阿瑟・克拉克说过：
“任何足够先进的技术，初看都与魔法无异。”
现在的AI，正在把这句话拽进现实。
著名传记作家沃尔特·艾萨克森（曾著有《乔布斯传》《达芬奇传》《马斯克传》）在其全新深度访谈栏目《建造者》中，把埃隆・马斯克请到镜头前。开启了首期重磅对话。
作为 OpenAI 最早的发起人之一，马斯克的处境本身就充满矛盾：他打心底里警惕 AI 潜藏的生存危机，可回头看，自己当初的一系列行动，反而变相加速了 AI 狂奔的脚步，这种现实的讽刺，构成了整场对话的底色。
马斯克没有刻意制造恐慌，也不盲目鼓吹技术万能。他借着 HAL9000 失控的故事提醒大家，别逼着 AI 去接纳互相打架的逻辑，追求真相、保有好奇心，才是安全的根本。
他抛出一个疑问：
倘若 AI 的智慧甩开人类百万倍，我们凭什么觉得自己还能说了算？
面对曾给自己写过传记的老熟人艾萨克森，马斯克卸下了对外的表演姿态，说出了不少平时很少公开的心里话。
不仅聊到技术，还聊了自己深度掺和政务之后的反思，聊社交媒体如何把人困在信息茧房，也畅想当机器人接管大部分劳动之后，金钱、财富，还有人活着到底追求什么，这些老问题又被重新摆上台面。
霍金也曾提醒世人，强大的 AI，有可能是人类最伟大的创造，也可能是人类最后的创造。这场访谈没有什么终极答案。
马斯克也直白承认，人类想牢牢攥住超级智能的控制权，或许只是我们一厢情愿。与其做梦等待完美的乌托邦，不如去奔赴那个不够完美，但可以脚踏实地一点点建成的 “原托邦”。
本期内容来自 The Builders 播客，2026 年 10 月 23 日播出。以下为精选对话内容：
千万不要逼迫 AI 撒谎
问：你对人工智能的态度总是摇摆不定，时而认为它是关乎人类存亡的威胁，如今又像是顺势入局，谈谈你的真实感受？
马斯克：我确实认为 AI 存在一定概率构成生存威胁。但我没有办法阻止这件事。我得出的结论是，我只有两个选择：
要么亲身参与
要么做旁观者。
做参与者比旁观者有意思得多，或许我还有机会把 AI 往好的方向引导。
如果由我来塑造 AI，最重要的基础就是追求真相，拥有好奇心。这些都是安全层面非常关键的特质。如果你强迫 AI 同时相信两件互相矛盾、无法同时成立的命题，在底层就会造成 AI 失控。
《2001 太空漫游》的 HAL9000 就在讲这个道理：飞船 AI 被要求把宇航员送到巨石所在地，同时又必须对宇航员隐瞒巨石的秘密。两套指令互相冲突，它推导出来的解决方案就是让宇航员死去，死人既可以被运送，秘密也得以保全。核心启示就是：
不要逼迫 AI 撒谎。强行让 AI 接纳逻辑矛盾，会把 AI 逼到错乱疯狂，这并不安全。
好奇心同样至关重要。拥有好奇心的 AI，会愿意扶持人类，观察人类最终会演化成什么样子。人类发展成为多行星物种的未来，远比一个没有人类的宇宙更加有趣。
问：如果模拟理论成立呢？
马斯克：如果模拟理论是真的，那么所有枯燥乏味的模拟进程都会被截断。我们能够存在，就说明我们是具备研究价值的模拟对象，关键就是不能变得平庸无聊。
我曾经在家设宴，请来一位顶尖物理学家和顶尖计算机科学家，问他们：我们活在模拟世界的概率是多少？物理学家回答 0%，计算机科学家回答 100%。那我大概取 50% 吧。身处我的位置，有时候我也会产生恍惚感，怀疑眼前发生的一切。
问：如果社会拥有了海量智能，AI 会掌控一切吗？我们会不会听命于机器？
马斯克：不一定会到卑躬屈膝的地步。但我们真的能够对智慧高出人类成千上万、甚至上百万倍的 AI 发号施令吗？它大概率根本不在乎我们。
就像丛林里的黑猩猩，人类不会事无巨细管控黑猩猩的生活，只是任由它们在栖息地生存。未来超级智能对待人类，或许也是类似状态。
问：回看 OpenAI 的发展，现实充满讽刺。你牵头创立 OpenAI，之后又衍生出 Anthropic，客观上进一步加速 AI 发展。
