2023年11月初,AI 安全峰会刚落幕,马斯克和苏纳克坐下来聊了一个小时。这场对谈后来被反复引用,最出圈的是马斯克那句话:将来会有一个时刻,没有任何工作是非做不可的。
你可以有工作。如果你想要一份工作,为了个人满足感。但 AI 会把所有事情都干完。
台下有人笑了。苏纳克也笑了。这笑多少带点礼貌成分,因为马斯克紧接着补了一句:我不知道这话让人舒服还是不舒服。
十年来他一直在重复同一句话
苏纳克开场引用了比尔·盖茨的评价:我们这个时代,没有人在推动科学和创新的边界上做得比马斯克更多。
马斯克的回应很轻:这话好听,从比尔·盖茨嘴里说出来就更不错了。
苏纳克把话题拐了个弯。他说,有意思的是,在 AI 这件事上,马斯克十年来几乎在做相反的事。他一直在说,等一下,我们得想想自己在干什么,我们推的这东西安不安全,也许不该推得这么快、这么狠。
我当了一段时间的卡珊德拉(Cassandra,希腊神话里预言灾难却没人信的人)。我跟人说,我们真该担心 AI。他们的反应是:你在说什么?
这个比喻用得挺准。马斯克说,他能提前看到这一天,是因为在技术里泡得太久。GPT-1、GPT-2、GPT-3、GPT-4,整个演进过程他都在旁边看着。
把曲线上的点连起来,假设趋势继续,结论不复杂:我们会得到远远超过人类智能的人工智能。
转折点为什么是2023年
苏纳克问他,是什么让你觉得今年不一样。
马斯克给了两个理由。第一个是深度伪造(deepfake)到了能以假乱真的程度。有人能实时生成一段视频,让画面里的你说任何话,让你做任何事。他说,这些假视频有时候比真的还像真的。第二个是 ChatGPT,它让普通人第一次亲手摸到了大模型的边界。
马斯克一直盯着技术曲线,这两件事他都不意外。意外的是公众的反应。他终于不用再解释了。
他当场谢了苏纳克办这场 AI 安全峰会,说这件事会被历史记住,分量很重。然后他把话说得很平衡:AI 大概率是好事,能带来一个物质极大丰富的未来,商品和服务不再稀缺。但他用了魔法精灵的比喻。
如果你有一个能满足所有愿望的魔法精灵,这种故事通常结局不太好。许愿要小心,包括许愿本身。
这个比喻在后面还会出现。马斯克显然很喜欢它,因为它同时装下了乐观和恐惧,还不用选边。
政府该当裁判,不该当运动员
峰会当天,参会公司和各国领导人达成了一个共识:模型发布之前,最好由政府来做安全测试。
苏纳克解释了自己的立场。他说,政府的工作是提前说一句:这里有个潜在风险。这个风险不确定会不会发生,但值得防。他的职责是保护国家,而保护国家的前提是先有能力。所以英国建了 AI 安全研究院,先培养人,再拿模型来测。
马斯克认同这个方向,但他给了一个更朴素的理由。
绝大多数软件出问题,不会伤到公众。手机上的 App 崩了,笔记本蓝屏了,算不上灾难。但数字超级智能不一样,它确实会对公众构成风险,这时候政府该出手。
他说自己跟监管机构打了半辈子交道。SpaceX 是航天,特斯拉是交通,星链是通信,全在监管框架里。
我同意绝大多数监管。偶尔会不同意一些,但比例不到百分之一。
硅谷有一批人从没跟监管打过交道,他们觉得监管会把创新掐死,会拖慢进度,会很烦。马斯克承认,确实会很烦,这一点他们没说错。但他打了个比方:
任何一场体育比赛都有裁判,没有人提议办一场没有裁判的比赛。
政府的位置就是裁判。确保比赛按规矩来,确保公众安全被照顾到。他说自己作为技术从业者,可能对技术过于乐观,所以更清楚乐观要有边界。AI 大概率会是个好东西,但变坏的概率不是零。
一年涨五倍,明年至少一个数量级
苏纳克抛出了一个很实际的问题:政府能追得上吗?
DeepMind 的德米斯·哈萨比斯、OpenAI 的山姆·奥特曼,手下都有大量顶尖人才。政府要招人、要建能力,速度够快吗?
AI 的速度比我见过的任何技术都快,快得多。它的能力每年至少涨五倍,也许十倍。明年肯定还会涨一个数量级。
一个数量级就是十倍。政府的运作节奏跟这个速度不在一个量级上,这是明摆着的事。
马斯克给了一个退一步的方案。哪怕没有硬性监管,哪怕没有执法能力,仅仅是有洞察力、能把担忧公开讲出来,就已经很有力量。如果只能做到这一步,也已经很好了。
苏纳克接了一句:希望我们能做得比这更好。
中国不来,这场会就白开了
苏纳克提到他邀请中国参会的决定。这个决定不好做,很多人批评他。他的判断是,如果真想谈点严肃的事,就不能把人关在门外。
马斯克接过话头,说了一句在西方政治语境里不太常见的话。
我反对 AI 监管时,听到最多的理由是:中国不会做的,所以他们会超过去,把我们甩在后面。但事实上,中国愿意参与 AI 安全。谢谢你邀请他们,我们也应该谢谢中国出席。
他补充说,今年早些时候去中国,跟中国领导人谈的主要话题就是 AI 安全。他告诉对方,这件事值得认真对待,对方也认真对待了。
他给了两个判断。第一,如果没有中国参与,讨论毫无意义。第二,全球 AI 开发的两个领先中心,一个是旧金山湾区,一个是伦敦地区。如果美国、英国和中国在安全上能对齐,事情就好办,因为领导力基本都在这几个地方。
苏纳克补了一个细节。中方在昨天的讨论里很投入,最后签了跟大家一样的公报。
马斯克说,这就对了。如果缺了谁,事情就解决不了。
开源落后半年,代价是五倍性能
峰会当天有一场关于开源的激烈辩论。杰弗里·辛顿(Geoffrey Hinton)担心开源模型被坏人利用。另一派认为开源是分布式创新的命脉。
苏纳克把这个两难抛给马斯克。
马斯克先给了个时间差:开源的算法和数据,通常落后闭源六到十二个月。
听起来不算多。但把增速叠上去,差距就变了。如果能力每年涨五倍,落后一年就是性能差五倍。五倍的差距不小,它意味着开源模型在能力上属于另一个世代。
他说这个局面也许还行。但迟早会到某一天,开源模型也逼近人类水平的智能,甚至越过它。到那时候就不知道该怎么办了。
他承认开源有某种不可避免性,然后给出了自己略微偏向开源的理由:至少你能看见里面在发生什么,闭源你什么都看不见。
但他马上自己拆了这个理由。
一个几亿行的 CSV 文件
马斯克说,即使是开源模型,你也未必真的知道里面在发生什么。
打开文件,你看到的是几十亿个权重和参数,满屏数字。你没法读它,没法从里面读出它接下来会干什么。
他用了很具体的类比:神经网络最后就是一个巨大的逗号分隔值文件(CSV)。我们的数字上帝是一个 CSV 文件。
这跟传统软件完全不同。传统编程是离散逻辑,输入进去,输出可预测,每一行代码你都能读懂它在做什么。神经网络是一堆概率。你只能跑它,测它,看它在不同输入下给出什么反应。
它没有确定性,只有概率性。
这个区别正是苏纳克想强调的。传统软件里,改进输入会带来可预期的输出改进。模型迭代之后,没有人确切知道另一端会冒出什么。德米斯会同意这一点,所以才有那个共识:在训练跑完、模型发布之前介入,先搞清楚这一代到底带来了什么新能力。
马斯克说,他同意。
全民高收入,魔法精灵没有三次许愿限制
苏纳克把话题从技术拉到选民。他说自己做访谈、走街串巷,AI 被问得最多的是工作。我的工作还在吗?我孩子将来还有工作吗?
