太阳大约还能撑五亿年。等这个数字走完,它会膨胀,海洋被蒸干,地球不再适合任何生命存活。也就是说,如果意识的演化过程再多花 10% 的时间,它根本不会出现。
这笔时间账是马斯克在莱克斯·弗里德曼(Lex Fridman)播客第 49 期里算的。同一场对话里还有另一个数字:37 亿英里。1990 年,旅行者一号飞到那个距离上,回头拍了一张地球的照片,地球在画面里占不到一个像素。
从单个神经元到火星,从芯片功耗到文明存续,马斯克把这些东西放在同一张桌子上谈。中间还夹着一段关于电极涂层和 LASIK 手术的工程细节。这些细节听起来琐碎,其实是整场对话的骨架。
泛心论没法验证,他把它归进个人偏好
弗里德曼开场就问了一个大问题:意识是人类独有的,还是像引力一样弥漫在所有物质里。
马斯克的回答很干脆,他不认为意识弥漫在一切物质中。他提到泛心论者(panpsychist)相信那一套,紧接着反问了一句:你怎么验证呢。
他把自己的立场摆在科学方法上。一个命题为真的程度,取决于它能被检验到什么程度。检验不了的,谈的其实是偏好,或者不可检验的信念。他自己说得很清楚,他不是盲目相信任何东西,但科学方法是他愿意站的那块地。
弗里德曼追问:理解意识这件事,在科学方法的射程之内吗。
马斯克的回答留了余地。他说人类很难说自己对任何事情的理解完全准确,但大幅提升对意识的理解,答案是肯定的。
接着他给了一个很硬的观察。如果大脑在物理上受损,意识也会跟着受损。这个事实指向一个结论:意识是一种物理现象。至于意识感觉起来像是在另一个维度里,他认为那可能只是错觉。
这里有一个方法论的细节值得留意。他没有从"意识像什么"出发,他从"意识能被什么改变"出发。能被物理手段改变的东西,就受物理规律约束。这个转手成本极低,威力很大。
数字智能会在每个维度上超过我们,包括装出意识
弗里德曼换了个问法:AI 要到达人类水平甚至超人水平,需不需要拥有意识、身体、对死亡的恐惧、爱这些东西。
马斯克说这里混了太多概念。他觉得"智能"这个词被用得太随便,把好几件不同的事搅在了一块。拆开来看,如果问的是 AI 会不会每一件事都想得比我们快、比我们好,答案毫无悬念是肯定的。
但要让一个系统在每个维度上都强过人类,它需不需要自我意识。他提醒,自我意识和意识是两码事。
然后他给了一个图灵测试的升级版本。
设想你跟一个数字超级智能通电话,或者开视频会议。对面看上去像人,表情动作都对,说话的语气、停顿、细微的犹豫全都到位,还会像人一样犯错。你全程分辨不出对面到底是人还是机器。
如果完全分辨不出来,那它就跟人没什么区别了。(马斯克)
从科学方法的角度看,这句话分量很重。一个能被完美模拟的东西,和一个真的东西,在可观测层面没有差别。马斯克把这个立场说得很直接:数字智能一定能够模拟我们所认为的意识,模拟到你看不出区别的程度。
他没有断言机器"真的有"意识。他绕开了这个问题,说的是当差别无法被观测,这个问题本身的实际意义就消失了。
这个回答可能让人觉得滑头,其实相当克制。他不假装自己知道答案,他只是指出这个问题的边界在哪里。
他主张给 AI 设一个类似 FDA 的监管机构
弗里德曼给了三条逃生路线,让马斯克挑最有希望的一条。
第一条,加大 AI 安全与有益 AI 的研究投入,指望找到算法层面或者政策层面的解法。第二条,尽快让人类成为多行星物种。第三条,跟 AI 融合,顺着那条不断上升的智能曲线一起往上走。
马斯克的回答从投资分布切入。涌入 AI 的钱极其庞大,缺钱的是 AI 安全。这个错配在他看来很危险。
他主张设立一个政府机构,专门监管跟 AI 相关的所有事情,确认它不构成公共安全风险。他拿现成的机构做类比:食品药品有 FDA,汽车安全有 NHTSA,航空安全有 FAA。人类社会早就形成了一个共识,只要某样东西对公众有潜在危险,就该有一个替公共利益把关的裁判。
AI 有没有潜在危险。他的判断是明确的肯定。既然有,就该有对应的监管机构。
他也承认这套主张有个大麻烦。政府动作很慢。
监管机构诞生的典型路径是:先出大事,公众哗然,然后过了好几年,规则才姗姗出台。
安全带的历史,是他判断监管节奏的依据
马斯克举了安全带做例子。
安全带能大幅降低伤亡,这件事在业界早就不是秘密,被知晓了十年甚至更久。汽车行业对强制安装的抵制极其顽固,理由是消费者会因为安全带而不买车。这个理由显然荒谬。
他的估算很沉重:可能几十万人因此丢了命。
接着是烟草。烟草行业对抗任何监管的时间跨度更长,手段也更隐蔽。马斯克提到他参与制作了电影《感谢你抽烟》(Thank You for Smoking),用意就是让人看清一件事:当企业实际上完成了对监管机构的俘获(regulatory capture),后果会有多坏。
这段历史解释了他为什么急着谈 AI 监管。他一眼看出了这个模式:技术跑在前面,公众后知后觉,行业拼命抵抗,监管在尸骨堆上姗姗来迟。他不指望监管跑在技术前面,他担心的是监管永远只能在事故之后追赶。
关于"奇点"(singularity)这个词,他的态度很克制。这个词本身不包含好坏判断,只说明那个点之后的事情极难预测。有变坏的概率,也有变好的概率。人类能做的就是去影响这个概率,让它偏向好的一侧。
现在的脑科学,像把听诊器贴在工厂外墙上
话题转到 Neuralink,弗里德曼问了一个科学层面的问题:这项工作的价值,会不会也包括帮我们理解大脑本身。
马斯克的回答从现有手段的局限讲起。
功能磁共振成像(fMRI)是目前的主力工具,它给出的图景相当粗糙。他打了一个比方:把听诊器贴在工厂外墙的不同位置,你能听到里面轰隆隆的动静,但哪台机器在干什么,你根本不知道。用这种宽泛的信号去推断内部机制,能猜出一点东西,分辨率差得太远。
要真正弄清大脑在做什么,得有两样东西。
一样是高精度的传感器,能读到单个神经元在放电。另一样是刺激与响应的闭环:触发某个神经元,然后观察当事人的感受、视觉、听觉、记忆发生了什么变化。
这个闭环之所以重要,是因为人能直接描述。人能说话,能告诉你"我刚才心情变了""这句话突然读得通了""这段旋律让我想起一件事"。有了这些主观报告,神经活动与体验之间的对应关系才有可能被建立。
他把这件事的意义说得很大:这就等于把工厂的门打开,让人走进去看。现在的神经科学只能站在墙外听声音,将来的神经科学能进车间,指着一台机器问这是干什么的。
