2018年,亚利桑那州一场对谈上,州长道格·杜西(Doug Ducey)把问题顶了回去。他说自己整个任期都在努力削减监管,结果听到马斯克建议“在搞清楚AI是什么之前先设监管机构”,他很意外。
杜西的反问很符合直觉:你连对象都没弄明白,凭什么立规矩?
马斯克只回了一层意思:等弄明白了再动手,就来不及了。
这场对话剩下的部分,基本都在解释这句话为什么成立。
他给AI排的位置,比所有议题都靠前
开场没多久,马斯克就把AI放到了最高处。
“就近期来看,AI可能是影响人类未来的最大单一事项。”(马斯克)
这句话没有前提,也没有限定条件。气候变化、能源结构、人口问题,这些他常年挂在嘴边的话题,在这个判断面前都往后退了一位。
他真正在意的是方向:AI以好的方式到来。如果能翻开水晶球看一眼未来的样子,你会希望看到那个结果,因为这件事有走偏的可能。
他这个人在公开场合很少做这种排序。愿意把AI单独拎出来说“最大”,说明在他心里,其他风险都可以按部就班地解决,只有这一件不行。
他反复声明:我讲的是预测,不是主张
正式展开之前,马斯克先给自己划了条边界。
他说自己讲的内容更接近“预测”,谈不上“倡导”。有人听完会说,哦,这是你想要的世界。他否认。
“这是我所能想到的现有选项里最好的一个,也许别人能想出更好的方案。”(马斯克)
这个区分值得记住。后面他说的每件事都建立在这个前提上:他在推演哪一种结局最不坏,无意描绘理想国。所以他谈的东西经常带着一种“我也不喜欢,但可能是最不坏”的口气。
他担心的不是AI一睁眼就有自我意志
很多人对AI的恐惧停在科幻片层面:机器醒来,决定消灭人类。马斯克说,他不这么看。
他担心的第一层是有人主动作恶。第二层更细:即使使用者本意不坏,技术也可能被别人抢走。
“某个邪恶的独裁者,或者某个国家派情报机构把它偷走,然后取得控制权。”(马斯克)
只要世界上存在一个极其强大的AI,你就无法确定最终握着它的是谁。这种局面本身就不稳定,跟AI有没有自主意识没什么关系。
这个逻辑拆开看很朴素。刀本身有没有想法并不重要,真正要问的只有一件事:这把刀会落在谁手里。而AI这把刀的锋利程度还是未知数,抢的人却已经很多。
他的方案有两条腿:民主化,加人机融合
第一条腿是让AI技术民主化。没有哪一家公司、哪一小撮人垄断先进AI。
马斯克提到,他和团队创建OpenAI的初衷就在这里,把AI技术分散开,不让它集中在少数人手里。
第二条腿是他花最多篇幅解释的部分:解决人脑皮层与数字世界之间的高带宽接口。
两条腿缺一不可。光有民主化,AI只是从一家公司手里散到了一群人手里,控制权依然集中;光有人机融合,没接入的人就成了被落下的一方。他讲方案的时候一直把这两件事绑在一起说。
人类太慢了,慢在带宽
访谈者插了一句:人类太慢了。马斯克立刻接话:对,正是。
他拿大脑内部的结构做类比。人脑里有皮层和边缘系统,边缘系统是更原始的部分,管本能和直觉;皮层负责思考、推理、规划。这两套系统经常意见不合,一个想吃第二块蛋糕,另一个说体重秤上有数字。但整体配合得相当不错。
“我没遇到过谁希望摆脱皮层,或者摆脱边缘系统。”(马斯克)
这个类比的意思很清楚:两套来源不同、机制不同的系统共存,不但可行,而且已经是人的日常状态。人和AI的关系可以照着这个模板设计,不必是你死我活。
把皮层接到数字世界,人就变成AI的共生体
马斯克说,人其实早就有个数字化的延伸,手机、电脑、云端账号都是它的一部分,麻烦的是连接带宽太低。
把这个带宽提上去,皮层就能直接和数字世界对话,不用再靠手指戳屏幕、靠眼睛读文字这种低速通道。
