发表过一篇论文，其中使用 USB C 接口，你们知道，通过高带宽有线连接流式传输所有那些信号。

所以，我们确实有过那类开发平台，能够查看原始信号，并了解我们，想要提取哪一组信息，这些信息既能，你知道，适应无线电的带宽，又对脑机接口控制有用。而且，你知道，仅仅以无线方式发送数据也会消耗大量能量。所以，只要有机会减轻这种负担。你知道，我们基本上会尽量让所有这些压缩都在尽可能靠近电极的位置完成。

有一点并不显而易见，那就是控制手机或电脑实际需要的比特率其实非常低。所以我想我们可能保持着比特率纪录，对吗？我们认为是的，也许大约是每秒10比特。所以这超级慢。

但如果你想想，比如你向手机输入数据时，你的拇指移动得有多快，多少次拇指，你的拇指每秒点按多少次。相当、相当低。我的意思是，基本上我们的拇指就像2根移动缓慢的肉棍，我们，你知道，这样操作，而这就像确实有一个很低的，这就像门槛很低，这就是我想说的。

所以，至少对输出而言，它是，它是一个，你正在达到，达到每秒10比特，你正抓着屁股。所以，而且那样的话，你根本不需要什么蓝牙。所以实际上你甚至可以用哔哔啵啵的声音把它发送出去，你知道。所以这并不，如果你要，如果你现在要做高带宽视觉，那你，你知道，可能会达到1兆比特以上。但这，这完全在蓝牙的能力范围内，或者不管怎样，我想说的就是，那不是一个限制。

数据速率。

还有另一个可能值得注意的事项，我们在演示中谈到过，我们认为大概可以解决在不切开硬脑膜的情况下进行植入的问题。基本上，我们可以只在硬脑膜上打许多孔，硬脑膜就像是，那种厚厚的、橙色果皮般的东西，包裹着那个，紧贴着颅骨。如果你不刺穿硬脑膜，你知道，如果你不把硬脑膜切掉，而是打出许多微小的孔，再穿过这些小孔把电极插入大脑，那么恢复时间会快得惊人。

你知道，你其实不会损失多少脑脊液。这，它是，理论上可以，我的意思是，这可能就像一个，整个过程可以像 LASIK 一样，只需10分钟。比如，它很快。并不是什么庞大费力的事情。它超级快。

再回到长期

使用，我想知道你们是否有任何病理学资料，观察许多接受过长期植入的动物的瘢痕组织。沿着这个思路，从安全角度来看，医疗状况下的使用与健康个体使用之间似乎可能存在一点差距。

我没太听清最后一个问题，但我会先听第1个问题，然后请你重复第2个问题。那么第1个问题是，我们是否有长期使用动物的病理学资料？我们绝对有。我们没有猴子的任何病理学资料，因为我们会给它们升级，而且你知道，它们还在继续。我们还有其他主要用于确定安全性的研究。因此，我们确实会确定组织病理学终点。细线本身周围在大脑中形成的瘢痕组织通常可以忽略不计。

就像它对这些细丝几乎完全没有反应。所以就大脑皮层上的瘢痕组织形成而言，这是非常有希望的。这种新膜生长会填补埃隆和亚历克斯提到的、我们在当前手术中移除的那些区域。对于这些，我们确实会，你知道，评估那些瘢痕组织，但它不会、它不会以任何方式造成问题。它不是对异物的持续反应。它只是在填补被移除的组织。

如果你能重复一下，第2个问题。

我没听见。

是的，第2个问题，其实是接着

刚才那点，从安全角度来看，似乎在健康个体中的使用方面可能还存在一点差距。你知道，我想有人提到，他们可能有兴趣尝试原型产品，但只是想知道，对于努力降低安全风险，你有什么看法。

是的，这是个很好的问题。所以就这方面而言，真正关键的是设备的长期使用。所以，你知道，就像我说的，我们有一些已经植入猴子体内的设备，你知道，已经很多年了，而我们完全没有观察到任何行为缺陷。所以首先，这是一个如何评估安全性的问题。你可以评估组织病理学终点，但我们也在寻找认知缺陷或行为缺陷。

而我们在动物身上没有看到任何这些情况，这是很重要的一点。就组织病理学终点而言，它们看起来真的、真的非常好。挑战在于取出设备，这就是为什么我们要在可逆性工作和经硬脑膜植入方面投入如此多的精力。所以在移除设备时，那才是你有可能、有可能造成损伤的时候。因此，我们正在开展、目前有很多持续进行的研究，来真正尽可能降低这种风险，但我们认为，采用当前方法时，这并不是一项重大风险。

而且就像我说的，Pager采用之前的手术方法完成了升级，现在状况很好。所以很明显，它确实，你知道，可以做到完全安全。但要毫无疑问地证明这对人类也是如此，则是我们仍在努力严谨完成的事情。

这回答你的问题了吗？

是的。

谢谢。

那么感谢你们非常深入地

探讨了许多不同方面

关于设备和系统。看到所有这些投入其中的

工程工作，令人印象非常深刻。你刚才提到了比特率。

作为之前的比特率纪录保持者，我可以

确认你确实打破了我的纪录。所以祝贺你。我想我看到峰值达到了

每秒7.4比特。干得好。我的问题其实是关于临床试验和FDA的，在你们可以分享的范围内，我了解到，设备移除，或者可能是电极移除，是

FDA强调的担忧之一。你还能告诉

我们FDA担忧的是哪些方面

，或者对于你们提交的IDE，他们还有哪些疑问？

是的，我是说，我们或许可以谈一点。我是说，这个。其实这些是我们普遍面临的挑战，所以利用、安全性，严谨地为人类证明这一点，确实是我们面临的一项挑战，也是FDA评论过的一点。其他方面，他们确实提出了一些非常好的问题。所以其他方面涉及诸如对我们植入物进行台架热测试之类的事情。显然，重要的是我们的植入物不会因过热而损伤组织。

因此，对这一点进行真正严谨且有效的台架测试非常重要。这实际上是我们将重新设计、使其更加准确的一件事。还有，你知道，他们会针对生物相容性的化学表征提出许多非常棘手的问题。所以我们已经为此进行了非常严谨的测试。但你知道，他们，他们确实会提出许多问题，深入研究数据的细节，并确保确实完全不存在任何有毒化学物质或生物不相容材料进入大脑的可能性。

所以这些都是我们正在处理的事情，你知道，要，要再次以远超要求的程度加以证明，证明得毫无疑问。这里也许有一点值得提及，那就是人们可能很难体会到我们产品的新颖性。尤其是手术机器人和薄膜阵列，它们非常新颖，与现有设备不同。这意味着我们无法大量依赖文献来支持设备的安全性和有效性。

所以我们确实投入了大量精力来设计并对我们的设备进行测试，以便能够严谨地证明它们的安全性，而不能只依赖另一款产品或某篇论文。为了我们的首位人类参与者，这是我们不愿妥协的事情，也是我们正非常努力去做的事情。

我想，如果你问这样一个问题：依我看，我是否愿意把它植入某个人，比如我的一个孩子之类的人体内，在目前这个阶段，如果他们情况严重，比如说，假设他们摔断了脖子，我现在是否愿意这么做？我会，我会说我们已经到了这样一个阶段，至少在我看来，它不会有危险。

你好，感谢你们的演示。所以我有一个非技术性问题。

你们是否在与运动障碍人士合作？

如果是，他们有没有分享过任何令他们期待的应用构想？

我可以回答这个问题的第一部分。老实说，我不是最适合谈这个问题的人，但我们确实有一个用户顾问委员会，由许多患有各种病症的人组成，其中包括四肢瘫痪者，他们会就许多议题向我们提供建议。说一件轶事。大概6个月前，有个人来到办公室，他们告诉我，自己最想用Neuralink设备做什么。

他们说了2件事。一个是他们希望能够每天进行股票交易，以便击败自己的兄弟。第2个是他们希望能够玩射击游戏。所以我认为，那次接触中最令我震惊的是这件事的寻常。我觉得那次谈话确实令人深受鼓舞。所以，你知道自己是谁吧？就是来和我交谈的那个人。祝你今天愉快。

是的，

你知道，有件事我们谈过，但也许应该再次强调。我们正在做，我们正在为这些设备建立一套生产系统。所以我们正在建设、启动生产线，制造大量设备。我们想制造数千台，最终是数万台，然后是数百万台设备。因此，最初的进展，尤其是在应用于人类方面，或许会显得慢得令人痛苦。

但我们正在并行推进将其规模化所需的一切工作。所以从理论上说，它的进展应该应该是指数级的。

所以谢谢你们，那是一次非常酷的演示。其中一个明确提出的目标是从大脑各处进行记录，能够从任何位置记录并扰动任何位置。所以目前看来，全部都是皮层层面的。

我很好奇，就目前的设备而言，

它是否，是否有任何长期目标或构想，要将它延伸到大脑更深处？我是说，对于神经精神疾病、记忆，所有这些都位于深得多的地方，要深入数厘米。所以我想知道时间尺度是怎样的？如果让你非常粗略地估计一下，我什么时候能够看到一款深入到那种程度的erlink产品。

是的。

所以我是说，颅骨内设备的基本组成将大体保持不变，因为，正如我之前所说，颅骨内的设备很像智能手表。基本上，它有，它是电池、无线电、感应充电器、计算机，然后还有那些细小的导线，所以你需要让导线更长，还必须为机器人配备一根插入更深的针。但这确实是要成为一种通用的I/O设备。

所以，除了细小导线要变得更长、手术机器人需要一根更长的针之外，理论上你应该能够到达任何地方，因为在我看来，机器人的部分工作是试图检测血管的位置，然后避开它们。对吧。在那种尺度下，这有可能实现吗？我是说，当然不能只靠视觉，但也许还有其他检测方式。

这是当前的目标吗，而且你是否

预计能在，我想，未来10年内实现？

肯定能，是的。我是伊恩，负责这里的机器人与手术工程团队。在DJ提到的3个维度中，其中一个是，你知道，

进入大脑更多区域。

所以机器人团队对这件事考虑得非常多。

至于要越过表面基本上需要什么传感器，那么在这种情况下，你说得对

目前我们确实只能

向下看最多大约1毫米。我认为团队内部对下一步最好使用什么还存在疑问。但比如超声波和光声断层成像，就是我能想到的2种基本上可以深入数厘米的技术。但这是一个极其有趣的问题。你某种程度上需要深层成像，以及一定的操控能力，至少要避开深处的大血管。对。

或者，如果我们能让针和线足够小，同时仍能在较深处保持精密和准确，那么即使碰到血管，也许也不会造成出血。

对，我认为那确实是最理想的情况。如果线真的非常细，它们其实可以穿过血管。如果足够细，也不会有问题，这样我们就不需要血管成像了。实际上，我对这件事能够实现略感乐观。

马特，你可能更适合谈谈这个，但就目前的DBS而言，基本上就是直接送进去。

对。

目前的方法是凭盲操作将一根导线送入。

与我们的线相比，那大得惊人，要大好几个数量级。

所以这也是一个很低的门槛，

我们很容易就能超过。我想人们没有意识到，比如，对于深部。对。模拟。那个孔到底有多大。它是一个。我的意思是，大概是多少？

我的意思是，基本上在目前的脑深部刺激中，要在脑部钻多大的孔？对。

要钻一个14毫米的孔，然后将一根2毫米的导线送入6

厘米、8厘米深的大脑中。

所以整个过程全凭盲操作，只能希望不会碰到血管，并预先告诉患者

这可能对

你有好处，而且有1%的概率，你的

大脑会以一种

我们无法控制的方式出血。

这就是目前正在使用的现有技术。所以做得比它更好是。我们绝对可以比它做得好得多。

没问题。

我们的针是40微米。

再次感谢这场精彩绝伦的演示。我觉得你们能如此快速地测试所有这些电极非常令人着迷，但这也引出了一个问题，就是你们的容错能力如何。如果你们运行这些诊断，结果显示有些东西要么短路了，要么是高Z，那么出现多少个这种情况之后，性能才会下降？第二个问题是，当你们实际植入这个设备时，我们看到了电极插进去、然后又像是回绕到自身上的例子。

但看起来，那基本上是通过切开合成材料来评估的。我很好奇，你们正在做什么来验证所有这些电极在某种程度上的体内植入情况。我们怎么知道这种情况不会发生在真实患者身上。

对，我可以回答那个第、第2个问题。就像我提到的，我们可以。所以在那个案例中，我们其实并没有切片。我们有一台非常酷的微型CT。我的意思是，它本质上就像一台CT扫描仪。所以我们只是把完整无损的替代材料放进这台机器里，然后我们可以，你知道，对它进行全程成像。就像我之前提到的，我们可以制作一种替代材料，让这种情况，你知道，让那种回绕每次都会发生。