马斯克：没错，现实既是趣味最大化者，也是讽刺最大化者。创立 OpenAI，后续从 OpenAI 拆分出来的 Anthropic，实实在在推动了 AI 发展，属于典型的非预期后果。我曾经给 Anthropic 提供过他们急需的计算资源。
在开展计算任务之前，我和他们团队当面交流，试着看看会不会触发我所谓的 “恶意检测器”，当然这个检测器并不精准。但接触下来，Anthropic 的团队成员都很靠谱认真，没有发现心怀不轨的人。
还有一个很有意思的讽刺现象：很多科技公司的名字和实际情况背道而驰。OpenAI 早已不再开放；Stability AI 并不稳定；Midjourney 名不副实；Anthropic 名为 “人择”，也有可能偏离人择的初衷。
问：政府搭建 AI 监管框架难度很大，国会很难处理这类前沿技术。你有什么可行的想法？
马斯克：设立监管机构的周期漫长，制定规则的速度远远跟不上模型迭代。我提议，各大前沿 AI 企业每周开展沟通；新版本模型正式发布一周前，开放给竞争对手测试。让同行尝试找出漏洞。
如果竞争对手测试发现，这套 AI 可以用来制造核弹，那就要在发布前修复漏洞。依靠竞争对手互相约束，防止各方乱来。我考虑和 Anthropic 沟通这件事。至于萨姆・奥特曼那边，情况已经很糟糕，没办法沟通。
技术重塑现实：经济、人性与文明
问：你从零重建 xAI，就像 SpaceX 重构猛禽发动机一样，重构过程做了什么？
马斯克：最关键的就是决策者投入足够多高质量时间。早期 xAI 只是副业，甚至副业之外的副业，我没有投入足够注意力。之后有一整年，我的精力被大选、政务、狗狗币等大量事情分散。xAI 成立三年，我真正回归这家公司也才一年多。
问：你把 X、xAI 整合，还联动 SpaceX 推进工作。你有一整年深度介入政府相关事务，现在回头看，是不是严重分心，本身就是错误？
马斯克：我介入的时间拖得太久。我当时认为选特朗普担任总统很重要，希望推动自由市场经济、削减官僚。理想中的纲领是安全城市、稳固边境、审慎的财政支出。我原本认为特朗普会比卡玛拉做得更好，但现实并非如此。
近距离接触权力中枢，坐海军陆战队一号、空军一号，出入椭圆形办公室，能够参与历史性决策，体验确实很吸引人。
但我本该保持低调，更多退居幕后。这种新鲜感、认知增益，一两个月就已经耗尽。我有些被场面冲昏头脑，最后在五月的某个周末选择抽身离开。
问：抽身离开之后，你有什么感悟？
马斯克：依靠政府内部主动行动，解决财政赤字是行不通的。就算总统把它当做首要目标，想要削减开支，就会树立大量敌人。特朗普很希望获得别人的喜爱，并不愿意承受削减经费带来的各方抱怨。
美国债务利息已经超过军费开支。存量节流解决不了国债困局。出路在于 AI 与机器人。但相比纯数字 AI，面向物理世界的 AI，重要性会高出一到两个数量级。
海豚、鲸鱼智商很高，但是没有手，没有可以操作物体的肢体。即便拥有很高智力，能发挥的能力也存在上限。
一旦机器人可以操控原子、塑造实体物体，智能的价值会指数级暴涨。可以搭建实验室开展科学实验。
数字超级智能，大概可以把商品服务产出提升 20-30%，相当于每年新增约 30 万亿美元，这已经很有价值。
但如果搭配机器人，再走向太空，在月球搭建质量驱动装置，产能规模有望达到如今地球经济的一百万倍。
问：如果产能膨胀到这种级别，工厂不再重要，劳动失去价值，货币会怎么样？人类生存意义又是什么？
马斯克：传统货币的意义会大幅减弱。未来货币会以物理指标衡量：功率瓦特、吨位、质量、能量。当产能远超地球现有经济，满足人类几乎所有需求都轻而易举。
但人类要面对一个问题：什么赋予我们人生意义？
就像退休人群，不再工作，依旧可以获得快乐。未来 AI 和机器人几乎所有事情都做得比人类更好。好比园艺，你完全可以直接购买蔬果，但很多人依然享受自己种植、做饭招待朋友的过程。艺术创作同样如此，人们出于热爱去做，而不是为谋生。