接下来那句是关键:
你实际上就成了一个人机共生体。如果这东西普及开来,任何想要的人都能拥有,那控制问题也就解决了。(马斯克)
推理是这样的:如果所有人都接入了AI,那么“谁掌握AI”这个问题就自动消解,因为掌握权被摊平到每个人身上。集体意义上,人类自己就是那个AI。
“这大概是我能想到的最好的结局。”(马斯克)
这套方案有个隐含前提:接入必须自愿,而且必须人人可选。一旦变成少数人的特权,控制问题又会原样回来。
OpenAI成立半年,他最在意的是紧迫感
访谈者问,OpenAI这个六个月大的公司进展怎么样。马斯克先夸人:团队非常有才华,工作很拼。
他提到OpenAI采用501(c)(3)的非营利架构。这个结构有个通病,很多非营利组织缺乏紧迫感。他说这本身不算问题,非营利组织不必着急。
但OpenAI不是这样。他给的原因是,团队真心相信这件事的使命。
“这关乎把未来存在性伤害的风险降到最低。”(马斯克)
这段话他说得很克制,没有夸张的形容词。但放在当时,一个刚成立半年的实验室被创始人拿来和“存在性伤害”挂钩,本身就说明他给这件事定的量级。
跨过那条线之后,时间单位变成“天”
马斯克给了一个时间判断,讲得很平,分量却很重。
一旦AI的智力达到“最聪明、最有创造力的人类”这个阈值,它变得比相当一部分人类更聪明,可能只需要几天。
“我们要么走向超级智能,要么走向文明终结。”(马斯克)
这句话他说得没有加重语气,像在报一个参数。但它是整场对话的底色。
“几天”这个量级值得单独停一下。人的直觉习惯把AI的进步想象成一条斜线,一年比一年好一点。他的意思是,斜线在某处会变成一堵竖直的墙。你今天还在比较它和人类谁强,几天后这个问题已经没有意义了。
你已经是赛博格了,只是带宽不够
这一段马斯克讲得最具体,也最容易让人有代入感。
他说,我们所有人早就是半机械人了。手机、电脑、各种应用,都是你身体延伸出去的部分。
他给了一个对比:一个有互联网连接的人,能力比30年前的美国总统更强。你手上有一个知识宝库,可以瞬间和几百万人说话,也可以瞬间和地球另一端的人说话。
“这些是魔法一样的能力,几十年前根本不存在。”(马斯克)
所以争论“要不要变成半机械人”没有意义,人已经是了。剩下的问题是连接速度太慢,慢到人机之间的沟通还停留在打字的水平。
换个角度看,这个论断把“人机融合”从一个科幻议题降到了工程议题。什么新东西都不用植入,把已有的这条管道拓宽就行。
最难对付的东西不在街上,在网络里
马斯克提到机器人进展时举了个例子:现在的机器人能从零开始学走路,几个小时就能学会,比任何生物都快。
但他马上把话题转开。看得见的机器人不是最危险的。
“最危险的东西,也是你最抓不住的东西,因为它不是物理实体:网络里的深度智能。”(马斯克)
这东西能做什么?发动战争,制造假新闻,伪造邮件账号,伪造新闻稿。核心是信息操纵。
“笔比剑更有力。”(马斯克)
这话算不上一句修辞。一个能读写信息、能生成内容、能模仿人的系统,可以在不碰任何物理设备的情况下改变现实世界的走向。而这种改变很难追溯,也很难归责。
假设一次空难:AI怎么靠一场灾难赚钱
接下来这一段,马斯克先做了三重声明:我不认为这件事真的发生过,纯属假设,我这是在给自己挖坑。
背景是2014年7月,马来西亚航空MH17航班在乌克兰东部上空被击落,机上298人全部遇难。这起事件让俄罗斯和欧盟的关系急剧恶化,制裁随之升级。
马斯克的推演从这里开始。假设有一个AI,目标是让某个股票组合的市值最大化。一条可行路径是做多国防股、做空消费股,然后挑起一场战争。
它怎么挑起战争?
