内容来源：https://www.wired.com/story/uncanny-valley-podcast-move-aside-chatbots-ai-humanoids-are-here/

内容总结：

近日，人工智能研究机构OpenAI被曝正加速布局人形机器人领域，重点招募相关人工智能系统研发人才。这一动向被视为其迈向通用人工智能（AGI）的关键一步。

人形机器人因外形与人类相似且能执行日常任务而备受关注。尽管数年前这类机器人仍以动作笨拙著称，但近年来机器学习技术与硬件设计的突破正推动其快速进化。行业专家指出，人形机器人的核心优势在于能适应人类生活环境——无论是上下楼梯、操作车辆还是完成家庭任务，其设计逻辑与人类活动场景高度契合。

目前，特斯拉、波士顿动力、Agility Robotics等企业已在该领域取得显著进展。亚马逊等企业开始测试人形机器人执行物流搬运任务，现代汽车工厂也计划今年部署波士顿动力的电动人形机器人Atlas。然而行业普遍认为，实现家庭场景的复杂任务（如照料宠物、整理家务）仍需长期技术积累，当前演示效果与实际可靠性间仍存在差距。

尽管市场研究预测人形机器人产业规模可能从目前的50亿美元增长至2050年的5万亿美元，但技术瓶颈不容忽视：硬件灵活性不足、环境适应性有限、安全可靠性待验证等问题仍是产业化的重要挑战。与此同时，人工智能与机器人技术的结合也引发了关于军事应用、劳动力替代等伦理与社会议题的讨论。

（注：本报道基于公开行业信息及技术分析，不代表任何企业立场。）

中文翻译：

本周《连线》获悉，OpenAI正在加强其人形机器人领域的研发力度——具体而言，正在招募从事人形机器人人工智能系统的研究人员。人形机器人（即模仿人类外形并能执行日常任务的机器人）在几年前还因动作笨拙而闻名。在本期由迈克尔·卡洛尔与资深记者凯莉·罗比森共同主持的节目中，高级撰稿人威尔·奈特将为我们解读这一领域正在发生的飞速变革。

本期提及内容：
威尔·奈特《OpenAI加速机器人研发角逐通用人工智能》
威尔·奈特《人形机器人迎来成熟期》
拉塞尔·布拉多姆《2025年将成为人形机器人工厂工人的元年》

您可以在Bluesky上关注迈克尔·卡洛尔（@snackfight）、威尔·奈特（@willknight）和凯莉·罗比森（@kylierobison.com）。欢迎来信至uncannyvalley@wired.com。

收听方式
您可通过本页音频播放器收听本周节目，若想免费订阅获取每期更新，请参考以下方式：
使用iPhone或iPad用户请打开播客应用，或直接点击此链接。您也可下载Overcast或Pocket Casts等应用搜索"uncanny valley"。我们亦在Spotify同步更新。

转录稿
注：此为自动生成转录稿，可能存在误差。

迈克尔·卡洛尔：嗨凯莉，最近好吗？
凯莉·罗比森：不错，你呢？
迈克尔·卡洛尔：挺好。很高兴在演播室见到你——
凯莉·罗比森：是啊。
迈克尔·卡洛尔：...这个位置平时都是劳伦·古德坐的。
凯莉·罗比森：我的救世主劳伦·古德啊，能替班我很开心，但实在太想念她了。
迈克尔·卡洛尔：我们都想。本周节目还有《连线》人工智能专家威尔·奈特。欢迎回来，最近如何？
威尔·奈特：大家好。很荣幸受邀。我很好，谢谢。