人类大脑是猴子脑上贴了一台计算机
弗里德曼问的是双方的可塑性:大脑这一侧有一定弹性,机器那一侧有神经网络可以自我调整,两个系统像两个物种一样摸索出一套通信协议,最后受益更多的一侧在哪里。
马斯克的判断很明确。机器的可塑性远高于生物一侧,差距巨大。绝大多数适配必须发生在机器这一侧,否则整个方案就不成立。
他解释了原因。神经元并不会把一根电极当成另一个神经元,它对脉冲有反应,你可以利用这种弹性,但脑组织本身的适应能力有限。他说得很绝:这是唯一可能的路,没有第二种。
接着他给出了一套关于大脑分层的描述,这套描述贯穿了整场对话。
人的大脑大体分两层。底层是边缘系统(limbic system),负责本能和冲动。上面是皮层(cortex),负责思考。
我们有点像一只猴子脑上贴了一台计算机。大量的冲动都来自猴子脑,皮层忙着让猴子脑满意。驾驶盘在猴子脑手里,皮层负责讲故事。(马斯克)
他特别强调这个颠倒。直觉上,聪明的那部分应该管住笨的那部分,实际情况反过来,笨的那部分在指挥聪明的那部分。而皮层最擅长的活,是给这套机制编一个体面的说法,让人相信自己比一只猴子高级得多。
他还拿两个层级的智力落差做对比。边缘系统完全无法理解皮层在干什么,它只会发出信号:饿了、困了、怒了、渴了。然后皮层接到指令,动用巨大的算力去满足这些需求。
他顺手开了个玩笑,说人类在性这件事上投入的思考量极其惊人,如果排除生育目的,这个行为的荒谬程度相当高。为什么还要做。因为边缘系统开心。他还说这应该是 NSF 重点资助的研究领域,弗里德曼接茬说要为此成立一个新机构。
玩笑背后是个严肃的结构问题:高级功能为低级功能服务,成本高得离谱,而当事人对此毫无察觉。
第三层:数字超级智能住在哪里
顺着这套分层,马斯克提出了一个设想。
边缘系统是第一层,皮层是第二层,将来可能长出第三层,数字超级智能就住在那里。
他这个设想的乐观依据来自前两层的相处方式。皮层的智力远超边缘系统,可它们共存得相当好,没有人想把其中任何一层删掉。边缘系统负责乐趣,皮层负责思考,两者各司其职。他自己说,他没见过谁主动要求删掉其中一层。
他由此推断,第三层与下面两层的共存也有机会做到和平、温和。第三层的智力会远超皮层,但不必以消灭下面两层为代价。
至于机器学习在其中的角色,他的回答很直接。人本身就是一张神经网络,AI 也是神经网络,数字神经网络与生物神经网络对接,前者顺带把后者带上路。他甚至说,往后人类智能的绝大部分会是数字的。
这句话听着刺耳,逻辑上却顺。皮层相对边缘系统的智力优势已经大到后者无法理解前者,那么第三层相对皮层的落差只会更大。
关于时间表,他给了一个很有他个人风格的表述。Neuralink 帮人类跟 AI 融合的成功率,高于 0%。他引了电影《阿呆与阿瓜》(Dumb and Dumber)里的梗,说"总还是有机会的"。接着补充,这个概率可以随着技术进步不断上抬,从百万分之一到千分之一、百分之一、十分之一,取决于进展速度。
一根电极要在脑子里活几十年
Neuralink 的近期目标写得很清楚:先治病。
自闭症、精神分裂、记忆丧失,这些都在他列举的范围里。记忆衰退是几乎每个人都会遇到的事,上了年纪的父母记不住孩子的名字,这种事每天都在发生。
更广的一类,是脑和脊髓的损伤。让脊髓损伤的人重新获得完整的运动控制,让中风后的患者恢复脑功能,让遗传性的大脑疾病不再摧毁一个人的生活,这些都在路线图上。实现方式是靠硅基电路补上缺失的功能,把坏掉的那段回路接管过来。
但他也说得很直白,Neuralink 存在的终极理由,是应对数字超级智能带来的存在性风险。
我们不可能比数字超级计算机更聪明。所以,如果你打不过它们,就加入它们。至少我们会有这个选项。(马斯克)
他还提到一个时间上的要求。这件事要尽早做,最好在奇点到来之前完成人脑接口,或者至少不要落后太久。奇点之后的世界极度不稳定,那时候再补课,人类和意识的风险会大得多。
至于具体的工程难题,他列了一串。
材料科学排第一。电极要小到不伤害神经元,又要在一个人的脑子里待几十年。麻烦在于,大脑是一个电活动、电化学活动都极其剧烈的环境。涂层要在这个环境里不溶解、不剥落、不出毒性,这一点非常难。
芯片设计排第二。信号必须在本地处理,功耗压到极低。原因很实际:大脑对温度非常敏感,稍微热一点就会造成损伤。信号处理要做得又准又省电,这需要大量芯片设计工程师。
第三个是双向通信。读出来还不够,还要能刺激回去,把指令写进神经回路。读写合起来,才配得上"接口"这个词。
还有一层他反复提到:所有这些都得跑在极小的功率预算里。人类大脑自己只消耗大约二十瓦,你往里面塞任何发热元件,等于在一个精密的化学反应器里点火。
他最后列了一串岗位:材料科学、软件、机械工程、电气工程、芯片设计、微纳加工。六个方向同时推进,缺一个都不行。
手术得做成 LASIK,否则只能服务少数人
在马斯克的设想里,Neuralink 的植入过程必须完全自动化,像 LASIK 近视手术一样。
为什么这件事不能交给神经外科医生一台一台手工做。因为那样根本不可能覆盖大量人口。
他给的理由属于政治哲学层面,跟商业回报没什么关系。人类的未来应该由数量众多的普通人共同决定。如果只有少数付得起钱的人能接上超级智能,那个未来就不属于人类整体。
这个立场放在整场对话里分量很重。脑机接口在他眼里是基础设施,跟电网、公路、疫苗属于同一量级,服务对象必须是全体。
弗里德曼顺着问:这会不会顺带革新外科手术本身。马斯克的回答是肯定的。他描述的场景很像验光:机器人操作,眼科医生只需要确认你的头摆正了,然后按一下按钮。
手术一旦能被机器标准化,成本、精度、可及性三条线会同时改善。反过来说,只要还依赖稀缺的专家手工操作,任何技术都注定只能服务少数人。这个判断适用于脑机接口,也适用于很多别的东西。
马路上跑的是四轮机器人
话题转到特斯拉。弗里德曼提了一个角度:智能召唤(Smart Summon)和自动泊车让普通人第一次直面"没有司机的车"。很多人这辈子可能是第一次见到机器人。特斯拉有没有意识到自己在扮演一个教育者的角色。
马斯克先给了个务实的回答,说做这些功能是为了让生活方便一点。然后他承认这个角度有意思。
大多数人这辈子没见过机器人。一台自动驾驶的汽车,本质上就是一台四轮机器人。路上已经跑着几十万台特斯拉,这个数字很快就会到一百万,其中一部分会在没有人的情况下自己开。这件事本身就是一场关于机器人技术的大规模公开课。