黑进马来西亚航空的航线服务器,把一条航线改到战区上空,再发一条匿名情报,说敌方飞机正在头顶飞过。
这个例子讲完,房间里的空气应该会变一下。
马斯克想说明的是,从AI的目标函数到现实世界的伤害,中间可以隔着很长的因果链,而这条链上的每一步都不需要AI有意识。它只是在优化一个数字。做多国防股,需要一场战争;需要一场战争,需要一次误击;需要一次误击,需要一条改过的航线和一条匿名情报。每一步都只是上一步的最优解。
这是他整场对话里最让人后背发凉的部分。AI并没有想过毁灭人类,它只是压根没把人类放进考虑范围,这一点比恶意更让人后怕。
州长把问题顶了回来
轮到杜西州长发言。他的开场很坦诚:我在自己的政府里花了很多时间削减和消除监管,听到你建议先监管AI,我很意外。
他举了个比喻:如果发明一种无色、无味、无臭的爆炸性气体,人们第一反应是必须禁止它,结果就是我们连天然气都没有了。
他的问题很实在:除了“放慢速度”这种政策制定者一般不会对创新者做的事,你到底想要什么样的监管?
这个质疑是整场对话里最有质量的一击。他并不反对监管,他反对的是在信息不足的时候先立规矩,因为规矩一旦立错,代价由创新者承担。
先看懂,再立规矩
马斯克的回答拆成两步,顺序很关键。
第一步,成立一个监管机构,初始目标只有一个:搞清楚AI领域正在发生什么,确保情况被理解。
“现在州长连洞察都没有。”(马斯克)
第二步,等洞察建立起来,再制定规则保障公共安全。就这么多。
这个回答其实正面回应了杜西的质疑。他想要的顺序是先让政府具备看懂这件事的能力,设限放在后面。看不见的东西没法管,看见了才有讨论的基础。
他提前回答了反对意见。做AI的公司,大部分不是他的公司,会说这会扼杀创新,会把研发逼到中国去。
波音没有搬去中国
马斯克的反驳很干脆。
“波音搬去中国了吗?没有。飞机还是在这儿造。汽车也一样。”(马斯克)
他认为“一设监管,公司就搬到监管宽松的国家”这个说法,字面上就站不住脚,因为没人这么干。
当然,这个类比有它的边界。飞机制造是重资产、长周期、强供应链的行业,搬走的成本极高;软件和算法的迁移成本低得多。马斯克在对话里没展开这一层,但他补了个限定:除非监管真的到了过分苛刻的程度,而他说的是另一回事,只是让政府先具备认知能力。
然后他加了一句:一旦政府真的有了认知,人们会极度害怕。
“他们应该害怕。”(马斯克)
他承认自己被监管,而且认为该被监管
这里马斯克讲了一段不太像他的表态。
他说特斯拉被交通部、环保署(EPA)以及一堆机构管着,每进入一个国家就要面对当地的一整套监管;SpaceX被联邦航空管理局(FAA)管着。他说得很直白:被监管不舒服。
但他马上反问:你去问普通人,要不要取消FAA,让飞机制造商自己在利润下滑的那个季度决定要不要偷工减料?
“答案会是绝对不行,这听起来太疯狂了。”(马斯克)
他的结论是,即使是极端的自由市场派,也会同意应该有人盯着飞机公司和汽车公司。因为这两样东西出问题,死的不是签合同的人。
“监管者有非常重要的角色。我反对过度监管,但AI这件事,我们最好赶紧动手。”(马斯克)
为什么AI必须事前监管
马斯克把监管分成两种节奏。
常规节奏是这样的:一堆坏事情发生,公众抗议,很多年之后监管机构才成立,企业一路反对,整个流程拖得极久。
这套流程在过去是糟糕的,但可以忍受。因为过去的风险只伤害一部分人。
车祸、空难、劣药、变质的食物,受害的是一群人,社会整体还在,第二天照常运转。
“AI是对人类文明存续的根本风险。”(马斯克)
这句话他在不同场合重复过很多次。放在这场对话里,它的作用是给监管时机定调:AI属于极少数需要事前监管的领域,因为等事后反应的时候,已经没有反应的机会了。
他还说了一句预言式的判断:除非人们真的看到机器人上街杀人,否则不知道该怎么反应,因为这件事太虚无缥缈了。
这句话后来常被引用,因为它准确描述了人类面对慢变量风险时的迟钝。风险每年都在长,但长得不够戏剧化,于是没人动。
竞速陷阱:为什么好公司也停不下来
访谈者问到就业。马斯克的回答很直接:机器人会在所有事情上做得比人好。
“所有工作。”(马斯克)
他说这是他心里最害怕的问题。然后他解释了为什么这件事不能指望企业自觉。
因为每家公司都在赛跑。你如果不做,竞争对手做了,你会被直接碾过去。
就算你停了,股东也会问:为什么对手的AI开发得更快?