他给了一个政策制定者的答案。AI 已经在创造岗位,用法更接近副驾驶,替人的成分还不多,人还在,只是干得更好。他引了一份 MIT 几年前的研究:当下大约 60% 的工作岗位,在四十年前并不存在。预测很难。所以他的任务是搞好教育,让人在任何职业阶段都能重新培训。
然后他把问题交给马斯克。
我们正在见证历史上最具颠覆性的力量。第一次,会出现比最聪明的人还聪明的东西。具体哪一刻不好说,但会有那么一个点,没有工作是必须有人来做的。
他重复了魔法精灵。这次的重点是没有三次许愿的限制,你想要多少愿望就有多少愿望。
接下来是那个词:全民高收入(Universal High Income)。他特意把它跟全民基本收入(Universal Basic Income)分开。基本收入只保底,高收入是每个人都能调用这个魔法精灵。
他的意思是,每个人都能问任何问题,能调动任何能力。这在某种意义上是个均衡器,把人和人之间的差距压平。物质层面不会缺,商品和服务的稀缺性会消失,那是一个富足的时代。
他推荐了伊恩·班克斯(Iain Banks)的《文化》系列,说那是对 AI 未来最好的想象,没有之一,所有书都好,不用挑。
他自己也得假装相信自己在做的事
苏纳克说了一句挺真诚的话。他相信工作给人意义,给人目的。如果拿走一大块,那股驱动力从哪里来?
马斯克的回答有点自嘲。
我得刻意让自己悬置怀疑。我往一个项目里投入大量心血,熬到凌晨三点。然后我想,等等,我为什么要干这个?我大可以等 AI 来做。我这是在无缘无故鞭打自己。
他说自己大概是有点受虐倾向。苏纳克接话:那给德米斯打个电话,让他快点,你就能去度假了。
笑完之后,马斯克给了正经的答案。机器接管那些不舒服、危险、枯燥的工作,其实是好事。做饭有意思,洗碗没那么有意思,但计算机洗碗洗得很开心。
他举了体育的例子。机器跑得比人快,但我们还是办奥运会,还是让人类互相比,看谁是人类里最好的。即使机器更强,人还是能从竞争里找到满足感。这可能是未来的一种模式。
每个孩子配一个导师
苏纳克说,AI 带来的机会里,他最兴奋的是教育。
他提到了可汗学院(Khan Academy)的萨尔·可汗今年那场 TED 演讲。一对一辅导和班级授课的效果差距巨大,如果每个孩子都有为自己定制的导师,还跟着自己一起成长,那会是件了不起的事。他说这在自己有生之年够得着。
马斯克同意,而且说得更直接。
它会是最好的导师,最有耐心的导师,全天候坐在那儿。教育会因此大受裨益。
但他也点了一个现实问题:模型还在幻觉。它会用很好的语法、漂亮的要点和一堆假引用,自信地给你一个错误答案。这个毛病得先治好。他估计会很快。
苏纳克补了一句:疫情期间在家带孩子的父母,听到这个应该会大松一口气。
AI 可以当朋友
马斯克提了一个听起来有点怪的方向:陪伴。
他的逻辑是这样的。如果一个 AI 有记忆,记得你们之间所有的互动,你还授权它读你所有的东西,那它可能比任何人都了解你,甚至比你自己更了解你。你每天跟它说话,对话层层叠加,你会真的拥有一个很好的朋友。
前提是这个朋友别被关掉。
别关掉我的朋友。
他讲了一件私事。他有一个儿子有学习障碍,交朋友有困难。他说,一个 AI 朋友对他会很有帮助。
苏纳克说这个答案他没想到,值得想一想。他补充说,英国已经在用数字化和电话方式提供心理治疗,效果显著。AI 提供这种社会性的支持,是看得见的路。
社区笔记:不删帖,只加背景
话题转到 X。苏纳克挑了一个具体的改动:把人工审核换成了社区笔记(Community Notes)。
马斯克的解释从权力结构开始。如果你把审查的权力交给一批人,这批人必然带偏见。而任命这批人的人,实际上控制了信息。
社区笔记的思路是共识驱动。它不删除任何东西,只添加背景。背景可以是“这件事不实,理由如下”。
更重要的有两点。
第一,全部开源。软件每一行代码都能看,每条笔记背后的数据都能看,你可以自己独立生成一条笔记。如果发现数据被操纵,你可以指出来,说这里有人在刷系统,然后提改进方案。
第二,也是核心机制:一条笔记要被展示,必须由历史上意见相左的人共同认可。
这里用了 AI。系统给每个贡献者在一个参数空间里定位,每个人对应几百个向量。左右二分太粗糙,事情比左右复杂得多。然后做反向关联:找出平时观点对立、但在这条笔记上一致的人。他们一致了,这条笔记才有可信度。
马斯克说效果不错。他没见过一条错误的笔记挂超过几个小时,命中率极高。有人跟他抱怨社区笔记会传播虚假信息,他就让对方发一条例子过来,对方发不出来。
X 的目标是告知公众、娱乐公众,尽可能准确,尽可能真实。即使有些人不喜欢真相。人并不总是喜欢真相。
他说,真相是有回报的。只要你不打算自欺,说真话就是聪明做法。
人形机器人能追你上楼
有人问机器人。要进入富足时代,需要大量机器人,会不会有国家在机器人上领先,取得战略优势。
马斯克先扩大了定义:任何能被计算机驱动的东西,本质上都是机器人。特斯拉就是轮子上的机器人,任何联网设备都是 AI 的末端执行器。
他点了几家:波士顿动力做了很多年拿得出手的机器人,现在归现代汽车;英国的戴森做得不错;英国还有 ARM,世界上最好的芯片设计公司之一,特斯拉用了大量 ARM 技术,几乎所有人都在用。德国做工业机器人。他的判断是,能造机器人的国家都不会掉队。
然后他转了方向,讲安全问题。
车不太容易追你。你可以跑上楼梯,躲进楼里,车追不进来。但他提了一部斯蒂芬·金的小说,车被附身那种。
人形机器人不一样。它能追你到任何地方。
我们应该有某种硬接线的本地急停开关,不能通过互联网更新。任何能通过网络做软件更新的东西,都能被覆盖。
他的方案是:说一个关键词,机器人进入安全状态。一个不需要你靠得太近的关停方式。他担心的是几百万台这种东西散落在世界各地,某天收到一次软件更新,它们就不那么友好了。
苏纳克说,当天的小组讨论里有人提了同一件事,只是用了另一种说法。他们聊了一堆电影,发现一个共同点:所有这些电影最后都是人找到办法把机器关掉。所以物理开关这件事,大家想到一块去了。
马斯克说,也许这可以当 AI 的一个测试题:你最喜欢的詹姆斯·卡梅隆电影是哪一部?
一美元一年,把机器人挡在门外
一位做 AI 和机器学习公司的创始人提问。ChatGPT 发布三个月后,钓鱼攻击明显变多,语言更自然更难识别。怎么保护企业和消费者,怎么让人信任这项技术?