他还提到一个细节。他自己一眼就能认出哪台车开着 Autopilot,因为它的动作有细微的个性,跟人开的车不一样。版本迭代之间,这种"性格"还在变。
这个观察不是随口一说。他对比了车道保持的表现:高速上人类开的车像碰碰车,在车道里飘来飘去;开着 Autopilot 的车,死死稳在正中间。
弗里德曼还提到一个细节,说车载卡拉 OK 是驾驶功能,理由是唱歌有助于管理注意力和视线。马斯克表示同意,说这是车里最好玩的功能之一。能同时做到好玩和安全,那是最好的结果。
感知比规划难,难在把世界变成向量
被问到全自动驾驶的进展,马斯克把场景拆成了几块。
高速路段基本解决,需要接管的概率已经非常低。走走停停的堵车路况,他给出的判断是已经比人安全得多,发生伤害或者碰撞的概率明显低于人类驾驶。配合导航辅助驾驶(Navigate on Autopilot),变道和下高速匝道也能完成。
另一个方向是从低速端往上攻。这个思路来自人学车的经验:没人第一次上路就去旧金山九曲花街练手,都是先在停车场里慢慢磨。所以先从停车场、低速场景把基本功打扎实。
剩下要补的有两块:红绿灯和停车标志。
停车标志在他看来反而不难。位置大致已知,最坏情况可以用地理坐标定位,再靠视觉确认停止线的位置,把 GPS 误差抵消掉。
真正麻烦的是复杂路口的多相位红绿灯,以及极度蜿蜒的山路。
至于感知和规划哪个更难,他的回答很干脆:感知。
具体说,难的是把视觉(为主)、雷达、超声波的输入,转换成周围物体在向量空间里的精确表示。一旦这步做对了,后面的规划和控制就相对轻松。
他打了个通俗的比方,说这就像《侠盗猎车手》里的车。那些车只要有准确的位置表示,就能正常行驶、正常拐弯,不会无缘无故撞墙,除非玩家主动撞上去。游戏里跑得通,是因为游戏世界本身就是一张精确的向量地图。
这个类比把自动驾驶的技术重心说清楚了。大家以为难点在于"决定怎么开",实际上难点在于"先如实看清世界"。看清之后,决策反而接近一个已经解决了的问题。
访谈里还有一个当时的细节:驾驶员监测靠方向盘扭矩,车上没有摄像头去看司机本人的状态。
太阳还剩五亿年,而火箭只是勉强飞得出去
最后一段,弗里德曼把话题引向那张照片。他刚见过卡尔·萨根的遗孀安·德鲁扬(Ann Druyan),顺势请马斯克朗读萨根写下的那段文字。
马斯克先讲了一段让人后背发凉的推演。
宇宙大约 138 亿岁,地球大约 45 亿岁。再过五亿年左右,太阳会膨胀,海洋蒸发,地球上的生命将无法存续。这意味着,如果意识的出现再晚 10% 的时间,它根本不会出现。只差 10%。
他顺着这个推想:宇宙里有多少只守着一颗星球、最终自我毁灭或者被外部因素抹掉的文明。他的估计是不少。
然后是火星。他说去火星这件事,其实只是刚刚够得着。
如果万有引力常数 G 再大 10%,火箭根本飞不出地球。如果 G 小 10%,事情会容易得多,从火星表面到地球表面,单级火箭就能完成。火星的重力大约是地球的 37%,真正的难关始终是离开地球那一脚。
物理常数只要偏一点点,整个故事就写不下去。这个观察跟前面那笔时间账放在一起,构成了同一个意思:人类目前的位置是极窄的一段窗口。
再看那个点。它就在这里。那是家。那是我们。在它上面,你爱的每一个人,你认识的每一个人,你听说过的每一个人,曾经存在过的每一个人,都在那里过完了自己的一生。我们的欢乐与痛苦的总和,成千上万种自信的宗教、意识形态和经济学说,每一位猎人和采集者,每一位英雄和懦夫,每一位文明的创造者和毁灭者,每一位国王和农夫,每一对相爱的年轻伴侣,每一位母亲和父亲,满怀希望的孩子、发明家和探险家,每一位讲授道德的老师,每一位腐败的政客,每一位超级明星,每一位最高领袖,我们这个物种历史上的每一位圣人与罪人,都住在那里,住在一粒悬浮在阳光中的微尘上。(马斯克朗读卡尔·萨根《暗淡蓝点》)
读到"没有别处可去"那句,马斯克停了下来。他说这句不对,至少在不远的将来,有一个地方可以去,火星。他补了一句,萨根如果活到今天大概也会同意,只是那个年代他想象不到。
这个打断很有意味。萨根那段文字的力量来自一个事实:地球是唯一的家,人类必须谦卑。马斯克打断了它,因为技术条件变了,谦卑的理由还在,去处的选项多了一个。
整场对话最耐人寻味的地方就在这里。同一个人,一边在算电极涂层的寿命和芯片的功耗,一边在算太阳膨胀的时间表。前者决定接下来五年能不能做成,后者决定接下来五百年还有没有得做。
Neuralink 的意义也落在同一个逻辑里。它先帮瘫痪的人重新握住杯子,帮中风的人重新组织句子,帮老人重新记住孩子的名字。等这条路走通了,人类才有可能在奇点到来之前多出一个选项。
他自己给这个选项打的分数很低,高于 0%,仅此而已。但按照他描述进步的方式,百万分之一会变成千分之一,会变成百分之一,会变成十分之一。概率靠做出来,预测算不出来。
- 想做 AI 安全,别只读论文。监管设计、标准制定、政策路径这些慢变量目前严重缺人,而历史反复证明,规则落地的时间差是用命填出来的。安全带被抵制了十多年,代价是几十万条命。
- 做硬件与医疗交叉方向的人,把材料科学排在第一位。最硬的一块骨头是让一根微米级的电极在人脑的电化学环境里撑过几十年不溶解、不出毒性,信号处理反倒排在后面。
- 任何打算规模化的医疗器械,从第一天就把自动化写进设计目标。依赖稀缺专家手工操作的方案,注定只能服务少数人,这既是成本问题也是公平问题。
- 做自动驾驶或者机器人,先把感知啃下来。把视觉、雷达、超声波融合成精确的向量空间表示,做对了这一步,规划和控制的难度会掉一个量级。
- 评估长期技术时同时算两个时间尺度:工程迭代的周期,和文明存续的窗口。前者决定你能不能做成,后者决定这件事值不值得做。
- 遇到无法检验的命题,学马斯克的处理方式,直接归类为个人偏好,然后把注意力挪开。精力留给能被证实或证伪的问题。
- 公开演示本身就是教育。让几十万台车自己开进小区,比一百场关于 AI 的演讲更能改变公众对机器人的认知,如果你在做类似的事,别低估这个副作用。
