2019 年 7 月 16 日,旧金山的加州科学院。马斯克上台后先花了半分钟澄清一件事:刚才那段开场视频不是素材库里的空镜,是 Neuralink 办公室的日常。
画面里有人穿过实验室,机械臂在显微镜下移动,工程师盯着屏幕上跳动的波形。看上去太像科技公司的宣传片,所以他专门说明了一句。
然后他说了句更实在的话:今晚这场发布会,主要目的是招人。
“事实上,今晚这场演示的主要目的就是招聘。”(马斯克)
一场后来被全球媒体解读成“人类正式向赛博格迈进”的发布会,开场定调是招人。这个顺序挺有意思:先把人凑齐,再谈要造什么。
他说 Neuralink 会一步一步扩大要解决问题的范围,直到做出完整的脑机接口。今晚摊开讲的是第一版,做什么、为什么做、怎么做,全讲。
AI 可能是最后一个发明,脑机接口可能是下一章的第一个
马斯克给的理由跟 AI 绑在一起。他的判断是,AI 很可能是人类最后一个发明,而脑机接口有机会成为下一个篇章的第一个发明。
他自己的说法更直接:高带宽的脑机接口,给人类留了一个和 AI 融合的选项。
“如果有高带宽的脑机接口,我们其实可以搭上这趟车,有效地拥有和 AI 融合的选项。”(马斯克)
“搭上这趟车”这个说法,背后是他一贯的框架。如果 AI 的智能远远超过人类,人类只有两个位置:被甩在后面,或者合流。他选了后者,并且认为这件事极其重要。
他说这些话听起来会很怪。确实怪。但他紧接着就把话题拉回工程:要做到这一步,得先解决一堆具体问题,从一根线有多粗开始。
一千亿个神经元,和一根比头发细十倍的线
发布会放了一段科普动画,讲神经元怎么工作。这部分信息量大,值得慢慢看。
人脑里大约有一千亿个神经元。形状各不相同,但基本结构一致:一团树突组成的树冠,一个叫胞体(soma)的细胞主体,一条叫轴突(axon)的长尾巴。
神经元之间靠轴突和树突的连接点沟通,这些连接点叫突触(synapse)。
沟通用的信号是化学的,叫神经递质。轴突末端释放递质,触发下游细胞。
当一个细胞收到足够多、足够对的递质输入,就会引发连锁反应,产生一次放电,叫动作电位(action potential)。这次放电会继续传给下游的突触。
动作电位会在神经元周围产生电场。把电极放在附近,就能检测到这个电场,从而读出这个神经元携带的信息。这就是所有脑机接口的物理基础。
Neuralink 的目标是记录和刺激这些放电,规模要比迄今做过的任何事大几个数量级,还要安全到算不上“大手术”。马斯克给的参照物是近视激光手术。
对比数据在这:FDA 批准过的最好的系统,是帕金森病的深部脑刺激,大约 10 个电极。
Neuralink 第一版系统,1000 个电极,而且全部可读可写。他现场用的词是“一千倍”。按他给出的两个数字算,10 个对 1000 个,其实是一百倍。这种口误在马斯克的演讲里不算罕见,结论不变:量级差得非常远。
电极做在线上。这些线极细,横截面积大约是人发的十分之一,粗细和神经元本身差不多。
“如果你要往大脑里插东西,你不会希望它是个庞然大物。你希望它很小,和已经在那里的东西大致相当。”(马斯克)
现场给了尺度对比。一根手指,还是小指,线上的电极比汗毛还细,而这样的线有上千根。
为什么必须由机器人来扎针
线细到这个程度,人手没法操作。
机器人透过显微镜看,一根一根地把电极插进去,每根都单独定位,绕开所有血管,确保插入时不造成创伤,或者只造成最小的创伤。
“每根线,机器人基本上是通过显微镜看,然后单独插入每个电极,绕开任何血管,确保插入不造成创伤。”(马斯克)
刺破血管是件麻烦事。脑组织里的微血管密布,一次误伤就可能带来出血和后续的免疫反应,电极周围的疤痕组织会把信号吞掉。绕开血管,是这套系统能不能长期工作的前提。
机器人整机看起来相当复杂,但所有复杂度都收敛到一个极小的点上。针头直径大约 24 微米,小到在那张跟一分钱硬币的合影里几乎看不见。
手术流程被压缩到很简单:2 毫米切口,撑开到 8 毫米,器件放进去,切口缩回 2 毫米。他说基本上不用缝针。
接口是无线。头上不会有线伸出来,这点他强调“非常非常重要”。
数据通过蓝牙传到手机。他顺口开了个玩笑:得盯着应用商店的更新,别哪天出个驱动问题。
时间表当场给了:希望在明年年底之前,用在人类患者身上。2019 年说的明年,是 2020 年。
一个被放弃的 3072 电极版本
接着上台的是 Neuralink 总裁马克斯·霍达克(Max Hodak)。他说做脑机接口是他这辈子最核心的目标,从记事起就想做。
“你的所有体验、想法、记忆,全都在大脑里,用动作电位的发放统计来表示。”(马克斯·霍达克)
他讲的是产品哲学。设计约束在第一天就定下了:必须完全无线,必须能长期使用,不能是那种两三年就得取出来的东西。
他们走过四代原型,照片排成一排。
最左边是一块全被动电路板,64 个电极,接两个连接器,连到体外的大体积放大器。那时候脑袋上还得挂一堆设备。
第二代加了定制芯片,还是 64 通道。
第三代是个大跃进:3072 个电极,全植入封装,只留一个 USB-C 口露在外面。
第四代退了一步,1536 通道。霍达克的解释很实在:通道数不是唯一指标,安全性、寿命、带宽要一起优化,还要好制造。为了后面几项,他们选了更容易量产的那条路。
第三和第四代,是当天同步发布那篇论文的主角。这两代产品教给团队的东西,构成了第一代人体产品 n1 的架构基础。
n1 的核心是一颗传感器。8 毫米直径、4 毫米高的圆柱体,完全组装好的状态。现场图里少了一层外壳,打开看,里面是薄膜,带着那些线伸向一侧,下面是密封基底,之后会用激光焊上一个盖子,再接上定制电子。
这些线没法用手碰。顶部那块东西只是手术包装的背衬材料,机器人要把线一根一根从上面剥下来,再送进大脑。
霍达克讲了两个动因。第一个很朴素:线必须放进去,手放不了,所以需要机器人。第二个动因是做着做着长出来的:得知道血管在哪,得对组织成像,还得处理脑表面随呼吸和心跳起伏的问题。用他的话说,是在一个熵极高的基底上做微米级的活。
植入数量:一个半球最多可以放 10 个。首批患者先用 4 个,3 个在运动区,1 个体感区。
这 4 个传感器无线穿过皮肤,连到一个可穿戴设备上,他们叫它 Link,里面装蓝牙和电池。整套东西用 iPhone 应用控制,不需要跑去医生办公室,让医生用一台专用编程器给你配置。
第一批患者要解决的三类控制
首批产品聚焦三类控制。现场讲得最细的是第一类:让患者自己操作移动设备。
这个需求来自患者团体的反复反馈。
“如果必须有个护理员在旁边帮你按按钮,那有什么意义?你不如让他直接做这件事。”(马斯克)
逻辑很硬:一台手机如果还得别人替你点,那这台手机就不属于你。所以第一优先级是让使用者能独立操作设备,恢复基本的自主性。
支撑这件事的是那台手术机器人。它能在不到一小时里,把几百根线、几千个独立电极插进大脑皮层。
精度是微米级的。外科医生用它对准血管之间的空隙,覆盖脑表面的方式可以精确到微米。机器人在术前规划好的位置放线,重复性极高。
手术体验要对标近视激光
发布会花了不少篇幅讲传统神经外科是什么体验,这段对比很有说服力。
传统脑外科手术,患者听到就发怵,除非万不得已没人愿意做。流程通常是:一个头架夹住颅骨,把脑袋死死固定在手术台上;头发全部或大部分剃掉;术后留下明显的疤痕。
Neuralink 想做的是完全不同的体验,对标 LASIK(准分子激光近视手术)。他们甚至希望能在清醒镇静下完成,这样就避开了全麻的复杂度和风险,也避开了全麻的一堆副作用,比如恶心、插管后的喉咙痛。
简化后的流程有五步:注射局部麻醉;皮肤上开一个小口;在颅骨上开一个无痛的孔;快速精确地把线放进皮层;用传感器把这个孔堵上,头皮缝回去。
没有现成器件能满足全部要求,不管研究用的还是商业用的。所以他们自己造。
材料选的是微加工薄膜聚合物。一根头发的直径大约 100 微米,在这个尺寸里,要塞进电极、导线和绝缘层。这个设计叫 linear edge,是他们为研发做过的 20 多种设计之一。
迭代方向很清楚:每根线上的电极数不断增加,线本身的宽度基本不变。电极更密,单根线的侵入性不增加,这是整个设计的核心取舍。