迈克尔·卡洛尔：既然今天聊机器人，我想问问二位：有特别喜欢的机器人电影吗？
凯莉·罗比森：我最爱《铁巨人》。刚查了上映年份，我小时候绝对是用录像带看的。超爱这部，虽然结局很伤感——机器人最终消失了。
迈克尔·卡洛尔：剧透也没关系，都上映几十年了。
凯莉·罗比森：1999年上映的，应该没问题。
威尔·奈特：我推荐《银翼杀手》，这部经典视觉震撼。初看时只觉得酷，后来读了原著才发现剧情刻意模糊了人与机器的界限。这种对"何以为人"的探讨令人震撼，虽然离现实还很远，但概念很有趣。
迈克尔·卡洛尔：我选《机械战警》。
凯莉·罗比森：不错。
迈克尔·卡洛尔：和《铁巨人》完全相反。但这部80年代范·霍文风格的杰作历久弥新，对警力私有化和公共服务的资本化预言堪称先知。
凯莉·罗比森：很迈克尔的选片风格。
迈克尔·卡洛尔：谢谢。这里是《连线》"奇幻山谷"节目，聚焦硅谷的人物、权力与影响力。今天我们将探讨机器人技术如何成为角逐通用人工智能（AGI）的新战场。近日我们获悉OpenAI正通过招募人形机器人AI研究员加速布局。几年前还因笨拙闻名的人形机器人，正如我们同事威尔所报道的，随着机器学习与硬件的进步正重新引发关注——尤其是在AI行业。我们将深入解析OpenAI等企业的战略意图，探讨其对AGI研发路径的影响，以及可能为人类带来的变革。我是迈克尔·卡洛尔，《连线》消费技术与文化总监。
凯莉·罗比森：我是凯莉·罗比森，负责AI商业领域的资深记者。
威尔·奈特：我是威尔·奈特，《连线》高级撰稿人。

迈克尔·卡洛尔：威尔，首先需要明确：什么是人形机器人？与其他机器人的区别何在？
威尔·奈特：好问题。直观答案是它们拥有人类形态——双腿、双臂和头部。但更关键的是，行业关注它们是因为其设计初衷是适应人类环境：爬楼梯、坐交通工具，完成人类能做的任何事。若将AI定义为类人智能，那么物理世界的行动能力就是核心要素。ChatGPT擅长高等数学，但无法为你冲咖啡，而它们正在突破这个界限。

迈克尔·卡洛尔：如你和其他《连线》作者所述，人形机器人以往常被视作噱头。波士顿动力机器人的跳舞视频超4000万点击，但已知功能仅止于此。最糟时它们连基本任务都难以完成：摔倒、碰撞。现在这种认知如何改变？
威尔·奈特：2012年DARPA机器人挑战赛时，这些人形机器人移动缓慢、频繁摔倒，场面甚至滑稽。但这些年来硬件大幅进步，尤其是电机技术。波士顿动力开创的爆发性平衡动作所需硬件现已更普及，这正是人形机器人开始商业化的原因。

迈克尔·卡洛尔：OpenAI对人形机器人显露出强烈兴趣。你最近从调查中发现了什么？
威尔·奈特：多家实验室都在研发人形机器人，因为这是探索人类物理智能的关键。OpenAI不仅招募相关人才，其招聘清单和其他信号都表明正在加大投入。这很合理，因为其他AI公司也都在转向实体智能领域。

凯莉·罗比森：读到报道时我立刻想到杨立昆的观点：没有物理智能，没有现实世界学习，就不可能实现超越人类的AGI。OpenAI在2021年关闭机器人部门后重启研发很有意思。他们曾用仿生手算法解魔方，现在不仅回归更聚焦AI融合。今年2月OpenAI与专注人形机器人的Figure AI分道扬镳，你认为其内部项目如何与长期深耕该领域的公司竞争？顺便我必须提一下"挂钩机器人"事件。
威尔·奈特：什么是挂钩机器人？
凯莉·罗比森：你们没看过那个视频？天啊，希望听众知道我在说什么——简直骇人听闻，像场行为艺术。（解释后）总之OpenAI应该不会用挂钩固定机器人吧？
威尔·奈特：若研发硬件总需要防护措施。他们的优势在于算法——大语言模型已展现出对物理世界的惊人理解。但要真正实现肢体操控等未解难题，还需突破。