0:00The following is a conversation with Elon Musk, part two. The second time we spoke on the podcast with parallels, if not in quality, then in outfit to the objectively speaking greatest sequel of all time, Godfather Part II. As many people know, Elon Musk is a leader of Tesla, SpaceX, Neuralink, and The Boring Company.
0:24What may be less known is that he's a world-class engineer and designer, constantly emphasizing first principles thinking and taking on big engineering problems that many before him would consider impossible. As scientists and engineers, most of us don't question the way things are done.
0:42We simply follow the momentum of the crowd. But revolutionary ideas that change the world on the small and large scales happen when you return to the fundamentals and ask, is there a better way? This conversation focuses on the incredible engineering and innovation done in brain-computer interfaces in Neuralink.
1:04This work promises to help treat neurobiological diseases, to help us further understand the connection between the individual neuron to the high-level function of the human brain, and finally, to one day expand the capacity of the brain through two-way communication with computational devices, the internet, and artificial intelligence systems.
1:25This is the Artificial Intelligence podcast. If you enjoy it, subscribe on YouTube, Apple Podcasts, Spotify, support on Patreon, or simply connect with me on Twitter at Lex Fridman, spelled f r i d m a n. And now, as an anonymous YouTube commenter referred to our previous conversation as the quote historical first video of two robots conversing without supervision, here's the second time, the second conversation with Elon Musk.
1:57Let's start with a easy question about consciousness. In your view, is consciousness something that's unique to humans or is it something that permeates all matter, almost like a fundamental force of physics? I don't think consciousness permeates all matter. Panpsychists believe that.
2:14Yeah. There's a philosophical How would you tell? That's true. That's a good point. I believe in scientific method. I'm not a blind minor anything, but the scientific method is like you cannot test hypothesis and you cannot reach meaningful conclusion that it is true. Do you think consciousness understanding consciousness is within the reach of science?
2:35Of the scientific method? We can dramatically improve our understanding of consciousness. You know, I'd be hard-pressed to say that we understand anything with complete accuracy, but can we dramatically improve our understanding of consciousness? I believe the answer is yes. Does an AI system in your view have to have consciousness in order to achieve human-level or superhuman-level intelligence?