接下来是组装。电子件装好,再用激光焊上一个带线的盖子。这两步内部开发了很久。
模拟像素迭代了三代,体积和功耗一步步降,性能保持住。最新一版的像素,比学术文献里同类架构的最好水平至少小 5 倍。所有功能集成进一颗 4 乘 5 毫米的芯片。
让杂乱波形突然出现结构的那个小动作
信号长什么样?现场放了一段。
一根线上取下来的多个电极,每条轨迹是一个随时间变化的电压波形。算法在实时检测尖峰,尖峰发生时立刻抓下来。
抓下来的数据画成一张图,叫尖峰栅格图(spike raster)。每一行是一个记录通道,时间从左往右走,每个小刻线代表一次动作电位发生的时间点。
第一眼看,这图很乱。看不出什么名堂。
然后演讲者做了个小动作:把这些神经元按它们的调谐顺序重新排列。调谐,指的是细胞对某个特定运动方向的偏好程度。
排完之后,结构出现了。原本杂乱的点阵里,显出一条条清晰的斜带。
这说明信息一直在,只是埋在乱序里。运动的方向、速度、力度,都能从这片放电模式里读出来。他们要做的,是把这种“魔法”自动化。
工具叫解码算法。用这类数据训练出来后,输入尖峰栅格,输出这个人想做的动作。
现场用生成的假数据做了演示,一个极简的解码器,成功还原了意图动作。
这里有个关键细节:即使人没真的动,只是在想这个动作,或者只是在看别人做这个动作,运动皮层的细胞也会以类似方式放电。这给瘫痪患者留出了空间,他们动不了,但想得出来。
基于这一点,使用者应该能自然地控制电脑,不只是鼠标,还有键盘和游戏手柄,以及其他可以接入的设备。
往远看,同样的原理可以用来给失去说话能力的瘫痪者恢复语言。原则上,这套东西可以覆盖整个运动皮层,那样任何被想到的动作都能被解码,跑步、跳舞,甚至功夫。
霍达克最后把话题收回到人本身:让人们接入这些表征,修复断裂的脑回路,最终更好地连接世界、连接彼此、连接自己。
把时间表和现实放在一起看
这场发布会的信息密度很高,但有几处需要打个折。
最明显的是时间。2019 年说“明年年底之前用在人类患者身上”,实际的首例人体植入发生在 2024 年 1 月,中间隔了四年半。马斯克的时间表一贯乐观,这不是秘密,看他的发布会要把这一点算进去。
第二个是规模。1000 个电极听起来很多,FDA 批准的深部脑刺激只有 10 个,差距确实大。但要读出复杂意图,需要的通道数可能还要往上走一个量级。当天发布的论文里,3072 电极版本已经被放弃,退回到 1536 通道,原因是安全性和寿命。这个取舍比任何愿景陈述都更能说明工程难度在哪。
第三个是手术机器人。整场发布会最容易被忽略的部分,恰恰是最难的部分。绕开血管、在随呼吸起伏的脑表面做微米级定位、一小时内完成几百次插入,这些难度不亚于芯片本身。芯片可以迭代流片,机器人要在活体上一次性成功。
还有一点值得说:他们把“无线”和“长期植入”当成硬约束,这两个约束决定了产品的天花板。有线方案能做得更简单,短期方案能做得更激进,但都不构成产品。
这场发布会真正回答了什么
回到开头那个问题:为什么一场技术发布会要以招聘开场。
因为马斯克清楚这件事的难度不在某一点,而在跨学科的缝合处。神经科学、材料、芯片设计、机器人、临床手术,每一块都需要顶尖的人,而且必须凑在一起工作。他在台上说,团队的人才密度让他印象极深。这话不像客套。
发布会上没有回答的问题同样清晰:长期植入后电极周围的疤痕组织怎么处理,信号能稳定多少年,免疫反应怎么控制,监管路径怎么走,成本最终落在什么区间。这些在 2019 年都没有答案,今天的部分答案也是用一次次临床试验换来的。
但第一版系统把几个关键数值摆在了台面上:1000 个电极、每根线比头发细得多、24 微米的针、2 毫米的切口、无线传输、一小时以内完成植入。这些数字比任何关于人机融合的宏大叙事都更有说服力。
马斯克在台上说,这些东西听起来会很怪。他说的没错。但怪的不是目标,是目标被拆解成这些具体数字之后的样子。
- 看脑机接口的新闻,先找四个硬指标:电极通道数、单根线的尺寸、植入创伤、是否无线。这四个数字决定产品能不能用,愿景不能。
- 看到“读心”“意念控制”这类标题,把关注点换到手术机器人上。放着几百根线绕开血管这件事,比解码算法更能反映真实难度。
- 判断时间表时,按公开承诺往后推:马斯克 2019 年说 2020 年底进人体,实际是 2024 年 1 月。以后看到类似承诺,先在心里加三到五年。
- 想搞懂神经解码,从尖峰栅格图和调谐曲线入手,这是最直观的入口,一张重排前后的对比图胜过十页综述。
- 读原始论文,别只看发布会视频。Neuralink 每场发布会同步发白皮书,里面写着他们放弃了什么、为什么放弃,这些信息比台上讲的更有价值。
- 关注监管与临床路径,不只看技术参数。一项植入式设备能不能普及,取决于审批、术式标准化和成本,这些通常不在发布会的聚光灯下。
0:03[Applause] [Music] [Applause] [Music] hello everybody so that that video was not shutterstock that was actually your link so that that's actual video from the company so if you want to get a sense for what it's like to work near a link that video is indicative of the atmosphere of your link it's an incredibly talented team and you're gonna hear a lot from from them tonight so we're gonna actually go quite into depth on what we're doing why we're doing how we're doing it and I'm just incredibly impressed with the caliber of of talent at your link and the in fact
2:22the the main reason for during this presentation is recruiting and this will be a slow process where we will gradually increase the issues that we solve until ultimately we can do a full brain machine interface yeah those are going to sound pretty weird but achieve a sort of symbiosis with artificial intelligence but I think with a high
2:50bandwidth brain machine interface I think we can actually go along for the ride and we can effectively have the option of merging with AI and this is extremely important of nearly 100 billion cells called neurons neurons come in many complex shapes but generally they have a dendritic Arbor a cell body called a Selma and an axon the neurons of your brain connect to form a large network through axon dendrite junctions called synapses but these connection points neurons communicate with each other using chemical signals called neurotransmitters neurotransmitters are released from the end of an axon in response to an electrical spike called an action potential when a cell receives