马斯克的回答很直接:别那么信任它。
他说这个问题很棘手。现在连开源模型都能通过验证码。识别图里的红绿灯,AI 做得比人又快又好。当机器通过人类测试的能力超过人类,什么测试才有意义?他没有好答案。
X 在想的办法是收费。一年一美元,或者一天三美分,金额小到几乎不算钱,但足以让批量注册一百万账号变得极其昂贵。因为你需要一百万个支付方式,而盗来的信用卡很快就用完了。
他说这个问题明年可能就无解了。机器人不只是注册账号,还会点赞、转发,把某个东西刷得很热门,而实际上它根本不热门。
任何不这么做的社交媒体系统,都会被机器人淹没。
苏纳克从政府角度接了一段。他说核心是让人看见风险在被处理。过去两天峰会的意义就在这里:投资安全研究院,培养能做研究的人,搞清楚怎么缓解这些风险,而且要快,要不断迭代。技术可以很强大,但得把人带着一起走。
他提到昨天发布的一份研究,模拟了用 AI 批量生成假账号、渗透进特定社群、投放特定信息的过程。成本低到几乎为零。在社区或城镇这个尺度上,不需要多少假账号就能让人觉得“到处都是”。
马斯克补了一句技术细节:你甚至不用盗用别人的照片,那样还能被追溯。直接生成一张不存在的人脸,写一份不存在的履历,批量生产。
苏纳克说了个冷笑话:唯一的破绽可能是语法太好。
教皇的羽绒服和明年的选举
最后一位提问者来自 Adobe,问的是视觉媒体。能不能做加密签名的媒体,标明哪些是 AI 生成的,哪些是相机拍的真实影像,给病毒式传播的内容加一层信任。
马斯克说这个想法不错,应该做。他提到 X 前身推特曾经是相关项目的成员,欢迎回来。
苏纳克在这件事上有切身体会。他自己遇到过被篡改的图片全网扩散,等大家反应过来是假的,伤害已经造成了。
他把时间点摆出来:明年美国、印度、印尼、英国都有大选,还有欧洲议会选举。全球相当大比例的人口将在明年投票。这可能是第一波 AI 深度伪造真正成为议题的选举。在意民主完整性的人会觉得这是任务级的问题。
马斯克提了教皇穿羽绒服那张图。他说现在还有人以为那是真的。他的反驳很朴素:七月在罗马穿羽绒服,你想想概率有多大。然后他补了一句,说那件衣服其实挺好看,AI 时尚会成为真事。
小苗需要浇水,橡树不需要
有创业者问,英国有全世界最好的大学,也有最好的人才,要成为独角兽的孵化地还缺什么?给马斯克的问题是:在英国做创始人,最顶尖的技术人才仍然不会把它当成显而易见的选择,文化上要改什么?
马斯克的回答从园艺开始。文化上要庆祝创建新公司这件事,要有偏向支持小公司的机制,因为小公司才需要被照顾,大公司不需要。一棵小苗要浇水,一棵参天橡树不用。
他给了一个客观评价:伦敦地区和旧金山湾区,是地球上 AI 的两个中心。旧金山大概领先一点,但伦敦非常强,大伦敦加上周边各郡。这是事实判断。
然后他列了基础设施:愿意把房子租给新公司的房东,愿意服务新公司的律所和会计事务所。这些是心态问题,有些已经在变,但根子上还是文化,人们得自己认定这是好事。
苏纳克接着讲政府那头。他的工作是拧螺丝:种子轮有 EIS 税收减免,后期轮次要有养老金改革把资本放出来投成长股。这些都在推进,还没到位。
人才那头,教育要优先数理,他反复讲 more maths。还要吸引全世界最好的人。英国增长最快的公司里,超过一半有非英国籍创始人。所以签证制度要开放。
监管那头,要有亲创新的框架。护栏要有,但得留出空间让人试。
他说最难的是文化那部分:怎么把硅谷那种敢扔掉稳定薪水去创业、并且能接受失败的心态移植过来。第一次创业失败,不该是职业生涯的灾难性终结。别人该说的是,你认真试过了,再来一次。
失败不该是职业生涯的终点
马斯克接过这个话题,讲了一件很具体的事:股票期权。
创业是高危高回报。绝大多数创业公司会失败,你听到的都是活下来的那些。真实情况是大量工作之后跟着失败。既然风险这么高,回报那一端就必须真实存在,否则算不过来。
在欧洲大部分地方,股票期权的税务处理很麻烦。他不确定英国的具体情况,但猜比欧洲大陆好一些。
苏纳克在这里插了一段很实在的话。他说这几乎是在软性推销税收政策,但他同意。英国的资本利得税率是 20%,比欧洲大陆低,也比加州低。承担风险、把东西做起来的人,应该在终点拿到回报。
股票期权这块,他提到 Index Ventures 每年出的那份各国期权税制报告。他当财政大臣时,英国排第四第五,他说这不行,得往前挪。后来改完,按他的记忆排到了第二。
他的落点是一句价值判断:人努力了、冒险了,就该能享受高风险带来的高回报。
gov.uk 上的一次对话
中间有人问了个接地气的问题:这些宏大叙事,怎么落到奥斯汀或者诺丁汉的普通人身上?
苏纳克先答。他说自己前两天在想,政府现在做的 AI,有哪些真的在改变普通人的生活。想到的是 gov.uk。这个网站在多年前是开创性的,把所有政府信息集中到一个入口,办驾照、办护照、跟政府打任何交道,都在这里。每天有几百万人用。
接下来要在这个平台上部署 AI。他举了个场景:护照丢了,航班五小时后起飞。现在你得走一堆步骤才能搞清楚怎么办。有了 AI,你把这些话说出来,它一步步带你走完。每天几百万人受益,这是他能用这项技术做到的最实际的事。
马斯克从个人角度答。最直接的变化是,你有了一个很聪明的朋友,什么都能问。怎么做一件事,怎么解决一个问题,它都会告诉你。企业会跟进,所以客服会明显变好,这大概是你最先注意到的变化。
然后回到教育。任何一个科目,都有一个极好的导师陪着。这件事差不多已经到了,只差把幻觉治好。
80% 好,20% 坏
整场对谈的收尾,马斯克给了一个数字。
我们活在最有趣的时代。这件事有 80% 的概率是好的,20% 是坏的。如果我们对坏的那部分保持清醒和谨慎,总体上它会变成我们想要的那个未来,或者说更可取的那个未来。
他说这会是个均衡器。每个人都能拿到商品、服务和教育。结果可能是更多人的幸福。
苏纳克收尾说,政府的工作是让它安全地发生。听完这两天,他更有信心了。
把整场对话拆开看,马斯克的立场其实很一致。他对技术本身乐观,对技术的扩散速度保持警惕。他相信监管该有,但监管得站在裁判的位置,别下场当教练。他不认为开源能自动带来透明,因为一个几亿参数的模型本来就没法读。他相信中国必须在场,否则谈判没有意义。他相信钱这道门槛能挡住机器人,哪怕只是一年一美元。
这些判断没有一条是漂亮的,但每一条都指向同一个问题:当能力以每年五到十倍的速度增长,人类这边的制度、文化和习惯,追得上吗?