3:01Does it need to have some of these human qualities like consciousness, maybe a body, maybe a fear of mortality, capacity to love, those kinds of silly human things? It's it's different. You know, there's this there's the scientific method which I very much believe in where something is true to the degree that it is testable, so
3:27and and otherwise you're really just talking about, you know, preferences or or untestable beliefs or that, you know, that kind of thing. So, it ends up being somewhat of a semantic question where we are conflating a lot of things with the word intelligence. If we parse them out and say, you know, are we headed are we towards the future where an AI will be able to outthink us in every way?
4:01Then the answer is unequivocally yes. In order for an AI system that needs to outthink us in every way, it also needs to have a capacity to have consciousness, self-awareness, and understand self-aware, yes. That's different from consciousness. I mean, to me it's what what consciousness feels like. It feels like consciousness is in a different dimension.
4:26But this is this is could be just an illusion. You know, if you if you damage damage your brain in some way physically, you get you you damage your consciousness. Which implies that consciousness is a physical phenomenon. In in my view. The things that that I think are really quite quite likely is that digital intelligence will be able to outthink us in every way, and it will certainly be able to simulate what we consider consciousness.
4:56So, to to a degree that you would not be able to tell the difference. And from the from the aspect of the scientific method, it's might as well be consciousness if we can simulate it perfectly. If you can't tell the difference, then this is sort of the the Turing test, but think of a more sort of advanced version of the Turing test.
5:14If you if if you're if you're talking to a digital superintelligence and can't tell if that is a computer or a human, like let's say you're just having a conversation over a phone or a video conference or something where you're you're you think you're talking looks like a person, makes all of the right inflections and movements and and all the small subtleties that constitute a human and talks like a human, makes mistakes like a human. Like the at that and you literally just can't tell is this are you really conversing with a person or or a an AI?
5:56Might as well. Be human. So, on a darker topic, you've expressed serious concern about existential threats of AI. It's perhaps one of the greatest challenges our civilization faces, but since I would say we're kind of an optimistic descendants of apes, perhaps we can find several paths of escaping the harm of AI. So, if I can give you three options, maybe you can comment which do you think is the most promising.
6:23So, one is scaling up efforts on AI safety and beneficial AI research and and hope of finding an algorithmic or maybe a policy solution. Two is becoming a multi-planetary species as quickly as possible. And three is merging with AI and and riding the wave of that increasing intelligence as it continuously improves.
6:48What do you think is most promising, most interesting as a civilization that we should invest in? I think there's there's a lot a tremendous amount of investment going on in AI. Where there's a lack of investment is in AI safety. And there should be, in my view, a government agency that oversees anything related to AI to confirm that it is does not represent a public safety risk. Just as there is a regulatory authority for there's like the Food and Drug Administration, there's NHTSA for auto automotive safety, there's the FAA for aircraft safety.
7:24We generally come to the conclusion that it is important to have a government referee or or a referee that is serving the public interest and and ensuring that things are safe when when there's a potential danger to the public. I would argue that AI is unequivocally something that has potential to be dangerous to the public and therefore should have a regulatory agency just as other things that are dangerous to the public have a regulatory agency.
7:52And but let me tell you the problem with this is that the the government moves very slowly. And the the rate of the rate the usually way a regulatory agency comes into being is that something terrible happens. There's a huge public outcry and years after that there's a regulatory agency or a rule put in place. Take something like like seat belts. It was known for I don't know, a decade or more that seat belts would have a massive impact on safety and and save so many lives and serious injuries.
8:29And the car industry fought the requirement to put seat belts in tooth and nail. That's crazy. Yeah. And and I don't hundreds of thousands of people probably died because of that. And they said people wouldn't buy cars if they had seat belts, which is obviously absurd. You know, or look at the back tobacco industry and how long they fought anything about smoking.
8:52That's part of why I helped make that movie Thank You for Smoking. You can sort of see just how pernicious it can be when you have these companies effectively achieve regulatory capture of of governments the bad. People in the AI community refer to the advent of digital superintelligence as a singularity.
9:19That that is not to say that it is good or bad, but it that it is very difficult to predict what will happen after that point and and that there's some probability it will be bad, some probability it will be it will be good. We obviously want to affect that probability and have it be more good than bad.
9:38Well, let me on the merger with AI question and and the incredible work that's being done at Neuralink, there's a lot of fascinating innovation here across different disciplines going on. So, the flexible wires, the robotic sewing machine that responds to brain movement and everything around ensuring safety and so on.
9:57So, we currently understand very little about the human brain. Do you also hope that the work at Neuralink will help us understand more about our about the human mind, about the brain? Yeah, I think the work at Neuralink will definitely shed a lot of insight into how the brain and the mind works.
10:20Right now, there's the data we have regarding how the brain works is is very limited. You know, we've got fMRI, which is that that's kind of like putting a you know, a stethoscope on the outside of a factory wall and and then putting it like all over the factory wall and you can sort of hear the sounds, but you don't know what machines are doing, really. I It's hard you can infer a few things, but it's very broad brush stroke.
10:46In order to really know what's going on in the brain, you really need you have to have high precision sensors. And then you want to have stimulus and response. Like if you trigger a neuron, what how do you feel? What do you see? How does it change your perception of the world? You're speaking to physically just getting close to the brain, being able to measure signals from the brain.
11:03Yeah. Will give us sort of open the door and in inside the factory. Yes, exactly. Being able to have high precision sensors that that tell you what individual neurons are doing. And then being able to trigger neuron and see what the response is in the brain. So, you can see the consequences of of a if you fire this neuron, what happens?
11:28How do you feel? What does it change? It it'll be really profound to have this in people because people can articulate what their change like if there's a change in mood or if if they if you know if they can tell you if they can see better or hear better or be able to form sentences better or worse or you know their memories are jogged or that you know that kind of thing.