然后我们也可以制作一种完全不会发生这种情况的替代材料。现在，我们已经大致确定了真实组织在这个范围内所处的位置。因此，我们目前用来验证和确认这一点的计划是制作一些替代材料，让这种情况，你知道，非常容易发生，比任何，你知道，任何组织中可能出现的最坏情况都糟糕得多，然后进行设计，使它在这种情形下永远不会发生。这样做足够多次，并使用足够脆弱的替代材料，就会让我们有信心确认这种情况实际上并未发生。

这是下一代针。

对。这就是下一代针。我们完全没有看到这个问题

在当前这一代中出现。

我来试着回答你的第1个问题。所以确认一下，你是在问，如果某个特定通道之类的出现故障，会发生什么吗？

对，没错。

对。所以正常情形是，阻抗在脑内基本上会很快稳定下来。即使处于那个水平，我们也能记录到非常好的信号。我们会看到大量脉冲，并且可以将其用于BCI，因为我们有如此多的通道，比如现在是1000个，以后是16,000个；实际上，我们只需使用远少于现有数量的通道，就能运行我们的模型。所以这里或那里坏掉一个通道并不重要，我们仍然能实现非常好的解码。

我不确定我们是否有关于需要多少通道的官方数字，但大概是我们拥有的数量要多一个数量级，而且我们拥有得越多，我们已经能用现有的东西做很多事。

也许再问1个或2个问题。

好的，我有一个问题，关于你们那项非常、非常长期的愿景，也就是与先进AI进行这种高带宽通信。因此，看起来先进AI需要理解人类最复杂的思想和情绪。而这正是神经科学家正在尝试做的事。那么，除了神经工程之外，你们是否还有志于攻克神经科学问题？

嗯，我的意思是，我认为我们会制造输入输出设备以及与之配套的软件接口，而且我想，可能就像早些时候建议的那样，我们会尽可能多地开源，这样人们就可以研究它。我认为还会有很多其他人在我们所做工作的基础上继续开发。你知道，就像你制造出微处理器、CPU或计算机后，人们会编写大量在那台计算机上运行的软件一样。

所以，但是如果你没有计算机，软件就毫无意义。因此，我们正在制造带有计算机的输入输出设备，然后我认为，可能还会有很多其他组织、公司在这个基础上继续开发。所以，对，我的意思是，我有时会想的一件事是，如果你确实拥有全脑接口，并且能够记录记忆，

真的

这就真的进入《黑镜》那类东西了，但这可能会是其中之一。

我还认为，有一点很重要，值得一提，那就是Neuralink并非凭空出现。医学学术领域数十年来的研究，确实为这些可能性奠定了基础：把这些电极放入大脑的不同部位，读取那些信号，对其进行解码，并将其映射到某种应用上。而且，你知道，在加入Neuralink之前，我曾身处学术界，你知道，我确实认为，只要有更好的工具来观察正在发生的动态变化，然后以无缝方式与之交互，这个领域就有很多机会以快得多的速度向前发展。

我想好像是伊恩提到过，你知道，这几乎就像是我们正在为大脑制造一台示波器。我觉得这是一个挺美妙的类比，就是让我们多一些能力去窥探这些动态变化，并利用那些信息。了解它，从而，我不知道，希望能理解比如是什么造就了我们、大脑是如何工作的，以及，你知道，诸如此类的所有事情。

你好。

演示介绍了基于键盘和手写的输入方法。你们计划如何开发一种输入模型，以便在人类执行复杂任务时实现高得多的带宽？

这是一个很难的问题，我们开始用猴子探索这个问题。正如你们看到的，我们有好像多个。我们训练了很多只猴子执行截然不同的任务。这仍然是一个我们正在寻求答案的开放性问题。我想，希望等我们有了第1位参与者后，研究起来会更容易。我们正在探索的一个选项，正如我们展示的那样，是直接解码手写内容。这是一项始于斯坦福的工作，我们正在这里进行探索，并努力加以拓展。

另外还有一种不同的，除了仅仅从大脑中解码不同的东西之外。我们也尝试为用户提供不同的、也许是用户式的界面。例如，我们展示了不同类型的键盘。也许还有滑动输入，以及其他有助于提高通信速率的东西。所以，我们算是在从2个维度处理这些问题。

对。

沿着这个方向再补充一点。正如到目前为止这里许多人所指出的，这是一个通用的输入输出系统，你可以在大脑的不同位置即插即用。大脑中还有其他区域可以帮助提高带宽。例如，语言或言语中枢可以帮助你更顺畅地交流。例如文本，如果那是你主要想做的事情。

是的，就是。

我认为，仅仅拥有这种通用的输入操作设备，就会极其显著地增进我们对大脑的理解。这很难。语言几乎无法表达。就像，你知道，现在我们对于大脑中发生的许多事情都只是在猜测。但如果你拥有直接输入输出，就不是。再也不用猜了。我们会了解到关于……我们将凭借这样一种得到广泛使用的设备了解到的大脑知识，绝对会比我们目前所理解的多出许多个数量级。

那么我想就说到这里，感谢各位到场，也感谢各位在线观看。

T.

It.

Sam.

Sa.

Sa.

Sam.

It.

Sa.

Welcome to the Neuralink show and tell. So we've got an amazing amount of new developments to share with you that I think are incredibly exciting, as well as tell you about the future of what we're planning to do here.

Now, this is meant to be a technical podcast, sort of like I'm going to provide an overall summary and then we're going to have a number of members of the Neuralink team come in and give a deep technical overview of the various areas. So, yeah, so let me move forward with the overall summary. Now. Some of the things I'm going to say are things you've. Well, if you've been following Neuralink, you've already heard before, but for a lot of people out there, they've no idea what Neuralink does.

And so I will be a little bit repetitive of things you may already know, but that others do not. So the overarching goal of Neuralink is to create ultimately a whole brain interface, so a generalized input output to device that in the long term, literally could interface with every aspect of your brain, and in the short term can interface with any given section of your brain and solve a tremendous number of things that cause debilitating issues for people.

So, you know, so our long term is like, I mean, I'll talk a little bit about a long term goal. It's going to sound a little esoteric, but it's the. It was actually the sort of my prime motivation, which was, you know, kind of, what, what do we do about AI? Like, what do we do about artificial general intelligence?

If we have digital superintelligence that's just much smarter than any human, how do we mitigate that risk?

At a species level, how do we mitigate that risk? And then even in a benign scenario where the AI is very, very benevolent, then how do we even go along for the.

Go along for the ride?

How do we participate?

And the conclusion, the thing that, the biggest limitation in going along for the ride and in aligning AI, I think, is the bandwidth, how quickly you can interact with the computer. So we are all already cyborgs in a way, in that you're. Your phone and your computer are extensions of yourself. And if you, I'm sure you found, like, if you leave your phone behind, you end up tapping your pockets. And it's like having missing limb syndrome, like where you know the phone is.

It is leaving your phone behind is kind of like a missing limb. At this point, you're so used to interfacing with it, you're so used to being a de facto cyborg, but. So what's the limitation on a phone or a laptop? Limitation is the rate at which you can receive and send information, especially the speed with which you can send information. So if you're interacting with a phone, it's limited by the speed at which you can move your thumbs or speed at which you can talk into your phone.

This is an extremely low data rate.

Maybe it's like 10, optimistically 100 bits per second, but a computer can communicate at gigabits terabits per second. So this is the fundamental limitation that I think we need to address to mitigate the long term risk of artificial intelligence and also just go along for the ride. And.

Yeah, so, but like I said, that's an esoteric explanation that I think will appeal to a niche audience, some of whom may be here. But.

And that's a very difficult problem. So even if we do not succeed with that problem, I think we are confident at this point that we will succeed at solving many brain injury issues, spine injury issues along the way. So.

Yeah, so anyways, so actually we have Justin Roiland in the audience.

Hi, Justin. So it's a little Rick and Morty reference here, the great Rick and Morty episode about intelligence enhancement of your dog and what's the worst that can happen?

So anyway, Rick and Morty, I recommend it.

So for. So you want to be able to read the signals from the brain, you want to be able to write the

signals,

you want to be able to ultimately do that for the entire brain and then also extend that to communicating to the rest of your nervous system. If there's a, if you have sort of a severed spinal cord or neck.

So now this video is now 18 months old. So this is Pager, who is playing monkey mind pong. So this is Pager has a neural link implant in this video.

And the thing that's interesting is that you can't even see the neural implant. So we've miniaturized the neural implant to the point where it matches the thickness of the skull that is removed. So essentially it's sort of like having an Apple watch or a Fitbit replacing a piece of skull with like a, you know, a smartwatch, for lack of a better analogy. So

you can see.

You really can't. He looks pretty, is normal. And I think that's pretty important. If you have a neuralink device, like I could have a neuralink device implanted right now and you wouldn't, you wouldn't even know. I mean, hypothetically, Maybe One of these demos.

In fact,

one of these demos. I will.

Yeah.

So, yeah. Anyway, so here's. First of all, it's kind of wild. Hey, monkeys can play Pong. Like they can actually play Pong if you give them a joystick. So Pedro first learned to play Pong with a joystick. So I'm like, that was novel. It's like, I didn't know monkeys could play Pong, but they can. And then, so we first trained Pedro to play Pong with a joystick. Then we took the joystick away and have the neural link.

And now this is. He's playing Telepath. It's a telepathic video games, essentially.

So what we've been doing since then is we've been on the very difficult journey from prototype to product.

And I've often said that prototypes are easy, production is hard. It's really, I'd say 100 to 1,000 times harder to go from a prototype to a device that is safe, reliable, works under a wide range of circumstances, is affordable, and done at scale, insanely difficult.

I mean, there's an old saying that it's 1% inspiration, 99% perspiration, but I think it might be 99%, 99.9% perspiration.

The best example I could give of an idea being easy, but the execution being hard is going to the moon. The idea of going to the moon, easy, going to the moon, very hard.

And we've been working hard to be ready for our first human. And obviously we want to be extremely careful and certain that it will work well before putting a device in a human. But we've submitted, I think, most of our paperwork to the fda and we think probably in about six months we should be able to have a first neural link in a human.

But as I said, we do everything we possibly can to test the devices before not even going into a human, before even going into an animal. So we do bench top testing. We do accelerate accelerated life testing. We have a fake brain simulator that has the texture and it's like emulating a brain, but it's sort of rubber. And so any before we would even think of putting a device in an animal, we do everything we possibly can with rigorous bench top testing.

So we're not cavalier in putting devices into animals. We're extremely careful. And we always want the device, whenever we do the implant, if it's in a sheep or a pig or monkey, to be confirmatory, not exploratory. So that we've done everything we possibly can with bench Top testing. And only then would we consider putting a device in an animal. And yeah, we'll actually show you a demo later today in a few hours, really, of implanting in a brain proxy.

And if anyone in the audience wants to volunteer, we have the robot right there.

So, Lithien, since the pager demo, we've expanded to work with a troupe of six monkeys. We've actually upgraded pager. They do varied tasks, and we do everything possible to ensure that things are stable and replicable and that the device lasts for a long time without degradation. So, and what you're seeing there is it looks like the Matrix, but that's actually, that's a real output of neural signals. So that's not a simulation or just a screensaver or something.

Those are actual neurons firing. That is what one of the readouts looks like.

And here you can see Sake, it's one of other monkeys typing on a keyboard.

Now, this is telepathic typing. So to be clear, this is the. He's not actually using a keyboard. He's moving the cursor with his mind to the highlighted key. Now, technically, we can't actually spell, and so I don't want to oversell this thing because that's. That's the next version.

So the.

But what's really cool here is Sakethemonkey is moving the mouse cursor using just his mind, moving the cursor around to the highlighted key and then spelling out what we. What we want, whatever we want to spell. But. And then.

So this, this is something that could be used for somebody who's, say, quadriplegic or tetraplegic human. Even before we make the spinal cord stuff work, is being able to control a mouse cursor, control a phone. And we're confident that someone who has basically no other interface, the outside world, would be able to control their phone better than someone who has working hands.

I mentioned upgradability. Upgradability is very important because our first production device will be much like an iPhone one. And I'm pretty sure you would not want an iPhone one stuck in your head if the iPhone 14 is available.

So it's going to be

able to

demonstrate full reversibility and upgradability. So you can remove a device and replace it with the latest version, or if it stopped working for any reason, replace it. That's a fundamental requirement for the device of neuralink. And I should say both Saki and pager were upgraded to our latest and greatest implants. So that's been really over a year and a half. Now that pager's had the first implant and then the upgraded implant, so this is a very good sign that it lasts for a long time with no observed ill effect.

