2020年8月,Neuralink发布会的第一个画面是一块硬币大小的芯片,最后一个画面是三只猪。中间两个多小时,马斯克把"来投简历"这句话说了至少五遍。
开场第一句他就把话说死了:这场演示的主要目的是招人,他们没打算融资。这话在开头讲了一遍,结尾又讲了一遍。
一场给你看活猪脑子里实时神经信号的发布会,主要功能是招聘广告。这个反差本身就说明了很多事:技术路径已经想清楚了,缺的是人。
神经元像电线,电的问题得用电子设备解
马斯克给Neuralink定的目标只有一句话:用无缝植入的设备,解决重要的脊柱和大脑问题。
他接着讲了一件他自己觉得很多人没意识到的事。人的视觉、听觉、触觉、痛觉,全都是神经元发给大脑的电信号。既然是电信号,就有被读取、被修正、被重写的可能。
顺着这个逻辑往下推,能解决的问题清单长得出奇:记忆衰退、听力丧失、失明、瘫痪、抑郁、失眠、剧痛、癫痫、焦虑、成瘾、中风、脑损伤。
神经元像布线。既然是电的问题,就得用电子设备去解。
他还补了一个更冷的判断:如果活得够久,几乎每个人都会碰上某种大脑或脊柱的问题。从轻微的记忆衰退到严重的脑损伤,这件事只是时间问题。所以植入式设备迟早不是给少数病人准备的。
用气锤往脑子里打钉子
为了说明自己到底在改进什么,马斯克先把现有的天花板摆了出来。
医学研究里有个东西叫犹他阵列(Utah array),大约100个通道。它是一块刚性尖刺组成的床,植入方式是拿气锤直接打进大脑。马斯克评价说,这听着确实有点不舒服。
用的时候头上要连线和盒子,有感染风险;走在街上顶个盒子,外观也说不过去;操作必须有专业医生在场;全世界只在几十个人身上做过。
但它证明了一件事:读取人脑神经元,能做到。
普通人目前能接触到的方案叫深脑刺激。往脑子里放少量电极,用电流刺激。马斯克打了个比方,这有点像踢电视机,有时候真能修好,但不是每次都灵,而且能做的事有限。
它读不了也写不了高带宽信息。尽管如此,这套笨办法已经帮到了超过15万人。
Neuralink要做的是把这个数量级往上抬。先抬100倍,再1000倍,再10000倍。
一块硬币,替换你的一小块头骨
一年前他们的设备还有好几个部件,其中一个要卡在耳朵后面。复杂,而且做不到"看起来完全正常"。
现在的版本简单到只剩一个东西,大小接近一枚大硬币。它嵌进颅骨,替换掉取出的一小块骨头。导线从几厘米外接进来,连到附近的皮层。
马斯克自己也知道这个类比不严谨,但他还是说了:本质上,这就是一块装在你头骨里的Fitbit,带一堆极细的线。
0.9版原型有1024个通道,尺寸23毫米乘8毫米。人类颅骨大约10毫米厚,设备装进去和骨面齐平,外面看不出来。恢复之后只有一个小疤,藏在头发下面完全看不见。
我可能现在就有一个Neuralink,你不知道。也许我还真有。(马斯克)
除了神经接口,它还装了智能手表上那些传感器:运动、温度、压力。这意味着它同时是个健康监测设备,能在心脏病或中风之前给出预警。当然也能做点日常的事,比如放音乐。
充电方式是无线感应,和充手表、充手机一样。白天用,晚上充,第二天满电。马斯克说这叫彻底无缝,一根线都不用。
手术一小时内做完,不用全麻
设备做得好不够,还得有人把它装进去。Neuralink的答案是机器人。
流程大概是:切开一小块头皮,取出一枚硬币大小的颅骨,机器人插入电极,设备填回取骨的位置,缝上。缝合用的是医用胶水,很多伤口都这么处理。做完就能下地走。
马斯克说他们有信心把整个植入过程压到一小时以内。早上进医院,下午出院,不需要全身麻醉。
机器人的目标是整台手术全包,从切口到缝合。目前所有植入手术都是这台机器人完成的。马斯克给的理由很直接:想达到所需的精度,只有先进机器人这条路,人工做不到。
插入电极的过程有个反直觉的地方。按理说,拿一根线扎进脑子,肯定会出血。
实际不会。电极极细,机器人先对大脑成像,避开静脉和动脉,然后以极慢极准的方式插进去。在足够小的尺度上,没有可见的出血,也没有可察觉的神经损伤。
马斯克在台上专门点了这个现象:你拿铁丝扎东西肯定会流血,但在极小的尺度上,它不会。
三只猪,分别回答三个问题
演示环节被马斯克叫作"三只小猪",用了童话的梗。三只猪各自回答一个问题。
第一只叫Joyce,没植入任何东西。作用是参照组:健康、快乐、正常的猪长什么样。
现场出了状况。Joyce缩在猪圈角落,专心对付地上的食物,怎么叫都不肯过来。马斯克站在台上说,这才说明是真直播。
第二只叫Dorthy。她曾经植入过Neuralink,后来取出来了。
这一只回答的是可逆性。装进去能取出来吗?取出来之后还是健康的吗?Dorthy的表现是:健康、快乐,和一只普通猪没有区别。
这件事在长期看比什么都重要。设备会迭代,第一代用几年,第三代第四代出来的时候人总得能升级。不能取出的植入物是死路。
第三只叫Gertrude,植入已经两个月。她一出来,现场响起了连续的蜂鸣声。
那是她脑子里Neuralink的实时信号。设备连接的是她鼻吻部的神经元。她拱来拱去、碰到东西的时候,神经脉冲就会爆发。屏幕上一千多个电极对应的点会同时亮起来,底部的蓝色柱状图记录该区域的脉冲累积量。喂她点吃的,反应更明显。
现场还展示了一件更激进的事:双植入。当时已经有三只猪身上同时装了两个Neuralink,都健康,都能正常输出信号。也就是说,一个脑袋里放多个接口,技术上成立。
她是一只健康快乐的猪,植入两个月,工作良好。(马斯克)
然后是一段跑步机视频。研究人员让猪在跑步机上走,读取神经元信号,预测四肢关节的位置,再和实际位置对比。两条曲线几乎完全重合。
一个无线的脑植入设备,能以很高的精度预测全身四肢的位置。
读之外,还要能写
读信号只是半件事。另一半是往大脑里写,也就是刺激神经元。
难点在于对电场在空间和时间上的精确控制。不同脑区需要的电流差别很大,有的要极细微的刺激,有的要大电流,同时不能在长期造成损伤。
Neuralink用双光子显微镜观察刺激效果。红色闪动的是电极在放电,绿色的是神经元在响应电极电流。不同脑区被点亮的过程能直接看到。
这里有个数字值得记住:精确控制电场之后,一个电极可能影响到一千到一万个神经元。也就是说,植入1024个电极,实际能影响的神经元数量可能是几百万。
设备的能力清单:1024个通道全部可读、全部可写,全天续航,过夜充电,无线连接手机,通过低功耗蓝牙通信,应用程序跑在手机上。
FDA给了突破性设备认定
2020年7月,Neuralink拿到了FDA的突破性设备认定(Breakthrough Device Designation)。这个认定不批准上市,但意味着审评通道更快、和监管沟通更密切。
马斯克在这一点上表了态:会显著超过FDA的最低安全标准。他拿特斯拉做类比,法律允许造一星车,但特斯拉只造五星。安全这一项,他们打算拉满。
然后是招聘的部分。他要的人包括机器人工程师、软件工程师、机械工程师、电气工程师、芯片设计,以及一类他特别强调的:做过产品、把产品真正交付出去的人。做过智能手表、手机、复杂电子设备或者高级医疗器械的都算。
他特意说了一句:你不需要有大脑方面的经验。很多人以为不懂脑科学就来不了Neuralink,其实那些东西可以学。
问答:芯片里到底发生了什么
第一个问题,脉冲检测(spike detection)是在ASIC上做的吗,能不能软硬件升级。
回答的是Paul,做过数字芯片设计,现在做信号解码算法。
检测神经脉冲,传统做法是离线录数据,事后找特征波形。Neuralink要的是在线检测。最简单的办法是设阈值,超过就记一次。他们做得更细,在芯片上直接找特定的波形特征。
每个电极都做了带通滤波。可以把这些电极想成一群小麦克风,实时过滤,只挑出想要的形状。
原始数据量太大,蓝牙根本传不出去。屏幕上那些漂亮的脉冲曲线,是芯片直接输出的检测结果,不是原始波形。
顺带解释了什么是脉冲。神经元之间通信的主要方式,叫动作电位(action potential),可以理解成一个在时间上发送的数字信号,1或者0。一个神经元会把信号发给成千上万个下游神经元。
先做低带宽,先救人
怎么让植入过程更简单。答案有两层:设备做得更小,机器人接管更多。人做手术会累会出错,机器人不会。
神经外科主任Matthew McDougall补了一句,机器人目前只负责操作线程,要扩展到整台手术,基础科学层面没有障碍。过去没人做,只是因为没这个需求。要服务几十万上百万患者,整台手术必须自动化。
最先做哪些低带宽应用。神经科学家Joe Odori说,恢复运动和沟通是最直接的价值,对象是脊髓损伤患者。学术界已经有不少好看的演示,Neuralink要做的是把它变成能带回家的产品。
中间夹了个玩笑问题:Neuralink能不能让你用心灵感应召唤特斯拉。马斯克的回答是"绝对可以,100%"。这种问题他一向配合。
设备API怎么演进。Max回答了这个问题。
1000个通道的时候,神经脉冲数据可以全传到手机上处理。通道数涨上去之后,无线电会成为瓶颈,更多计算得放回头上的芯片里,编程模型也会跟着变。
开发者拿到的东西,要么是打包好的脉冲,要么是原始波形,用起来像接一路音频流。1024个通道的无损压缩数据流,大约100 kbps。
会不会用于游戏。回答很干脆:会,100%。
马斯克给了个有意思的功能基准:判断设备在人身上是否好用,一个标准是让四肢瘫痪的人能不能顺畅地打星际争霸。这个目标真实可测,也能拿出来检验。
能进多深,能撑多久
能不能进到大脑更深处。目前只做皮层表面。计划是改设备和机器人,做到任意深度。
皮层能解决的事情已经很多:失明、瘫痪、听觉。很多低级处理都在皮层完成,运动意图、感觉信息、听觉、视觉。实际插入深度是皮层下3到4毫米,已经跨越多层。
更深的地方有下丘脑这类结构,和抑郁、成瘾、焦虑有关,迟早要接。技术上主要差在电极长度和机器人调整。现在的线程43毫米,可以调长。
机器人团队负责人Ian说,执行器本身不受限于6毫米,可以更深。难点在于深层血管的规避,这是个有意思的问题。
最难的问题是什么。马斯克说,材料科学,尤其是电极的长期植入。
微加工团队负责人Felix解释得更细。线程由导体和绝缘体组成。脑内环境腐蚀性极强,电极要在里面待几十年。