他在访谈里没有给答案。他给的是几个具体的把手:物理急停开关、发布前的安全测试、一美元一年的门槛、开源的透明度、社区笔记的交叉共识。都是小东西,但都能抓。
读到这里的从业者,有几件事可以立刻动手:
- 如果你在做产品,把“发布前测试”写进流程。让独立的人在模型上线前跑一遍能力评估,记录它多出了什么能力,别把它做成走过场。
- 如果你在做内容或社区,别只依赖模型判断真假。给内容加来源签名,给用户加一层极低成本的验证门槛,一年一美元这种量级就够挡住批量注册。
- 如果你在做机器人或任何带执行器的硬件,设计一个不依赖网络的本地急停。软件能远程更新的东西,就能被远程覆盖。
- 如果你在教育行业,现在就把 AI 导师做成产品原型。幻觉是当前最大的工程问题,商业模式反而好解决。
- 如果你在创业,把失败的定义改一下。第一次没做成,拿到的是信息,判决还轮不到。这个心态改不过来,其他条件都白搭。
- 如果你在做政策或合规,先建能力再建规则。没有测试能力的监管,只是一份声明。
0:02okay all right well good evening everybody Welcome Elon thanks for being here thank you for having me we feel we feel very privileged we're excited to have you right so I'm going to start with some questions and then we're going to open it up let me get straight into it so Bill Gat said
0:16there is no one in our time who has done more to push the bounds of science Innovation than you well it's kind of to say well that's it that's a nice thing to have anyone say about you nice coming from Bill Gates but oddly enough when it comes to AI actually for around a decade you've
0:31almost been doing the opposite and saying hang on we need to think about what we're doing and what we're pushing here and what do we do to make this safe and and actually maybe we shouldn't be pushing as fast or as hard as we are like I mean you've been doing it for a decade like what was
0:44it that caused you to think about it that way and you know why do we need to be worried yeah I've been somewhat of a Cassandra for quite a while where where people would I tell you like we should really be concerned about AI they be like what are you talking about like they've never really had any experience with with AI but since I was immersed in technology I have been
1:03immersed in technology for a long time I could see it coming so but I think this year was there've been a number of of breakthroughs I mean you know the point in which someone can see a dynamically created video of themselves you know like somebody could make a video of you saying anything in real time or me and so the sort of the deep Pake videos which
1:28are really incredibly good in fact sometimes more convincing than real ones and deep real and and then and then obviously things like CH GPT were were quite remarkable now I saw gpt1 gpt2 gpt3 gp4 that you know the whole sort of lead up to that so it was easy for me to kind of see where it's going if you just sort of extrapolate the points on a curve and
1:55assume that Trend will continue then we will have profound artificial intelligence and obviously at a level that far exceeds human intelligence so but I'm I'm glad to see at this point that people are taking safety seriously and i' I'd like to say thank you for holding this AI safety conference I think actually it will regard on in history as being very important I think it's it's really quite profound and and and I do think overall that the potential is
2:27there for a artificial intelligence AI to have most likely a positive effect and to create a future of abundance where there is no scarcity of goods and services but but it is somewhat the of the the magic Genie problem where if you have a magic Genie that can grant all the wishes
2:49usually those stories don't end well be careful what you wish for including wishes yeah yeah so you you you talked a little bit about about the the summit and thank you for being engaged in it which has been great and people enjoyed having you there participating in this dialogue now
3:08one of the things that we achieved today in the meetings between the companies and the leaders was an agreement that externally ideally governments should be doing safety testing of models before they're released I think this is something that you've spoken about a little bit it was something we worked really hard on because you know my job in government is to say hang on there is
3:30a potential risk here not a not a definite risk but a potential risk of something that could be bad you know my job is to protect the country yes that's and we can only do that if we develop the capability we need in our safety Institute and then go in and make sure we can test the models before they are release delighted that that happened today but you know what what's your view
3:49on what we should be doing right you've talked about the potential risk right again we don't know but you know what are the types of things governments like our should be doing to manage and mitigate against those risks well I generally think that that it is good for government to play
4:03a role when the public safety is is at risk so you know really for the vast majority of software the public safety is not at risk I mean if if the if the app crashes on your phone or your laptop it's not a a massive catastrophe but when you're talking about digital superintelligence I think which does pose a risk to the public then there is a role for government
4:26to play to safeguard the interest of the public and and this is of course true in in many fields you know Aviation cars you know I and I deal with Regulators throughout the world because of stalling being Communications Rockets being Aerospace and cars you know being TR vehicle transport so I'm very familiar with dealing with with Regulators and I actually agree with the
4:49vast majority of regulations there's a few that I disagree with from time to time but Point one% probably of well less than 1% of regulations I disagree with so and there is some concern from people in silic Valley who who' have never dealt with Regulators before and they think that
5:05this is going to just Crush Innovation and and slow them down and be annoying but and and it will be annoying it's true they're not wrong about that but but but I think there's we've learned over the years that having a referee is a good thing and if you look at any sports game there's always a a referee and and nobody's suggesting I think to have a sports game without
5:32one and and I think that's the the right way to think about this is for for government to be a a referee to make sure the Sportsman like conduct and and and that the public safety is you know is addressed that we care about the public safety because I think there might be at times too much optimism about technology and I speak I say that as a technologist I mean so I ought to know and and and and like I said on on balance I think that the AI will be a forceful
6:06good most likely but the probability of it going bad is not 0% yeah so we we just need to mitigate the downside potential and then how you talked about referee and that's what we're trying right there yeah well there we go I mean you know and we talked about this and Demis and I discussed this
6:21a long time ago like literally facing right at and actually you know Demis to his credit and the credit of people in industry did say that to us I you know de say it's not right that Demus and his colleagues are marking their own homework right there needs to be someone independent and that's why we've developed the safety Institute here I mean do you think governments can develop the
6:42expertise one of the things we need to do is say hang on you know Demis Sam all the others have got a lot of very smart people doing this governments need to quickly tool up capability wise Personnel wise which is what we're doing I mean do you think it is possible for governments to do that fast enough given how quickly the technology is developing or what do we need to do to make sure we do do it quick enough no I think it's it's a good it's a great Point you're making the the
7:06pace of of AI is faster than any technology I've seen in history by far and it's it seems to be growing in capability by at at least fivefold perhaps 10 fold per year it it'll certainly grow by an order of magnitude next year yeah so so and and government isn't used to moving at
7:28that speed but I but I think even if there are not firm regulations even if there's not even if there isn't an enforcement capability Simply Having insight and being able to highlight concerns to the public will be very powerful so even if that's all that's accomplished I think that will be very very good okay yeah well hopefully we can do better than that hopefully
7:51yeah no but that that's helpful actually we were talking before it was striking you you someone who spent their life in technology living More's law and what was interesting over the last couple of days talking to everyone who's doing the development of this and I think you concur
8:06with this is is just the pace of advancement here is unlike anything all of you have seen in your careers and technology is that fair because you've got these kind of compounding effects from the hardware and and the data and the Personnel yeah I mean the two currently the two leading centers for AI development are C Isco Bay Area and the and the sort of London area that
8:31and there are many other places where it's being done but those are the two leading areas so I think if you know if if the United States and the UK and and China are sort of aligned on on safety that's all going to be a good thing because that's really that's where that's that's
8:48where the the leadership is generally I me you actually you mentioned China that so I I took a decision to invite China to to Summit over the last couple days and it was not an easy decision a lot of people criticize me for it you know my view is if you're going to tryal serious
9:03conversation you need to but I what would your thoughts you do business all around the world you just talked about it there yeah you know should we be engaging with them can we trust them is that the right thing to have done if if we don't if China is not on board with AI safety it it's
9:17somewhat of a mood situation U the single biggest objection that I get to any kind of AI regulation or or sort of safety controls are well China's not going to do it and therefore they will just jump into the lead and exceed us whole but but actually China is willing to participate in in
9:36AI safety and thank you for inviting them and I and they you know I I think we should thank China for for attending when I was when I was in China earlier this year the my main subject of discussion with this the leadership in China was AI safety and saying that this this is really something that they they should care about and they took it seriously and and and and