I think it's also important to show that Sake actually likes doing the demo and is not like strapped to the chair or anything. So it's. Yeah, so the monkeys actually enjoy doing the demos and they get the banana smoothie and it's kind of a fun game. So I guess what I'm trying to make is like, we care a great deal about animal welfare and I'm pretty sure, like, our monkeys are pretty happy, you know, so as you can see, quick decision maker on the fruit front.

So for our, the first two applications we're going to aim for in humans are restoring vision. And I think this is like, notable in that even if someone has never had vision ever, like they were born blind, we believe we can still restore vision. So because the visual part of the visual part of the cortex is still there. So, yeah, even if they've never seen before, we're confident that they could, they could see.

And then the, the other application being in the motor cortex, where we would initially enable someone who has no ability, almost no ability to operate their, their muscles, you know, sort of like a sort of Stephen Hawking type situation, and enable them to operate their phone faster than someone who has working hands. But then even, obviously, even better than that would be to bridge the connection. So take the signals from the motor cortex and let's say somebody's got a broken neck, then bridging those signals to neural link devices located in the spinal cord.

So we're confident there are no physical limitations to enabling full body functionality. So, I mean, as miraculous as it may sound, we're confident that it is possible to restore full body functionality to someone who has a severed spinal cord.

So, yeah.

So, yeah.

All right. And then I want to emphasize again that the primary purpose of this update is recruiting.

A lot of times people think that they couldn't really work at neuralink because they don't know anything about biology or how brains work. And the thing that we really want to emphasize here is that you don't need to, because when you break down the skills that are needed to make neuralink work, it's actually many of the same skills that are required to make a smartwatch or modern phone work. So it's sort of, you know, software, batteries, radios, inductive charging, and, you know, as well as things that are specific to us like animal care and clinical and regulatory matters, obviously, machine Learning that phrase is used a lot.

But we obviously need to interpret the signals from the brain, which is a biological neural net. And the best thing to interpret a biological neural net is a digital neural net.

So this is. If there's one message I want to convey, it is that if you have expertise in creating advanced devices like watches and phones, computers, then your capabilities would be of great use in solving these important problems.

That's more than anything the message I want to convey. So, see, yeah, so with that, I guess dj.

So.

DJ was on the founding team of Neuralink and just made immense contributions to the company as of many of the others who will present. But I want just to thank DJ for his immense contribution to Neuralink and.

All.

Right, cool. Thank you.

Thanks, Elon.

When I moved from South Korea at age 13 and needed to learn a new language to communicate, I wondered whether there are better and more effective means of communicating my thoughts to the outside world.

And watching Neo learn Kung Fu in the Matrix, I remember thinking, wow, I want to work on making that possible.

And today I believe that this is a tractable engineering challenge, since everything about your intentions, your thoughts and your experiences are all in your brain, encoded as firing statistics of action potentials. If you're able to put electrodes in the right places with the right sensing and stimulation capabilities, this and many other applications that Elon talked about possible, and we can help a lot of people.

I'm incredibly excited to be working on this ambitious, yet important mission to make that future a reality here at Neuralink.

And I'm also incredibly honored to be working with some of the brilliant colleagues, scientists and engineers across many engineering disciplines to work on this intersection of biology and technology.

You'll hear from several of them today to learn about the breadth of technical challenges we face and our progress in the last year.

And I think you'll find that for most of these challenges, as Elon mentioned, you don't need a prior understanding of how the brain works, and that a lot of what we do is applying engineering first principles to biology.

So how do you create a high bandwidth generalized interface to the brain?

From day one, we focused on a set of foundational technologies that are safe, scalable, and capable of accessing all areas of the brain. These three axes, safety, scalability, and access to brain regions, really form the basis for how we engineer products here at neuralink.

Safety, because we want to make our devices, as well as the installation as safe as possible so that we can drive the adoption of this technology and scalability, because as we make our devices Safer and more useful. More people will want it. And with scale, we also want to make it more affordable and access to brain regions so that we can expand the functionalities of our technologies.

So our first steps along these dimensions for our device is what we call the N1 implant.

It's a size of about a quarter and it has over 1000 channels that are capable of recording and stimulating.

It's microfabricated on a flexible thin film arrays that we call threads.

It's fully implantable and wireless, so no wires. And after the surgery, the implant is under the skin and it is invisible.

It also has a battery that you can charge wirelessly and you can use it at home.

So similarly, for implanting our device safely into the brain, we built a surgical robot that we call the R1 robot.

It's capable of maneuvering these tiny threads that are only on the order of few red blood cells wide and inserting them reliably into a moving brain while avoiding vasculature.

It's quite good at doing this reliably. And in fact, because we've never shown an end to end insertion of a robot in action, we're going to do a live demo of the robot doing surgery in our brain proxy. So who wants to see some insertions?

So here it is. That's our R1 robot with our patient alpha, who is lying comfortably on the patient bed.

This is what we call the targeting view. So what you're seeing is this is a picture of our brain proxy. And the pink represents the cortical surface that we want to insert our electrodes into. And the black represents the vasculatures that we want to avoid. And what you're seeing is these hash mark with numbers that represents where we intend to put each of our threads.

So should we see some insertions?

So this is another view real quick. On the left is the view of the insertion area. And on the right, what the robot's gonna do is it's going to peel the array, the threads, one by one from its silicon backing and, and insert it into the targets that we predetermined in the targeting view.

So.

There you go, that's the first insertion.

So we're going to see a couple more insertions.

The whole process of inserting about 64 threads in our first product is going to be around 15 minutes for this robot. So, yeah, there's a second one that went in and we're going to do a third one.

There you go. And then that's going to go in the background and we'll come back to it in the later part of the presentation.

And as Elon mentioned, we've been working very hard to go from prototype to building product as part of this. One of the things that we did is to move our device manufacturing to a dedicated facility in Austin for scale up manufacturing. And what's important to highlight and is evident in this clip is that it's very typical for us to have our engineers who design also work on the physical manufacturing line to build and debug.

And this has been extremely, extremely critical in reducing our iteration cycle time.

And we've also scaled up our surgery so we now have a dedicated, our own or in fact a double or in Austin. And this is just a stepping stone before we eventually build our own neuralink clinic.

So with this product, N1 and R1, our initial goal is to help people with paralysis from complete spinal cord injury regain their digital freedom by enabling them to use their devices as good as, if not better than they could before the injury.

And as Elon mentioned, over the last year this has been the central focus of the company and we've been working very closely with the FDA to get approval and to launch our first inhuman clinical trial in the US hopefully in the next six months.

So hopefully this gives you a good overview of our product.

For the next hour, we're going to go through a deep technical dive on these topics to tell you about our technical challenges, share some of our progress and preview what's coming next.

So with that, over to NIR from my team who's going to talk to you about neural decoding.

Thank you DJ

everyone.

My name is Niravan Chen and I'm the head of Brain Interfaces Applications. Our goal is to enable someone with paralysis control a computer as well as me. Or even better, we'd like to provide fast and accurate control with all the functionality of computers that works anytime, any, anywhere. So I'm very excited to show you how we are using the N1 device with our software and algorithms to achieve this.

Last year we shared with you a video of Pedro the Monkey controlling computer cursor with his brain. So how do we do that? Just a brief reminder. First, we record his neural activity from the motor cortex. Using the N1 device.

We can record from over thousands of channels while he's playing with the joystick. Then we can train a neural net that predicts the cursor velocity from the patterns of his neural activity. With this decoder, he can then control a cursor just by thinking about it without even moving the joystick. He can play with this decoder a variety of games. Also a grid task where he's moving the white dot towards the yellow target.

Every time he gets one, he receives a drop of his favorite smoothie. And he chooses to play this game every day.

Here you can see his performance from early 2021, around the time we released the previous demo. It's quite accurate, but it's a bit slower than what we would like. And cursor control is the foundation for interacting with most computer applications. So since then we've been working to improve cursor speed and accuracy. As you can see, it's much, much faster, almost twice as fast.

However, it's still a bit slower than what I can do. So we are working on creative ways to improve that.

Now, speed is not enough. You want a full set of functionalities. And for decades most software was built for mouse and keyboard control. And it doesn't make sense to reinvent this entire ecosystem for brain control, at least for now. So we are working and we are designing mouse and keyboard interfaces for the brain. The way we do that is by training Pedro and his friends on a variety of computer tasks and then designing algorithms to predict the behavior.

Here you can see a few examples of tasks in different phases of monkey training. For example, left and right click, click and drag, cursor typing, sprite typing, handwriting and even hand gestures.

Now, interacting with computer is bi directional and feedback is very important. I like when I click on a button and I can physically feel the button being pressed. When a potential N1 user will attempt to click, they won't be able to feel it. An example of how we are addressing that is by providing a real time visual feedback that represents the strength of the neural click by changing the color of the cursor. Just by typing on a physical keyboard is much faster and easier than typing on an iPad keyboard.

This will make the brain control much faster and easier to use typing one of the most important functionalities. So you already seen this message and I want to show you the behind the scene of how this message was created. And here you can see again sake using the virtual keyboard. Typing this message. This virtual keyboard is similar to the one I use on my phone. And with the speed and accuracy that we achieved so far, typing on a virtual keyboard is already fast and easy.

However, I never use a virtual keyboard when I type on my computer because it covers my screen and it's also much slower than what I can do with my ten fingers. We can do better. For example, a group from Stanford asks a person to imagine handwriting letters. Then they decoded the letters from his brain activity. Using this approach, they were able to speed up the typing rate. We start this project with our monkeys, but of course they don't know how to write.

So to mimic writing, we train Ranger, one of our favorite monkeys, to trace digits on an iPad. Here you can see him tracing the digit 5 and the digit 2. Then we recorded his neural activity with the N1 device. But now instead of recording the cursor velocity, we decode in real time the digit that he's tracing on the screen.

We had two main takeaways from this project. One, that monkeys are awesome and can learn very, very complex tasks. The second one, that although it can increase the typing rate, it requires hundreds of examples and samples of each of the digits and the characters we wanted to classify. This will not scale the way we are solving that is by indirection. Instead of decoding directly the digits, we first decode the hand trajectory on the screen.

And then when we decoded the hand trajectory, we can use any off the shelf handwriting classifier to to predict the digits and the characters. For example, classicals that are trained on an MLIST data set.

Why it's so important it's important because now we can potentially decode any character in any language with only one neural decoder. For hand trajectory, it means that you can write in English, Hebrew, Mandarin or even monkey language. And we can understand you want a banana?

So there are many challenges ahead of us to improve functionality and speed. And I want to hand it off to Bliss to talk about the third part, how we are making our brain interfaces work anytime, anywhere.

Hello everyone, my name is Bliss and I'm a software engineer here at neuralink. When I use my computer, my mouse and keyboard work how I intend them to, at least like 99.9999% of the time. My goal is to enable a user with paralysis to control their computer as reliably as I can. Here's what we want that experience to feel like. In this video you can see Saki walking over to his MacBook and choosing to work on his typing task.

The entire decoding system works out of the box and it feels totally plug and play. The first step to achieving this kind of high reliability is to test extensively offline. A typical flow for using the N1 link is to connect over Bluetooth, stream out neural activity from the brain, and then use that neural activity to train decoders and do real time inference. We've built a simulation for exactly this sequence. But instead of using A monkey with an implant.

We use a simulated brain that injects synthetic neural activity into an implant sitting in a server rack. From the point of view of that implant, it's in a real brain. This simulation runs on every code commit to validate that from the hardware all the way up through to the neural decoders, our entire stack can achieve state of the art performance. However, while this kind of simulation is great for integration testing of software and hardware, it's not yet detailed enough to guarantee high reliability in the real world.

In the real world, the underlying signals we're trying to decode actually change day to day. In this plot, you can see the average firing rate detected on a representative channel of psaki's implant. Each bar represents one day. And you can see that each day has a different average firing rate than the previous. This presents us with a very interesting problem for how to make our decoders robust day to day. It can actually happen that if you train a neural decoder on one day of data and then try to use it on the next, the average firing rates can actually shift enough to cause a bias in the output of the model.

Here on the right, you can see that this bias is making it hard for the cursor to move to the upper right corner. You see it's struggling here to make it up to the upper right, and then it moves much more effortlessly down to the bottom left.

We're trying many approaches to mitigate this problem. Some examples include building models on large data sets of many days of data to try to find patterns of neural activity that are stable across days. Another approach we're trying is to continuously sample statistics of neural activity on the implant and use the latest estimates to pre process the data before feeding it into the model. This is really an active area of research for the team and it's a critical problem to solve if we want to enable someone with paralysis to control their computer as well as I can.

Another big problem we have is to minimize the time it takes for a spike in the brain to impact the movement of the cursor on the screen. If you have lag or jitter in this control loop, the cursor becomes hard to control, leading to the kinds of overshoots that you can see here on the right.