这里有个两难:要读电信号,要写电信号,又不能被腐蚀。绝缘层必须极薄,又必须长期稳定,厚度不能随时间变化。
马斯克认为长期看碳化硅是最合适的绝缘材料,但它难加工。另一个方向是降低电极和神经元界面的阻抗,让电极做小之后还能清楚地说话。
多薄。一根头发直径约100微米,可以分成20份。现在的厚度是5微米,未来能到亚微米。但线越细,采集信号和施加刺激越难,因为电流通过的截面积变小了。
读取和写入的速度。芯片设计负责人DJ Seo给了参数。
1024个通道全部支持记录和刺激。记录端有片上算法做压缩和特征提取。信号以20 kHz数字化,10位分辨率,也就是1024级。神经脉冲的宽度大约1毫秒,采样速度是它的20倍。
脉冲检测在不到900纳秒内完成。刺激端可以生成任意波形,时间分辨率7微秒。
他还说了一句,这些速度比大脑本身处理信息的速度快得多。
团队规模。现在大约100人,长期看可能是10000人甚至更多,两个数量级。
有人开玩笑说,按通道数和人数的超线性关系,等做到百万通道,公司得有一百万人。回答是:招聘配额由通道数决定。
抗干扰。现在用低功耗蓝牙,2.4 GHz,和市面上的蓝牙设备共存。演唱会这种人多的场合信号会变差,频段太挤。
新一代无线电在换频段,目标是传更多数据,支持百万级电极。电磁兼容性上会按FDA的公开指南做大量测试。
第一批病人是谁
第一批临床应用针对截瘫和四肢瘫,也就是颈段脊髓损伤。计划招募少量患者,先验证安全和有效性。
马斯克描述了严重脊髓损伤患者的使用场景:连面部肌肉都几乎无法控制的人,靠想就能输出文字、打字、控制电脑和手机。
长期方案更激进。在脊柱损伤位置的下方再放一个植入,做成神经分流,把上面的信号跳过去。脊髓断了等于线断了,把信号跨过断点送过去,人就能重新走路、重新用手。
他说自己对长期恢复全身运动有信心。断掉的线,只要能有办法跨过去,功能就能回来。
身体对植入物的攻击怎么防。机械工程师Robin负责封装。
身体不是个友善的地方。他们的办法是造加速老化测试舱,模拟体内条件,用尽可能快的速度跑。目前受电子元件耐温限制,但子系统可以单独测。植入前,同一架构的设备都在测试舱里跑过将近一年。
要防的东西包括化学腐蚀、压力、机械冲击和振动。他们把所有环节放在内部做:镀薄膜金属、玻璃焊接、激光微加工,需要什么封装就能设计什么封装。
能不能保存和回放记忆。马斯克说可以。
他自己也承认这听起来越来越像《黑镜》,然后补了一句,那部剧预测得挺准。
逻辑是:如果有了完整的脑接口,记忆里编码的一切都能读取,就能备份、恢复,最终可能下载到新的身体或者机器人身体里。
未来会很怪。(马斯克)
猪为什么这么乖。动物照料项目负责人Autumn说,核心是让动物自愿参与。她指了指不远处的猪圈,说有时候它们就是选择不参与,那也没关系。
所有行为研究都用正向强化,动物有选择权。兽医Sam补充说,整个公司从管理层到一线都参与照料,目标是让动物能表现出物种特有的行为。
所以你能看到猪在稻草里拱来拱去,一脸满足。马斯克说,猪脸上天生带着小小的微笑,这点确实讨喜。
用什么编程语言。答案是很多种。芯片开发用Verilog,做寄存器传输级的底层设计,往上走是C、C++、Python。
真正的回答是:会哪种语言不重要,重要的是懂不懂让系统跑起来的原理。机器人团队用的是实时底层平台C++,加一些Java和Python脚本,都是实用选择。
有人问能不能跑《孤岛危机》。回答是:能。
意识、安全和成本
最哲学的一个问题:这个设备能不能解释意识。
马斯克的回答比较长。他说,物理学的任何问题都是X到Y的映射。X是神经相关物,物理上正在发生的事情;Y是现象状态,也就是主观体验。
问题在于,历史上我们没法很好地观察X,也没法直接观察Y,除非那个体验发生在你自己身上。
一旦神经科学家能同时拿到工具、看到相关物、又亲身体验,这个难题会消失得很快。
然后他跑了一段宇宙演化史:宇宙从夸克和轻子开始,氢慢慢演化,某处氢开始说话,并且认为自己有意识。
他喜欢一个玩笑:往地球上轰足够久的光子,它会发射出一辆特斯拉。
设备安全。回答者强调隐私和安全是Neuralink的首要事项,尤其是脑数据的敏感程度。
脑数据的交互会加密和认证。Neuralink的特殊之处在于从芯片到源码全部自己做,可以在设计之初就把安全嵌进去,避免单点故障。
比如把蓝牙无线电这类敏感模块在硬件层面完全隔离,把通往大脑的接口和外部攻击面隔开。
公司内部有安全专家,也请外部做审计和渗透测试。
有人补了一个没人问但值得说的问题。微加工团队的Zach说,整个设计更像一个平台。电极尺寸、数量、设计都能改,不同脑区可以用不同方案。
Dan解释了为什么用猪。最初是因为猪的颅骨解剖和人接近,厚度和硬脑膜都相似。后来发现猪还有别的优势:能训练上跑步机,能做各种动作,鼻吻部在皮层占据的表征面积很大,容易刺激。
另一个原因是猪容易满足。给点稻草、给点玩具、给个同伴、给好吃的,它就很快乐,福利水平容易做高。
还有一点:猪是皮实的动物,像小坦克。它们拱来拱去,撞墙,互相顶头,动作比人大得多。设备在猪身上撑了两个月还正常工作,对人的鲁棒性就不用太担心。
成本。马斯克说发布时的价格不具代表性,初期会很贵,然后快速下降。目标是把包含自动手术在内的总价压到几千美元,和做一次激光近视手术(LASIK)差不多。
器件本身不会很贵,因为它大量使用手机、手表、可穿戴设备里已经海量生产的零部件,那些东西的年产量是几千万台。
内部架构。DJ Seo说数字部分是全定制的。
最难的问题是能量密度。电极越多,耗电越大。市面上找不到现成方案能同时做两件事:把微伏级的微弱信号放大到可数字化,再从里面找出要的东西。
所以只能自研ASIC。市面上没有任何东西是专门为记录脑信号设计的,用别的方案就是浪费电。
另一位补充说,除了神经放大器和算法是从零做的,周边很多系统借用了可穿戴行业已经量产的部件:蓝牙无线电、低功耗微处理器、小型传感器。
拿掉一块骨头换成设备,结构强度够吗。
回答是从第一性原理出发。骨头的力学性能是已知的,设备会承受的应力也是已知的,设计时可以覆盖所有工况。
更有说服力的证据还是猪。它们动得比人剧烈得多,翻滚、撞墙、互相顶。你不能跟猪解释"你头上有个植入物,注意点"。所以设备必须扛得住反复的头部冲击。
还有一个细节:植入体固定在颅骨上,但电极线很长,脑和颅骨之间有相对运动时,线不会把应力传到插进皮层的电极上。线程本身极细极软,会跟着脑组织的搏动一起动,这也减少了脑组织里的应力。
线程当前长度43毫米,可调。
每人说一个最想要的未来
最后一个问题,团队每人说一件最希望Neuralink实现的事。这段回答比前半场所有技术细节更能看出他们的动机。
做视觉神经科学的Dan说的是视觉假体。视网膜或眼部受伤导致失明的人,可以把摄像头直接接到视觉皮层,用几千甚至几万个电极重建图像。再往后,没失明的人也能用,做出某种抬头显示,像终结者那样。
马斯克把这个想法推得更远:可以给人超视觉,看到紫外线、红外线、雷达,任何频段,传感器可以动态切换。
Joe说的是心灵感应。他的理由是:把想法压缩成词很费力,而且压缩损失巨大。词的数据率太低,说话太慢。
马斯克接过这个话题,讲了他写那篇文章时的措辞。他说自己写的是"概念性心灵感应"(conceptual telepathy),有人看成了"共识性"(consensual)。他两个都要:非语言的、双方同意的、概念级的心灵感应。
逻辑是:大脑花了大量算力把概念压缩成词,再用更慢的语速说出来。如果能直接传输未压缩的概念,沟通效率会高出一个量级。
还有人说,设备的附带价值是把大脑变成一块能看的电路板。有了示波器,你会顺便学到大量关于大脑如何工作的知识,这对神经疾病研究本身就有用。
有人提到"被困住的创造力":闭上眼睛能想象出达利式的画面,但要把它画出来得练好几年。电极足够多、位置足够准,也许能直接读取那些原始的概念向量,把它展示给别人。艺术、音乐、工程三维模型都算。马斯克给它起了个名字:心理艺术(mental artistry)。
有人希望设备越来越不像技术,越来越像生物,长期稳定地和大脑融合;刺激越来越精确、越来越多维,最终大脑分不清信号来自内部还是外部,人就彻底融合了。
DJ说,他们的设备能在每个通道上同时读写,也就是同时刺激和记录。同时做这两件事比听起来难得多,但打开的可能性也大得多,这是和神经系统打交道的全部游戏。
有人想做焦虑、抑郁,甚至消除恐惧。他自己是攀岩运动员,说没有恐惧大概挺爽。马斯克补了一句:还是留一点比较好。
马斯克自己的答案有两个。一个是把人类的痛苦压缩到今天的一小部分,疼痛是痛苦的本质,也许终于能控制它。另一个,也是他认为最重要的:实现和人工智能的良好共生。
他的说法是,在边缘系统和皮层之上再加一层AI扩展,让世界的未来由地球上所有人集体意志的总和来控制。从存在性威胁的角度看,这可能是这类设备最重要的成就。
还有人说,这是把意识问题从哲学拉进物理学。一千年来关于意识的哲学写得很多,也很荒唐,真正卡住的是工具。工具变好,问题就会变成可测量的问题。
有人想要一个没有疾病的未来。设备不只读取电信号,还能读取大脑里的化学线索,在疾病发生之前就预警和干预。
有人想做的是帮那些遭遇不幸的人拿回一些功能:脊髓损伤、脑部疾病,一些彻底改变人生的事情。
有人提到自闭症谱系,想看看这个设备在他们自愿的前提下能提供什么支持。
最后一个人说的是记忆。每个人都会失去记忆,年轻时候的事已经记不清了,以后只会更糟。如果有一个记忆仓库,随时能进去,心情低落的时候调出好的记忆,想念某个人的时候把他调出来,这件事本身就足够改变人生。
这场发布会真正在说什么
回到开头那个反差。一场发布会的目的是招聘,说明在2020年这个时间点,Neuralink认为最难的部分已经想清楚了。
设备形态定了:硬币大小,嵌进颅骨,无线充电,蓝牙连手机。通道数定了:1024。手术路径定了:机器人全自动,一小时内,不用全麻。可逆性验证了:能装能取。长期稳定性在猪身上跑了两个月。FDA给了突破性认定。
剩下的全是工程问题。材料、封装、无线电、算法、量产、成本。这些问题需要人,而且是大量的人。