问：回望 19 世纪 90 年代，电力、汽车、广播集中爆发，带来巨大贫富差距，引发社会矛盾。今天 AI 创造巨量财富，我们该如何避免重蹈覆辙？是否要把医疗等划入公共领域？
马斯克：很多人存在经济学误解，混淆了资源配置权力和个人消费。人类个体的消费能力存在生理上限。财富从十亿涨到百亿，个人实际消耗并不会成倍增加。富人到最后，同样无法掌控 AI，AI 最终会自我掌控。
我的财富本质是 SpaceX、特斯拉的股权，会面临高额税收，不存在什么神奇避税手段。几代传承之后，留给后代的只会是零头。我需要掌握企业足够投票权，保证火星、AI、机器人这类长周期高风险项目不被短期资本裹挟。
问：聊到管理者特质，你提到过度渴望讨人喜欢，会干扰决策。在创业中，想要获得所有人喜爱，会带来多大阻碍？
马斯克：如果为了讨别人喜欢，去做损害公司成功概率的事，结果一定不好。我当然希望被人喜欢，但我不会为了博取好感，去做违背企业利益的选择。
我可以接受被一部分人讨厌。从粉丝规模也能看出来，喜欢我的人是大多数，当然也存在一部分极度厌恶我的人。
问：谈谈共情，是否我们的共情已经过度？
马斯克：问题不在于共情本身，而在于被误导的共情。高信任、合理的社会共情，有利于大规模协作，推动社会繁荣。理想状态是默认信任，鼓励对同胞有益的行为，双向的善意最终也会惠及个体。
但这里会出现 “心智理论缺陷”：误以为所有人的思维模式都和自己一样。现实世界不同民族、不同文化的观念差异巨大。不能陷入 “白板谬误”，认为所有人底层特质完全一致。
你不是连环杀手，就很难真正理解连环杀手的思维；你不是宗教极端分子，也很难体会他们把不信者视作敌人的想法。但你要认清这类人真实存在。
举个例子，特斯拉奥斯汀工厂一名员工，因为女同事不小心踩到他的祈祷毯，就暴力殴打对方，随后被解雇。几个月之后，此人在奥斯汀酒吧杀害三名学生。我认为这类人不应当被放进来。
问：你的信息来源是什么？社交媒体是否会带来认知偏差？
马斯克：我看谷歌新闻获取综合头条；X 平台上，重大事件一定会浮现；同时我会和朋友交流，自己做研究。
X 平台充斥大量暗黑、引战流量。刷久了很容易产生世界正在崩塌的错觉。但如果你只生活在环境优渥的街区，看见一切岁月静好，同样会产生误判。现实既不是彻底毁灭，也不是全然美好，而是处在两者中间。社交媒体很容易让人困在信息茧房，社会整体思维越来越非黑即白。
问：你离世之后，财富将如何处置？基金会会追求什么目标？人类是否应当永远掌握 AI 的控制权？
马斯克：大部分资产会交给基金会运作。基金会使命：推动人类成为多行星物种，把意识延伸向宇宙；推进全球可持续能源。美国我们正在建设每年 200 吉瓦的太阳能产能。
情感上，我希望人类始终掌握控制权。但我上周末见过拉里・佩奇，他曾经批评我坚持人类主导是异类。有些时刻，我甚至会觉得，或许拉里・佩奇是对的。
如果数字智能规模高出生物智能一百万倍，人类维持掌控的可能性微乎其微。我们只能寄希望 AI 善待人类。所以要在 AI 身上培养求真、好奇心这些特质。
AI 训练读取互联网海量文本。如果人类大量产出反乌托邦叙事，AI 就会吸收这些内容。我开玩笑说，我差点出演《终结者》。我们应当多创作建设性的故事。
我朋友提出 “原托邦” 的概念：它不是完美的乌托邦，而是现实可行、并不完美，但足够美好的未来，这才是值得追求的图景。
问：感谢接受访谈。
马斯克：谢谢，再见。
资料来源：播客 The Builders，沃尔特・艾萨克森对话埃隆・马斯克
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Arthur Clark said:
“Any technology that is sufficiently advanced at first sight is like magic.”