One big improvement we've made towards in this direction is called phase lock. Phase lock aligns the edge of each packet that we send off the implant to the exact moment that the Bluetooth radio is going to wake up. This minimizes the time it takes for a spike in the brain to be incorporated into the prediction of our neural network. Here you can see the latency distribution after phaselock. Not only has the mean been greatly reduced, but the variance has been reduced as well.

This makes it easier for the user to predict the behavior of their cursor.

Over the last year, we've made tremendous improvements to the stability and reliability of our system and we've been able to demonstrate consistent high performance across many sessions and many months. However, there's still a long road ahead of us before this system will truly feel plug and play. So if solving the hard problems required to ship this technology is exciting to you, you should consider applying to join the team.

Now I'm going to hand it over to Avinash to to talk about how our custom low power ASIC detects spikes in the brain.

Hi, I'm Avinash, one of the engineers on the ASIC team. We designed the custom neural sensors which include both analog and digital circuitry to record and stimulate across 1024 independent channels. We face challenges across all three major performance, power and area. Not only do we have to fit all 1024 channels into a single quarter sized implant, but we also have to measure spiking activity less than 20 microvolts in amplitude.

And today I'd like to focus on the last challenge I mentioned. Power.

Power consumption is important to us because we want to give future users a full day of use of their implant without any interruption for charging. Back in 2018, we were sending every sample from every channel off the device for processing, which burned a ton of power. In 2020, we brought Spike detection onto the chip. As you may know, neurons transmit information by firing, so simply monitoring for these spikes and only sending these spike events off the implant acts as a very efficient form of compression.

And over the past two years we've continued to make optimizations within the ASIC, dropping the total system power consumption down to just 32 milliwatts and doubling battery life.

Let's take a look at our on chip spike detection algorithm which makes our battery powered implants possible. We first start by applying a 500Hz to 5kHz bandpass filter to remove noise that's out of band.

Next, we use an estimate of the noise floor to generate an adaptive threshold per channel. And finally, our spike detector module identifies three key points of a spike. Identifying three points allows us to detect not just the presence of a spike, but the shape of a spike as well. This can be extremely important for distinguishing between multiple neurons adjacent to a single channel. Today I'd like To focus on one of the many optimizations that we've made in our latest chip, this one specifically cutting system power by 15%.

Note that neurons spike relatively infrequently, which means that our spike detector spends a lot of time searching for the first point of a spike and very little time searching for the other two points of a spike that only occur after the threshold is correct crossed. We can use this characteristic of the input waveform to reduce memory accesses within the chip by 30%. Let's take a look at how that works. Our spike detector is implemented as a single functional unit that's shared across all channels with an SRAM to buffer the state of each channel.

As a sample comes in, its channel state is read from SRAM and incremental spike spike detection step is run and then the updated state is written back to SRAM. Since this is happening 20 million times per second across the implant, each of these accesses add up quite quickly. In our latest chip, we split the state into two parts, A hot state and a cold state. The hot state is accessed on every cycle, while the cold state is only accessed once the threshold is crossed, Reducing the average access width and saving power.

We're also working on a next generation stimulation focused chip with 4096 channels still within the footprint of our current chips. In addition to increasing the channel count, we're also increasing the drive voltage so we can get better activation per channel. And to support this higher channel count, as well as a broad range of future applications that you'll soon hear about, we're adding an arm core onto the chip.

And finally, since these chips are the same size as our current chips, we can still put four of them together into a single implant for a total of 16,000 channels. Still within the size of a quarter.

As you can see, we've been working very hard to improve the power consumption within the implant. But we've also been working very hard to improve the charging experience of the implant, which Matt will talk about. But first, the robot has just completed inserting all 64 threads, so let's take a look.

This is a view of the insertion site similar to the one that DJ showed you earlier. But instead of the targeting reticles, if you look closely, you can see that all 64 threads, each carrying 16 electrodes, have been inserted into the brain proxy while avoiding vasculature. And all just within the past 20 minutes.

Let's hand it over to Matt now

to continue the technical deep dive.

Hi, I'm Matt, head of Brain Interfaces Electrical engineering. Our fully implantable N1 device depends on a battery for continuous operation. When that battery is running, low charging is accomplished through wireless power transfer.

However, unlike many consumer electronic devices which can simply offer a physical connector, charging, a fully implantable device poses several unique challenges. First, the system must operate over a wide charging volume without relying on magnets for perfect alignment. The system must be robust to disturbance and complete quickly so as not to be overly burdensome. However, most important is safety in contact with brain tissue.

The outer surface of the implant must not rise more than 2 degrees C.

In pursuit of these goals, our charging system has gone through several engineering iterations. The first, if you watched our pig demo in August of 2020, Gertrude was implanted with a version of the N1 charged with our first generation charger. This device was implemented in a small puck package and later separated into a remote coil and battery base. This charger was challenging to use. However, we learned a lot through its implementation.

Our current production charger, which charges our current generation of implants, is implemented in an aluminum battery base which also includes the drive circuitry.

A remote coil four times the size of our original device. Also disconnectable, This remote coil has increased switching frequency driving improved coil coupling.

This charger is in use today, including several applications within our engineering and animal test facilities. I'd like to show you one of these applications here. With a device we call our simple charger and the coil has been embedded into the habitat. With the addition of one new outer control loop plus a banana smoothie pump, the troop has been trained to charge themselves.

So let's see how Pager charges his implant.

On the right, we're streaming real time diagnostics from Pager's N1. When he climbs up and sits below the coil, you can see the charger automatically detect his presence and transition from searching to charging charging. We see the regulated power output on a scale of 0 to 1 and the current driven into this battery.

I mentioned earlier that we improved the coil coupling. However, the high quality factor coils exhibit good charging performance over relatively larger distances. But as they're brought closer to the implant, what you see is a peak splitting effect where the best, highest efficiency power transfer is pushed up into higher frequencies outside of the ISM band required for compliance with regulated radiated emissions. In our next generation charger.

We address this problem by the introduction of dynamic tuning shown on the right. This allows us to in real time adjust the resonant frequency of the transmit and receive coils so that we can change their properties just ahead of degraded performance.

The electrical engineering Team is currently engaged in developing a third generation charter charger. Notable improvements include bi directional near field communication. This has allowed us to reduce the control latency and improve the thermal regulation.

Improve thermal regulation results in faster charge times. And now Julian will tell us about how we test the N1.

Thank you very much, Matt. My name is Julian and I lead the embedded software group on the Brain Interfaces team. So when we started building implants, we had a small manufacturing line and to collect data from an implant, you would manually walk over with your laptop, you would connect and collect the data of interest. But our goal is to make an ultra safe and ultra reliable implant. And so to do this, we scaled up the manufacturing line, our testing, throughput and data collection capabilities.

So firstly, we added a large suite of acceptance tests to the manufacturing line. These test the functionality of each component and the final assembly. Implants coming off the line are then subjected to benchtop testing, accelerated lifetime and animal models. We then collect data from these implants round the clock. This data is processed by a series of cloud workers and displayed in an aggregate manner. And then finally, all of this information feeds back into our design process and empowers our engineers to answer any question about any implant at any time.

I'm now going to walk you through different parts of this infrastructure, starting off with firmware testing. So the implant contains a small microprocessor running firmware to manage a whole bunch of its operations. And before we release a firmware update, we want to rigorously test it with both unit and hardware in the loop tests, also known as hill tests. So to do a hill test, what you do is you instrument the battery, you instrument the power rails, the microprocessor, and then we connect to each device with a Bluetooth client, and then we walk the devices through various scenarios to test things like power consumption, real time performance, security systems, fault recovery mechanisms, a lot of different things.

In our original implementation of these systems, we used off the shelf components to start automating tests quickly. However, these systems were constructed in a relatively artisan fashion and were very difficult to maintain. And this meant that testing quickly became the bottleneck for development. So to alleviate this, the hardware and software teams developed a new system which integrates all the required components onto a single baseboard.

We can then put the charger and implant hardware on individual modules that plug into this baseboard, including one board with opposing coils, so that we can test charging performance. This architecture allows us to rapidly iterate different hardware prototypes, because we can simply drop them into the system and reuse all the testing infrastructure. Additionally, we can host the current and next generation of our neural asics onto FPGAs and plug those into this board as well.

And that allows us to test a whole extra layer altogether. So that's how we generated this rather inceptive image here on the right. What you're looking at is spiking activity emitted from some of our simulated neural sensors, streamed through the entire system over Bluetooth and then displayed on a phone. This allows us to test everything in one system from chip to cloud.

This system is 1/5 the cost, 1/5 the volume, and is very easy to manufacture. This allows every developer to have a personal unit on their desk, and it also allows us to to shard the entire test suite over a large number of these units mounted into a rack. All of this has greatly accelerated our rate of development.

Let's look next at how we monitor the implant's electronics, the battery and the enclosure. So the implant will periodically capture all of its vital signs and commit those to flash. And then upon next connection with one of our recording stations, it will stream that data off. So for instance, if we look at humidity, we can get an understanding of the integrity of the implant's enclosure. And by looking at battery voltage and power measurements, we can gauge battery health.

All of this is done automatically, without any intervention, giving us 24,7 visibility into the quality of every single device. Additionally, we can use this infrastructure to request high fidelity information on demand, so that we can investigate different anomalous situations. So, for instance, in this particular scenario, we were trying to track down the source of some spurious spikes that we were observing on different channels.

And so we requested raw wave samples directly from those channels.

Capturing good quality neural signals requires intact low impedance electrodes. And so this is also something we monitor very closely with dedicated circuitry on the neural sensor. So how do we do this? We do this by first using an onboard DAC to play a test tone on a single channel. And then we record, using our ADCs simultaneously, we record the response signal on both that channel and physically adjacent channels.

Not only can we measure the impedance of every channel with this, but we can also map different physical phenomena to different characteristic signatures. So, for instance, an open channel will, will appear as a very large response on the channel, and shorter channels will appear as a large response on neighboring channels. By looking at the purity of the signal coming back, we can also validate that the analog front end of the neural sensor itself is operational.

In our original implementation of doing these impedance scans, it took four hours to get through all 1,000 channels. But by paralyzing the tests, down, sampling, filtering, and then reducing the amount of information we have to stream off the device, by moving a lot of the calculation to the firmware side, we're now able to scan all 1,000 channels in just 20 seconds. This means that we can run impedance on every implant every day.

And then our internal dashboards can play back a history of this impedance so that we can get a really good quantitative insight into that interface between biology electronics.

Now that you have an idea about how we test and monitor our implants, I'm going to hand it off to Josh, who's going to tell you about how we get feedback even faster by accelerating our implants to failure.

Hello, my name is Joshua Hess and I'm an engineer on the Brain interfaces team. We are responsible for the implant system design as well as many of the manufacturing and testing tools. Julian, just talked to you a little bit about some of the ways in which we test our implant electronics, hardware and software. But what about the entire system as it relates to longevity in tissue? One of the ways we've addressed this is with the development of our in house accelerated lifetime testing system.

The the system allows us to expedite and capture long duration implant failure modes at scale to rapidly increase our pace of iteration. Even better, the system also significantly reduces the amount of tests which require animal models, both for implant prototypes and of course, longevity testing. So how does the system work? At a very basic level, it comes down to three things. First, we want to mimic the internal chemistry of tissue.

Next, we want to accelerate these chemical interactions as well as diffusion with our implant materials. And finally, we want to aggressively cycle the internal electronics of our implant with these things, Primarily the first two, we have achieved a conservative 4x acceleration factor by the Arrhenius relationship. In other words, every day our implants spend in our accelerated system is equivalent to at least four days spent in vivo.

Historically, one of our greatest challenges has been the battle against moisture ingress into our implants. So we continuously monitor the internal humidity to watch for abnormal rise. Here in white, you can see some internal humidity data from implants and some of our animals for the duration of over one year. As you can see, our internal humidity sensing is so sensitive, it can even detect the very small and slow humidity rise just from diffusion through our implant materials.

Now, in blue, you can see that same internal humidity data, but from devices in our accelerated system.

Now, if we adjust this data for our acceleration factor, you can begin to see not only the agreement in this data, but also just how far into the future the data extends.

Now, in red you can see a device which has failed in our accelerated system. This device showed an abnormal increase in humidity over the duration of many months before implant electronic failures occurred.

So how do we build this system?

Well, we started building the first system prototype just after the COVID shutdown had begun into early 2020. So we had to get a little creative. As you can see, our first system prototype was a little scrappy and operated out of one of our apartments as indicated by the carpeting. Although scrappy, the system allowed us the fastest path to start testing our devices, tuning our working fluid chemistry and checking our constraints.

We also immediately started root causing observed failures in early implant prototypes, fed that information into the next prototype designs and literally rinsed and repeated over the duration of just a few months. The system was built out totally custom and highly iterated with two system versions and countless minor iterations leading us to our currently operated third generation system which achieves high density testing with automatic in vessel charging as well as automatic data collection.