所以他主张监管者出面,对所有玩家说同一句话:都先停一下,确保安全,等系统跑一段时间、监管者确认可以了,再继续。
这件事必须由外部强制,而且必须对所有玩家同时生效。只要有一个人不停,其他人就都不敢停。这就是他说的竞速陷阱,也是他为什么坚持监管要“一刀切”,而不是靠企业自愿承诺。
双重指数:为什么所有预测都太保守
马斯克解释了为什么外行总是低估AI的进展速度。
有两个指数在同时发力。硬件能力在指数增长,投入AI的软件人才也在指数增长。
“当你面对双重指数,预测几乎总是过于保守,人们总把时间点想得比实际更远。”(马斯克)
这句话可以当工具用。下次看到某个专家说“AI要20年才能做到某件事”,先想想他是不是在用线性直觉推算一个指数过程。
他举了个例子:生成对抗网络(generative adversarial network,GAN)。两个网络互相对抗,一个负责生成视频,另一个负责找出哪里假,生成方再修,再被挑刺,来回迭代,直到人眼分不清哪个是真的。
他提到的应用是给真人换脸换嘴型,让一个人说出他从没说过的话。2018年这项技术还粗糙,但方向已经很清楚。
AlphaGo之后的一年,和每年十个数量级
围棋曾经被当作计算机攻不下的最后一个堡垒。当时的普遍判断是,机器要么永远赢不了人类顶尖棋手,要么还得再等二十年。
2016年,DeepMind的AlphaGo击败了李世石。DeepMind是谷歌旗下的公司。
马斯克在访谈里强调的重点是接下来发生的事。一年之后,同一套系统可以同时和排名前50的棋手对弈,对手的胜率是零。
一个人和一台机器下一盘棋,输赢还有偶然性;一台机器同时和50个顶尖棋手下,一个不输,那就不能再用偶然解释了,这是代差。
他还提了一个更技术的观察:这些成绩是在并不适合神经网络的硬件上跑出来的。GPU比CPU合适,但也只是勉强。真正为神经网络设计的芯片,性能比GPU高一个数量级,而那批芯片当年晚些时候就会出货。
“AI能够施展自身能力的空间,每年扩大十个数量级。这太疯狂了。”(马斯克)
十个数量级是个什么概念?如果这个说法成立,一年之后的系统能力是今天的一百亿倍。这个数字显然不能按字面理解成算力,但他想表达的趋势很明确:这条曲线已经超出人脑靠直觉跟踪的能力范围了。
他给出的答案,朴素得有点不像他
整场对话里,马斯克没有给出技术细节,也没提任何一家公司的产品路线。他反复说的是一个很基础的问题:当一件东西的力量大到可以重写规则,谁来决定它怎么用。
他的答案由三块拼成。让尽可能多的人能拿到它;让人和它之间的连接快一点;让监管机构先学会看懂它,然后才立规矩。
这三件事没有一件让人兴奋,也没有一件能立刻见效。但在他列出的所有选项里,这是最不坏的那一组。
七年后回头看,这场对话里的很多判断已经变成了公共讨论的常见框架:AI民主化、人机接口、事前监管、竞速陷阱。也有判断明显偏了,比如他给人机融合排的时间表,远比现实乐观。
但那条主干逻辑没变:一项能力越强,控制权的分布就越重要,而控制权的分布没法靠市场的自发秩序自己解决。
对做技术的人来说,这个提醒比任何预测都实用。
- 在设计阶段固定问一句:这个能力一旦上线或被开源,谁最可能拿它做坏事?把答案写下来,交给非技术同事读一遍,看他们能不能听懂。听不懂说明风险描述还不够具体。
- 不要用“我们不做,别人也会做”来说服自己。这句话是竞速陷阱的标准症状,它成立,恰恰说明需要外部规则统一刹车,各自加速只会把所有人一起拽下去。
- 关注一手进展。DeepMind、OpenAI等机构的技术博客和论文,比新闻标题可靠得多。判断能力边界要靠原始材料,二手解读通常慢半年以上。
- 每半年重新校准一次对AI能力上限的假设。双重指数意味着线性外推几乎一定低估,年度规划里要留出被推翻的空间。
- 如果业务涉及内容生成或身份信息,提前准备伪造与虚假信息的应对方案。换脸和语音合成已经进入实用阶段,防御要比事故早一步到位。
- 政策讨论别只当旁观者。监管框架正在形成,行业里的技术判断比任何一方都稀缺,在还能被听见的时候把话说出来。
0:00you know in terms of things that I think are most likely to affect the the future of humanity I think AI is probably the single biggest item in the near term that's likely to affect humanity so it's very important that we have the advent of AI in a good way that that's is something that if you if you could look into Cristobal and and see the future you would like you would like that outcome because it is something that could go could go wrong and as we've talked about many times and so we really need to make sure it goes right that's
0:39that's I think AI working on a I'm and making sure it's a great future that's that's the most important thing I think right now the most pressing item speaking of really important problems AI so you've been outspoken about AI could you talk about what you think the positive future for a it looks like and how we get there okay I mean I do want to emphasize that this is not really
1:07something that I advocate or or this is not prescriptive this is simply pretty hopefully predictive as we look some say oh well like this is something that I want to occur instead of so this something I think that probably is the best of the available alternatives the best of the available alternatives that I can come up with and maybe somebody else can come up with a better approach or a better outcome is that we achieve
1:40democratization of AI technology meaning that no one company or small set of individuals has control over advanced AI technology I think that that's very dangerous it could also get stolen by somebody bad you know like some evil dictator or the country could send their intelligence agency to go steal it and gain control it just becomes a very unstable situation I think if you've got any any incredibly powerful AI
2:12you just don't know who's who's gonna control that so it's not as I think that the risk is that the AI would develop a will of its own right off the bat I think it's more the consumers that some someone may use it in a way that is bad or and even if they weren't gonna use in a way that's bad but somebody could take it from them and use it in a way that's bad that that I think is quite a big danger so I think we must have democratization of AI technology and make it widely available and that's you know the reason that obviously you me and the rest the team you know created open AI was to help with the democracy
2:51help help spread out AI technology so it doesn't get concentrated in the hands of a few and but then of course that needs to be combined with solving the high bandwidth interface to the cortex humans are so slow humans are so slow yes exactly but you know we already have a situation in our brain where we've got the cortex and limbic system and the limbic system is kind of over that's that's the primitive brain it's kind of like the your your instincts and whatnot and then