迈克尔·卡洛尔：竞争对手方面，除了Figure，还有Agility Robotics、Apptronik等知名企业。哪些值得关注？
威尔·奈特：现在有一批公司涉足该领域。Agility、Apptronik较知名，波士顿动力是先驱，Figure AI的演示很突出。当然不能漏掉特斯拉——他们通过电动汽车和制造业积累的物理世界AI经验，加上马斯克的专注，使其成为重要竞争者。中国公司优必选刚申请IPO，是全球最大人形机器人制造商。我见过他们的产品：成本低，能跳舞、打拳击、练武术，令人印象深刻。美国公司Apptronik也与谷歌合作。

迈克尔·卡洛尔：AI公司的终极目标似乎是让机器人在工厂、物流及家庭环境中替代人类。机器人公司也有相同商业目标。但OpenAI除AGI竞赛外的主要动机是什么？
威尔·奈特：值得注意的是，原本被寄予厚望的GPT-5表现未达预期。他们与其他企业一样在寻找AI新方向。构建对物理世界的理解（即"世界模型"）对AI至关重要，这对智能眼镜等消费硬件也意义重大。机器人与人类世界的交互将催生更精准的环境认知模型，带来商业应用潜力。

凯莉·罗比森：马斯克曾在内部会议中称特斯拉的人形机器人是xAI的最大优势——这是OpenAI所缺乏的底层能力。但炒作确实存在，就像当初的AGI热潮现在转向人形机器人。奥特曼称不久后人们就能在街头看到成群机器人，感觉像科幻成真。你报道称该行业现值50亿美元，到2050年达5万亿美元。如何看待这种炒作？
威尔·奈特：炒作确实存在。但与ChatGPT不同，机器人演示视频可能经过操控或百次尝试才成功一次。让机器人在任意环境中可靠工作极其困难——自动驾驶研发20年仍只能应对特定路况，而人形机器人需面对完全非结构化的人类世界，实际落地还很远。

迈克尔·卡洛尔：好的，我们先进入休息环节。欢迎回到"奇幻山谷"。今天讨论硅谷对人形机器人的重新关注。我们刚解析了OpenAI等企业为何认为机器人是实现AGI的关键。但多年来人形机器人也被寄予协助日常任务甚至替代工作的期望。我想知道这承诺是否仍有效？若创造能替代人类的机器人，我们离现实还有多远？
威尔·奈特：已有许多机器人从事人类工作。历史上机器人能力从高度特化逐步扩展：工业机器人精度高但零灵活性，近十年移动机器人开始适应简单环境，最近甚至有人形机器人测试。亚马逊投资Agility公司并在其设施中测试搬运箱子的机器人。亚马逊在电商仓储自动化方面有丰富经验，他们正探索人形机器人在人力短缺时的替代方案。但机器人始终面临非结构化环境的挑战，目前应用仍受限。

凯莉·罗比森：这些公司声称要解决劳动力短缺问题。但想到亚马逊的机器人，它们不需要上厕所也不会组织工会——这对资方倒是好事。制造业、运输和物流领域需要更多工人，人形机器人有望接管危险或不理想的工作。波士顿动力今年计划在现代工厂部署全电动人形机器人Atlas，其机器狗Spot和仓储机器人Stretch已投入工业应用。这是人形机器人首次进入商业制造场景。
迈克尔·卡洛尔：这些公司的宣传视频都展示机器人在工厂操作：安装保险杠、扫描箱内物品码放货架——听起来诡异但已是现实。
凯莉·罗比森："机器人眼睛"这名字挺酷。
迈克尔·卡洛尔：确实。许多人想象中的人形机器人是管家型，像Figure AI演示的叠衣服、浇花、喂狗等功能。但我认为家庭应用比工业环境遥远得多。威尔刚才提到的非结构化生活场景是主因。
威尔·奈特：没错。所有演示都需要问：为完成喂狗动作需要多少编程和预设？人类能进入陌生家庭找到狗碗并操作，而机器人极难做到这点。非结构化任务挑战巨大，观看演示时需注意其与人类能力的实际差距。因此制造业和电商才是更可能先落地的场景。