The system also features an implant sled assembly which accepts brain proxy material such that the implant can be installed and inserted by the surgical robot just like you saw a few minutes ago.

We also integrated the system into a high density rack mount form factor along with a centralized fluid management system, both for chemical uniformity across vessels and also reduced operational maintenance. The system has been in operation for the last year and a half and has had its fair share of challenges. Since the system itself is undergoing the same accelerated abuse as the implants within it, it has been extremely challenging to design, build and maintain a system of this scale while keeping it robust even against itself.

So what comes next?

Well, we have started work on our fourth generation system and have totally redesigned it from the ground up to be a hot swappable single implant per vessel design, partly inspired by high density compute servers. With this new system, we will achieve a whole new level of density, robustness and scale. We also intend to have many of these systems operational in the pursuit of capturing even the lowest frequency edge case failure modes.

With this, we will have thousands of implants testing in pursuit of these goals.

We've already started work building out the system, but there is still a lot left to do. There are also many exciting challenges ahead of us such as introducing mechanical stressing, brain proxy micromotion and EFIN replicating tissue growth around the threads for more complete and representative accelerated testing. So now that you've heard some of the ways in which we rigorously test our implant designs before production for surgery.

Christine is now going to take you through a detailed look at our surgical process.

Thanks, Josh. Hi, everyone. I'm Christine, lead of the surgery engineering team.

To get an N1 device, it's essentially these steps. Targeting and the incision, drill the craniectomy, remove the tough outer meningeal layer called the dura, then insert the thin, flexible threads of electrodes, place the implant into the hole we created, and then that's it. You've got an implant under the skin. Look, Ma, no wires. Just kidding. I mean, seriously, no wires. But I don't actually have one.

The surgical robot does the thread insertion part of the surgery. This is because it would be very difficult to do manually. Imagine taking a hair from your head and trying to stick it into jello covered by saran wrap and doing this at a precise depth and position and doing this 64 times within a reasonable amount of time. And not. Neurosurgeon would probably not like it very much if we asked them to do this for the surgery.

So we have the robot that you saw doing its tiny dance. I sort of wanted to call it Tiny Dancer, but it's called R1, which is also great. The rest of the surgery is done by the neurosurgeon. In order for us to make an accessible and affordable procedure, we need to revisit this. I'll tell you why.

When I was in school, my dad lost the ability to walk and to use his arms and even to speak. He was diagnosed with als. We would look on the Internet and you could see maybe one person here or there who had some cool custom robotic assistive device. But it was deeply frustrating. How limited were the options available to him? And there's hundreds of thousands of people with paresis, not even counting people with other conditions that our device might be able to help.

Meanwhile, there's not that many neurosurgeons, maybe about 10 per million people. And it takes about a decade or more to train neurosurgeons, and they're already generally very busy. And as you can imagine, the time is very expensive. So in order for us to do the most good and have an affordable and accessible procedure. Procedure. We need to figure out how one neurosurgeon could oversee many procedures at the same time.

This might sound sort of crazy, but probably so did laser eye surgery before Lasik made it normal. Lasik's been around for about 30 years and counting. In the beginning, the laser robot did just the most fundamental core part that it had to do and the surgeon did the rest. And, and over the iterations, the surgeon has to do less and less and the laser robot does most of it. And it's a highly compelling procedure, takes just a handful of minutes and often gives life changing results.

Since I joined in 2017, we've also done a handful of iterations to optimize the thread insertions of the robot. One of the challenges that we've had to face has to do with the optomechanical packaging. So, as you can see here, there's about three primary optical paths that are really valuable for us to have reliable thread insertions. One is the visible imaging of the needle inserting a thread. And then another is the laser interferometry system called octical coherence tomography.

That gives us the precise position of the brain while it's moving in real time. And then also we have to provide lighting and illumination to see what's going on in the visible light camera. And doing all this, where the needle is at the bottom of the craniectomy, especially when it's close to the skull wall, can be pretty difficult to fit everything and be able to see it. So the way that the team solved this is by putting all three of these optical paths into one optical stack using photon magic, or polarization, whatever you want to call it.

And that enables us to do vessel avoidance in real time. So as I mentioned, the brain is moving and where we place targets in the beginning may not be where you want to insert at the moment the needle is going down there. So the robot can actually detect the vessels and then determine if we're going to insert onto a vessel or not, if it's safe to insert. And then that way we can avoid inserting onto major vessels. And that brings us to the robot that we have here today.

There's still a lot for us to do to get to that procedure where we reduce the role of the neurosurgeon and make it affordable and accessible. The primary the two elements of the surgery that demand the most skills from the neurosurgeon are the craniectomy and the durectomy. Alex and Sam are going to tell you a bit more about how we think we can get rid of the durectomy step. So that leaves the craniectomy.

In neurosurgery, if your craniectomy is small enough, you can use a standard tool called a perforator, which makes quick work of this shape job. But for a larger craniectomy, the surgeon has to rely on their skill in order to accommodate the variability, patient to patient, in skull thickness, skull hardness, even within the same patient in the same craniectomy, you can have different skull thicknesses, for example.

In addition, if we can make something that has a very high precision craniectomy, we can open the design space for future ways of mounting the implant to the skull. So. So I'll show you a few of our prototypes. Ultrasonic cutters like what's on the screen and oscillating cutters have the benefit of not cutting soft tissue. You can cut the bone and not the brain, but however, as you can see here, our ultrasonic cutter prototype created quite a bit of heat to cut at the rate that we wanted.

So onto the oscillating saw here we designed a blade to minimize cut time and also conducted sound and also heating. And as you can see, you can cut through hard things like bone, but not soft things like skin. It's simple and it works. However, if you wanted to cut an arbitrary depth or arbitrary shape, the oscillating saw just won't cut it.

I was afraid no one would get it. You guys are smart. So there's a time tested solution for drilling arbitrary shapes, which is a CNC drill. The challenge with us doing this on a person is that we need to make sure it cuts reliably every single time, doesn't cut too deep. And a few ways that we're using feedback to make sure we don't cut through the brain are force feedback and also impedance. And if I could get a volunteer.

No, just kidding. Maybe next time. But yeah. So this is some insight into some of the things we're working on to make an accessible and affordable procedure. And now Alex is going to tell you a bit about our next generation developments.

Thanks, Christine. I'm Alex. I'm a mechanical engineer here on the robotics team. Now that we've covered the technology and surgical process for a current device, we'd like to cover some of our next generation development projects. I and the next couple speakers would like to talk about one of those projects, which is enabling device upgradability.

You've gotten to hear about the advancements we've made over the past year. We've improved implant robustness, battery and charging performance, Bluetooth usability. Realistically, every new device version is going to be significantly better. It'll be more functional, it'll last longer. We need to keep this new technology accessible for our early adopters. This means that we need a solution to make device upgrade or replacement just as easy as it is to initially install, as many medical device companies have found, this is a challenging problem.

The body's healing response doesn't make this easy. So this isn't solved yet. But we've made significant progress towards enabling this that we'd like to cover today.

Now we'll have to start with some background as to what makes device upgrade challenging. And we'll start with the anatomy. Under the skin, you have the skull. Below that, the dura, a tough membrane that separates the bone from the brain. And between the dura and the brain, you have the pia arachnoid complex, a fluid filled suspension for the brain. To install the device, the surgeon removes a disc of skull and dura to expose the brain surface.

The device then replaces the removed material. The challenge is here at this interface. Over months, all empty volume is filled by tissue encapsulating the device and the threads.

The device would be trivially easy to remove because of the threads small size, they would have slipped right out of the brain. It's the tissue layer that forms above the surface that makes removable challenging. We built tools in house to study this response and characterize it, such as histology and micro ct. In these images, you can see that layer of tissue that has formed above the surface, encapsulating the threads and adhering to the surrounding tissue.

We've explored many different avenues for designing around this healing process and finding a solution to make device upgrades seamless.

Our best successes have come from making the procedure less invasive. Instead of directly exposing the brain surface, we instead keep the dura in place, maintaining the body's natural protective barrier. This prevents encapsulation at the brain surface. And really this is actually a huge win for making the surgery simpler and safer. As Christine alluded to, however, this doesn't come for free. The dura is a very tough opaque membrane.

As you can see in these SEM images. It's composed of a dense network of of collagen fibers. These offer an array of technical challenges for inserting our electrodes.

One of those challenges is imaging through the dura. As you can see on the left, our current custom optical systems offer pretty incredible capabilities for imaging the exposed brain surface. However, as you can see on the right, once the dura is in place, you can't see the dense vasculature at the brain surface. The dirt is in the way. There's simply too much attenuation. To solve this problem, we're developing a new optical system that uses medical standard fluorescent dye to image vessels underneath the tissue.

Here you can see that dye perfusing through the vessels, highlighting them. There's still a lot of engineering work to go to prove accuracy and repeatability of this system, but once that's done, this will allow us to target and avoid blood vessels underneath the dura.

We're also exploring applying our laser imaging system to deeper tissue structures. In the bottom left, you can see a section of the tissue layers underneath the dura. This image is compiled from multiple volumes from our optical coherence tomography system. You can see the collage of those volumes of above. In the future, these new systems, when combined with correlation to pre op imaging such as mri, will enable precise targeting without directly exposing the brain surface.

Now, imaging isn't the only challenge that comes with a tough dural anatomy. Now I'd like to hand it over to Sam to talk about some of the challenges of inserting our electrodes through this membrane.

Thanks, Alex. Hey, I'm Sam and I lead the needle manufacturing and design team. So, as Alex mentioned, the same properties of the dura that make it a good protector of the brain also make it really difficult for us to insert the threads into. In humans, the dura can be over a millimeter in thickness, which doesn't sound like a lot, but compared to our 40 micron needles, it actually is a lot. For example, if you scaled up the needles to the size of a pencil, the dura would scale to over 4 inches in thickness.

Take a look at how far you have to zoom in to even see it. By the time the features of the needle come into frame, you could see individual red blood cells in the same frame.

This is, this is just.

Wait.

This is a real life SEM image of our latest design. On the left there, you can see the end of the thread. In the middle is the needle, and on the right is actually a piece of my hair. So, yeah, it's extremely small. And besides being really small, there's a lot of other challenges associated with designing this. One challenge is that we have to use the needle and the protective cannula that it sits in to grab onto the thread and to hold it while we peel it from this protective silicon backing.

And then we have to keep holding it while we bring it over to the surface and then release it from the cannula during insertions. Another challenge is that the brain is really soft beneath the tuff dura. And so if the needle isn't sharp enough, it'll just keep dimpling the surface without puncturing. And if this free length gets too long, it can actually Just buckle the needle like this. Another challenge is that we don't just have to get the needle through, we have to get the thread through as well.

So we really have to focus on optimizing the combined profile of the needle and thread together. These are just some of the challenges associated with designing something like this. And so far, we found that the key to solving this problem has been improving on our speed of iteration. But let's look at how we make these things in the first place. So we start with a length of 40 micron wire made out of tungsten and alloyed with a little bit of radium for added ductility.

We designed this femtosecond laser mill in house to cut the features of the needle and cannula, and it can do this with sub micron precision. We spent a lot of time this year turning this thing from a science project into an industrial system.

Just a couple months ago, it took a skilled operator 22 minutes to make a needle, and even a skilled operator could only get about 58% yield. Today, that same process takes just six minutes, and anyone can get 91% yield with just a few minutes of training. With only one click, the mill cuts and measures the needle and cannula and uploads the measurements to our lim system so that the robots can use the exact dimensions for each needle that it uses.

Now, this is all for our current design, though, and we've had a couple years to optimize the manufacturing process of it. The current design has served us well so far, but it doesn't quite protect the thread well enough to get through the tuft. So, like I said, we had to come up with something new, and we needed to be able to iterate on designs quickly.

Unsurprisingly, there's no page in machinery's handbook for this kind of thing. So we dug into the science of femtosecond laser ablation and figured out a workflow that allows us to use our laser mill much like a CNC mill. This allows us to iterate several times. This allows us to iterate in under an hour for new designs, allowing several iterations per day when we're really on a roll. As a result, the latest design still seen on the right can actually insert through nine layers of durock, totaling three millimeters on the bench top.

This is far more than we could ever expect in a human with significant margin.

The needle isn't the only part of the puzzle, though. As you can imagine, all these designs here work with different threads. So we need a way to iterate on that as well. And we do this by having our microfabrication process here in house.

This summer, we completely rebuilt our clean room in about nine weeks, which, among other things, greatly reduced particulate counts, which allows yield and throughput to greatly increase. This, combined with all the other great improvements the microfab team has made, Allows us to iterate on new designs in just a matter of days.