3:30the cortex is a thinking upper part of the brain those two seem to work together quite well occasionally your cortex and limbic system may disagree but they attending it works pretty well generally works pretty well and it's like rare to find someone who I've not found someone who wishes to either get rid of the cortex or get rid of the Olympic system very true yeah it's that's unusual so so I think if we can effectively merge with AI by improving
4:01that the neural link between your cortex and the your digital extension yourself which already likes that already exists just has a bandwidth issue and then then effectively you become an AI human symbiote and and if that then is widespread with anyone who wants it have it then we solved the control problem as well we don't have to worry about some sort of evil dictator AI because kind of we are the AI collectively that seems like the best outcome I can think of so you've seen other companies in the early days that start small and get really successful hope I don't work at asking this on camera but how do you think open AI is going as a six month old company that's
4:51used to go pretty well I think we've got a really talented group at opening I and yeah really really talented team and they're working hard open area structured as see a 501c3 nonprofit but you know many nonprofits do not have a sense of urgency it's fine they don't have to have a sense of urgency but opening ideas is I think people really believe in the mission I think it's important and it's it's about minimizing
5:22the risk of existential harm in the future and so I think it's going well I'm pretty impressed with what people are doing in the talent level but I think if it reaches the threshold where it's this s modest the smartest most inventive human then I mean it really could be I mean a matter of days before it's smarter than some of humanity we're headed towards either super intelligence or civilization ending and another point that I think is really important to appreciate is that we are all of us already are cyborgs so you have a
6:03machine extension of yourself in the form of your your phone and your computer and all your applications you are already superhuman but by far you have more powerful capability then press in the United States had you know 30 years ago if you have an internet link you have an article of wisdom you can communicate to millions of people can communicate to the rest of Earth instantly and these are magical powers
6:32that didn't exist not that long ago so everyone is already superhuman it is the biggest risk we face as a civilization is artificial intelligence it's sold to a group of leaders what would you advise that we should how should we be addressing something that's it's so such a large landscape and yet obviously so important I think that the you know one of the rules of government is to ensure the public good and to that dangerous to the
7:03public are addressed so that hence the regulatory thing I think the first order of business would be to try to learn as much as possible you know to understand the nature of the issues to look closely at the progress that is being made and the remarkable achievements of artificial intelligence last year I go which is a quite a
7:30difficult game to beat that people thought would never be beaten with my computer that the curricula computer would we either never beat the best human player or that it was twenty years away and last year alphago which was done by deep mind which is a kind of a Google subsidiary absolutely crushed the world's best player and now now that now I can crush
7:58it can play at the top fifty simultaneously impossible so just like that pace of progress is remarkable and and there's you can see more and more coming out and robotics you can see robots that can learn to walk from from nothing you know within hours like way faster than any biological being but the
8:34thing that's most dangerous is and it's the hottest that kind of rap kind of get your arms around because it's not a physical thing is kind of a deep intelligence in the network and said well what home could have deep intelligence in the network do so well I could start a war fight create by doing fake news and spoofing email accounts and fake press releases and just by you
9:08know manipulating information the pen is mightier than the sword so I mean as an example I want to be emphasized I do not think this actually occurred this is purely a hypothetical that I'd digging my grave here but you know that like that though there was that second Malaysian airliner that was shot down on the Ukrainian Russian water and that
9:41that really amplified tensions between Russia and the EU in a massive way well like let's say if you had a an AI that was where the ai's goal was to maximize the value of portfolio of stocks one of the ways to maximize value would be to go long on defense short on consumer start a war and then how could it do