马斯克说公司现在100人,长期可能10000人。这个数字听着夸张,但考虑到目标是把一个植入式设备做到像激光近视手术一样普及,几千美元、一小时以内、当天回家,这个规模不算离谱。
他在结尾又强调了一遍:从今天到广泛可用、可靠、可负担的产品,中间有巨大的工作量。
四年后再回看这场发布会,很多承诺还没兑现,临床试验的进度比当时的乐观预期慢。但这场演示给出的信息密度依然很高:一个团队在公开场合把技术路线、未解难题、可选方案和商业化目标一次性摊开,连猪不肯配合的尴尬都留着没剪。这种坦率本身,比任何宣传语都更能说明他们在做什么。
如果你想认真跟进这件事
- 去读Neuralink那篇技术论文《An Integrated Brain-Machine Interface Platform With Thousands of Channels》,1024通道的芯片架构、线程材料、手术机器人设计在里面写得比发布会细得多。
- 把Neuralink和Synchron、Blackrock、Paradromics放在一起对比。Synchron走的是血管内路径,从颈静脉把支架电极送到脑血管附近,不用开颅;Blackrock的犹他阵列已经在人身上用了近二十年,通道数停在百级;Paradromics在做高通道数的穿刺式方案。这四家各自卡在不同的地方:Synchron胜在手术风险低,输在信号质量和位置精度;犹他阵列胜在数据积累,输在刚性和植入方式。看清这些差异,才能理解Neuralink为什么把赌注押在"柔性线程加机器人植入"上。
- 盯住clinicaltrials.gov上Neuralink的注册信息。发布会上的时间表是愿景,注册信息里的入组人数、主要终点、随访周期才是能核对的硬约束。判断一家脑机接口公司走到哪一步,看这个比看发布会靠谱。
- 想确认"柔性电极长期留在脑子里到底行不行",去查在体生物相容性的纵向研究,重点看胶质细胞包裹(glial scarring)的十年期数据。这块最不性感,也最要命,发布会里马斯克自己承认材料科学是最难的。
- 别把"演示"和"获批"混为一谈。FDA的突破性设备认定只是给了一条更快的审评通道,不构成任何安全或有效的结论。看到任何"已获批"的说法,先去查它是哪一种认定。
- 如果你是工程师,认真考虑一下简历。这场发布会把招聘条件说得异常直白:不需要懂大脑,需要的是把复杂电子产品真正交付出去的经验。做过手表、手机、医疗器械、机器人、芯片的人,门槛比想象中低。脑科学可以学,把东西从图纸做到量产,这个学不来。
- 普通读者最实用的判断标准只有一条:当这项技术开始进入临床,看它第一批解决的是什么问题。如果先解决的是瘫痪患者的沟通和运动,那是医疗;如果先讲的是记忆备份和超视觉,那是叙事。这两件事的时间尺度差着几十年。
最后说一句可能不太讨喜的话。
"三只小猪"和"能打星际争霸"都会被人记住。马斯克在讲材料科学时那句判断,含金量其实更高:电极要在脑子里待几十年,而脑子是个腐蚀性极强的环境。
所有关于意识、记忆、超视觉的想象,最后都要回到这一层:一根五微米粗的线,能不能在盐水和免疫细胞中间,安安静静地待上二十年,还不失效。
这个问题没有发布会。有的只是实验台、加速老化测试舱,和一批愿意在这上面耗掉职业生涯的人。马斯克那天反复讲招聘,讲的大概就是这个意思。
0:52All right, welcome to the Neurolink product demo. I'm really excited to show you what we've got. I think it's gonna blow your mind. so the the the primary preface of this demo is actually recruiting. So I'm going to emphasize this at the beginning and then again at the end. we're we're not trying to raise money or do anything else. The the main purpose of this is to convince great people to come work at Neurolink and help us bring the the product to fruition. make it affordable and reliable and and such that anyone who wants one can have one and so I want to emphasize the the purpose of neural link like what do we what's our goal our goal is to solve
1:35important spine and brain problems with a seamlessly seamlessly imp implanted device so you want to have a device that you can basically put in your head and feel and look totally normal but it solves some some important problems in your brain or spine. And the reality is that almost everyone over time will develop brain and spine problems. And these range from minor to very severe, but if you live long enough, you you everyone's going to basically have some kind of neurological disorder. And these range, you know, from memory loss to brain damage. But the the thing that's important to appreciate is that is is is that a an implantable device can
2:25actually solve these problems. and I think a lot of people don't don't quite realize that. but all of these the all of your senses, your sight, hearing, feeling, pain, these are all electrical signals sent by neurons to your brain. And if you can correct these signals, you can solve everything from memory loss, memory loss, hearing loss, blindness, paralysis, depression, insomnia, extreme pain, seizures, anxiety, addiction, strokes, brain damage. These can be insult these all can all be solved with an implantable neur neural link. This is an
3:09extremely fundamental thing and I think a lot of people don't quite understand that. the neurons are like wiring. and you kind of need an electronic thing to solve an electronic problem. So current medical research it we'll just go through what is the state-of-the-art in medical research and then and then what's the state-of-the-art in what consumers or people in general can get. So the current medical research has shown that you can read neurons in a human's brain. So there something called the Utah array which has about 100 channels per array. but it's it's kind of like a it's a bed of rigid spikes that's literally inserted with an air hammer. so you know that's
3:54slightly discomforting I think. and there's a big there's there's wires and a box on your head and so it's some infection risk. and obviously it will look pretty weird if you walk around with boxes on your head. and in order to use it, you have to have an expert medical profession professional there.