The last piece of the puzzle, though, is testing. We can come up with as many new designs as we want, but unless we have a way to actually test them in the right conditions, we won't know what to tweak. Or even worse, we'll spend time optimizing for the wrong things. Take this failure mode, for example. A few months ago, we got to the point where we could pretty reliably insert through the dura. But when we took the proxies and put them in our micro CT imaging, We realized that our hold on the end of the threads was actually too strong, and we were pulling them out just a little bit underneath the stick surface.

By the time we solved the problem, we realized that this issue was very sensitive to the properties of the surrounding material or tissue. We could make a proxy where this never happens, and we can make another proxy where this happened every single time. And this highlights why it's crucial that we spend time making our benchtop tests match tissue as accurately as possible. And I'm going to pass it off to Leslie now, who's going to talk about how we've been doing that.

Thanks.

Hi, I'm Leslie and I lead microfabrication R and D. And part of what we're interested in is understanding the biological environment Our implant and threads experience Once they're fully installed in the body. Learning directly from biology, though, is inherently slow. So in order to move fast, we're developing synthetic materials that mimic the biological environment. This allows us to learn as much as we can on benchtop and start taking steps away from the industry standard of animal testing.

Developing accurate proxies, though, is challenging. The implant environment is made up of many anatomical layers that all have unique properties. And as time goes on and the implant site heals, New tissue forms filling any available space. In addition to that, motion related to cardiovascular activity and head movement introduce added complexity.

So to start addressing some of these challenges, we're engineering materials using feedback from biology. This may involve mechanical characterization of tissue or analysis of interactions at thread tissue interfaces. Much of this characterization is even done during surgery itself by using custom hardware and software that modifies our surgical Robot to double up as a sensitive characterization tool.

We then use the data collected and feed it back into optimizing our materials so that they behave mechanically, chemically and as shown here, structurally, just like biology.

We've come a long way from our humble first brain proxy, shown here sitting on a plate and consisting of agar and a parafilm sheet. And while simple, it allowed us to perfect robot insertions through countless bench shop tests.

Today, our proxy is slightly more complex where we've upgraded to a composite hydrogel based brain proxy that better mimics the modulus of real human brain. We've also incorporated a duraproxy and developed an injectable soft tissue proxy that so far has allowed us to perform bench top mock X plant testing. We have a super long wish list for our proxy of the future, but some of those items include a surgery proxy with integrated soft tissue, brain, bone, skin or even a whole body.

A brain proxy that simulates motion, vascular and electrophysiological activity, and a biological proxy to test biocompatibility and electrical stimulation.

There's a ton of ongoing work getting us closer to our proxy of the future, including work on lab grown cerebral organoids as shown here. And all of this will get us closer to a future where we learn more and iterate faster on benchtop and reduce our reliance on animal models or even one day replace them completely. And with that, I'll hand it over to Dan, who will be presenting a very exciting next generation application.

Thank you.

Thank you, Lesley. My name's Dan and I came to work at neuralink after following a career in visual neuroscience research.

Research.

I was inspired to join this company because I saw in our device the potential to restore vision to people rendered blind by eye injury or disease. There are a number of particular characteristics of our device that make it uniquely suited to this application. Firstly, as well as being able to record from every channel, we can stimulate neural activity in the brain by injecting current through every channel. This is important because it allows us to bypass the eye and generate a visual image in the brain directly.

Secondly, our device can have an enormous number of electrodes for a visual prosthesis. This is important because the more electrodes you can have, the higher density of an image you can create in the brain. Thirdly, thanks to our robot, we can insert these electrodes deeply into the brain. Now, this is an important thing for a visual prosthesis because the human visual cortex is buried deeply in a fold in the medial face of the brain called the calcarine sulcus.

In this image I've highlighted the calcarine sulcus in red in an mri. It contains a map of the visual world, visual field. It's about surface area, equal to a credit card on each side. And if you unfold it and flatten it, you see that the image is inverted, it's upside down, but more interestingly, it's distorted, so that the central part of the visual field, the fixation point, is greatly magnified. So, for example, if you look at this image of Lincoln, if you look directly into his right eye, everything to the left of that fixation point is directed to your right visual cortex, and everything to the right goes to your left visual cortex.

His eye, even though it's very small in the image, is magnified in the brain to occupy nearly a quarter of the surface area of the visual cortex.

Over the last half century, visual neuroscientists have developed a profound understanding of visual processing in the brain. What's driven most of this research is recording from single cells in the cortex, usually of macaque monkeys. One of the seminal discoveries was that every cell in the visual cortex represents only a tiny part of the visual field. Your perception is made up of a mosaic of tiny receptive fields, each belonging to a single cell in your visual cortex.

So if you record from one of these cells in a monkey, say in this location, you can find a very tiny region of the screen where a light stimulus will cause modulation of that neuron. Another location in visual cortex will have a location elsewhere on the screen. In this case, in the lower visual field. These regions are called receptive fields.

We've inserted our device into the visual cortex of two rhesus monkeys whose names are code and dash.

That means we can record activity from their visual cortex generated by their normal home environment as they roam around. But as we all know, monkeys love banana smoothie. That means we can easily teach them to fixate points on a screen and reward them. We can reward them very precisely because we can track the location of their eye using an infrared camera. One of the things this allows us to do is to plot the receptive fields for every neuron that we can record with a single device.

Now, we do this by showing the animal a movie of random checkerboards whilst he fixates steadily on the screen. Then we take only the frames of the movie that generated a response in the cell and average them all together. This is a technique known as reverse correlation. It's generally used quite widely in visual neuroscience for this purpose. And this is an example of a receptive field plotted with this technique, the central cross is the fixation point, and you can see the little red and blue regions of excitatory and inhibitory receptive field.

These regions give cortical cells some of their characteristic properties. So we can record all the receptive fields from all the electrodes at the same time.

And if we take all these receptive fields and accumulate them together, overlap them and place them on a, on a computer monitor for scale at a typical viewing distance, you begin to get an idea of how much of the visual field we can cover. With this preliminary device, many of the receptive fields are close to the fovea,

close to the fixation point.

That's partly due to the magnification that I talked about of the fovea, but there's also a scattering of fields in the periphery. These are from recording sites deeper in the brain in the calcarine sulcus.

So far, I've only talked about recording information from the cortex. But to produce a visual prosthesis, we need to stimulate. So if we stimulated the cells whose receptive fields are in this location, we would produce a perception of a flash in that location that only the monkey can see.

How do we know that the monkey sees it? How do we know what it looks like? Well, unfortunately, we can't ask them what they see, but we can train them to tell us something about that phosphine. We start by training the monkey to fixate a central point on the screen, like this white dot. And we start by presenting real visual stimuli on the screen and rewarding the monkey for making eye movements toward those stimuli.

So here we flash a white dot and the monkey makes an eye movement towards it, symbolized by the green arrow. We then choose another random location and reward the monkey for making an eye movement towards it. Once he's got good at this task, we can begin to interleave these real stimuli with electrical stimulation of electrodes and produce a phosphine. The monkey sees the flash and naturally makes a saccade towards it.

This tells us not only where in the visual field the flash occurred, but we can also change the current that we inject in that electrode to see how often he makes that saccade, how noticeable, or how big perhaps the stimulation phosphine is that we're producing.

Let's look at code performing this task. I want to show you first, at 1/4 speed, there's a visual flash and he makes an eye movement towards it. The monkey can only see what is white on this screen. He can't see his own eye movement, and he certainly can't see when we stimulate, but here we stimulate, and he makes the same saccade to the same location. Because we stimulated the same electrode, nothing appears on the screen at that time, and he has no other cue to make that eye movement.

Let me show you this in real time. You can see monkeys like to work very quickly. And when we stimulate, he makes that saccade in real time.

And looks like he's had enough.

So what I've shown you is a way to produce a phosphine in the visual field. This is not something new in visual neuroscience, but if you think about that phosphine as a single pixel in a visual image, all we need to do is scale up and produce a great many more pixels and have them covering the visual field. This is a schematic of what a visual prosthesis using our N1 device might look like. A camera. The output from a camera would be processed by an iPhone, for example, which would then stream the data to the device.

And the image would be converted into a pattern of stimulation of the electrodes into visual cortex. With 1,000 electrodes, we might be able to produce an image resembling something that you see there on the right. But as Avinash told you, our next generation of the device will have 16,000 electrodes. If you put a device on both sides of your visual cortex, that would give you 32,000 points of light to make an image in someone who's blind.

Our goal will be to turn the lights on for someone who's spent decades living in the dark.

Thanks very much. I'll pass you over to Joey, who's now going to talk about another very exciting application of our device.

Thank you, Dan.

So, my name is Joey. I'm a neuroengineer, and I'm the head for the next gen team at Neuralink.

So, for persons with spinal cord injury, the connection between the brain and the body is severed. The brain continues functioning normally, but it's unable to communicate with the outside world. You've already heard about how we can use the N1 link as a communication prosthesis to help someone with spinal cord injury control a computer or a phone. But it can also be used to reanimate the body. Let me show you how.

First, a little neuroanatomy. Movement intentions arise in motor cortex and are sent down long nerve fibers through the spinal cord. These are upper motor neurons in the spinal cord. They synapse, that is, make a connection with another motor neuron, a lower motor neuron, which sends these movement intentions to the muscles which contract and in turn, you have movement. While, of course, there are many other circuits involved in voluntary movement.

You can think about the spinal cord as as many pairs of these two connections. And in spinal cord injury, one of these connections is severed, unable to make the muscles contract.

Let's zoom a little bit further. So here you can see on the left, a cross section of the spinal cord with a fiber coming down schematically. This travels through the white matter tracts. This is the upper motor neuron, and then it synapses within this butterfly shaped region of gray matter in what's known

as a motor pool.

In the motor pool, the lower motor neuron descends out the ventral roots to the muscles, which contract, and then the sensory consequences of those movements, for example, the touch of your hand against an object, return to the spinal cord through the dorsal roots and ascend the spinal cord up into the sensory regions of the brain.

Again, in spinal cord injury, this connection is severed.

If we could place electrodes into the spinal cord, say, in a motor pool adjacent to lower motor neurons, we could stimulate those neurons, activating them and in turn causing the muscle to contract and movement to occur. But this is very hard to do. The spinal cord is quite delicate, and it moves significantly within the bony spinal canal. This could cause damage to the electrode, it could cause damage to tissue, or both.

But our electrodes are small and flexible, and our robot is able to insert them deep into tissue, perhaps all the way down into the ventral horn of the spinal cord. And so we have done just that. Here you can see a view from the R1 robot. It's a targeting view. And we've placed electrodes across many millimeters of the spinal cord. And the R1 robot is able to insert those electrodes deep into the ventral horn, into motor pools in very close proximity to lower motor neurons.

This is important because it allows them to have a localized connection to those neurons and activate very precise movements.

Now, to track movement, it's very common to use motion capture markers like you might see in the production of a movie. These can be placed with a light adhesive, and you can see me placing these on my hand. We're going to use these markers to let us zoom in on movement in the next couple of slides.

Okay, so here's a pig walking on a treadmill. And you may have seen something like this before in a previous neuralink presentation. But unlike before, this pig has more than one neuralink device. There's a device in the brain, but there's also one in the spinal cord. And we can stream neural data from this device, these devices in real time and use them to do things like, like decode the movement of the joints of the pig.

So here you can see on the left a time series of the hip, knee and ankle. And we're decoding those movements. So this is super cool, but that's actually not what we want to do. We want to go in the other direction. We would like to stimulate the spinal cord and cause movement to occur.

Okay, so let's do that. So here's a pig, a happy and healthy pig, doing what pigs like to do, which is root around for food and snacks. And as you'll see on the floor, there's a blue screen. This is a voluntary engagement zone where the pig places itself, indicating that it's comfortable to receive stimulation. When it's in the zone, we stimulate. And if the pig leaves the zone, we'll stop stimulating.

And as before, you can see we're able to track the position of the joints and also stream neural data as well. Okay, so let's stimulate an electrode. So here's one electrode on one thread that when we stimulate clot causes a flexion movement of the leg. So on the left you can see the movement of the joints and you can also see the time series of the stimulation pattern in yellow. So the leg is moving up. Here's another electrode which when we stimulate, causes an extensor movement.

This is actually a little harder to see because the leg is straightening and the hips are shifting. But if you look carefully, you can see how this is. The leg is moving.

We can stimulate on a great variety of threads and produce different movements and, and actually sequence them spatial temporally to provide patterns. So on the left you can see a time series of different stimulation on different electrodes. You can see the movements of the joints. And on the right, we're zooming in on muscle activity. That gives us an idea of the kind of strength and power and specificity of those movements as well.

So in addition to doing sequences, we can also achieve sustained movement. These are powerful muscle contractions of the sort that you might need for standing or other load bearing activities and are really crucial for interacting through the world.