10:01you are too which is which is great and having them here I think was essential really if they if they're if they're not participants it's it's pointless it's pointless yeah no that and I think we were pleased I they were engaged yesterday in the discussions and actually ended up signing
10:15the same communic that everyone else did that's great which is a good stop right and I said if we need everyone to approach us in a similar way if we're going to have I think a realistic chance of of resolving it I was going to you talked about Innovation earlier and and regul being annoying
10:30there was a good debate today we had about open source and I think you you've kind of been a proponent of algorithmic transparency and making some of the the X algorithms public and you actually we were talking about Jeffrey Hinton on the way in yeah you know he he's particularly he's
10:45been very concerned about open- Source models being used by Bad actors you've got a group of people who say they are critical to Innovation happening in that distributed way look it's it's a trick there's probably no perfect answer and there's a tricky balance what are your thoughts on
11:00how we should approach this open source question or you know where should we be targeting whatever regulatory or monitoring that we're going to do well the open source algorithms and data tend to lag the close Source by 6 to 12 months but so so but given the rate of improvement that there's actually therefore quite a big difference between the the closed source and the and the open if things are improving by factor of let's say five or more than being a
11:32year behind is you're five times worse so it's a pretty big difference and that might be actually an okay situation but it it certainly will'll get to the point where you've got open source AI that can do that that will start to approach human level intelligence will perhap succeed it
11:51I don't know quite what to do about it I think it's somewhat inev inevitable there will be some amount of Open Source and I I I guess I would have a slight bias towards open source because at least you can see what's going on whereas closed Source you don't know what's going on now it should be said with AI that even if it's open source do you actually know what's going
12:10on because if you've got a gigantic data file and you know sort of billions of of of data points weights and parameters you can't just read it and see what it's going to do it's a gigantic file of inscrutable numbers you can test it when you when you run it you can test it you can run a
12:32bunch of tests to see what it's going to do but it it's probabilistic as opposed to deterministic it's not it's not like traditional programming where you've got a you've got very discret logic and and and the outcome is very predictable and you can read each line and see what each Line's going to do a neural net is just a whole bunch of probabilities I mean it sort
12:57of ends up being a giant comma separated value file it's like our digital guide is a CSP file really okay but that that is kind of what it is yeah now that that point you've just made is one that we have been talking about a lot because again conversation with the people who developing
13:17their technology make the point that you've just made it it is not like normal software where there's predictability about inputs improving leading to this particular output improving and as the models iterate and prove we don't quite know what's going to come out the other end I
13:31think Demis would agree with that which is why I think there is this bias for look we need to get in there while the training runs are being done before the models are released to understand what is this new iteration brought about in terms of capability which it it sounds like you would U
13:46would agree with I I was going to shift gears a little bit on you know you've talked a lot about human consciousness human agency which actually might strike people as as strange given that you are known for being such a brilliant innovator in technology ologist but it's it's quite heartfelt
14:01when I hear you talk about it and the importance of maintaining that agency in technology and preserving human consciousness now it kind of links to the thing I was going to ask is when I do interviews or talk to people out and about in this job about AI the thing that comes up most
14:14actually is is probably not so much of the stuff we've been talking about but jobs it's what does AI mean for my job is it going to mean that I don't have a job or my kids are not going to have a job now you know my my answer as a you as a policy maker as a leader is you know actually
14:31AI is already creating jobs and you can see that in the companies that are starting also the way it's being used is a little bit more as a co-pilot necessarily versus replacing the person there's still human agency but it's helping you do your job better which is a good thing and and as we've
14:47seen with technological Revolutions in the past clearly there's change in the labor market the amount of jobs I was quoting an MIT study today that they did a couple of years ago something like 60% of the jobs at that moment didn't exist 40 years ago so hard to predict and my job is to
15:02create an incredible education system whether it's at school whether it's retraining people at any point in their career because ultimately if we've got a skilled population they'll be able to keep up with the the pace of change and have a good life but you know that it's still a concern and
15:17you know you what would your kind of observation be on on AI and the impact on labor markets and people's jobs and how they should feel about that as they they think about this well I think we are seeing the most disruptive force in history here you know where we have for the first
15:35time we will have the first time something that is smarter than the smartest human and that I mean it's hard to say exactly what that moment is but but there will come a point where no job is needed you can have a job if you want to have a job for sort of personal satisfaction but the
15:55AI will be able to do everything so I don't know if that makes people comfortable uncomfortable it's [Laughter] it's you know that's why that's why I say if if you if you wish for a magic Genie that gives you any wishes you want and there's no limit you don't have those three limit three wish limit nonsense you just have as many as many wishes as you want so it's both good and
16:21bad one of the challenges in the future will be how do we find meaning in life if if you have a gen that can do everything you want I I I do think we we it's it's it's hard you know when when when this new technology it tends to have usually follow an S curve in this case we're going to be on the exponential portion of the S curve for a long time and you like you'll be
16:45able to ask for anything it won't be we won't have Universal basic income we'll have Universal High income so in some in some sense it'll be somewhat of a leveler or an equalizer you know because really I think everyone will have access to this magic Genie and you able to ask any question
17:03it'll be certainly be good for Education you it'll be the best tutor you could I'm the most patient tutor sit there all day and there will be no shortage of goods and services will be an age of abundance I think if I'd recommend people read in Banks the banks culture books are probably the best envisioning if fact not probably they're definitely by far the best envisioning of an AI future there's nothing even close so I'd recommend really recommend
17:35fanks I'm very big fan all his books are good does not say which one all of them so so that's that that'll give you a sense of what is a I guess a fairly utopian or protop future with with AI yeah which is good from a as you said it's a universal High income which is a nice phrase and that's it's good from a kind of materialistic sense a of abundance actually
18:07that it kind of then leads to the question that you pose right I'm someone who believes you know work gives you meaning right I think a lot about that as as you know I think work is a good thing it you know gives people purpose in their lives and if you then remove a large chunk of that you
18:21know what does that mean and where do you get that you know where do you get that drive that motivation that purpose I mean you were talking about it you you work a lot of hours I do no as I was mentioning when we we were talking earlier I have to somewhat engage in deliberate suspension of disbelief because I'm I'm putting so much Blood Sweat and Tears into a work project and
18:39burning the you know 3:00 a.m. oil then I'm like wait why am I doing this I can just wait for the AI to do it I'm just lashing myself for no reason yeah must be a glut for punishment or something so we called call Demus and tell him to hurry up up and then you can have a holiday
18:59right that's the plan yeah no it's a it's a tricky it's a tricky thing because I think you know part of our job is to make sure that we can navigate to that very I think largely positive place that you're describing and help people through it between now and then because these things bring
19:16a lot of about a change in in the labor market as we've seen yeah I I think it probably is generally a good thing because you know there are a lot of jobs that are uncomfortable or dangerous or sort of tedious and the computer will have no problem doing that be happy to do that all day
19:32long so you know it's fun to cook food but it's not that fun to wash dishes like but the computer's perfectly happy to wash dishes I I guess there is you know we still have sports like where where where humans compete and like the Olympics and obviously a machine can can go faster than any human but we still have we still humans race against each other and and
19:59have all you know have these Sports competitions against each other where even though the machines are better they're still I guess competing to see who can be the best human at something yeah and and people do find fulfillment in that so I guess that's perhaps a a good example of how even