迈克尔·卡洛尔：传统机器人编程需逐步示教，而AI机器人需自主决策抓握力度、触感感知、倾倒角度等变量——这正是机器智能的用武之地。这些公司招募人才研发家用机器人的目标正在于此吧？
凯莉·罗比森：想到十年前AI还在识别猫狗，现在竟讨论喂动物，感觉很奇妙。
威尔·奈特：关键点在于，尽管这类任务困难，但我们开始看到通用模型的出现。未来可能出现物理世界的ChatGPT，在有限情境中将所学迁移至新任务。虽然离完全陌生厨房操作还很远，但已初见端倪，这令人兴奋。

凯莉·罗比森：所有技术都有挑战，尤其是大语言模型和机器人这类新兴领域。大语言模型有幻觉和谄媚问题，乐观者称人类也如此。但聚焦技术本身，人形机器人的主要挑战是什么？
迈克尔·卡洛尔：别杀人。
凯莉·罗比森：第一条：不伤害。
威尔·奈特：语言模型的幻觉若体现在物理世界，后果更严重。安全性、可靠性问题将浮现。开发通用模型并非必然成功——我们没有涵盖人类所有物理行为的数据库，这是根本挑战。硬件也与人类不同：我们尚未造出能精细操作的人类双手。特斯拉等公司的硬件创新值得关注。

凯莉·罗比森：能讨论机器人与军事应用吗？奥特曼曾对塔克·卡尔森说"不了解军方如何使用我们的模型"，被追问时表示"我没有机器人大军"。有消息源曾开玩笑说如果发现奥特曼在追梦中心藏着机器人大军会通知我。我们正走向未知的军事应用未来，这令人担忧。
迈克尔·奈特：我在语言模型前就报道过AI在国防领域的应用。这些模型因可靠性问题应用受限。危险在于炒作压力可能导致滥用，但目前系统在多数场景下仍不可靠。
凯莉·罗比森：谢天谢地。
迈克尔·卡洛尔：确实。感谢二位以人类最佳状态参与讨论。我们稍作休息后继续。感谢精彩对话，现在进入人类推荐环节。凯莉作为演播室嘉宾先请。
凯莉·罗比森：我沉迷播客《扶手椅专家》，达克斯·谢泼德主持的名人访谈有时也请专家。他们做过囤积癖、多维空间（我们活在模拟中）等诡异主题。最近常聊AI，作为听众多年的记者我常对着手机喊话。今早他们讨论机械可解释性时没说术语，只说"AI有我们无法破译的内部语言"，我急得大叫"说机械可解释性啊！"。推荐收听，当然要在听完我们节目之后。
迈克尔·卡洛尔：虽有缺陷仍值得听？
凯莉·罗比森：缺点很多，有时甚至让人听着来气。但旧金山AI泡沫之外的观点很有价值，所以仍推荐。
威尔·奈特：我推荐真正实用的家用机器人。我的猫常叼动物回家（包括活蹦乱跳的大兔子）。这款叫OnlyCat的智能猫门用计算机视觉检测宠物是否叼着老鼠之类，会显示"违禁品 detected"并拒绝开门。网站有叼青蛙的录像，很有趣。
迈克尔·卡洛尔：我推荐美利奴羊毛T恤。旧金山九月的实际夏天很热，这种面料透气排湿，野外活动时保持干爽。还有天然抗菌特性，露营时穿几天也不臭。虽然单件60-80美元较贵，但无需频繁洗涤。斯科特·吉尔伯森有季度选购指南，他最爱Proof牌72小时款，我穿Smartwool。这是种经典创新。
凯莉·罗比森：就像火人节推荐的美利奴羊毛袜突然合理了。
威尔·奈特：我正穿着Icebreaker牌美利奴T恤，确实舒服。
迈克尔·卡洛尔：绝非广告。感谢二位探讨机器人话题，我们这些被谄媚大语言模型驱动的冰冷机械心由衷感激。威尔谢谢参与。
凯莉·罗比森：很荣幸。
迈克尔·卡洛尔：感谢收听《奇幻山谷》。若喜欢请关注并评分。来信请至uncannyvalley@wired.com。本期节目由阿德里安娜·塔皮亚与马克·莱达制作，宏声之声的阿马尔·拉尔混音，马克·莱达任旧金山录音师，凯特·奥斯本任执行制片人，凯蒂·德拉蒙德任《连线》全球编辑总监，克里斯·班农任康泰纳仕全球音频主管。