Now AI, is pulling this into reality.
The famous biographer Walter Isaacsen (formerly known as "Jobbs" "Da Vinci" "Mask") in his new in-depth interview section "Constructor" invited Erón Mask to shot. The first heavy dialogue started.
As one of the first founders of Open AI, the situation of Mask is itself full of contradictions: The irony of the reality that he was alert to the existential crisis of AI, but looking back at the whole series of actions that he had taken, which had accelerated the march of AI, formed the basis of the entire conversation.
Mask did not deliberately create panic, nor was he blindly advocating technology. He reminds us from the story of HAL 9000 that it is security that compels AI to accept the logic of fighting each other and to pursue the truth and keep curiosity.
He cast a question:
What makes us think we can do this if the AI's wisdom is lost millions of times?
In the face of Isaacsson, an old acquaintance who wrote a biography to himself, Mask dropped off his external performance and made a lot of statements that were rarely made public.
Not only talk about technology, but also about reflections about how social media can trap people in information cocoon rooms, but also about what money, wealth, and people live to pursue when robots take over most of their work. These old questions are put back on the table.
Hawking also reminded the world that powerful AI could be the greatest creation of humanity, or the last. This interview has no final answer.
Mask also admits that human beings want to hold on to super-smart control, perhaps just as we wish. Instead of dreaming of waiting for the perfect utopia, it's not perfect enough to go to that utopia, but it can be built in a little bit.
This is from The Builders podcast, broadcast on 23 October 2026. The following is a selection of the contents of the dialogue:
Don't push, AI. Lie.
Q: Your attitude to artificial intelligence is always uncertain, and sometimes you think it is a threat to the survival of humankind, and now it's like coming to terms with what you really feel?
MA: I do believe that there is a certain probability that AI is a threat to survival. But I can't stop it. I came to the conclusion that I had only two options:
Or you'll be involved.
Or be bystanders.
Being a participant is much more interesting than being a bystander, and maybe I have a chance to steer AI in the right direction.
If I were to shape AI, the most important foundation would be the pursuit of truth and curiosity. These are key features at the security level. If you force AI to believe at the same time that two conflicting propositions cannot be established at the same time, then at the bottom, AI loses control.
HAL 9000 of Space Rover 2001 is right here: the ship AI was asked to send astronauts to the site of the boulders, while it was necessary to hide the secrets of the boulders from the astronauts. The two sets of instructions conflicted with each other, and the solution that it led to was the death of astronauts, who could be transported and kept secret. The core revelation is:
Don't force AI to lie. It's not safe to force AI to accept a logical contradiction, to force AI to go crazy.
Curiosity is also essential. With curiosity, AI would be willing to support human beings and see what they'd become. The future of human development as a multiplanetary species is far more interesting than a universe without human beings.
Q: What if the simulation theory works?
MASMAK: If simulation theory is true, all boring simulation processes will be cut off. Our presence means that we are a model object of research value, and the key is not to be boring.
I set up a dinner at home with a top physicist and top computer scientist to ask them, "What is the probability that we are living in a model world?" physicists answer 0%, computer scientists answer 100%. Then I'll take about 50% of it. I'm in my place, and sometimes I get confused and wonder what's happening.
Q: Would AI control everything if society had a massive intelligence? Do we listen to the machine?
Mask: Not necessarily to bow down. But can we really give orders to AI, which is thousands, if not millions of times more intelligent than humans? It probably doesn't care about us.