3:40enough of the right kind of neurotransmitter input a chain reaction is triggered that causes an action potential to fire and the neuron to in turn relay messages to its own downstream synapses action potentials produce an electric field that spreads from the neuron and can be detected by placing electrodes nearby allowing recording of the information represented by a neuron a goal is to record from and
4:09stimulate spikes in neurons and do so in a way that is orders of magnitude more than anything has been done to date and safe and good enough that you can it's not like a major operation it's sort of equivalent to just sort of a LASIK type of thing so this is in contrast to the the best fda-approved system which is like a Parkinson's deep brain stimulation a thing which would have on the order of 10 electrodes so the system even in version 1 that we're going to unveil today is capable of a thousand times more electrodes than the the best
5:00system out there and they're all read and write so this is this is really quite I think I mean for something to be a thousand times more than what is public approved is quite a big difference so that there's there's very tiny threads that are about about 1/10 roughly of the cross sectional area of a human hair so there are extremely tiny threads in fact the threads that we have it likes it even in version one are about the same size as a neuron so if you're gonna go stick something in your brain you wanted to not be giant you want to be tiny and to be approximately on par with the things that are already there the neurons you really need this to be
5:44done with the robot because it's very tiny and it needs to be very precise so you don't and you don't want to pierce a blood vessel so when you're in so each thread that the robot looks look sort of basically through a microscope and put a inserts each electrode specifically bypassing any vasculature you know any
6:07kind of blood vessel and and making sure it's like inserted without causing trauma or minimal trauma so just give you a sense of scale this is how tiny the threads are that is not even a big finger that is a small finger so the the these threads are just like like I said where smaller than hair and there's a thousand of them and this is what what the robot looks like it's sort of quite
6:41quite a complex device but it I it all comes down to a very tiny tiny point so just like you see the robot the robots on the left and and then the what looks like the needles for insertion next to a penny but in fact that the actual needle that gets inserted is way way tinier it's that little tiny thing at the where the arrow is pointing that's actually the size of the the needle it's about 24 microns in diameter it's so small you can't really even see it within the picture with the penny you can get a sense for the robot doing the electrode
7:26insertion but that's a very zoomed in view so they're all very very tiny and the robot is very selectively very doubt very delicately and and then this is what the gym looks like so action potentials so the each one of those represents one electrode so there would be up to ten thousands of these that these lines the the operation on a
