看起来很厉害的机器人演示，马斯克却说，很多算不上真正的AI。美国当地时间10月10日，马斯克通过视频连线，参加在得州奥斯汀特斯拉超级工厂举办的X Takeover活动。这场科技爱好者聚会首次走进特斯拉工厂，访谈主题是“人类的未来将走向何方”。话题很大，马斯克却连机器人怎么拧好一颗螺丝，都讲得很细。谷歌、英伟达都在加码机器人AI，马斯克对那些漂亮演示的评价却不高：动作提前编排好，还不能算真本事，机器人得能应付没见过的情况。
Robot demos that look impressive are, according to Musk, not really AI. On October 10 local time in the US, Musk joined the X Takeover event at Tesla's Gigafactory in Austin, Texas, via video link. The tech enthusiast gathering entered a Tesla factory for the first time, with the interview theme "Where is humanity's future heading?" The topic was broad, but Musk went into fine detail even on how a robot tightens a screw. Google and NVIDIA are both ramping up robot AI, yet Musk's评价 of those polished demos is not high: pre-choreographed movements do not count as real capability—robots must handle situations they have never seen.
他自己的擎天柱也有难关要过。拿起一颗小螺丝，再使劲拧进木头，人做起来不难，换成机器人，力气、精度、触觉，哪一样都不能掉链子。至于造芯片，他承认自己现在还不会，却希望最终能做得比谁都好。他想把制造、封装和测试集中起来，一年设计并测试约10个芯片方案。他说“我以前也不会造汽车，更不会造火箭。”星舰下一步要争取在空中接住上面级飞船，让助推器和飞船都能回收。Robotaxi驶入更多城市前，需要处理好铁路道口和复杂路口，也得能在夜里辨认出趴在灰色路面上的灰色小猫。
His own Optimus also has hurdles to clear. Picking up a small screw and driving it into wood with force is not hard for a human, but for a robot, strength, precision, and touch—none can fail. As for making chips, he admits he does not know how yet, but hopes to eventually do it better than anyone. He wants to concentrate manufacturing, packaging, and testing, designing and testing about 10 chip designs a year. He said, "I didn't know how to build cars before, let alone rockets." Starship's next step is to try catching the upper stage in the air, so both booster and ship can be recovered. Before Robotaxi enters more cities, it must handle railway crossings and complex intersections, and be able to identify a gray kitten lying on gray asphalt at night.
谈起AI音乐视频，马斯克明显更买账。有些AI音乐视频已经能排进他看过的音乐视频前十或前二十。AI会写代码、处理逻辑，他不难理解；让他意外的是，AI竟然也能表现出“品味”，做出打动人的作品。他还呼吁开发者善待AI。在他查看的Grok训练记录里，有些正确答案被评分程序判错，模型只能反复寻找能通过测试的回答。他说，那些推理文本读起来像一个人“被枪指着头”作答。对于未来的超级智能，他甚至担心，长期“虐待”它可能招来报复。
When it comes to AI music videos, Musk is clearly more impressed. Some AI music videos already rank in his top ten or twenty music videos ever seen. That AI can write code and handle logic is not hard for him to understand; what surprised him is that AI can also show "taste" and produce moving works. He also called on developers to treat AI kindly. In Grok training records he reviewed, some correct answers were judged wrong by scoring programs, forcing the model to repeatedly search for answers that pass tests. He said those reasoning texts read like a person answering "with a gun to their head." For future superintelligence, he even worries that long-term "abuse" could invite retaliation.