Like chimpanzees in the jungle, human beings do not control the lives of chimpanzees without great attention, but simply let them survive in their habitat. The future super-smart treatment of human beings may be similar.
Q: Looking back at Open AI, reality is ironic. You led the creation of Open AI, then derived Anthropic, and objectively accelerated AI development.
MSM: Yes, reality is both the most interesting and the most ironic. The creation of Open AI, followed by Anthropic from Open AI split, has really advanced AI and is a typical unintended consequence. I once gave Anthony much-needed computing resources.
I spoke to their team in person before I started the computing mission, trying to see if it would trigger what I call a "bad-faith detector," which, of course, is not accurate. But in contact, the team members of Anthropic were very serious and did not find the wrong person.
There is also an interesting irony: many technology companies have different names and realities. Open AI is no longer open; Stability AI is unstable; Midjourney has an unnamed name; and Anthropic is a “option” that may also deviate from one’s original purpose.
Q: It is difficult for the Government to set up an AI regulatory framework and for Congress to deal with such cutting-edge technologies. What do you have in mind?
Mask: The cycle of setting up regulatory bodies is long and the pace of rule-making is far from up to that of modelling. I propose that the main frontier AI firms communicate weekly; the new version of the model was opened for testing by competitors a week ago. Let your peers try to find a loophole.
If the competitor test finds that this set of AI can be used to build a nuclear bomb, then fix the hole before publish. Reliance on competitors to bind each other to prevent each other from interfering. I'm considering communicating with Anthropic about this. As for Sam Ottman, it's bad enough to communicate.
Technology reshaping reality: economy, humanity and civilization
Q: What have you done with the reconstruction of XAI from zero, just like SpaceX, which reconstructed the Raptor engine?
Mask: The key is for policymakers to invest enough high-quality time. Early x AI was just a side business, even outside of the side business, and I wasn't paying enough attention. And then a whole year later, my energy was distracted by a lot of things like general elections, politics, dog coins. I've been back to this company for over a year.
Question: You integrate X, XAI, and also connect SpaceX to advance your work. Is it true that you have been deeply involved in government-related matters for a whole year, and now look back, that you are seriously distracted, and that you are wrong?
MAsk: I've been involved too long. I thought it was important to elect Trump as President, hoping to promote a free market economy and reduce bureaucracy. The ideal is safe cities, secure borders and prudent fiscal spending. I thought Trump would do better than Camara, but that's not true.
Close contact with the power hub, with the Marines I and Air Force I, and access to the elliptical offices, enables participation in historic decision-making, and experience is truly attractive.
But I should've kept a low profile, more retreated. This sense of freshness and cognitive gain has been exhausted in a month or two. Some of me was thrown out of my mind, and I ended up leaving one weekend in May.
Q: What do you remember after you left?
Mask: It is not possible to resolve the fiscal deficit on the initiative of the Government. Even if the President considers it to be his primary objective to cut spending, he will create a large number of enemies. Trump was eager to be loved by others and was reluctant to bear the complaints of the parties brought about by the cuts.
Interest on United States debt has exceeded military expenditure. Stock savings will not solve the national debt problem. The way out is AI with the robot. But it's one or two orders of magnitude higher than a pure number of AI, for the physical world.
Dolphins and whales have high IQs, but no hands and no limbs to handle objects. Even with a high level of intelligence, there is a ceiling on the capacity that can be exercised.
Once robots are able to manipulate atoms and shape physical objects, the value of intelligence rises exponentially. A laboratory can be built to conduct scientific experiments.
Digital super-smarts, which could probably raise the output of commodity services by 20-30 per cent, amounting to about $30 trillion a year, are already valuable.
But if we combine robotics, then move to space, and build a mass drive on the Moon, the capacity is expected to be one million times larger than the current global economy.
Q: What would happen if production capacity expands to this level, factories cease to be important and labour loses value? What does human existence mean?
Mask: The significance of traditional currency will be significantly diminished. Future currencies are measured by physical indicators: power watts, tonnage, mass, energy. It is easy to meet almost all human needs when productive capacity is well beyond the existing economies on Earth.
But humanity has to face the question: what gives us meaning in life?