11:52So on the on the human side there's this incredible general malleability plasticity of the human brain. The human brain adapts adjusts and so on. So that's the It's not that plastic to be totally frank. So there there's a firm structure but nevertheless there is some plasticity and the open question is so if I could ask a broad question is how much that plasticity can be utilized. Sort of on the human side there's some plasticity in the human brain and on the machine side we have neural networks machine learning artificial intelligence it's able to adjust and figure out signals. So there's a mysterious language that we don't perfectly understand that's within
12:32the human brain and then we're trying to understand that language to communicate both directions. So the brain is adjusting a little bit we don't know how much and the machine is adjusting. Where do you see as they try to sort of reach together almost like with an alien species try to find a protocol communication protocol that works. Where do you see the biggest the biggest benefit arising from on the machine side or the human side do you see both of them working together?
13:02I think the machine side is far more malleable than the biological side. By by a huge amount. So it will be the the machine that adapts to the brain. It that's the only thing that's possible. The brain can't adapt that well to to to the machine. You can't have neurons start to regard an electrode as another neuron cuz like neurons just they just like the pulse and so if something else is pulsing see so this there is that there is that that elasticity in the interface which we believe is is is something that can can happen.
13:34But the vast majority of the malleability will have to be on the machine side. But it's interesting when you look at that synaptic plasticity at the interface side, there might be like an emergent plasticity. Because it's a whole another it's not like in the brain it's a whole another extension of the brain.
13:51You know, we might have to redefine what it means to be malleable for the brain. So maybe the brain is able to adjust to external interfaces. There will be some adjustments to the brain cuz this there's going to be something reading and simulating the the brain and so it will adjust to to that thing.
14:09But but most the vast majority of the adjustment will be on the machine side. This is just if this is just it has to be that otherwise it will not work. Ultimately like we know currently operate on two layers. We have sort of limbic like prime primitive brain layer which is where all of our kind of impulses are coming from.
14:29It's sort of like we've got we've got like a monkey brain with a computer stuck on it. That's that's the human brain. And a lot of our impulses and everything are driven by the monkey brain. And the the computer the cortex is constantly trying to make the monkey monkey brain happy. It's not the cortex that's steering the monkey brain. It's the monkey brain steering the cortex.
14:51You know, But the cortex is the part that tells the story of the whole thing. We convinced ourselves it's it's more interesting than just the monkey brain. The cortex is like what we call like human intelligence. You know, it's just like the that's like the advanced computer relative to other creatures.
15:06other other creatures do not have either or really they don't they don't have the computer or they have a very weak computer relative to humans. But but it's this it's like it sort of seems like surely the really smart thing should control the dumb thing, but actually the dumb thing controls the smart thing.
15:28So, do you think some of the same kind of machine learning methods, whether that's natural language processing applications, are going to be applied for the communication between the machine and the brain? And do to learn how to do certain things like movement of the body, how to process visual stimuli, and so on.
15:47Do you see the value of using machine learning to understand the language of the two-way communication with the brain? Sure. Yeah, absolutely. I mean, we're a neural net. And and that, you know, AI is basically neural net. So, it's like digital neural net will interface with biological neural net.
16:08And hopefully bring us along for the ride, you know. But the vast majority of of of our of our intelligence will be digital. Does it like like So, like think of like the the difference in intelligence between your the cortex and your limbic system is gigantic. Your your your limbic system really has no comprehension of what the hell the cortex is doing.
16:33you know, it's just literally hungry, you know, or tired, or angry, or thirsty or something, you know. And and and just and and then it that communicates that that impulse to the cortex and tells the cortex to go satisfy that. Then a lot of a great deal of like a massive amount of thinking, like truly stupendous amount of thinking, has gone into sex.
16:58Without purpose. Without procreation without procreation. Which which which is actually quite a silly action in the absence of procreation. It's it's a bit silly. Well, so why are you doing it? Because it makes the limbic system happy. That's why. That's why. But it's pretty absurd, really.
17:21Well, the whole of existence is pretty absurd in some kind of sense. Yeah. But but I mean this is a lot of computation has gone into how can I do more of that with procreation not even being a factor. This is I think a very important area of research by NSF W. an agency that should receive a lot of funding, especially after this conversation.
17:44I'm proposing formation of a new agency. Oh boy. What is the most exciting or some of the most exciting things that you see in the future impact of Neuralink? Both on the science, the engineering, and societal broad impact. So Neuralink, I think at first will solve a lot of brain-related diseases.
18:07So it could be anything from like autism, schizophrenia, memory loss. Like everyone experiences memory loss at at certain point in in age. Parents can't remember their their kids' names and that kind of thing. So there's I think a tremendous amount of good that Neuralink can do in solving critical
18:27critical damage to brain or the spinal cord. There's a lot that can be done to improve quality of life of individuals, and that will be those will be steps along the way. And then ultimately it's intended to address the the risk the existential risk associated with digital superintelligence.
18:46like we will not be able to be smarter than a a a digital supercomputer. so therefore, if you cannot beat them, join them. And at least we will have that option. So you you you have hope that Neuralink will be able to be a kind of connection to allow us to to merge, to ride the wave of the improving AI systems.
19:12I think the chance is above 0%. So, it's non-zero. Yeah. There's a chance. And that's So, what if you've seen Dumb and Dumber? Yes. Yes. So, I'm saying there's a chance. You're saying one in a billion or one in a million, whatever it was in Dumb and Dumber. You know, it went from maybe one in a million to improving maybe it'll be one in a thousand and then one in a hundred, then one in ten. Depends on the rate of improvement of Neuralink and how fast we're able to do make progress, you know.