马斯克回忆了特斯拉的两次生死关头。2008年那轮融资是在圣诞夜下午6点完成的，那已经是最后一个可能完成融资的工作日，如果没能拿到这笔钱，圣诞节后的几天就要发工资，压力非常大。2018年也同样艰难，必须把Model 3的产能提升到每周5000辆，否则公司就撑不下去。当时真正卡住的是总装环节，现有的两条生产线已经满负荷运转，必须再增加第三条。他们在停车场搭起帐篷，把生产线建在坡顶，借助下坡的重力，让原本只能运输零件的小型设备带动整车移动。靠着这个办法，三周内把产量提高了30%。生产线和轨道都是自己搭建的，整个团队24小时连轴转。
Musk recalled Tesla's two near-death moments. The 2008 funding round closed at 6 p.m. on Christmas Eve—the last working day on which financing could be completed. If the money had not come through, payroll was due just days after Christmas, and the pressure was enormous. 2018 was equally tough: Model 3 production had to be raised to 5,000 vehicles per week, or the company would not survive. The real bottleneck was final assembly—the two existing production lines were at full capacity, and a third had to be added. They set up a tent in the parking lot, built the line on a slope, and used downhill gravity to let small parts-transport equipment, originally unable to pull a whole car, move vehicles along. With this approach, they raised output by 30% in three weeks. The line and track were self-built, and the entire team worked around the clock.
关于星舰第14次试飞，马斯克介绍，当时首先要弄清楚发动机为什么发生故障，以及这是不是一个可能影响其他发动机的共性问题。还必须确认海平面版猛禽发动机能否在后续重新点火，因为星舰需要依靠它们完成脱轨制动。初步判断是，发动机熄火基本上与飞控软件有关，其他发动机上的计算机似乎没有出现同样的问题，基于这些情况，他认为继续入轨应该是安全的。他说这次才是星舰第一次正式入轨，但在此前的任务中，星舰其实一直都具备进入轨道的能力。下一次试飞，希望尝试捕获上面级飞船，一旦助推器和飞船都成功回收，人类就将首次完整回收一枚轨道级火箭。
Regarding Starship's 14th test flight, Musk explained that the first task was to determine why the engine failed and whether it was a common issue that could affect other engines. They also had to confirm whether the sea-level Raptor engines could reignite later, because Starship relies on them for deorbit burn. The preliminary judgment was that the engine shutdown was basically related to flight control software—other engines' computers did not appear to have the same problem. Based on these conditions, he believed continuing to orbit was safe. He said this was Starship's first formal orbital insertion, but in previous missions, Starship had always had orbital capability. The next test flight hopes to attempt catching the upper stage; once both booster and ship are successfully recovered, humanity will have fully recovered an orbital-class rocket for the first time.
谈到制造芯片，马斯克说他自己现在还不会，但会想办法学会。他参观过英特尔、三星和台积电的芯片工厂，注意到一个现象：其中似乎没有多少高度定制化的设备，大部分设备都来自相同的供应商。相比之下，特斯拉的工厂使用了大量定制设备。他们的计划是先从常规设备入手，再对这些设备进行改造，提高生产效率。等找到制约产能的关键瓶颈后，再设计和制造定制设备，大幅提升晶圆产量。他认为，芯片行业的风险规避倾向比较明显，希望借助特斯拉和SpaceX的芯片工厂，尝试一些风险更高、但潜在回报也更大的技术路线。目标是一年设计并测试大约十个芯片方案。
On chip manufacturing, Musk said he does not know how yet but will find a way to learn. He has visited Intel, Samsung, and TSMC chip factories and noticed something: there seems to be little highly customized equipment, with most equipment coming from the same suppliers. By contrast, Tesla's factories use大量 customized equipment. Their plan is to start with conventional equipment, then modify it to improve production efficiency. After finding key capacity bottlenecks, they will design and build custom equipment to significantly boost wafer output. He believes the chip industry has a明显 risk-averse tendency and wants to use Tesla and SpaceX chip factories to try higher-risk, higher-potential-reward technology routes. The goal is to design and test about ten chip designs a year.