8:07per chip basis it involves just a a two meal a two millimeter incision which is dilated to eight millimeters and then the the the Chavez placed placed through that and then we add it goes back to being two millimeters and you can basically good shot you're gonna need a stitch and then the the interface to the
8:29to the to the chip is what is wireless so you have no wires poking out of your head very very important so you it's it's basically bluetooth to your phone because we'll have to watch the App Store updates for that one make sure we don't have a driver issue and we hope to have this aspirationally
8:50and in a human patient before the end of next year so this is not not far I'm Mac so doc I'm the president of of neural link so I've wanted to build a neural interface has really been like a central goal of my life basically as long as I can remember this is I think like we talked about AI being potentially the last invention that we have I think that I've been with BMI might be like really the first invention in many ways of like the next chapter of us it's just really like as Elon alluded to earlier everything about your experience or thoughts or memories it's all in your brain and represented in the firing statistics of action potentials we knew as Ellen mentioned that whatever
9:27we built we wanted it to be completely wireless it had to be something that would last for a long period of time not something that you'd have to take out at two three or four years in this is a photo of some of the prototypes that we've gone through over that over that time so we started on the far left that's the entirely passive board that has 64 electrodes on it and connects two connectors that go to big external amplifiers and then we added integrator chronics with our first custom ship that's also 64 64 channels and then there is a big leap to the the device that Ilan showed a photo of earlier that has 3072 electrodes and a fully implantable package with just a USB C
10:04port coming out and then we we took a step back in channel count B's room we have to optimize safety longevity and bandwidth all together and so in order to optimize some of those other things we moved to an easier to manufacture system as 1536 channels in a USB C port and those last two are the focus of the paper that we released today and they taught us a lot about the architecture that we think is the basis for our first human product that we're calling n1 and the central component of that is the n1 sensor this is it's a little hermetic
10:35package it's about it's when it's fully assembled this is missing an outer mold it's into an 8 millimeter diameter force millimeter tall cylinder exploding it blowing like opening it up a little bit you can see there's there's the thin film which has the threads that heal I'm talked about which is the wisp going off to the side there's a hermetic substrate and then that gets welded later to a package that goes over top and that's mated to our custom electronics and you really can't manipulate these with your hands that that part at the top is just a backing material that's surgical packaging there they're peeled off the threads were peeled off that one at a