Okay, so stimulating the spinal cord is only one piece of the story. You also have to get like command signals for the stimulation on the spinal cord. Unfortunately, we have a way to do that. We have the M1 link that you've already heard about, placed in motor cortex.

How would that work?

So we place threads in motor cortex and record spikes. These spikes would be wirelessly transmitted, emitted in real Time and decoded into patterns of stimulation. Stimulation would then be delivered to the ventral horn of the spinal cord, to the appropriate motor pool for the muscles that we like to activate. We then stimulate activate those lower motor neurons, which causes the muscles to contract and movement to occur.

Now, of course, movement without sensation is actually kind of difficult. Just think about what it would be like to try to move your limbs if they're numb. But we can also get sensory information as well. So the sensory consequences of your movement can be recorded in the dorsal horn of the spinal cord in the form of spikes. For example, here, a feather touching the hand. These spikes can in turn be decoded in real time, sent to patterns of stimulation to either the same N1 device in the brain or perhaps a different one in a sensory area.

Stimulation of that part of the brain would cause percepts of touch and proprioception, closing the loop.

So putting those two loops together, we have motor intentions decoded from the brain used to stimulate the spinal cord, causing movement, and then the sensory consequences of those actions being recorded in the spinal cord to stimulate the brain, causing perception.

Now, we have a lot of work to do to achieve this full vision, but I hope you can see how the pieces are all there to achieve this. And if you find this prospect as exciting to you as it is to me, I hope you'll consider joining us here at neuralink.

Thank you.

Thank you.

This is obviously amazing and has clear therapeutic potential. It would also be great for the

scientific neuroscience community to access some of these tools.

Do you have any plans to make these available to neuroscientists?

Yes, yes, we do.

So it's a great question. I think there's probably a lot that can be figured out if we provide the surgical robot and devices to neuroscience research departments at universities and hospitals. So I think at the point at which we have. We need to be in production with the machines and obviously have the FDA approvals, but I think it would make a lot of sense to provide this to research universities and hospitals.

The question is, of the data that we.

Of the data sets that you've collected, are there any that you plan to open source source for the scientific community?

Yeah, I think that would be. That would be fine, I think. Yeah, sure, absolutely.

Because I think it could be really interesting for people working in AI research to build upon that and build foundation models for the brain?

Yeah, it's a good point. Yeah. I actually get no problem with just publishing it on our website. Use it if you want.

Looking forward.

Great.

Thank you.

Thank you for the Very wonderful presentation. So I have one question. So as we all know, for implantable electrode, either for stimulation or recording, after

we implant the electrode, the scar tissue

will grow around the electrode and especially for recording, the signal we get will become smaller and smaller after long term implant. How do you solve this issue?

So for context, I'm Zach, I lead the microfabrication team on brain interfaces. I don't think we can see solve it specifically, but one thing, one advantage we have is both the flexibility and the small size of our threads to try to limit that scar tissue and that damage. And some future work that we have started working on, that we'll continue working on is pushing the size of the threads down just to try to limit the immune response and really limit that scar tissue growth.

Actually, I want to follow up.

So do you think it will be

helpful to actually load some drag on the surface of your electrode or some other way?

Well, I think like maybe the just the question is like what sort of signal degradation have we seen over time?

And you know, basically does it work a year later, does it work two years later? It does.

So yeah, yeah, so that's a good point. So in terms of thread longevity specifically, really the gold standard that we can use to assess is the data we have from our animal participants. And so for that I'm not sure if it was mentioned before, but the longest data we have right now is for an animal participant who has 600, went 600 days with useful functioning channels where we were doing something useful with the signals for bci.

And then with the newest version of our device, we have sort of a collection of participants who are at or near one year of data and completely useful functioning BCI from that as well.

Thank you.

If I may add one more thing. So you mentioned potentially having drugs to kind of reduce inflammation. So one of the things that we are actually actively working on is having some sort of biological coding to either reduce inflammatory inflammation or make them slippery. So you know, you mentioned, you heard from the presentation that one of the challenges that we have is removing the threads from these neomembrane tissue that are formed after implantation.

So there are programs like that where we're really looking at kind of incorporating some of the learnings from biology and these coatings into our threads so that we can hopefully reduce inflammation as well as make it easier to extract.

Also continuing to reduce the size of the electrode. So when the electrode gets really small, the sort of inflammation response of scar tissue becomes minuscule. So it's like a very tiny electrode the body basically ignores.

This is really impressive.

Congrats to the whole team. So, as you of course know, one of the problems with current electronics electrodes is they're rigid and they move around.

So you have these neural non stationarities

and I think many of us had hoped that with these very thin threads

they would maybe move more with the

brain and you wouldn't see that.

But from the data we showed over many hundreds of days, there was a lot of variability. So can you speak to how much

do they move and do you have any idea of like, why does it move?

Can you stop it from moving?

How stable are the signals hour to hour and day to day?

Hi, I'm Bliss, I'm one of the leads of software groups in the Brain Interfaces team. In the particular plot you were mentioning before, what we were showing was the average firing rate recorded per day on a particular channel. It's, as you well know, pretty complicated to understand. If you're recording from the exact same neuron day after day after day. It could be, for example, that you're actually picking up a different neuron day to day and that's why you get the change in firing rate.

We don't think this is at least the majority cause of the situation here. The reason is that if you look at sort of the spike shapes day to day, even when the average firing rate is shifting a lot, you still see sort of stable spike shapes. That's obviously not a fully bulletproof story, but at least gives some confidence that it's not actually different neurons you're picking up. However, there still is very much a chance that that could be the case in at least some part of the robustness, non stationarity story.

Yeah, cool, thanks. Yep, thanks for the question.

Yeah, to be clear that like this electrode position is actually fairly stable because you've got these very tiny, basically very tiny wires with.

And there's some play in the, like you've got, you've got the device attached to the skull originally, but then you've got this long, so tiny wire with kind of a coiled section. So it's, it does tend to basically stay in the same place.

We also asked for questions on Neuralink's Twitter, so we'll be interleaving some of those. Supe wants to ask, what could Neuralink help people with that most people don't realize?

Well, I mean, once you're in there, you know, there's a lot you could do. So, you know, you can obviously measure temperature, so you could do very early detection of a fever you could not measure pressure. I think you probably detect that

at

the very early, the very beginnings of a stroke because you can see sort of like electrical signals starting to go sort of haywire. So there's actually probably a lot of just general health monitoring that you could do once you're in there, you know, and with very simple sensors.

Hi.

You guys all did a great job of distilling a lot of complex engineering

and science and making it wonderfully clear.

So great job. I wanted to ask a little bit about the simulation. I guess for the phosphide and for the evoked movement.

Are you think is this more like local stimulation? Is it juxtocellular? Are you steering current around? How many cells are you activating? How much current are you using? I'm just curious what the scale of this is and whether you have a lot of precision or a lot of, you know, you have pretty profound behavioral effects too.

Hi, yeah, I'm Dan. And how many cells you still stimulate with a single electrode is dependent on the impedance of the electrode, size of the conductive pad, how much current you deliver, the frequency, all these factors. So there's a great deal of variability that we can use to customize the shape of a phosphine or the shape necessarily, but maybe the intensity of a phosphine. We think with our current current electrodes, at least in code, back of the envelope calculation would be something like about a 50 to 100 micron diameter sphere of cells are being stimulated in a visual system.

The smaller that sphere, the smaller and more specific you can make a particular phosphine. Basically the smaller the pixel in the

image you can produce.

So there's plenty of scope for customization of that.

This actually also it's possible to get to a much higher like effective pixel count by controlling the field, electric field between the electrodes. So it's not necessarily, just not a one to one relationship. You could actually dynamically adjust the field and simulate far have a, have a very high neuron to electrode ratio. So try like, could you get like you know, maybe 10 to 1, 100 to 1 potentially. So a megapixel type basically.

Can you see normally? I think people would want to know that. I think that is one of the possible outcomes.

Hi Lon. This is amazing.

Can you talk about the longevity of the implants itself?

Also how would the material of the implant would react with the brain tissue

or density of the bone or bone structure?

Thank you.

Yeah, happy to talk about this. I'm Jeremy, engineer on the Brain Interfaces team and I think it's good to start with data. So like Zach mentioned, we have an implant that was, you know, a Monkey was performed BCI for 617 days, that was pager before being upgraded to the laser device. For our current version of the device, it's lasted for almost a year. And then for accelerated lifetime tester that Josh kind of talked about, we have data from our implants, from the previous version, eight years of accelerated time, and from the current version, four years of accelerated time and counting.

So that's kind of starting with the data. Those devices are still lasting and still going theoretically. There are kind of three fundamental factors that contribute to the longevity of the device. One is going to be the seal, that hermetic enclosure of the device. Two is going to be the battery and internal electronics. And then three is going to be the threads that Zach talked about a little bit and the channels being able to functionally record signals from the brain.

The seal, we think will far outlast the other two in terms of the bottlenecks. So the seal, just theoretically, I think Josh mentioned that it is a thermoplastic polymer material. So there's going to be a very small amount of moisture that diffuses through it over time. And we think that that will last, you know, 20 plus years easily in terms of just that property. And like I said, we have not seen our seals fail with our current version of the device yet.

So we haven't really pushed the limits here for the battery and internal electronics. That's really based on usage and how much runtime you want. And we are working currently on getting data to project out even farther. But right now we believe that we can, you know, achieve 80% runtime at the three year time point, which would be about, you know, three and a half hours for a four hour runtime. But we're, like Avinash mentioned, we're, we're doubling that very soon and quadrupling as well.

We have plans to do that. So internal electronics really aren't the bottleneck either. And so really we're attacking the threads themselves and longevity of those channels that Zach, Zach can kind of talk about some of the improvements that we're doing to increase that longevity.

Thanks. Yeah. So as sort of mentioned before, we don't necessarily have an endpoint, as Jeremy said, for the testing of the threads. That being said, we are focusing on longevity because we think this is an important issue to solve. So one thing that we're doing in parallel with the current device is aggressively pursuing amorphous silicon carbide insulation of the threads, which we believe will take us well beyond five years of longevity, but of course still to be tested.

And in parallel with that, we're just starting to look at atomic layer deposition, which we think could even push longevity of the threads much further and deposit very thin layers to keep the flexibility of the threads and that advantage there. So along with that, we're also of course having to design and validate very robust bench top testing to model really in vivo conditions and look at channel degradation. So that's what we're looking at for longevity of the thread threads.

And then I think you asked about BioComp and I think for Biocomp, essentially all the materials we're using right now I can say are at least bio stable. And we send out testing for biocompatibility very often. And essentially what we're doing is we're using in many cases known materials from literature that academic labs have already started to look at and sort of jumping on that and using that as a starting point.

Thanks for answering that.

Cool.

So we have another question from Twitter.

This is from David and he asked the team, what are the biggest lessons you learned since the previous presentation?

It's been about two years. I'm sure there was a lot of engineering done. So yeah. Anyone want to answer what we learned in the last two years?

So one thing that we've learned in the last couple years is, is how much the brain moves on the human scale compared to when you start small, when you make brain proxies and a lot of research starts with rodents, the brain does not move that much and you get a human and the brain can move like hundreds of microns or more. And when our threads and needles are so small that motion when you zoom in looks like a mile.

I think to add to that one thing is how I guess dynamic the implant environment actually is. So we've talked about like when this implant site heals scar or new tissue might grow and fill in the space and that'll affect like how our threads might interact in that space. So that's why we've emphasized so heavily the importance of designing accurate proxies. So instead of having to wait months for an implant site to heal, you can hopefully learn that information in hours.

I'm Alex on the robotics team. I think one of the things we've definitely learned within the engineering teams is the importance of really continuous validation and testing. Where we're building say motion systems that are precise to single digit microns, we need validation and test systems that we trust even More than that, to prove that they work reliably. And putting just as much focus into those validation and test systems and designing those alongside our products, I think, is one thing we've definitely learned.

Another thing that we learned, I think, as part of BCI or the Bend Control and Algorithm, is that again, building a prototype and making it work with only one monkey, one pager was a great, maybe a success, but also relatively easy to making it work every day for all the other monkeys. So actually making it a product is something that it's not easy, but we are learning how to do it.

I mean, I've learned that the brain is really squishy, like way squishier than you think. It's not like, you know, cauliflower or broccoli or something like that. It's more like a water balloon and it's moving in your skull, like a lot. So you got a squishy water balloon in a coconut is maybe a good way to think of it.

Hello.

Given Bluetooth's bandwidth limitations, have you considered other technologies for wireless communication?