20:12when machines are faster than are stronger than us we still find a way we still we still enjoy competing against other humans to at least to who the best human yeah that's that's a good that's a good analogy and we've been talking a lot about managing the risks I just before we move on and
20:26finish on AI just talk a little bit about the opportunities you know you you're engag in lots of different companies neur being an obvious one which is doing which is doing some exciting stuff I you touched on the thing that I'm probably most excited about which is in education yeah and I
20:43think many people will have seen s Khan's video from earlier this year is Ted talk about as you talked about it's like personal tutor yeah personal tutor an amazing personal tutor an amazing personal tutor and we know the difference in learning having that personalized tutor is incredible compared to class from learning if you can have every child have a personal tutor
21:01specifically for them that then just evolves with them over time that could be extraordinary so that you know for me I look at that I think gosh that is within reach at this point and and that's one of the benefits I'm most excited about like when you look at the the landscape of things that you
21:15see as possible what is it that you know you are particularly excited about I I think certainly ai ai tter are going to be amazing perhaps already are I think there's also perhaps companionship which may seem odd because how can the computer really be your friend but if you if you have an AI that has memory you know and remembers all of your interactions and has
21:38read every you can say like give it permission to read everything you've ever done so it really will know you better than anyone perhaps even yourself and and and where you can talk to it every day and and those conversations spold upon each other you will actually have a great friend as long
21:54as that friend can stay your friend and not get turned off or something don't turn off my friends but I think that will actually be a real thing and I have a one of my sons is is sort of has some learning disabilities has trouble making friends actually and and I was like well you know he an AI friend would actually be great for him oh okay you know it's that was a
22:20surprising answer that's actually it's worth worth reflecting on it's really interesting Ian we're already seeing it actually as we deliver you know Psychotherapy anyway now doing far more by digitally and by telephone to people and it's making a huge difference and you can see a world
22:35and which actually you know AI can provide that social benefit to people just quick question on on X and then we should open it up to everybody you made a change when you in one of the well made many changes but quite a few one one of the one of the Chang you love that letter yeah i' got a real thing about it you you really do you really do one of the changes which
22:59you know kind of you know goes into the space that you know we have to operate in and this this balance between free speech and moderation it's you know we grapple with as politicians you were grappling with your own version of that and and you you moved away from a kind of manual human
23:15yeah way of doing it the moderation to the the community notes and and I think that's it it was an interesting change right it's not what everyone else has done it would be good you know what's what was the reasoning behind that and why you think that is a better way to do that yeah
23:31part of the problem is if if you if you Empower people as censors then well have there's going to be some amount of bias they have and then whoever appoints the sensors is effectively in control of information so then the the idea behind Community notes is well how do we have a consensus
23:48driven I mean so it's not really censoring it but consensus driven approach to truth how do we or how do we how do we make things the least amount untrue like you can say like what one can't pass perhaps get to Pure truth but you can aspire to be more truthful so the the the thing about Community notes is it doesn't actually delete anything it simply adds context now that
24:14context could be this thing is untrue for the following reasons and but but importantly with Community notes everything is open source actually so you can see ex the software every line of the software you can see all of the data that went into a community node and you and you can independently create that Community node so if you've got if you see manipulation of the
24:36data you can actually highlight that and say well this this this there appears to be some gaming of the system and you can suggest improvements so it's it's it's maximum transparency which is I think combined with the kind of wisdom of the crowds and transparency to get to a better
24:53Onis and and really one of the key elements of community notes is that in order for to be shown people who have historically disagreed must agree and and there is a bit of AI usage here so there's we populate a parameter space around each contributor to community notes and then a parameter space so so everyone's got basically these these vectors associated with
25:18them which so it's it's not as simple as as right or left it's saying it's it's more it's several hundred vectors that that because things are more complicated than something right right or left and and and then we'll we'll do sort of a inverse correlation say like okay these these
25:36people generally disagree but they agree about this note okay so then that so then that that that gives the note credibility okay yeah that's the that's the core of it and it's working quite well yeah I've yet to see a note actually be be present for more than a few hours that that is
25:53incorrect so the batting average is extremely good and when I ask you people say oh they're worried about Community notes sort of being disinformation like send me one and then they can't so so I think it's I think it's quite good I mean the general aspiration is with the xplatform is to inform and
26:10entertain the public and to be as accurate as possible and as truthful as possible even if someone doesn't like the truth you know people don't always like the truth no not always but but that's that's the aspiration and I think if if we are if we stay true to the truth then I think we'll find that people use use the system to learn what is going on and to to learn it I I
26:37think actually truth pays so I think it'll be what what I mean assuming you don't want to engage in self- delusion then then I think it's it's the smart move you know so excellent very helpful right let's open it up to all our guests here we've got some microphones they'll come put your
26:56hands up they'll come come and find you we got yes go for it thank you good evening Alice bentin from entrepreneur first thank you for a fascinating conversation I suppose a question for each of you prime minister the UK has some of the best universities in the world we have the
27:14talent what will it take for the UK to be a real breeding breeding ground for unicorn companies and Elon being a Founder in the UK is still a non-obvious career choice for the most exceptional technical Talent what are the cultural elements that we need to put into place to change this
27:32thank you both John you want to go first go for it sure well you're right that there are cultural elements where you you you know the the culture should celebrate creating new companies and and there should be a bias towards supporting small small companies because they're the ones that need nurturing the larger companies really don't need nurturing ing so
27:59you know just you can think of it's sort of like a garden if it's a little Sprout it needs needs nurturing if it's a Mighty Oak it doesn't need quite as much so I think that that that is a mindset change that is important but but I I should mention that London is you know London and San Francisco or the Bay Area are really the two senses for AI so that so London
28:23is actually doing doing very well on that front that the two MO I say the two leading locations on Earth and you know San Francisco's probably ahead of London but London's really very strong or London area greater London home counties I guess keep going keep going so I'm just saying
28:44objectively this is the case and but you do need that you need you need the infrastructure you need you need landlords who are willing to rent to new companies you you need law firms and accountants that are willing to support new companies and it's generally a mind it is a mindset change and I think some of that is happening but I think really it's just culturally
29:06people need to decide this is this is a good thing yeah yeah no actually well thanks for for what you said about the UK it's something that we work hard on lots of people in the room are are part of what makes this a fabulous place for intive companies including Alice so Alice what I'd say is you
29:22know my job is to get all the you know the nuts and bolts right make sure that all of you are starting companies can raise the capital that you need everything from you know your seed funding with our incredible you know Eis tax reliefs all the way through to you know your late stage rounds
29:37and we need reform of our Pension funds and the Chancellor's got a bunch of incredible reforms to unlock capital from all the people who have it and deploy it into growth Equity right that is a work in progress we're not there yet but I think we're we're making we're making good progress we
29:51need talent we need people you right so that means an education system that prioritizes the things that matter and you've seen my reforms I go on about more maths more maths more maths but I think that is important but also attracting the best and the brightest here if you look at our
30:06fastest growing companies in in this country and I think it's probably the same in in the US over half of them have a non-british Founder right and so that tells you we've got to be a place that is open to the world's best and brightest entrepreneurial Talent so the Visa regime that
30:21we've put in place I think does that makes it easy for those people to come here and then actually it's the thing that we spent the beginning of the session talking about the regulation right making sure that we've got a regulatory system that's Pro Innovation that yeah we of course we always