4:12And it's only been done in a few dozen pe people. So but it is a served as an important important proof of concept that this can be done. So we did want to point this out and and show that this is actually does work. it's just not something that the average person could could use effectively. And in terms of what is currently available, there is something called deep brain stimulation where they put electrodes, a small number of electrodes in your brain and will actually zap your brain with an electric current. and it's it's it's valuable for its uses, but it can't read or write high bandwidth information. I would say this is sort of a bit like sort of
4:57kicking the TV, which does work u but not always and it has limitations. nonetheless, this has greatly helped over 150,000 people. and it's so it's it's actually just despite being somewhat of a brute force approach, it has been very effective for a lot of people. And this is what's currently available. So, we want to radically improve this by multiple orders of magnitude. Improve it by a factor of a 100, then a thousand, then 10,000.
5:26So going into the neural link architecture, what we've done over the past year is dramatically simplify the device. So we we we about a year ago we had a device which had multiple parts including a piece that it had to sort of sit behind your ear. and it was it was it was complex and you and you wouldn't still look totally normal. You'd have a thing behind your ear. So, we've simplified this to simply something that is, about the size of a large coin. and it it goes, in
6:05your skull, replaces a piece of skull. and the wires then then connect within a few centimeters or about an inch away from the device. and this is sort of what it looks like. So yeah, this is our little device. It is that that thing at the bottom is just to hold the threads in place because they're just like little fine wires.
6:40I mean, frankly, to to sort of simplify this, what what we're I mean, it's more complicated than this, but it's in a lot of ways, it's kind of like a Fitbit in your skull with tiny wires. So, and it's, yeah, so our current prototype version 0.9 has about a thousand channels. so that's about a hundred times better than the the next best consumer device that's available and it's a 23 mm by 8 mm. It actually fits quite nicely in your skull because your skull is about 10 mm thick. So it fits it's it goes flush with your skull. It's invisible and all you can see afterwards is that there's a tiny scar and if it's under your hair you can't see it at all. In fact, I
7:31could have a neural link right now and you wouldn't know. Maybe I do. So, and it's also got all the things that you would expect to see, the sensors you'd expect to see in a smartwatch, or a phone like, notional measurement, temperature, pressure. so there's actually a lot of functions that this device could do, related to monitoring your health and warning you about a possible heart attack or stroke or other, damage, as well as, sort of convenience features like playing music. could do a lot. sort of like if your phone went at your brain or something.
8:07yeah, maybe that's not a great analogy. anyway, so it's also inductively charged. So it's charged in the same way that you char you charge a smartwatch or a phone. and so you can use it all day, charge it at night and have full functionality. So you would really you it would be it would be completely seamless. and yeah no wires in terms of getting a link. So we you need to have the device a great device and you also need to have a great robot that u puts in the the electrodes and does the surgery. So you want the surgery to be as as automated and as possible and the only way you can achieve the level of precision that's needed is with an advanced robot.
8:53so we're really looking for great people who can help develop both the device and the robot. and we feel confident about getting the the link procedure, the the installation of a link done in under an hour. so you can basically go in in the morning and leave the hospital in the afternoon and it can be done without general anesthesia.
9:16So, in terms of getting a link, like I said, it's essentially you open a piece of sculp. you remove about a coin sized piece of skull. and then the robot inserts the electrodes. we'll talk more about that later. then the device replaces the portion of skull that was removed and we we basically close that up with actually super glue, which is how a lot of wounds are closed. And and then you can just walk around right after right afterwards. It's pretty cool.
9:51So this is our surgical robot. And we actually ultimately want this robot to do essentially the entire surgery. so in in everything from from incision removing the the skull, inserting the electrodes, placing the device, and then closing things up and having you ready to to leave. So we want to have a fully automated system.
10:21And we've to clear this this robot does actually work when we've used it for all of the implantations. so this this shows you at sort of a close-up view which I think is actually not too gruesome of the electrodes being inserted in the brain. And if you look closely, you'll see that that's it's a little counterintuitive that if the electrodes are inserted very carefully that there is no bleeding.
10:53and so the if you have very tiny electrodes and if they're inserted very carefully so the robot actually images the brain and make sure to avoid any veins or arteries so that the electrodes can be inserted with no noticeable damage. So you will have no noticeable neural damage in inserting the link.
11:17Yeah. Like you sort of think like if you stab something with a wire, surely it will bleed. But actually at a really small scale, it does not. So does it actually work? And what I'm excited to show you, I call like the three little pigs demo. and if our animal has this, we're bringing out the the pigs. And what we're going to show you is a well, I'll walk right over and show you.
11:56So, what we have in pen number one is Joyce. and she does not have an implant. Obviously healthy and happy. [laughter] we're trying to get go through it out. This is how you know it's a live demo. She's a little she's like trying to eat something in the corner of her pen. come on. Go. Let's go.
12:26Snacks are this way. Oh man. All right, we'll we'll wait. We'll give Grude a second and we'll move on to Darthy. Sometimes the pigs are a little shy. So, here's Darthy. and in the case of Darthy, Darthy used to have an implant and then we removed the implant. So, this is an very important thing to demonstrate is reversibility. So if you if you have a neural link and then you decide you don't want it or you want to get an upgrade and the neural link is removed is it removed in such a way that you are still healthy and happy afterwards and what Darthothy illustrates is that you can put in the neural link remove it and be healthy happy and indistinguishable from a
13:14normal pig. Oh thanks Darthothy man. Got are you serious? [laughter] Okay. well, should we bring them all out or something? Would that be better? How about everyone? Everyone just comes into this pan. How have a few? It'll be a little crowded, but whatever. [snorts] Okay. All right.
13:48Well, is is GO still back in the thing? Yeah. Okay, we need to bring it out. Yeah. [laughter] Oh, the beauty of live demos. This is real live demo. All right, maybe we can zoom in to go true or something. She's clearly very interested in something at the back of a pen.
14:26Can you see her or All right. All right. This might take a sec. well, this this worked earlier.
14:58All right. Well, what what if we if we lift the curtain and then zoom in? All right. Here we go. Great. Okay. Great. [applause] Okay. This is a high energy pig. all right. Go, thanks for coming out. so what you're the beats you're hearing are real-time signals from the neural link in Gree's head. So this neural link connects to neurons that are in her snout. So whenever she snuffles around and touches something with her snout, the that sends out neural spikes which are detected here.
15:56and so on the screen you can see each each of the the spikes from the,024 electrodes and and then if you if she yeah if she shuffles around touches this now in the ground or you kind of feed her some food. Pigs love food. then you you can see the neurons will fire much more than when you're not touching this now. And that's what's making the the beeping sound.
16:24All right. Cool. So, as you can see, we have a healthy and happy pig. initially shy, but obviously high energy and and you know, kind of loving life. And she's had the implant for 2 months. So, this is a healthy and happy pig with an implant that is two month old. Two months old and working well.
16:45Yeah. [applause] All right. Cool. and then we actually have I'm sure if this works is so we said well what if we do two neural link implants and we've been able to do dual neural link implants in actually I think three pigs at this point and we have a couple of them here and we've been able to show that you can actually have multiple neural links implanted and again healthy and happy and indistinguishable from a normal pig. So, so it's possible to have multiple links in in your head and have them all be sending out signals and you're working well.
17:32All right. All right. So, we just showed you a demonstration of reading brain activity and see probably see that as I was saying each of those dots represents a neural spike and the the blue chart at the bottom is showing an accumulation of neural spikes in that region. So in and in in terms of additional u brain reading activity when we have say one of our pigs on a treadmill pig on a treadmill it's funny funny concept really and we take the the readings from the neurons and we try to predict the pos position of the joints. and so we say we have the predicted position of the joints and then we we measure the actual position of the joints. You can
18:31see that they're almost exactly aligned. So we're able with a wireless neural neural implant to actually predict the position of of all of the limbs in the pig's body with with very high accuracy. Now in terms of of writing to the brain or stim stimulating neurons we obviously need precise control of the electric field in space and time. We need a wide range of current for different brain regions. some some regions require delicate simulations, some require a lot of current u and and you want obviously no harm to the brain over time.