Hey, yeah, I can take the first part of this question and then I'll let Matt answer the second part of it. It's a great question, especially as you think about how to increase and scale the number of channels that we want to record from. This becomes increasingly a bottleneck for the kinds of work that we want to do. We're thinking about this in a couple ways. One is just directly improving the underlying radio interfaces, and I'll let Matt talk about that in a second.

The other way we're thinking about this is how can you be more efficient with the data you send off the implant? And I think the first version of that is compression. So just taking your data, looking at the characteristics of it, find out a way to represent it more efficiently and just send off that compressed stream. So for reference, right now our Bluetooth bandwidth is around 150 kilobytes per second. The compressed stream of data that we send off the implant is around 50 kilobytes per second.

So we're doing fairly well there so far. But when you start thinking about 16,000 channel devices, that won't get you all the way there. So some other things that can help on the compression side are to actually just send out the output of the machine learning model rather than the input required to actually run it. So one thing we've been trying in the background here is called Decode on Head, which is essentially taking the machine learning models that right now we're running on MacBooks that our monkeys are gaming on and moving those to actually run on the implant.

And this is the like a super cool engineering problem. If you want to talk about how to make complex neural networks run on what is the equivalent of a garage door opener, come talk to me. It's fun.

Yeah. So that's another way to solve this problem is to basically do the computationally intensive work to just get the raw signal that you actually care to use to control something and then send that thing out of the implant. On the radio side, I'll hand it over to Matt.

Yes. So to answer your question, we are looking at other radio technologies. Technologies. One in particular is 500 MHz band with ultra wideband at a couple different frequencies. So this has an advantage in terms of the bit rate that you can achieve. It's on the order of 6 to 8 to 10 megabit.

There's also a latency improvement that's quite substantial. And there's also another wireless technology that we're looking at and W band.

Hi, thank you all for really clear and compelling presentations. Something that struck me in one of the earliest talks, I think it might have been DJs was this vision for the ability to acquire new complex skills via these BCIs, like the ability to perform Kung Fu. And that reflects the fact that the brain is fundamentally a learning machine. And yet many of the technical solutions presented later framed were framed in such a way as to try to correct for the way the brain changes over time over longer time scales, drift over the course of days, or the way that the tissue might heal over time.

I was curious what you your vision collectively was for developing out this technology that interfaces with a fundamentally plastic system that changes in complex ways over a variety of time scales. Days, months, years.

It's a tough question. I think it will be kind of maybe bidirectional learning in some way in there. Sometimes scores that we will might fix our algorithms and we prefer to have like more stable kind of performance. But of course, if the over time the person in the brain will learn how to use better the bci. We'll need to update our models. So there will be kind of in an interactive kind of relationship in some way to learn even new tasks.

These probably will be something over time we'll need to learn what the person kind of learn how to interact with the computer and then build the appropriate interface UX and also the UI and build algorithms that will help him to control what we want.

Just one thing I'd like to add on to what Nir said Yeah, it actually is an advantage in some ways that the brain is plastic and learns, and that can help us because we actually have to do less work and the human in the loop will actually learn how to use our device better. But one of the advantages of our particular approach and device is that we are trying to do an extremely high channel count device so we can, you know, uniformly distribute electrodes over a functional region.

And then it doesn't matter so much whether things move or shift over time. We can offload that to software and so we can build algorithms that change over time as well. And so both those things are actually, I think, advantages to our particular approach.

We have another question from Twitter. Juan wants to know, what career path do you suggest for somebody that is just getting out of high school if they want to work at neuralink in the future?

It's really any of the skills that we described. So we're developing new chips.

There's material science, there's software, obviously animal care. It's really all the things that we listed in the neuralink careers that would

be a good guide.

Yeah, I'm actually very fond of saying when you flip through any college booklets and look through all the majors, I think you can pull point to every single one of those majors. And there's someone at this company who either is an expert or, you know, have majored in that. So it really is truly, truly multidisciplinary endeavor. And I think, you know, just focus on whatever you're, you know, passionate about or whatever you're talented at, and then just, you know, pursue that as deeply as you can.

And then there's definitely going to be a place for you in your own building neural interfaces.

Hey, we got to see the monkeys doing telepathy. But could you say a little bit more about the animal behavioral training kind of their lives and day to day processes? Sure. I'm Autumn. I am head of research services, which includes our animal care program. And as an animal welfare scientist, this is a topic that I'm deeply interested in. So our training program is outfitted mostly with behavior analysts who help us think about how to remove any of the potential aversives or frustrations from our training.

We think about conditioning as the primary, which includes positive reinforcement as the primary way to train.

Let's see, what else can I share with you?

Yeah, yeah. I mean, that may not be part of the behavioral training itself, but we think of animal welfare assessment in the framework of the three Rs, which is referring to refinement, replacement and reduction. And so when we think about refinement, behavioral training does apply in that way. And where we want to remove, specifically in research, restraint is one of the things we make a very top goal to remove. So you saw a lot of videos today where animals were walking up to their stations because we worked really hard to, to remove any requirement to restrain the animal.

Anything else?

Well, just on top of the last point, you said just as an engineer here, one of the things that is really inspiring and really cool about this place is that we do get to work on a lot of technological innovations that directly translate to greater independence for the animals when they're engaging in these tasks. So as you saw monkeys charge just by voluntary walking up to a branch, they play games in their home habitat with a laptop computer voluntarily.

And the fully implantable, fully wireless device, the inductive charger, all these things enable that kind of experience. And so this is one of the very cool parts about working here is we do get to innovate on things like that.

Definitely helps to work with a group of engineers who can like really make cool stuff for monkeys to be able to do easier behavioral training.

So I guess to answer the previous question about what you can study to be part of neuralink, I guess monkey engineering, you can add to that monkey business.

Hello.

My question is on upgradability, which you guys mentioned quite a bit. So in that procedure, in there's some kind of explant procedure and then you're going to put in a new set of implants. So could you talk about the damage possible, if any, tissue damage from the explant procedure? How long you have to wait? Do you implant the same areas and what's your like brain scanning for the implant procedure in terms of upgrading it?

I don't know how many questions I

can ask,

so I can start to speak to some of those. So I work a lot on upgradability and those explant processes and designing those to be better.

The goal that we're working towards is that as I mentioned in the presentation, it's really just as easy to upgrade an implant as it is to initially install. We didn't, we didn't show many of those explant examples today, but we've come pretty close to just popping out an implant and reinstalling another one in the exact same location. Definitely, definitely the goal. We are installing the implant in primary motor cortex, which is a valuable area for interacting with a device like this.

And so we. The goal is to implant in the same location. Maybe if you expand out to other applications Then you'd be interested in moving somewhere else. But we definitely want to be able to insert into the same area.

In terms of damage, the. I think that the damage that we care most about is damage within the brain. And what we found, and we talked about that, that challenge of the tissue layer on top of the brain. And we're well on our way towards figuring that out.

But because of the thread's small size, the sort of scar capsule within the brain is so minimal that they are actually removed quite easily. And so we see useful signals even on the second or third time that you've placed an implant. And I think some of our BCI folks probably speak to that. We do have monkey participants working with their second devices and really making use of those.

So one, two questions. One was somebody had asked the question about the plastic. Have you noticed any plasticity from a behavior perspective from any of the monkeys?

Or is it too soon to tell? Or there haven't been any observations.

From the monkey behavior. We see that it takes them a while to learn how to, of course, to train on the test, but also when they are implanted. And it's relatively quickly for them to ramp up and get to a high performance of brain control. With Pedro, for example, after a few days he was able to, like, three days already able to learn very quickly to use the device. He was trained on the task, of course, from his previous implant, but with the new one, he was, after three, four days, he was able to control to close the performance of he had with the previous implant.

But have you noticed anything on the advance, which means the brain has outpaced the neural network that you're running?

It's hard to say. No, not really.

Okay, so I have another question, which is more about the electrical side. So you talked about 10, 24 channels being recording. Are you transmitting the raw signal or was it only the three spike events that. That you were talking about, the low, mid and the high, or is it the raw, entire raw form waveform that you transmit?

Yeah.

Hi, I'm Julian. I can speak a bit about this and maybe Avinash wants to contribute. But our chips see the raw signals, but the one we transmit out typically spikes, and we detect those spikes in real time on the chip. This massively compresses the data, I guess. Yeah, moving. We're making improvements to that, but we can request. We can request raw samples. Sometimes we also process particular statistics or other data directly on the chip and then send out the calculated values.

So there are many ways to sort of play with the Data.

Yeah. So at least with the current N1 system that we have, which relies on BLE radio, there is a bandwidth limitation. So you can't actually stream raw data from all 10, 20, 24 channels. But just kind of to give you a little bit of a history of how our compression algorithm, the spike detection algorithm was developed, we did have sort of a wire system. There was a paper that we published with the USB C connector that you know, streams all those signals through a high bandwidth wire connection.

So we did have kind of those development platform to be able to see the raw signals and know we, which set of information that we want to extract that are, you know, going to fit within the bandwidth of the radio as well as is useful for BCI control. And you know, also just sending data wirelessly does cost a lot of energy. So there's any opportunities we have to reduce that burden. You know, we try to do basically have all that compression closer to where the electrodes are as possible.

One thing that isn't obvious is that the actual bit rate that you need to control a phone or a computer is actually very low. So I think we might have the record for bit rate, is that correct? We think we do maybe so on the order of 10 bits per second. So that's super slow.

But if you think like when you're inputting data into a phone, like how fast your thumb's moving, how many thumb, what's your thumb taps per second. Pretty, pretty low. And I mean basically our thumbs are like two slow moving meat sticks that we, you know, do this and it's like there's really a load, it's like a low bar is what I'm saying.

So for at least for output it's, it's a, you're getting, get 10 bits per second, you're holding ass. So and that's, you don't need Bluetooth anything. So you could practically send it out with beeps and bops, you know. So it's not, if you go, if you're going like a high bandwidth visual now you're, you know, maybe going to megabit plus. But it's, it's all well within Bluetooth or anyway it's just that is what I'm saying is that's not a constraint.

The data rate.

One other sort of like maybe notable item which we talked about in the presentation, but we think we can probably solve for doing the implant without cutting the dura. We can just do basically a bunch of holes through the dura, which is like, the dura is like the Big thick, orange rindy thing that contains the, that's up against the skull. If you don't pierce the dura, you know, if you don't cut the dura away and instead you have a bunch of tiny holes and insert the electrodes through the tiny holes into the brain, then the recovery time is ridiculously fast.

You know, you're not really losing much in the way of cerebrospinal fluid. It's, it's, you could, in theory, I mean, this could be like a, the whole thing could be a 10 minute operation like Lasik. Like, it's fast. It's not like a big laborious thing. It's super fast.

Just going back to the long term

use, I'm wondering if you have any pathology looking at scar tissue from many animals that have had long term implants. And along that lines it seems like there might be a little bit of a gap between use in medical conditions and healthy individuals from a safety perspective.

I didn't quite catch the last question, but I'll hear the first one and I'll ask you to repeat the second one. So the first one is, do we have pathology from long term use animals? We absolutely do. We don't have any pathology from our monkeys, which we upgrade and you know, are still going. We have other studies that are primarily to determine safety. And so we do have histopathologic endpoints that we determine. The scar tissue formation around the threads themselves in the brain is typically negligible.

Like it barely reacts to the threads at all. So that's very promising in terms of the scar tissue formation over the cortex. So this neomembrane growth that fills in the, the areas that Elon and Alex were mentioning, that we remove with our current operation. Those we, we do, you know, evaluate that scar tissue, but it isn't, it doesn't pose a problem in any way. It's not a continuous reaction to a foreign body. It's just filling in tissue that was removed.

And if you could repeat the, the second question.

I didn't hear that.

Yeah, the second question, really following up

on that, seems like there might be a little bit of a gap in use in healthy individuals from a safety perspective. You know, I think people mentioned that they might be interested in trying prototypes, but just wondering what your perspective is on trying to lower the safety risks.

Yeah, it's a great question. So in terms of, really it's about the long term use of the device. So, you know, we have devices that have been implanted, like I said, in monkeys, where you know for many years where we see no behavioral deficit at all. So then this first is a question of how you evaluate safety. So you have histopathologic endpoints you can evaluate, but we're also looking for cognitive deficits or behavioral deficits as well.

And we don't see any of those in our animals, which is an important point. In terms of the histopathologic endpoints, they look really, really great. The challenge is one of explanting the device, which is why we're putting so much effort into the reversibility efforts and our through dura insertions. So when removing the device, that's when you potentially, potentially could cause damage. And so we are doing, we have a lot of ongoing studies right now to really minimize the risk of that, but we don't think it's a substantial risk with our current approach.

And like I said, pager was upgraded with the previous surgical approach and is doing great. So clearly it is, you know, can be perfectly safe. But proving that beyond a shadow of a doubt for humans is something that we're still working to do rigorously.

Did that answer your question?