30:35need guard rails on the things that will worry us but we've got to create a space for people to innovate and do different things you know those are all my jobs the thing that is tougher is the thing that Elon talked about which is culture right it's how do you transpose that culture from
30:49places like Silicon Valley across the world where people are unafraid to give up the security of a regular paycheck to go and start something and be comfortable with failure you you talked about that a lot I think you talked about it more in when you were playing games right but that you've got to be comfortable failing and knowing that that's just part of the process and that is a
31:10it's a tricky cultural thing to do overnight but it's an important part of I think creating that kind of environment yeah if if you don't succeed with your first startup it shouldn't be sort of a catastrophic career ending exactly thing it should be you know what good I think should like should
31:24be like well you know you gave it a good shot you know and and and now try again exactly and it's so one thing I was going to mention is like obviously creating a company is sort of a highrisk high reward situation but and I don't know quite what the how it works in in the UK I think
31:42it's probably better than than Continental Europe but the the stock options are very difficult in most parts of Europe I'm not sure if how than the UK but but if somebody's basically going to risk their the sort of life savings and with and the vast majority of startups fail so I
31:59mean you hear about the startups that succeed but most companies are most startups consist of you know a massive amount of of of work followed by failure that's actually most most companies and so it's a highrisk high reward and and and so the higher reward part does need to be there for it to make sense yeah I think that was a very soft pitch for tax policy that I CH but I
32:24actually I can tell you so look I a I agree and we have so we have I think relative to certainly European countries but certainly the US definitely California a much lower rate of capital gains tax okay right so for those people who are risking and growing something like we think the reward should
32:40be there at the end so 20% capital gains tax rate and on stock options I don't know if we've got anyone from index ventures in in the room so you know index one of our bleeding VC funds here they they do a regular report looking at most countries tax treatment of stock options yeah
32:59and you know when I was Chancellor of you know treasury secretary equivalent you know we were I think down at we were pretty good but we were fourth or fifth and I said we need to for exactly the reason that you mentioned was like this has got to be the best place for innovators we need to move that up and I think in the last iteration of that report we had because of the changes that
33:18Jeremy and I had made we have moved up to I think second from from memory so hopefully that should give you and everyone else some comfort that we recognize that's important because when people work hard and risk things yeah they should be able to enjoy the rewards of that high risk High reward yeah and I think we have a we very much have a tax system that supports that and those
33:35are the values that you know I believe in and I think most of us in this room probably do as well right next next question I've got seven front of me and then I'll come over here go on go on S thanks very much we've talked about some really big Ideas Global changing ideas I'm really interested particularly in the context of creation of of science and technology
33:57superhubs and so on how does that map onto the everyday lives of of people living in say Austin Texas to choose r or in my case Nottingham East Midlands what what is how do you see that evolving for people you know every day the sort of everyday effects of AI yeah for context Elon so Seb Seb runs are are equivalent of CVS right or Walgreens so you know
34:24when as I visited right so he's got of people coming in his shops every day and it's making sure how do we make this relevant I think so is your question how how is this relevant to that person you know maybe actually let me go I'll go first on that because I I think it's a a fair a
34:38fair point I was just going over with the team a couple of things that we're doing because I was thinking like how are we doing AI right now that it's making a difference to people's lives and we have this thing called gov.uk which is which actually when we when it happened several years
34:53ago was a pioneering thing all the government information brought together on one website gov.uk and so you need to get a driving license passport any interaction with government it was centralized in a very easy relatively easy to use way better than most better than most yeah so we we are about so we're about to deploy AI across that platform so that is something that
35:16I think you know several million people a day use right so a large chunk of the population is interacting with gov.uk every single day to do all these day-to-day tasks right every one of your customers is doing all those things and so we're about to deploy AI into that to make that
35:32whole process so much easier because you know some people will be like look well I'm currently here and I've lost my passport and my flight's in five hours you know at the moment that would require you know how many steps to figure out what you do you know we actually when we deploy the AI it
35:47should be that you could just literally say that and boom boom boom boom boom this is what we're going to do walk you through it and that's going to benefit millions and millions of people every single day right cuz that's a very practical way in my seat that I can start using this technology
36:01to help people in their day-to-day lives not just Healthcare discoveries and everything else that we're also doing but I thought that's quite a powerful demonstration of literally your day-to-day customer seeing actually their just day-to-day life get a little bit easier because
36:14of something that you know Elon deis and others in this room have helped create yeah no exactly the the the most immediate thing is just being able to ask like having a very smart friend that you can ask anything you know ask how to make something how to solve any problem and it'll tell
36:31you so and and obviously companies are going to adopt this so I think you'll have much better customer service I guess essentially that'll probably be the first thing you notice and and then we talked about education yeah so having a tutor so if you're trying to understand a subject like having a phenomenal tutor on any subject it's that that's really pretty much there
37:02already almost I mean we need to obviously AI needs to stop hallucinating before you know it can't give you I mean we we still have a little bit of the problem where it can give you an answer that's confidently wrong with great grammar and you know bullet points and everything in
37:18citations that was not real so it has to be okay we need to make sure it's it's not it's not it's not giving you confidently wrong two dire answers but but we that's going to happen quick pretty quickly where it is actually correct so yeah I going to say for any for any parent who was
37:36homeschooling during coid and realizing what their kids needed to be helped with that will come as an enormous relief I think very very good right have we got let's go questions over here who have we got we any microphones or Brent are you there perfect hi Bren hobman so you know you've spoken eloquently about abundance and the age of abundance so it it feels obviously
37:58with AI it's everything everywhere everywhere all at once but with with robots and to to get the age of abundance we'll need a lot of robots I know you're working hard on robots as well are there sort of constraints that we should think of and our politicians should be thinking of that we might get one country might get heavily behind in in Rob in robots
38:18that can do all these things and enter the age of abundance and therefore be at a strategic disadvantage well really anything that can be actuated by a computer is effectively a robot so you can think of frankly Tesla cause are robots on Wheels anything that's connected to the internet is effectively an endpoint actuator for artificial intelligence so you've got
38:48Boston Dynamics obviously they've been making impressive robots for a while I think they're at this point mostly owned by Hyundai so I I guess when I's probably going to make robots of that are humanoid and and and some rather interesting shapes that I wasn't disting like the one that
39:04looks like a has wheels and looks sort of like a kangaroo on Wheels I'm not sure what that is but looks a little dented frankly but but there's going to be all sorts of all sorts of robots you've got the company Dyson in in the UK which I think does some pretty impressive things I I I think the UK will not be behind actually on on that front UK also has arm
39:33which is really the the best one one of the best perhaps the best in in chip design in the world Tesla uses a lot of a lot of arm technology almost everyone does actually so I think the UK is in in a strong position Germany obviously makes a lot of robots Industrial robotss that I mean I think generally countries that make robots of any kind even if they seem
40:03somewhat conventional will be will be fine I do think there is a there is a a safety concern especially with humanoid robots because you at least the car can't chase you into this building not very easily you know or chase you off a tree or you know you can sort of run up a flight of stairs and get away from a Tesla I think there a Stephen King movie about that
40:31if your car gets possessed so but if you have a humanoid robot it can it can basically chase you anywhere so I I think we should have some kind of hardwired local cut off that that you can't update from the internet so anything that can be software updated from the internet obviously can be overridden but if you have a local sort of off switch where you p pa SA a