19:08and the way we part of the way we analyze the the stimulating stimulating neurons is with a two photon microscopy. I I always have trouble pronouncing that microscopy. and it's very impressive technology. You can actually literally see in real time how the neurons are firing. So the the red sort of things are the neurons red sort of flashing things are the neurons firing or I should say the the electrodes firing. So the red things are electrodes firing and then the green are the neuron bodies responding to the current from the electrode.
19:50So you can see them lighting up different brain regions. and then by carefully controlling the electric field, you can actually have one electrode influence possibly a thousand or 10,000 neurons. So although you might only have a thousand electrodes implanted, you could be influencing u millions of neurons.
20:12And this is just a similar chart showing stimulation at different power levels. So like I said for the initial device it's readr in every channel with about a,024 channels all day battery life recharges overnight has quite a long range so you can you could you can have the range the range being to your phone I should say that's kind of important thing this would connect to your phone and actually the so the the application would be on your phone and the and it would be communicating ating by by essentially Bluetooth low energy to the device in your head.
20:56that's why I say in a lot of ways it is like a foot in your skull with tiny wires. So and then like I said it you would not be able to see the device at all. It would you would look completely normal and just have a small scar u under your hair. And we're making good progress towards clinical studies. I'm excited to announce that we received a a breakthrough device designation from the FDA in July, thanks to the hard work of the Neuralink team. So, [applause]
21:27so I want to be clear, we're working closely with the FDA and we'll we'll be extremely rigorous. In fact, we we will significantly sign significantly exceed the minimum FDA guidelines for safety. We will make this as safe as possible. you know, just as with with Tesla, while it is legally possible to ship a one-star car, at Tesla, we the only cars we make are five stars in in every category.
21:53so we we actually maximize safety and we'll take the same approach here at NewLink. and then to emphasize again what the goal of this presentation is is recruiting. We want people who are great at solving problems to join the company and help us complete this device take care of the animals write the software create the chips and and and productionize everything. so we need like robotics engineers and we I think we also especially need people who have worked on on product worked on and shipped products. So if you've like shipped a smartwatch or a phone or you know any kind of complex electronics or complex device or advanced medical devices we'd love for you to contact
22:42us and consider working here. So and a very important point to emphasize is that you do not need to have prior experience on brains. So a lot of people think well I couldn't possibly work neural link because I don't know anything about how brains work. That's okay you can learn. but we need software engineering, we need mechanical engineering, electrical engineering, like I said, chip design, robotics, and all the things that a company needs to work. So please send in your resume. Actually, it's not just engineering obviously every everything at your link. So now let's actually move to questions. So, we've piled questions that have been asked over the internet
23:25and so we'll do live Q&A. so, bring in a bunch of people from the Neuralink team. All right.
24:04All right. So, if you have any questions, please submit your questions to the Neurolink Twitter account and we will try to answer as many questions as we can over the next hour and feel free to ask hard questions. It's no problem and we'll we'll do our best to answer them. all right, let's move over to the team.
24:23Perfect. Yes, we will continue monitoring questions until the end of the event. So, please keep sending them in over Twitter. Elen, do you want to come join us? All right, so the first question is, how is the spike detection implemented? Is it on the ASIC? Is it hardware and software updatable?
24:43Yes. yeah. I can answer that. So, I'm Paul. I work, I've had a few different hats. I've worked a little bit on digital chip design and more recently working on some of the algorithms, trying to decode the signals from the brain. And the spike detection algorithm, there's historically been many ways of detecting spikes.
25:06Typically, people will record data offline and then look for characteristic shapes. but here we're interested in detecting spikes online. And there's a number of simple algorithms that you can do for example like detecting just a threshold. And we're we've been interested in doing a little bit better than that. So we actually look for particular shapes character shapes that we think are spikes. And so we're doing this on the chip. it's for all 24
25:35electrodes. There's band pass filtering that's happening on the chip. So, if you think of these as like little microphones sending audio information, we're basically filtering all of that in real time and then looking for these characteristic shapes. you can configure what sort of shapes you're looking for and that information is what's being sent out. The raw information coming from electro is way too much to send out over Bluetooth. So, when you're seeing those beautiful spike raers on the screen, those are the detections that we we have coming straight from the chips.
26:04Got it. And just for everyone who doesn't know, can you just explain at a basic level what a spike is? Sure. So, so, traditionally, people think of spikes or action potentials as the electrical events that happen, in neurons and is the primary form of communication between neurons. and so this is a hysteresus event where you have currents that flow and generate this you can think of it as being a digital signal a one or a zero that's being sent in time where neurons will send that signal to often thousands of of recipient neurons.
26:38Awesome. Thanks, Paul. So, I got a question from Toby on Twitter. He asks, "What can be further done to simplify the device installation progress?" So, we're working on making the device as small as possible. Of course, with the robot taking more and more of the responsibility away from errorprone human surgeons, we're hoping to make the process faster and safer.
27:09Actually, it might be good for people to say, you know, who they are and what they do. I'm Matthew McDougall. I'm the neurosurgeon, the head neurosurgeon at Nurling. Thanks, Matt. Is there anything specifically from the robot that can be done to make this faster and safer and scalable to billions of people?
27:29yeah, I guess we've started with just implementing you know robot manipulating the threads. but we definitely need to expand to essentially doing the entire surgery with a robot. so there's nothing like really as far as I can tell fundamentally like stopping us from doing that. sort of like at a fundamental science level. it hasn't been done in the past I think probably just because like volumes of surgeries hasn't been needed but specifically for us to scale up to you know many hundreds of thousands or millions of patients we will need to automate like the entire surgery essentially awesome thank you Ian Garrett asks what are some of the lower bandwidth
28:09activities to target first is it muscle movement is it auditory signals what level of bandwidth is required for effective So there's a couple of just to say who you are and what you do. Hi, my name is Joey Odori. I'm a neuroscientist and neur engineer working on decoding from the brain. so there's some lowhanging fruit that I think can really be impactful to help many people's lives and that's restoring movement and communication in for example a spinal cord injured patient.
28:40and there's a lot of anticedants in the academic world where there have been very nice demonstrations of doing this and we think we can take our technology and really bring that to the home, something people can take home with them and improve their lives. Fantastic. Gil, there's a fun one next.
28:57All you Yeah. So, we're fielding questions from Twitter, so there's going to be some funny funny comments. first question is, who are you and what do you do? I'll lead with that. And the question is, can the Neuralink chip allow you to summon your Tesla telepathically? definitely. Of course, you heard it here first. That's a definite 100%.
29:18Carlos, that is the answer. just one bit of information. Yeah, it's very easy. That's an easy one. Actually, Max, this might be a question for you. Is yeah, like some of the questions we had on Twitter was how do you how do you see the Neurolink device and the essentially API growing over time and allowing developers to interact with the device?
29:43So, there's interesting questions longterm about whether you do decoding on head or on a phone or on a computer ultimately. So, at a thousand channels, it's possible to send all the spike data to a phone and do processing there. as this gets bigger and bigger, you're going to be more constrained by the radio. you're going to want to push more of this on head and that changes your programming model. but really the APIs you're getting are either bin spikes or you're getting raw waveforms and that's like you can consume it basically like an audio feed.
30:09yeah, it's it's worth learning like the lossless compression of the data stream is about 100 kilobits with,024 channels. Excellent. So, this one came up a lot and this particular question is from Yseph, but will this technology ever be used for gaming? Yeah, probably. 100%. Yeah, 100%. Yeah, I I think that a good benchmark of does it work well on humans is can a quadriplegic does it work well enough for them to play Starcraft?
30:38And it's a good functional target. Yeah, for sure. Yeah. Awesome. We'll pull out pull out back from the Twitter questions that are kind of funny. this is a question for the efficacy of the device. right now, is the device limited to surface layers of the brain only or can we go deeper? And if not now, what's holding us back from going deeper into the brain?
31:06We are planning on modifying the device and the robot to be able to sew into arbitrary depths of the brain. right now we're limiting ourselves only to the cortical surface because that simplifies many of the problems in involved with going a lot deeper and what what kinds of things can you solve on the cortical surface versus when you go deeper.
31:26Sure. So a lot of the low-level processing happens in the cortex in terms of motor intentions, sensory information that comes directly in so your hearing, your auditory percepts, your visual processing, a lot of that happens in the cortex. I mean you could you could solve blindness, blindness, you could solve paralysis, you can solve hearing, you can solve a lot with just by interfacing with the cortex. And to be clear, we actually do insert I guess about three or four millimeters into the cortex. So these these are these electrodes are sensing from multiple layers within the cortex.
32:10it's that there are deeper brain systems like that are underneath your cortex like the hypothalamus. and I think that is something we'd want to interface with for sure. because that's that's going to be important for curing things like I know depression, addiction, that kind of thing.
32:31anxiety. so, but yeah, like as like like I said in the presentation, overall we're aiming for a general purpose device. and really the thing that would would change between from interfacing with the cortex versus deep brain systems is kind of the length of the electrode. so we just need kind of longer wires and to adjust the robot in order to access deeper regions of the brain.