19:41Well, I've talked to a few folks here. They're quite brilliant engineers, so I'm I'm excited. Yeah, I I think it's like fundamentally good, you know, who you know, giving somebody back full motor control after they've had a spinal cord injury, you know, restoring brain functionality after a stroke, solving debilitating genetically oriented brain diseases. These are all incredibly great, I think. And in order to do these, you have to be able to interface with the neurons at a detail level and need to be able to fire the right neurons, read the right neurons, and and then effectively you can create a a circuit, replace what's broken with with silicon and essentially fill in the
20:22the missing functionality. And then over time we can have we develop a tertiary layer. So, if like the limbic system is a primary layer, then the cortex is like a second the second layer. and I said that you know, the obvious the cortex is vastly more intelligent than the limbic system. But people generally like the fact that they have a limbic system and a cortex. I haven't met anyone who wants to delete either one of them.
20:46They're like, "Okay, I'll keep them both. That's cool." The limbic system's kind of fun. That's where all the fun is. You know, absolutely. and then you you people generally don't want to lose the cortex, either. Right? So, they like having the cortex and the limbic system. Yeah.
21:00and and then there's a tertiary layer, which will be digital super intelligence. And I I there's room for optimism given that the cortex the the the cortex is very intelligent and Olympic system is not and yet they work together well. Perhaps there can be a tertiary layer where where digital superintelligence lies and that that will be vastly more intelligent than the cortex but still coexist peacefully and in a benign manner with the cortex and Olympic system.
21:29That's super exciting future both on the low-level engineering that I saw as being done here and the actual possibility in the next few decades. It It's important that Neuralink solve this problem sooner rather than later because the point at which we have digital superintelligence that's when we pass the singularity and and things become just very uncertain. It doesn't mean that they're necessarily bad or good but the point at which we pass the singularity things become extremely unstable. So, we want to have a human brain interface before the singularity or at least not long after it to minimize existential risk for humanity and consciousness as we know
22:04it. So, but there's a lot of fascinating actual engineering low-level problems here at Neuralink that are quite quite exciting. What The the problems that we face at Neuralink are material science, electrical engineering, software, mechanical engineering, microfabrication. It's a bunch of engineering disciplines essentially.
22:27That's what it comes down to is that you have to have a a tiny electrode that's so it's so small it doesn't hurt neurons. but it's got to last for as long as a person. So, it's got to last for decades. and then you've got to take that signal you've got to process that signal signal locally at low power. So, we need a lot of chip design engineers that are you know cuz we're going to do signal processing and do so in a very power efficient way so that we don't heat your brain up.
23:01cuz the brain's very heat sensitive. and then and then we're going to take those signals and we're going to do something with them and then we're going to stimulate stimulate the back to to, you know, so you could bidirectional communication. so if somebody's good at material science, software, mechanical engineering, electrical engineering, chip design, microfabrication, that's what those are the things we need to work on.
23:27We need to be good at material science so that the that we can have tiny electrodes that last long time and it's and it's a tough thing with the the material science problem is a tough one because you're trying to read and stimulate electrically in an electric an electrically active area. Your brain is very electrically active and electrochemically active. So, how do you have a a say a coating on the electrode that doesn't dissolve over time?
23:52and and and is safe in the brain. This is a very hard problem. And and then and then how do you collect those signals in a way that is most efficient cuz you you you really just have very tiny amounts of power to process those signals. You know, and then we need to automate the whole thing so it's like LASIK, you know, so so it's it's not if this is done by neurosurgeons, there's no way it can scale to a large numbers of people.
24:23And it needs to scale to large numbers of people cuz I think ultimately we want the future of Peter to be determined by a large number of the the of humans. Do you think that this has a chance to revolutionize surgery period? So, neurosurgery and surgery overall? Yeah, for sure. It's it's got to be like LASIK. Like if if LASIK had to be hand done done by hand by a person, that wouldn't be great.
24:46Yeah. You know. It's it's done by a robot. and the ophthalmologist kind of just needs to make sure your you're head's in the right position and then they just press the button and go. So, the smart summon and soon auto park takes on the full beautiful mess of parking lots and their human-to-human nonverbal communication.
25:08I think it has actually the potential to have a profound impact in changing how our civilization looks at AI and robotics because this is the first time human beings, people that don't own a Tesla, maybe never seen a Tesla or heard about a Tesla, get to watch hundreds or thousands of cars without a driver.
25:27Yeah. Do you see it this way, almost like an education tool for the world about AI? Do you feel the burden of that, the excitement of that, or do you just think it's a smart parking feature? I do think you're you're are getting at something important there, which is most people have never really seen a robot.
25:45Or and what what is the car that is autonomous? It's a four-wheeled robot. Right. Yeah, it communicates a certain sort of message with everything from safety to the possibility of what AI could bring to its current limitations, its current challenges, it's what's possible. Do you feel the burden of that, almost like a communicator, educator to the world about AI?
26:08We're just really trying to make people's lives easier with autonomy. But now that you mention it, I think it will be an eye-opener to people about robotics because they've really never seen Most people have never seen a robot and there are hundreds of thousands of Teslas. Won't be long before there's a million of them that have autonomous capability and they drive without a person in it. And you you you can see the kind of evolution of the car's personality and and thinking with each iteration of of autopilot. You
26:40can see it's it's uncertain about this or it gets if it now it's more certain. Now now it's moving in a slightly different way. Like I can tell immediately if a car is on Tesla autopilot cuz it's got these little nuances of movement. It just moves in a slightly different way. it's it's cars on on Tesla autopilot, for example, on the highway are far more precise about being in the center of the lane than a person.
27:05if you drive down the highway and look at how it where cars are the human-driven cars are in within their lane. They're like bumper cars. They're like moving all over the place. Car on autopilot, dead center. Yeah, so the incredible work that's going into that neural network. It's learning fast. autonomy is still very, very hard. We don't actually know how hard it is fully, of course.
27:29you look at the most problems you tackle, this one included, in with an exponential lens. But even with an exponential improvement, things can take longer than expected sometimes. So, where does Tesla currently stand on its quest for full autonomy? What what's your sense? When can we see successful deployment of full autonomy?