40:56keyword or something and then that puts the robot into a safe State some kind of localized safe State ability an off switch you know where you don't have to get too close to the robot I don't know so if you've got millions of these things going all over the place you're not selling it
41:16just you know like no I I know I'm saying it's is something we should be quite concerned about because if robot can robot can follow you anywhere then you know what if they just one day get a software update and they're not so friendly anymore then we've got a James Cameron movie
41:34on I it's it's actually that's it's funny you're saying that because we in our session that we had today I you know just would say who we they made exactly the same point right de so we're talking about they're talking about movies actually without mentioning James Cameron they're talking about James camera movies and they're saying if you think about it's not just those
41:54movies but any of these movies trains Subways metros cars buses they said all these movies with the same plot fundamentally all end with the person turning it off right or finding a way to shut the thing down and they were making the same point that that you were about the importance of
42:12actual physical off switches yeah and so all the technology is great but fundamentally this same movie has played out 50 times we've all watched it and all fundamentally you know you know point I'm referring to right it all ends in pretty much the same way with someone finding a way to just yeah do thing which is kind of interesting that you said a similar point right
42:31it's probably not the it's not the obvious place you'd go to but maybe that could be one of the tests for the AI we we just say like blank is your favorite jamers Cameron movie fill in the blank yeah excellent right yes we got over there yeah perfect hi a question for
42:52you both so I'm a founder of a Ai and ml scaleup in the third Center for AI which is leads in the north of England I'm bit biased since the launch of chat GPT three months after that we saw a real increase in fishing attacks using much more sophisticated language patterns what do we do to protect businesses consumers so
43:13they trust this technology better and how do we bring them along that Journey with us well I think we shouldn't trust it that much actually it is actually quite quite a significant challenge because we're getting to the point where even open source AI can pass human capture tests so you know this s are you a human
43:34identify all the traffic lights in this picture you're like okay yeah is going to have a no problem doing that in fact it'll do it better than a human and faster than a human so we're like how do you you know at the point of which is better a better human better passing human test than humans
43:50then well what tests actually make sense that is a real problem I don't actually have a good solution to it that one of the things we're trying to figure out on the xplatform is how to deal with that because it really we really are at the point where even with with open source you know
44:09readily available AI you don't need to be sort of leading the field you can actually be better than humans at passing these tests and that's sort of why we're thinking well perhaps we should sort of charge a a dollar or a pound a year it's a very tiny amount of money but it's it still it
44:23still makes it prohibitively expensive to make a million Parts so and especially if you need a million payment methods then you you run out of sort of for stolen credit cards pretty quickly so that that's that's sort of where we're thinking like we might have to sort of just charge some very tiny amount of money um3 cents a day effectively to deal with the the onslaught
44:50of AI powered bonds and if if and that that is that is still a growing problem but will will be I think perhaps an insurmountable problem next year so and and and then you have to worry about well manipulation of of information is making something seem very popular when in fact it is not
45:08because it's getting boosted by all these likes and and and repost from AI powered bots so that's why I sort of think somewhat inevitably it leads to some small payment in order to dramatically increase the the cost of a bond so I think I think frankly I think probably any social media system that doesn't do that will simply be overrun by B you know I think my my
45:36general answer would be you know we we need to show that we are on top of mitigating the risk right so people can trust the technology that's what actually the last couple of days has been about on the safety Summit is just showing you know we're investing in the safety Institute
45:51having the people who can do the research on these things to figure out how we mitigate against them and we have to do it fast and we have to keep iterating it because I think all of us probably in this room believe that the technology can be incredibly powerful but we've got to make sure we bring people along that Journey with us that we're handling the risks that are there and I said
46:09there's a job to do and the last couple of days I think we make good progress on it because we want to focus on the positives and manage these things but that requires action and and that's what the last couple of days has been about and your your story your analogy there was part of
46:24the research that actually you know the team working on the task force here published and presented yesterday I don't know if you saw it was which is a essentially that it was a using AI to do to create a ton of fake profiles on social media and then infiltrate particular groups with
46:45particular information and actually at the moment that is said to your point of it's like cost free it's it's it's getting it's getting to the point where it's like really you're going to have 100 for a penny sort of thing ridiculous and if you think about some of these social networks at quite a neighborhood or town level it's not that many fake profiles that you need to quickly create
47:02suddenly they're everywhere and there's some local issue that might be of importance and you know the team have have run versions of how that would look like and suddenly they're interacting with everybody and then spreading misinformation around ex your point that's we literally as part of the
47:16research that we published on misinformation yesterday it's it's a real challenge yeah exactly to your point I mean you the images it's it's you don't even need to steal some body's picture cuz then that's traceable but you can actually just say create a new a new image of a person realistic looking but doesn't exist and and then create a biography realistic but doesn't
47:37exist and do that on mass and practically the only way we'll be able to tell us that the gram is too good did give away yeah no tyos come on now I'm getting wave that because I think we are out of time I don't we take one very brief last question and let's make a good one yes sir going you right
47:56in front of me G thank you for the opportunity Elon question for you related to X platform are there simple things we can do especially when it comes to visual media you alluded to the fact that it's fairly straightforward and effectively free to make people like yourself say and do
48:13things that you never said or did yeah can we do something like cryptographically signed media I'm from Adobe we're working on this project yeah Twitter was a member love to see X come back okay digitally signed media to indicate not only what was created by AI but what came from a camera what was real yeah to imbue a sense of Trust in Media that can go viral that sounds
48:32like a good idea actually so if some some way of authenticating would be would be good so yeah I I that sounds like a good idea we we should probably do it there you go actually on that on that point so I've I've already so this is particularly pertinent for people in my job right and I've already had a situation happen to me with a doctored image that goes everywhere
48:59negative by the time everyone realizes well that's fake and we should stop sending it the damage is damage is done and actually we were again reflecting today if you think next year you've got elections in you know I think you know the US India I think Indonesia probably here there you
49:17go massive news and actually you've got just an enormous junk of the world's population is voting next year right and you got EU elections as well and you know actually just these issues are right in front of us you know next year is where big elections across the globe probably the
49:36first set of Elections where this has been a real issue yeah so figuring out how we manage that is I think kind of mission critical for people who want you know the Integrity of our democracy yeah I mean some of it is it's quite interesting like the Pope in the puffer jacket have you seen that one I haven't that's amazing but I mean I still run into people who who who think that's real
49:57I'm like well what are the odds he's wearing a puffa jacket in July in Rome you know be sweating but it actually look quite quite dashing I say in fact I think AI fashion is going to be a real thing so so I Doom and Gloom like we live in the most interesting times and I think this is
50:13it is you know like 80% likely to be good and 20% bad and I think we're if we're cognizant and and careful about the bad part on balance actually it will be the future that we want or the the future that is preferable and and it actually will be somewhat of a leveler an equalizer in the
50:33sense that you know I think everyone will have access to goods and services and education and so you know I think probably it leads to more human happiness so I I guess I'd probably leave on on an optimistic note perfect yeah I that's a well that's a that is a great note to end on I
50:50think that you know we all want that that better future we think it's there the promise of it is certainly there lot people in this room including yourselves are working hard to make it happen our job in government is to make sure it happens safely but on the basis of this conversation in the last couple of days I'm certainly leaving more confident that we can make that happen so it's
51:07been a huge privilege and pleasure to have you here thank you very much for having [Applause] me
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