33:00Awesome. Ian, did you have any say the robot is designed like the full length is like seven centimeters or 8 centimeters. Yeah. Yeah. So, right now we're designed for like six millimeters. The actuator itself isn't actually limited to six millimeters. You can go much deeper. that's sort of in addition to scaling out sort of to the rest of the surgery.
33:18also scaling to deeper which is mostly a sensing problem. So the ability to avoid deep vascule is sort of an interesting problem that we're working on in the robotics world here. So yeah lots of cool problems in the robot. And you're you're leading robot engineering. Yeah. Sorry. I'm Ian. I lead the robotics program here.
33:38Question. Robot is a man's best friend. all right. Next question is what is the most challenging problem that must be solved in order to meet Neuralink's ultimate goal? Well, I I think one of the hardest problems is kind of I'd say material science and the especially the installation of the electrodes and what would you say?
34:02Yeah, absolutely. So, you want to introduce yourself? So, I'm Felix Deu. I currently lead the micr fabrication team. my team makes up of a group of process engineers and material scientists and our job is to make the tiny wires as Elon referred to them what we call the threads that are implanted in a brain and of course it's the threads are made of you know conductors and insulators and choosing the right material that is aminable to the brain and compact to the brain is something that our team have expertise on and and we work seriously ly on interface engineering just to make sure that our layers are actually well thought through well designed and at the end of the day
34:47when we impact this device they'll be actually functional for either reading or for writing in the brain yeah I think it's it's it's going to be important from in terms of of material science problems is making sure that the the threads though the electrodes can last for decades in the brain which is a tricky thing because it's It's a very corrosive environment. and you're you're you're in a dichotomy where you you want to read and you you want to sense electrical signals and you want to generate electrical signals and
35:22but you don't want to corrode the electrodes over time. So you need to have an insulating layer that is very robust but also very very thin. it can't it has to be just the right amount of insulation and it has to stay that amount of insulation over time. So this is why I think we think probably like a silicon carbide type insulator is is probably the best long term from a material science standpoint. But it's m silicon carbide is a tough material to work with. but it's it's some that's probably the right choice long term would you say?
35:53Yeah. Yeah I I totally agree. and it's also apart from its electrical properties it also help with adhesion between the metal layers and the and the poly substrate. other thing that we also looking at is since we actually reducing the electro dimensions the electro site the actual side that actually speaks to the neuron also have what we call high resistance to the signal. So we actually look into like material engineering that will reduce the impedance at that interface so that we can actually have a very small electrode but still be able to speak clearly to the to the neuron.
36:24And when you guys say thin, how thin? H how thin? so one strand of your hair is about 100 microns in diameter. we can think about dividing that into 20. So you know our one of the thickness that we're working is about five microns. and we have this possibility to go even thinner in the nearest future. Yeah.
36:47yeah, we think we can probably go sub micron in in thickness. But obviously as the thinner you the thinner you make the wire, the harder it is to sense the signal and to do stimulation because there's less cross-sectional area for the current to traverse. and yeah, in terms of of upgrades, that's part of why I wanted to show the the expplanted plague is to show that you can actually take the the device out and there's no observable behavior change. The the pig is as happy as before and healthy. And I think this is going to be important for upgrades. because obviously if you get an early device then by the time let's say you get version one probably
37:26by the time we have version three or four you want to you want to upgrade you wouldn't want version one of a phone and you know 10 years later everyone's got version three or four. So it's going to be important to be able to remove the device and upgrade it over time. So so over time the your the
37:45neural link would would do even more than it did before. Awesome. So, here's a question about the device. We have been talking a lot about read speeds and write or we haven't talked about read or write speeds. DJ, what are the read and write speeds of the device? If you think of it like a computer, sorry, when you say speed, what do you mean?
38:10Like how quickly can you read information from the brain and how quickly can you write information in the brain? Okay, so let's see. So one of the things to highlight is that we What DJ? What do you work on? Yes. My name is DJ Sa. I'm leading implant electronics development like the chip design and a chip designer.
38:31so the prototype that we were showing it has 124 channels. All of those channels are capable of recording and also stimulating. And in terms of recording, as Paul mentioned, we do have onchip algorithms to do level of compression and extracting the signals of interest. In this case, spikes neural spikes.
38:50Sure. And those things are actually happening at a much faster speed than your brain even is processing those information. Well, do you want to say like things like time resolution like you know within how many milliseconds can you detect a spike and what's the you know just some technical data? So in terms of the signals that we're collecting you know we're digitizing them at 20 kHz generally the signals of interest are anywhere between about a millisecond in in width and so we sort of sample it at 20 times that speed and we have analog to digital converters that divides that into 124 levels so 10 bit resolution and the spike detection is done in less than 900 ncond which is
39:36really really fast time and it's only sort of determined by the speed of the clock and for low making sure that our implant is low power we're running them actually at a very slow speed and then for stimulation you know we can actually create any arbitrary waveforms with about seven microcond of resolution so basically whatever you want to draw we can stimulate and generate those pulses through any com
40:05version one. as we go to version two, three, four, these things will expand. I think ultimately by orders of magnitude, many orders of magnitude. Great. everyday Tesla asks, "How big is the Neuralink team and how much do you expect it to grow in the near future?" there about a hundred people right now. I think over time there there might be 10,000 or more people at Neuralink. so I think yeah, the couple orders of magnitude.
40:33Yeah, exactly. Couple orders of magnitude is my favorite phrases. yeah, just, you know, couple of magnitude. It's no big deal. so yeah, pro probably the the number of electros will grow with number of people at the company. That's a super linear relationship. Yeah. Well, we might need a million people at the company then. [laughter] Yeah. Super super. Yeah. Even more.
40:59Yeah. Yeah. Yeah. Ideally, like Yeah. Yeah. Super linear. Our hiring quotas are determined by our channel counts. I think it was 10 to1 ratio or something. Yeah. Cool. So, Twitter asks, "How does the implant system fare against various disturbances from the outside like electrical, magnetic, Wi-Fi,
41:23radio?" Yeah. So the current version of the implant that we have is using Bluetooth low energy radio. So similar to any Bluetooth devices that are out there you know they they are able to coexist with other devices that are using and sharing our same spectrum in 2.4 GHz. obviously as you imagine you know when you go to concerts or there's just a lot of people the signal quality does degrade because it is a pretty congested spectrum. So we are actually working on some new versions of the radio that are operating at different frequencies to be able to also send out a lot more data and have it be scalable to you know millions of electrodes. and then also
42:05in terms of sort of electromagnetic compatibility and interferences obviously it's very important for us to coexist with other systems and just disturbances out there. So, there are also welldocumented guidelines from FDA that we'll be following and doing a lot of testing for. Fantastic. we're going to take an audience question. Fred, take it away.
42:28So, we all want to play Starcraft using Neurolink, but what are some likely first applications? Well, I don't know if we all want to play Starcraft. there are other games. There are [laughter] other games to play. Exactly. But also Starcraft. Yeah. so our our first clinical trial is aimed at people with paraplegia or tetroplegia. so cervical spinal cord injury. We're going to enroll or plan to enroll a small number of patients to make sure the device is safe and that it works in that case.
43:02yeah. So actually just to elaborate on that if somebody is like a severe spinal cord injury you know to the degree that they they even they have very limited control even over their facial muscles. then but but with with this implant you can actually think just just by thinking you can output words and you can you can type and you can control a computer control a phone and which is pretty pretty wild and I think something that's very exciting as a long-term application is if you can if you can sense what somebody's trying to do with their limbs what they want to do with their limbs then you can actually do a second implant that's at the base of the
43:51spine or or wherever just after wherever the spinal injury occurred and you can you can create a neural shunt. so we I I think long term I'm confident that long term it will be possible to restore somebody's full body motion. So if somebody even has a severed spine they will be able to walk again. They will be able to use their hands. and
44:17when you have a severed spinal cord, you essentially have broken broken wires. and so if you can just jump over those wires and transmit the signals signals over those wires, you can give somebody the ability to walk again naturally. So DJ, you talked about outside interference.
44:39Can anyone talk about the inside interference? So, how to protect the implant from the body? Sure, I can talk about that. my name is Robin. I'm a mechanical engineer and I lead the implant mechanical packaging and assembly team. So, actually making the devices and designing the seal. So, yeah, the body is not a friendly place to be. So, what we do is basically create an accelerated test chamber that can mimic the conditions that the devices will see in the body. and basically test the devices at as fast as accelerated rate as we can. at this at this time we're limited temperature- wise by u some of the electronics components but basically we can sub
45:23test subsystems to mimic the environment in the body and by that we've brought some devices in the tester right now to almost a year and all the implanted ones prior to these implants we had had tested the same architecture in the tester to validate it and yeah so there's obviously chemical things attacking the implant. there's there's pressure, there's mechanical shock and vibration. So all of these things are part of it. and yeah, I think one of the things that really helps is having everything insourced in house. So we have the capability to deposit thin layers of metal. We have the capability to weld glass and machine micron
46:10scale with lasers and yeah. So basically we have all the tools in house to be able to design any sort of ceiling enclosure packaging that we need to do. Awesome. Thank you, Robin. another question from Twitter. Will you be able to save and replay memories in the future? yes, I think in the future you will be able to save and replay memories. I mean this is obviously sounding increasingly like a Black Mirror episode. but well I guess
46:44they're pretty good at predicting. but yeah, essentially if you have a whole RAIN interface, everything that's encoded in memory, you could you could upload you could basically store your memories as a backup and restore the memories. and ultimately you could potentially download them into a new body or into a robot body. The future's going to be weird. Cool.