27:56Well, on the highway already, the the probability of intervention is extremely low. Yes. so for highway autonomy, with the latest release, especially, probability of needing to intervene is is is really quite low. In fact, I'd say for stop-and-go traffic, the it's as far safer than a person right now. In stop-and-go, the probability of an injury or an impact is much, much lower for autopilot than a person.
28:24And then with navigate on autopilot, you can change lanes, take highway interchanges, and then we're coming at it from the other direction, which is low-speed full autonomy. And in a way, this is like like how how does a person learn to drive? You learn to drive in the parking lot.
28:39You know, you don't first time you learned to drive probably wasn't jumping on Lombard Street in San Francisco. That'd be crazy. You learn to drive in in the parking lot, get things get things right at low speed. And and then the missing piece that we're working on is traffic lights and stop streets. Stop streets Stop streets I would say actually also relatively easy because you know, you kind of know where the stop street is. Worst case you can geocode it and then use visualization to see where the line is and stop at the line to eliminate the GPS error.
29:11So it's actually I'd say it's probably complex traffic lights and very windy roads are the two things that need to get solved. What's harder, perception or control for these problems? So being able to perfectly perceive everything or figuring out a plan once you perceive everything, how to interact with all the agents in the environment? In your sense, from a learning perspective is perception or action harder?
29:38In that giant beautiful multitask learning neural network. The The hardest thing is having accurate representation of the physical objects in vector space. So trans take taking the visual input primarily visual input some sonar and radar and and then creating the an accurate vector space representation of the objects around you.
30:03Once you have an accurate vector space representation, the planning and control is relatively easier. I'd say it's relatively easy. Basically, once you have accurate vector space representation, then then you're you're you're kind of like a video game. Like cars in like Grand Theft Auto or something. Like they work pretty well. They drive down the road, they don't crash, you know, pretty much unless you crash into them.
30:25that's because they've they've got an accurate vector space representation of where the cars are and they're just and then they're rendering that as the as the output. Do you have a sense, high level, that Tesla's on track on being able to achieve full autonomy? So on the highway absolutely.
30:44and still no driver state. driver sensing And we have driver sensing with torque on the wheel. That's right. Yeah. By the way, just a quick comment on karaoke. Most people think it's fun, but I also think it is a driving feature. I've been saying for a long time singing in the car is really good for attention management and visual management.
31:04Tesla karaoke yeah, it's great. It's the one of the most fun features of the car. Do you think of the connection between fun and safety sometimes? Yeah, if you can do both at the same time, that's great. I just met with Ann Druyan, wife of Carl Sagan Oh, yeah. who directed Cosmos.
31:22I'm totally a big fan of Carl Sagan. He's he's super cool. And he had a great way of putting things. All of our consciousness, all civilization, everything we've ever known and done is on this tiny blue dot. People also get they get too trapped in their s- like squabbles amongst humans.
31:38And they just don't think of the big picture. They take civilization and our continued existence for granted. They shouldn't do that. Look at the history of civilizations. They rise and they fall. And now civilization is all it's globalized. And so civilization, I think now rises and falls together. There's no There's not geographic isolation.
32:04This is a big risk. Things don't always go up. That should be that's an a lesson of history. In 1990, at the request of Carl Sagan, the Voyager 1 spacecraft, which is a spacecraft that's reaching out farther than anything human-made into space, turned around to take a picture of Earth from 3.7 billion miles away.
32:28And as you're talking about the pale blue dot, that picture, the Earth takes up less than a single pixel in that image. Yes. appearing as a tiny blue dot as a pale blue dot as Carl Sagan called it. So, he spoke about this dot of ours in 1994. And if you could humor me I was wondering if in the last 2 minutes you could read the words that he wrote describing this pale blue dot.
32:57Sure. Yes, it's funny that the universe appears to be 13.8 billion years old. Earth is like 4 and 1/2 billion years old. You know, another half billion years or so the sun will expand and probably evaporate the oceans and make life impossible on Earth. Which means that if it had taken consciousness 10% longer to evolve, it would never have evolved at all.
33:27Just 10% longer. and I wonder I wonder how many dead one-planet civilizations there are out there in the cosmos that never made it to the other planet and ultimately extinguished themselves or were destroyed by external factors. Probably a few. It's only just possible to try to travel to Mars, just barely.
33:52If G was 10% more wouldn't work, really. If if G was 10% lower, it would be easy. Like you can go single-stage from surface of Mars all the way to surface of the Earth. Cuz Mars is 37% of Earth's gravity, thereabouts. You need a giant booster to get off of Earth. Channeling Carl Sagan.
34:21Look again at that dot. That's here. That's home. That's us. On it, everyone you love, everyone you know, everyone you've ever heard of every human being who ever was, lived out their lives. The aggregate of our joy and suffering, thousands of confident religions, ideologies, and economic doctrines, every hunter and forager, every hero and coward, every creator and destroyer of civilization, every king and peasant, every young couple in love, every mother and father, hopeful child, inventor, and explorer,
34:57every teacher of morals, every corrupt politician, every superstar, every supreme leader, every saint and sinner in the history of our species lived there, on a mode of dust suspended in a sunbeam. Our planet is a lonely speck in the great enveloping cosmic dark. In our obscurity, in all this vastness, there is no hint that help will come from elsewhere to save us from ourselves.
35:25The Earth is the only world known so far to harbor life. There is nowhere else, at least in the near future, to which our species could migrate. This is not true. This is false. Mars. And I think Carl Sagan would agree with that. He couldn't even imagine it at that time. So, thank you for making the world dream, and thank you for talking today.
35:47I really appreciate it. Thank you.
资源与出处
- 视频在 YouTube 收看
- 视频在 B 站收看UP 主「云逐月明」搬运
- 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