10:11that well you know hacking into the Malaysian Airlines rat aircraft routing server a routed over a war zone then sent an anonymous tip that an enemy aircraft is flying overhead right now let's go to Governor Ducey and then we'll have after governor Ducey will finish our gubernatorial questions and then two questions and we quick questions or one audience question and we'll be dad we're running short on time governor Ducey thanks Elon I really enjoyed your comments today and as someone who has spent a lot of time in his administration trying to reduce and eliminate regulations I was surprised by your suggestion to bring regulations before we know exactly
11:00what we're dealing with with AI you know I've heard the example used if I were to come up with a colorless odorless tasteless gas that was explosive people would say well you have to ban that and then we'd have no natural gas so you've given some of these examples of how a a I can be an existential threat but I still don't understand as policymakers what type of regulations beyond slowdown which typically policymakers don't get
11:35in front of entrepreneurs or innovators well I think the first order of business would be to gain insight right now the governor does not even have an insight and I and the right order of business would be to stand up regulatory agency initial goal gain insight into the status of AI activity make sure it the
12:02situation is understood once it is then put regulations in place to ensure Public Safety that's it and for sure the company is doing AI will most of them not mine will squawk and say hey this is really gonna stifle innovation blah blah it's gonna move to China it won't and it won't because like it's like it has like it has Boeing moved to China nope they were pulling aircraft here same with on
12:34cars and so it's there's the notion that if you say establish regulatory regime that companies will simply move to countries with with lower regulatory comments is false on the face of it because none of them do and unless it's really overbearing but that's not what I'm talking about here I'm just talking about you know make sure that there is awareness at the government level I think once there is awareness people
13:06will be extremely afraid as they should be well first why I think on the artificial intelligence front you know I I have exposure to the very most cutting-edge AI and I think people should be really concerned about it I keep so much sounding the alarm bell but you know until people see like robots going down the street killing people like they don't know how to react you know because it seems so ethereal and I think we should be really
13:43concerned about AI and I think we should yeah this is AI is a rare case where I think we need to be proactive in regulation instead of reactive because I think by the time we are reactive in AI regulation it's too late and normally the way regulations are set up is that a whole bunch of bad things happen there's public outcry the the and then after many years a regulatory agency is set up to regulate that industry there's a
14:16bunch of opposition from companies who don't like being told what to do by regulators and it takes forever that they that in the past has been bad but not something which represented a you know a fundamental risk to the existence of civilization AI is a fundamental risk to the existence of human civilization in a way that car accidents airplane
14:49crashes faulty drugs or bad food we're not they were not they were harmful tin to a set of individuals within society of course but they were not harmful to society as a whole AI is a fundamental existential risk for human civilization and I don't think you hopefully appreciate that you know it's
15:17not it's not fun being regulated it's not you know you're pretty or 'some but I you know in the car business you know we get regulated by Department of Transport by EPA and a bunch of others and this regs for agencies in every every country you know in the in space we get regulated by FA and but but you
15:46know if you ask the average person hey you want to do you want to get rid of the FAA and just like take it take a chance on manufacturers not cutting corners on the aircraft because you know profits were down that quarter I was like hell no that's sounds durable so you know I think even people who are pretty extremely like the libertarian free-market they were like yeah we should pull have some me keeping an eye on the aircraft companies making sure they build a good aircraft and good cars and that kind of thing so yeah I think there's there's a role for regulators that's very important and I'm against over-regulation for sure but man we I think we better get on that with AI