47:13So, I can't help but notice that the pigs back there are pretty well behaved and pretty darn cute. Can anyone talk about I'm looking at Autumn over here. First, can you tell everyone who you are and what you do here? And can you describe like how are they so well behaved and so smart?
47:32Well, I'm Autumn and I lead our animal care program and our philosophy here is to set up a system where the animals are able to volunteer to make choices to participate in our projects. As you see, sometimes they choose not to participate and that's okay. that's we want to make sure that our animals are happy and healthy. So, all of the
48:02behavioral research that we do is led by positive reinforcement and again that allows for them to choose to volunteer or not. Sam's also on our animal care team. Sam, do you have anything to add to that? Hi, I'm Sam. one of the veterinarians here work closely with Autumn taking care of the animals and yeah, it's just emphasizing what said the whole animal care program here. In fact, the whole company here is very dedicated to promoting the care of these animals.
48:29we, you know, everybody is involved. Everybody's very interested in making sure that they're taken care of. And yeah, the the program from top to bottom is designed to make sure that they can express their species specific behaviors, express you know, choose to do things. We don't force things upon them as much as possible. And yeah, that that's why you see the pigs happily rooting around the straw right now and generally being very content. They do have those natural little smiles on their faces which is ever delightful.
48:59[laughter] next question is what programming language are you guys using for devel developing the device? Several. Yeah, we use several different types of programming languages. so for chip development a lot of verilog for doing low-level programming at the kind of the register transfer level.
49:22as you go up higher and higher there are you know various C, C++, Python but to be honest like at the end of the day it's not you know which programming language do you know how to use it's do you know the principles behind you know coming up with just methods to enable these systems to work. so I think you know if you have any experience in programming and you've been building things since you're little you should really come join us.
49:50Yeah. But it can it play Crisis eventually. It can play Crisis. Cool. Crisis confirmed. And then Ian, what about [laughter] Yeah, Ian, what about the robot? What kind of programming languages and tech stacks are using over there? I mean sort of just reiterating what DJ said. sort of doesn't matter.
50:14like specifically for us it's there's real time low-level platform C++ some Java and Python scripts but really those are more just like pragmatic choices to essentially optimize for getting like productive work done so doesn't really matter what your background is or necessarily what the tools you use are more just can you like accomplish things really quickly
50:40awesome this next question is from this is Rex They're wondering, can this device be used to explain consciousness in the long term? Of course, I can I can certainly shed some light on on consciousness. So, this is this is a really interesting question. I I think the answer is yes.
50:56And I think one of the reasons that consciousness is so hard is because like anything in physics, you're looking at a mapping from X to Y where X is the neuronal coralates. It's the thing that's happening then physically and then Y is this phenomenal state. And historically, we've been unable to observe the neuronal coralates very well. And unless it's in you, we've been unable to observe the phenomenal state.
51:13So as soon as you we're able neuroscientists are able to personally get these tools where they can see the coralates and they can have the experience, I think the hard problem will vanish very quickly. What I find remarkable is that the universe started out as like quarks and leptons. we'll call it like you know hydrogen and you know hydrogen evolved and and and somewhere along the way that hydrogen started talking and thought it was conscious.
51:40I I like the joke that it turns out that if you bombard earth with photons for long enough it'll emit a Tesla. Exactly. [laughter] Yeah, that's exactly what happened. [laughter] just basic science. Yeah, they teach you in school. yeah. So, so there's a lot of Twitter comments coming in about a question that I think is on everyone's mind, no pun intended. What is the security of the device look like? What kind of precautions are being taked taken? And what does the future look like for the security of the system?
52:23So fir first and foremost privacy and security are top priorities at Neurolink especially given the sensitivity of the data that we're collecting and one of the things that we're ensuring is to make sure that a lot of the interactions with the brain data is going to be encrypted and authenticated properly and I think this has been kind of a sort of recurring theme but one of the things that we have the ability to do at Neurolink is that we work on every layer of the product from chip design to source code and it really gives us a unique opportunity to embed security as part of our design from the get-go and to make sure that there are no single
53:05points of failure. So as an example, we can actually completely isolate sensitive modules like the B radio by just segregating it out at the hardware level and and and really making sure that we protect the IO to the brain away from any potential attacks and really minimizing the attack surface. And you know we have in-house security expertise and we're also working with external parties to you know do audits and perform penetration testing.
53:38Excellent. Thank you DJ. Well actually is there is there any is there a point that anybody here wants to make that has not been m asked in a question but you think should be a question anyone? Yeah. So hey I'm Zach. I work on micr fabrication team. I help making implants and one thing that I think is really cool just in general is that you know this this is a platform essentially so we can change the design up in a variety of ways. So if we s find that a certain design or electrode size or number of electrodes works better in different areas of the brain we have the capability to do that. I personally think that's very cool.
54:16Fantastic. Anyone else? Dan, do you have something to add? Yeah, I think one of the things that's allowed us to make such fast progress in the last few months is the use of pigs as as a model and we started out using choosing pigs because of very similar anatomy of their skull to humans, same thickness and similar kind of membrane dural membrane.
54:35But then as time went on, we realized that pigs are actually have amazing other properties. You can train them to walk on treadmills. you can train them to do all kinds of tricks and and also that they have this large representation of the snout in the cortex which you can very easily stimulate. So the question would be why are we using pigs and I think we've even surprised ourselves at how how useful they are as a model in this respect. And another important point is that it's very easy to keep pigs happy.
55:03that they they have very low needs and so we can build an environment in which they're they have amazingly good welfare. Yeah. They seem very happy about food. Yeah. It's easy to make pigs happy. Basically they they love food. They really pigs really love food. This is a true thing about pigs. so you know
55:25that they're Yeah. give them like, you know, sort of some straw and some things to play with and some friends to hang out with and good food and they're they're they're happy as pigs. and and then yeah, they're pigs are actually quite similar to people. So that's you know, if we're going to figure out things for people, then pigs are a good choice. And they're al they're quite robust creatures like little tanks. and and then I think one of the questions was like what like is is the device itself robust? And it's like, well, you know, pigs, they bustle around quite a lot and they they bump into things and they they they they headbutt each other at times and
56:05they're pretty animated. So, if the device is lasting in the pig as it lasted there for two months and still going strong then that's a good sign that the device is robust for people. Fantastic. one common theme that's been coming up a lot on these Twitter questions coming in is that of availability. And so Matias has a specific question on this, which is any estimate of how much it will cost at launch and what price it will reduce to over time?
56:34Well, I I think at at launch it's probably going to be it it I I would say that's not really representative. because at first I think it's it's going to be, you know, quite expensive, but that price will very rapidly drop. and I think over time we want to get the the cost obviously down as low as possible. but I think inclusive
56:58of the automated surgery, I think we want to get the the the the price down to a few thousand dollars, something like that. and I I think that's possible. I think it should be possible to get it similar to LASIK and and then the device electronics itself I think will will not be very expensive because it actually does does use a lot of the parts that are made in extremely high volume in tens of millions of of units for smartphones and and smartwatches and wearables in general.
57:30Great. So here's another question about the device. What does the architecture look like? Is it a CPU, ARM CPU? What's going on inside the device that you can talk about DJ? I I can talk a little bit about the digital architecture. It's fully custom. one of the most difficult challenges I think of of building an implant is the energy density. So the more electrodes you record from, the more energy you're going to be consuming. And so there's just there was nothing commercial out there. So there's an analog front end that's able to amplify these really small signals like microvolt range so that we can actually digitize them and then take those
58:11signals and then find exactly what we're looking for. And there was nothing out there that could basically do those two things. And so that's why we had to build a full custom ASIC. and so there's really there's nothing out you know like it out there. It's designed specifically to record signals from the brain and anything else is is would just be wasting energy.
58:34Yeah. So the thing that I would add to that is so there are obviously elements that we customize from from ground up like our neural amplifiers and some of the algorithms that we develop but a lot of the other systems around it are really borrowing from parts that are already been productionized and available from the wearables industry. So B radio a lot of the low power microprocessors that are part of it a lot of small sensors.
58:59So there are obviously the neural sensors that we're developing that gives us kind of a unique data set for our applications but really a lot of that is getting packaged up and you know just not really resisting the industry that's been sort of been laid out to really create these devices with very short amount of time.
59:23So here's another question about the integrity of the device more from a mechanical perspective. So you're replacing a piece of bone and replacing it with an implant and then the person is going to take that home. How does the integrity of the device fare to compared to I don't know bone or or something similar?