11:12time by the robot to place it into the brain and the first impetus for this is just you have to place these threads you can't manipulate these threads you need a robot and then that turned out to that grew into understanding where the blood vessels are and imaging into the tissue and the surface of the brain moves because you're breathing and you have a heartbeat and there's lots of complexity of dealing with this incredibly high entropy substrate and so the n1 implant we can place as Elin mentioned many of these possibly up to 10 in one hemisphere for our first patients we're looking at four four sensors three in motor areas and one in a somatosensory
11:47area and that connect wirelessly through the skin to a wearable device that we call the link which contains a Bluetooth radio and a battery it'll be controlled through an iPhone app you won't have to go to a doctor's office and have them have an exotic programmer to to configure it and so the for the first product where we're really focusing on three distinct types of control the first is giving patients the ability to control their mobile device to be as we heard from over and over again from patient groups that if you have to have a care taker around the pressed buttons for you what's the point you might as well have them do the thing you have to
12:21get self sufficient using using the devices on your own but we are working as hard as we can towards our first in human clinical study next year we developed this robot that can rapidly and precisely insert hundreds of individual threads representing thousands of distinct electrodes into the cortex in less than an hour this tool allows the surgeon to aim between the blood vessels they'll cover the surface of the brain with micron scale precision here the robot is selecting individual electrode threads and placing them into the brain in the pre-planned location with remarkable accuracy and repeatability when you think of traditional neurosurgery you probably
13:02think of something very invasive traditional surgery on the brain isn't something that patients ever look forward to or are excited about except in the most dire circumstances usually a clamp is attached to the skull to keep it rigidly immobilized to the operating table we often shave all or most of the patients hair patients can end up with large visible scars at neural link we want to create an entirely different patient experience something more like LASIK we even want this to be possible under conscious sedation that means you can get rid of the complexity and the risk of general anesthesia as well as many of the unpleasant side effects nausea sore throat from a breathing tube
13:47but our aim is to simplify the procedure down to the injection of local anesthetic a very small opening in the skin a painless opening in the skull below quick and precise placement of threads into the cortex and then we fill that hole in the skull with the sensor allowing the scalp to be closed up over it currently there are no research or commercial commercial devices that meet all of our requirements so we built one out of microfabricated thumb film polymers and an average strand of hair is about a hundred microns yet in the small footprint we're able to fit our electrodes our wires and insulation for each of those wires this design is called linear edge it's one of over 20
14:32designs that we've made for our R&D work we progressively been increasing the number of electrodes per thread without significantly increasing the width of each of these threads at the base next we assemble the electronics and then also attach a wired lid using a laser welding process these two steps have required a lot of internal development as well the result is the sensor that's