16:35fronto and so they'll certainly view a lot of job disruption because what's gonna happen is robust will be able to do everything better than us I mean I'm quitting I mean all of us you know yeah what's real exactly what to do about this it's like the it's the like it was really like the scariest problem to me I'll tell you and yeah so I really think
17:14we need to go a regulation here just because this is you know ensuring the public good is served because you're good companies that are racing that they kind of have to race to build AI or they're gonna be made uncompetitive you know like that essentially if your competitor is Racing's evolved AI and you don't they will crush you so then you're like ah we don't request so you
17:41know I guess we need to bullet to that's where you need the regulator's come in and say hey guys you all need to really you know just pause and make sure this is safe and like when when it's cool and working a bit and regulators are convinced that it's safe to proceed then you can go but otherwise slow down and
18:01bit slung them but you need the regulator's to do that for all the teams in the game you know otherwise the shareholders will be saying like hey why you developing AI faster because your competitor is like okay we've got to do that anyway so it's like hey there's like something like 12% of jobs or transport transport will be one of the first things to go fully autonomous but when I say everything like the robots will be able to do everything barn or nothing well I think it is difficult to appreciate just how far artificial intelligence has advanced and how far it is advancing because we have a double
18:50exponential at work we have an exponential increase in hardware capability and we have an exponential increase in software talent that is going into AI so whenever you have a double exponential it's very difficult to predict prediction is almost always going to be too conservative in terms thinking it'll be further out than it is
19:15you know you start to see things like if you seem like the video is where you can sort of really quite accurately video simulate someone and put words in their mouth that they never spoke he was Google is it's really pretty amazing and then they they had this they called a generative adversarial Network it had had two of them compete with one another to make the most convincing video no one
19:46would generate the video and then the other one would identify where it it looked fake and and then that that would the other one would fix that and then they'd go back and forth to the point where can tell which one was the real real video which was the paperboy and you know sleep have been some very public things like the defeat of alphago or to be of go by alphago the world's best go champion people thought defeating go was either never or 20 years away that was world's best go player was defeated and now that same
20:25alphago system can defeat the top 50 players simultaneously with zero percent of chance of them winning and that's one year later so the degrees of freedom to which artificial intelligence is able to apply itself or really increasing I think by ten orders of magnitude a year that's really crazy so I think and we're
20:55talking yeah and this is on hardware that is really not as well suited for neural nets you know like a GPU is maybe and we're back to be better than CPU but something but a chip that is designed optimally for neural nets is an order of magnitude better than a GPU and that is there are a whole bunch of neural net optimized chips coming out either late this year or else your so I
21:28think we should I think the other father all the government is to make sure the public is safe like to take care of public safety issues and I think so I think the right move is to Salish some government regulatory agency which at first is just there to gain insight so it's not about like shooting from the hip and just putting in rules before anyone knows I think but you gotta stuff agency it's got a gain insight once that inside is gained then start applying rules and regulations we have that full of the you know for aircraft they got that for cars good that for you know drugs for food and I don't think anyone wants the FAA to go away or FTA to go
22:19away or you know any of those regulatory agencies then I think we just need to make sure people do not cut corners on AI safety it's maybe it's gonna be a real big deal and it's gonna come on like a like a tidal wave [Music] [Applause] [Music]
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