59:43sure. Well, I think Matt's probably replaced a lot of pieces of skull with plastic and other materials. So I don't know if you can talk about some of those. but yeah, basically I mean we can we know that what the mechanical properties of bone are and we know what the stresses on the device are going to be. So it's it's fairly easy to just go from first principles and design for any condition that it might see.
1:00:09I mean as you can see from the pigs as I was saying earlier the the implants are are obviously very robust because I pigs move around vigorously. They roll around. They they they bang into the wall. They bang into each other. they do like headbutting. and way more than people do. so well, sure. but you know, like it's not like the pigs are just walking around very delicately. they're they're high energy creatures. And I think it's it's clear that the the the implants are you know has been going for a few months still going strong despite a lot of vigorous activity from the pig and it's it's difficult. You can't just explain to the pig, hey you've got an implant on your head, why
1:00:52don't you take care of it? so it's got to be robust against you know a lot of head impacts essentially. So you know I think maybe one of the things that that isn't super obvious is that the the implant itself is attached to the skull but the the electrodes have long threads. So you can have quite a lot of movement relative movement between the brain and the skull without putting attention on the electrodes that are inserted in the in the cortex. So well let's see with that maybe a couple more questions or if anybody wants to make some comments and and then we'll call it a day.
1:01:29So I just want to add to what you just said. so the thread is actually also in the brain also. because of their thinness they actually very flexible and so as the pig is moving around and banging your head the the threads actually move with the brain even the normal pulsation of of the brain and that help minimize stress within the brain tissue also and that has a long-term impact on the functionality of the device.
1:01:56Zach, how long are the threads? Well, right now they're about 43 millimeters, but like I said, it's tunable. So, if we want to hit deep brain regions or or some other brain area, like we totally can do that. Just to follow up on what Phil said about the threads moving around with the brain. I mean, that that's great for protecting the threads, but it equally serves to protect the tissue around it.
1:02:17We have a lot of different ways to to image the tissue around the threads and iterate very quickly on on doing things to them that will improve the overall biompatibility. As a closing question, I thought we might go down the line starting with Dan. And what I'm wondering is what is the number one thing on your wish list that you're really hoping that the Neurolink device will do over time that you're working towards?
1:02:38so my my background is in visual neuroscience and one of the things I think has great potential for the Neurolink is to provide a visual prosthesis for people who have retinal injury or blindness through eye injury. You can essentially plug a camera directly into the visual cortex and stimulate with an enormous array of thousands or maybe tens of thousands of electrodes to recreate a a visual image.
1:03:03And in time perhaps you can use that same technology in people who haven't lost vision to produce some kind of heads up display something like Terminator or something like that. Wonderful. effects were saying that like over time we could actually give somebody supervision like you could have like ultraviolet or infrared or see and radar basically name your frequency you can just dynamically adjust the sensor or have sensors that feed into the visual cortex across a wide range of of frequencies and and actually have superhuman vision.
1:03:40Yeah. So for me telepathy so I think it's incredible amount of effort to put your thoughts into a set of words and you know it comes out completely compressed. So being able to do that seamlessly without being able to compress it with all the mechanisms would be great. yeah it's it's like just a sorry to add further to that
1:04:04in fact when I did the white but why article I I think Tim to thought I said consensual telepathy but I said conceptual telepathy [laughter] we presume it would be consensual because you definitely don't want just people you know sending stuff into your brain without your consent but a lot of our brain thought capacity is go goes into compressing our thoughts into words.
1:04:34and then you think of like the the the data rate of words. Words are a very slow very low data rate and and we're putting a tremendous amount of mental energy into compressing the concepts and thoughts in our head into words and then slowly talking. speech is so very very slow and we could actually send
1:04:54the the true thoughts that we can basically have far better communication because we can convey the actual concepts the actual thoughts uncompressed to somebody else. So non-linguistic, consensual, and conceptual telepathy. [laughter] Exactly. Non- linguistic, consensual, conceptual telepathy. There we go.
1:05:12Yeah. [laughter] I've actually been excited from the beginning sort of about the like side benefit of these devices. I sort of see them as essentially like an oscilloscope to a printed circuit board is our device to the brain where just by virtue of having this in there and being able to see what's actually going on, you'll end up learning a ton about how the brain works. and so sort of augmenting people but also just using that to learn a lot more about like neurological diseases is really exciting to me.
1:05:37Absolutely. So to to sort of follow up on Elon's thought, you I feel and I imagine a lot of other people feel the same way that there's a lot of sort of trapped creativity in your mind. You know, you can for example, you know, close your eyes and conjure up like an incredible like dolly-esque scene. But you know, if I wanted to actually show someone that, it would yeah, it would take years of cra, you know, honing a craft to be able to paint that. And so, you know, potentially with enough electrodes in the right places, you could begin to sort of tap into those raw concepts or thought vectors and be able to decode that and and show people. It could be
1:06:12for, you know, art. You imagine music or even for like engineering like a threedimensional model and like so mental artistry is a new field. [laughter] I I like to think about ways to interface devices with biology better. And so one of the things I'm looking forward to is getting this thing to be less look like less like technology and more like biology. So so that it really is you know seamlessly interface with the brain stable for a very long time. And then similar to that having stimulation be much more precise and multi-dimensional such that eventually the brain sort of doesn't really know if it's being stimulated from outside or inside. and
1:06:56you end up just sort of completely merging. Wow. Yeah. You know, following up on stimulation, you know, one of the things our device can do is simultaneously read and write on every channel, stimulate and record. And that's you know, both a more challenging problem than it may seem like at first, but also, you know, incredibly exciting the sort of the vistas that it opens up. I mean, it's really kind of the whole game in terms of interfacing with the nervous system and I'm I'm very excited to to use that.
1:07:27Yeah, I'm excited just because of the scalability of the device. We're doing everything in house and all of it can scale to more channels, more brain regions, etc. I think yeah, I am really interested about solving things related to anxiety or depression or even like removing fear. like I'm an athlete and to like rock climb without fear would be pretty maybe need a little bit of fear. Yeah, just a little. And also, it'd be great if we could make the pigs fly, but I [laughter] I think we have an incredible opportunity to limit human suffering to a tiny fraction of what it is today.
1:08:04in all kinds of different avenues, pain being the essence of suffering, we might be able to control that finally. and so many other diseases, so much other suffering in the world. I think the neural link device could help a lot with amazing. I I think all these things are are great u functions for a neural neural link. I think on a on a species level
1:08:32basis I think it's going to be important for us to figure out how we coexist with advanced artificial intelligence and you know I think achieving some kind of AI symbiosis where you have an AI extension of yourself like a tertiary layer above the lyic system and cortex and and having that having that symbiosis be good such that the future of the world is controlled by the combined will of the people of of earth.
1:09:05I think that that's obviously going to be the future that we want presumably is if it's the sum of our collective will and and so I think it's it's going to be important from a from an existential threat standpoint to achieve a a good AI symbiosis. And that's what I think is might be the the most important thing that a device like this achieves.
1:09:29I have in many ways a very basic science interest which is I'm really interested in in the nature of consciousness and that there's a lot of very silly philosophy that's been written about it over the last thousand years. but I think that it's really we've been very limited by the tools and our ability to interrogate and and measure the brain. And as these tools get better it will pull it into the realm of physics and it's really one of the last big great mysteries in in science.
1:09:57So for me, can you imagine a disease-free future? future where you know, you know what's going to happen to you before it happens so you can prevent it. with these devices, we'll be able to not just pick up electrical signals, but we can also pick up chemical clues in the brain. And if if you're successful, which I know we are, we were able to u kind of prevent ahead of time, you know, diseases and really the functionality of this device is widespread. so I'm looking forward to the future.
1:10:29I'm really excited about the opportunity to help people overcome challenges that they face through life circumstances, bad luck through no fault of their own, spinal cord injury, brain disease, some devastating things that completely change your life. Hopefully we can help them get some function back.
1:10:48I know and love many humans with autism spectrum disorder. So I'm really interested to see how the Neurolink might be able to support them if they chose to do that. Yeah. So I mean everyone else along the line has had, you know, amazing ideas and suggestions and you for me it's it's about you know memories and everyone loses those memories over time. You know I already can't remember what happened to me when I was younger and you know I it'll only get worse. And so having a repository of memories that you can access whenever you want. If you're feeling down, you can go access some good memories. You know, if you you miss something or you miss somebody, you can
1:11:26go and access those memories. And I think that would just be such a lifechanging experience to be able to just tap into that. Those are such a beautiful and diverse array of answers there. Yeah. All right. Well, thanks for tuning in and, as I said, please consider working at Neurolink and helping us solve these problems. there's a tremendous amount of work to be done to go from here to a device that is available, widely available and affordable and reliable. so please consider joining us and supporting us in our mission. thank you very much. [applause]
资源与出处
- 视频在 YouTube 收看
- 视频在 B 站收看UP 主「艺术资源素材基地」搬运,标题注明带中文字幕
- 档案silicon-valley.tech 档案页本页视频与文稿数据来自 Elon Musk Archive
- 精读本篇精读 Markdown 版