15:01ready for final assembly and implants into the body since the start of nor link we've gone through three major revisions to the analog pixel progressively improving both the size and power while maintaining performance and our latest pixel on the right is at least five times smaller than the known state-of-the-art of similar architecture with one pixel dedicated per electrode as published in the academic literature all of these functionalities that I outlined are integrated into a single four by five millimeter so it can die each this is in
15:38fact traces of a bunch of electrodes that came off of one of our devices a bunch of electrodes from a single thread and each trace shows you a voltage waveform in time as it's coming off of one of those threads we have algorithms that can detect these spikes in real time as they're happening and that allows us to collect data that looks something like this this is what we call a spike raster so each row there represents one channel of recording and time goes from left to right and each of those little tick marks is the time of a single spike in action potential if you look at that you might think that looks pretty messy and it's not clear what's
16:19going on but I'm gonna do a little trick I want to take those neurons I want to rearrange them so that they're in the order of the tuning that they have but just as I told you about those two neurons and if you do that look what happens now suddenly structure emerges and I think you'll agree looking at that but there's information in that stack of neurons that tells you about the movement and that's exactly what we want to do we want to do that kind of magic in an automated way to read out and to read out the movement the way we do that is by building something that we call decoding algorithms these are mathematical algorithms that we tune based on data like these to be able to
16:55take in just those raster's of spiking activity and output the movement that's that the person wants to make for these little fake data I built a very very simple decoder and sure enough it's able to to capture the intended movement this is what we want to do on a bigger scale but even if you're not actually making the movement even if you're just thinking about the movement or even if you're watching someone else make movement the cells and motor cortex respond in a similar way with that we think the people will be able to get naturalistic control over the computers not just a mouse but also keyboard game controllers and potentially other devices that's what we're trying to do
17:32so potentially with a device like this you could restore speech to a paralyzed person who's no longer able to talk but there's no reason in principle that we can't reach all of motor cortex and that would give us access to any movement that a person thinks about any movement at all a person could imagine running or dancing or even kungfu and we would be able to decode that signal but knurling wants to do is to give people the ability to tap into those representations to get act better access to that information both to repair broken brain circuits and also to ultimately give us better access to
18:20better connections to the world to each other and to ourselves [Applause]
资源与出处
- 视频在 YouTube 收看
- 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