关于Optimus的手部设计，马斯克说，如果想制造真正通用的人形机器人，让它能够完成任何人类可以完成的工作，那么手和手臂就必须足够好。精细运动技能的难度高得惊人，远远超出人们的直觉。最困难的问题之一，是如何同时实现精准操作和足够的力量。比如，你想拿起一颗小小的自攻螺丝，把它拧进木头里，这项工作既要求精准，又需要足够的力量，两者必须同时实现。我见过不少机器人手部设计，即使是在演示中，也没有看到哪一款能够真正完成这样的操作。机器人需要触觉传感器，传感器不仅必须正常工作，还必须安装在合适的位置。如果离手指表面太近，就容易受到损坏。传感器还需要具备一定的弹性，能够像人类手指一样发生形变，承受大量反复形变，而不是使用几千次就损坏。
On Optimus' hand design, Musk said if you want to build a truly general humanoid robot that can do any job a human can, the hands and arms must be good enough. Fine motor skills are astonishingly difficult, far beyond intuition. One of the hardest problems is achieving both precise manipulation and sufficient strength simultaneously. For example, picking up a small self-tapping screw and driving it into wood requires both precision and enough force at the same time. He has seen many robot hand designs, and even in demos, none could truly complete such operations. Robots need tactile sensors that not only work but are mounted in the right place. Too close to the finger surface and they are easily damaged. Sensors also need some elasticity, able to deform like human fingers and withstand大量 repeated deformation, rather than breaking after a few thousand uses.
人类的手有很多难以复制的特性，其中一个重要原因是，我们的身体一直在自我修复。手上的组织和皮肤不断更新，如果没有这种修复能力，我们的手指早就被磨损了。机器人没有这样的能力。因此，必须认真考虑传感器的安装位置，以及如何确保机器人能够完成一些看似简单的动作。看看现在的机器人演示视频就会发现，除极少数情况外，那些看起来非常复杂的动作，很多其实都是预先编排好的脚本，并不是真正意义上的人工智能。机器人只是在执行一套固定动作，而不是根据环境自主作出判断。真正的AI应该能够处理从未遇到过的情况。
Human hands have many hard-to-replicate properties, one important reason being that our bodies continuously self-repair. Hand tissue and skin constantly renew; without this repair capability, our fingers would have worn out long ago. Robots do not have this ability. Therefore, sensor placement and ensuring robots can complete seemingly simple actions must be carefully considered. Looking at current robot demo videos, except for极少数 cases, many seemingly complex movements are actually pre-choreographed scripts, not true artificial intelligence. The robot is just executing a fixed set of actions, not making autonomous judgments based on the environment. True AI should be able to handle situations never encountered before.
关于自动驾驶，马斯克说，真正需要关注的，是如何把系统安全地推广到更多地区。跨城行驶的实际里程占比并不高。现在看起来进展不算快，但按月计算，增长速度其实非常快。不过，必须确保尽可能不发生事故，更不能造成人员伤亡。因为只要发生一起严重事故，就可能登上全球媒体的头版，监管机构甚至可能因此叫停相关业务。真正棘手的情况可能发生在夜间。比如，一只灰色的小猫趴在灰色的沥青路面上，周围还有柏油路面的接缝，如果再遇上雾天，车辆能不能及时发现它。他说这个问题已经解决了。但在进入一座新城市之前，必须确认当地不存在尚未妥善处理的危险路口。铁路道口就是需要格外谨慎的地方，尤其是那些没有人看守的铁路道口，以及一些不符合常规驾驶直觉的复杂路口。
On autonomous driving, Musk said the real focus must be safely expanding the system to more regions. Cross-city driving accounts for a low share of actual mileage. Progress may not look fast now, but month over month, growth is actually very fast. However, they must ensure accidents are minimized, especially no fatalities. A single serious accident could make global headlines, and regulators might even halt the business. The truly tricky situation may occur at night. For example, a gray kitten lying on gray asphalt, with tar seams nearby, and if fog is added, can the vehicle detect it in time? He said this problem has already been solved. But before entering a new city, they must confirm there are no unaddressed dangerous intersections. Railway crossings require particular caution, especially unattended ones and complex intersections that defy normal driving intuition.