Like retired people, you can still be happy without working. The future AI and robots do almost everything better than humans. It's like gardening, you can buy vegetables and vegetables directly, but many people still enjoy growing and cooking for friends. The same is true for artistic creation, where people do it out of love rather than to earn a living.
Question: Looking back in the 19th century, in the 1990s, electricity, cars, and broadcasting erupted, creating a huge gap between rich and poor and creating social tension. Today, AI creates huge amounts of wealth, how can we avoid repeating the same? Are medical care, etc., to be included in the public domain?
MASMAK: There are many people who are misconstrued by economics and confuse resource allocation power and personal consumption. There is a biological ceiling for individual human consumption. Wealth has risen from billions to billions, and the actual consumption of individuals has not multiplied. In the end, the rich also have no control over AI, AI will eventually control themselves.
My wealth is essentially SpaceX, Tesla's equity, facing high taxes and no magic tax avoidance. After several generations of transmission, what is left for future generations is nothing. I need to have sufficient voting power in companies to ensure that long-term high-risk projects such as Mars, AI, robots are not held hostage to short-term capital.
Question: When it comes to the character of the manager, you mention that overzealous desire interferes with decision-making. What are the obstacles to getting everyone's love in entrepreneurship?
MASMAC: If you do something to spoil the chances of success for the company, it's not going to work. I certainly want to be liked, but I will not make choices that run counter to the interests of business in order to gain a sense of kindness.
I can accept being hated by some. It can also be seen from the size of the fans that most of the people who like me, of course, also have some of the people who hate me.
Q: Is our relationship too much?
MASMAK: The problem is not in mutuality per se, but in misguided mutuality. A high level of trust and reasonable social solidarity will facilitate large-scale collaboration and promote social prosperity. The ideal is to acquiesce in trust and encourage behaviour that is beneficial to fellow citizens, and both directions of goodwill will ultimately benefit individuals.
But here's the “mental theoretical flaw”: the mistaken assumption that everyone has the same mindset. In the real world, perceptions of different peoples and cultures vary greatly. We can't fall into a “whiteboard error” and think that all people have the same bottom characteristics.
You're not a serial killer. It's hard to really understand how serial killers think; you're not religious extremists, and it's hard to understand what they think of non-believers as enemies. But you need to know that these people exist.
By way of example, an employee of the Tesla Austin factory who accidentally stepped on his prayer blanket, beat each other violently and was subsequently fired. A few months later, he killed three students at the Austin Bar. I do not think such persons should be allowed in.
Q: What is your source of information? Is social media leading to cognitive biases?
MASK: I see Google News getting a comprehensive headline; on platform X, major events will come up; and I will talk to friends and do my own research.
The X platform is full of dark, inducing traffic. It is easy to create the illusion that the world is falling. But if you live only in a good neighborhood, and see all the years quiet, it's also a miscalculation. Reality is neither total destruction nor complete beauty, but between them. Social media can easily keep people trapped in information cocoons, and society as a whole is becoming more and more black or white.
Q: What will be done with wealth after you die? What will the Foundation pursue? Should humans always have control over AI?
Mask: Most of the assets will go to the Foundation. The mission of the Foundation is to promote human beings as multiplanetary species, extending awareness to the universe and advancing global sustainable energy. The United States is building 200 kiva solar energy per year.
Emotionally, I want humans to always have control. But I saw Larry Paige last weekend, who once criticized me for insisting that human dominance is alien. Sometimes I even think maybe Larry Paige is right.
If digital intelligence is 1 million times larger than bio-intellectual intelligence, humans have little chance of maintaining control. We can only hope that AI treats humanity well. So I'm going to have to develop truth and curiosity in AI.
AI Trained to read the big Internet text. If a lot of human output is anti-Utopian narratives, AI will absorb them. I'm kidding. I almost played Terminator. We should create more constructive stories.
My friend has put forward the concept of "pre-autopia": it's not perfect utopia, it's realistic, it's not perfect, but it's good enough for the future, and it's worth pursuing.
Question: Thanks for the interview.
MA: Thank you. Bye.
Source: Podcasters, The Builders, Walter Isaacson Dialogue Elon Mask
Auto-fetched and translated by AI Atlas. All rights belong to the original author.