英文来源：

This week, WIRED learned that OpenAI is ramping up its efforts in robotics—specifically, by hiring researchers who work on AI systems for humanoid robots. Humanoids, robots built to resemble us and perform daily tasks, were famous for their clumsiness just a few years ago. Senior writer Will Knight tells us about how that's rapidly changing on today's episode cohosted by Michael Calore and senior correspondent Kylie Robison.
Mentioned in this episode:
OpenAI Ramps Up Robotics Work in Race Toward AGI by Will Knight
Humanoid Robots Are Coming of Age by Will Knight
2025 Is the Year of the Humanoid Robot Factory Worker by Russell Brandom
You can follow Michael Calore on Bluesky at @snackfight, Will Knight on Bluesky at @willknight, and Kylie Robison on Bluesky at @kylierobison.com. Write to us at uncannyvalley@wired.com.
How to Listen
You can always listen to this week's podcast through the audio player on this page, but if you want to subscribe for free to get every episode, here's how:
If you're on an iPhone or iPad, open the app called Podcasts, or just tap this link. You can also download an app like Overcast or Pocket Casts and search for “uncanny valley.” We’re on Spotify too.
Transcript
Note: This is an automated transcript, which may contain errors.
Michael Calore: Hey, Kylie. How are you doing?
Kylie Robison: Good. How are you?
Michael Calore: Not bad. Good to see you here in the chair-
Kylie Robison: I know.
Michael Calore: ... that is normally occupied by Lauren Goode.
Kylie Robison: My Lord and Savior Lauren Goode, I'm so happy to fill in for her, but I miss her dearly.
Michael Calore: We all do. Also on the show this week, we have Will Knight, our AI expert at WIRED. Welcome back to the show, Will. How are you doing?
Will Knight: Hello. Good to be here. I'm doing well, thanks.
Michael Calore: Given the topic of today's episode, I want to ask you both. Do you have a favorite robot movie?
Kylie Robison: Yeah. Mine is The Iron Giant. I was just looking up when that came out because I swear I watched it on VHS as a kid. Yeah, I loved that movie. It's got a very sad ending. The robot is no longer. Oh.
Michael Calore: I guess we can probably spoil it, it's been decades.
Kylie Robison: Yeah, it came out in 1999. I think it'll be OK.
Michael Calore: Will, what is your favorite robot movie?
Will Knight: OK. I'm going to recommend Blade Runner, which is obviously a very famous movie and very beautiful. When I first saw it, I just thought it was just really cool. Then I later read the book and realized that it's ambiguous about whether some of the characters are robots or humans. I found it mind-blowing, this idea of blurring what it means to be human and asking whether you can build something that's human. Which I think we're a long way from, but it's an interesting idea.
Michael Calore: That's great. I'm going to go with RoboCop.
Kylie Robison: Nice.
Michael Calore: Which is the absolute opposite of The Iron Giant. But it's a great movie and it's timeless. It's a wonderful film. It's Paul Verhoeven going extremely hard in the '80s mode. And it's about the privatization of the police force and the corporatization of America's public services, it's very prescient.
Kylie Robison: Very classic Michael pick.
Michael Calore: Thanks. This is WIRED's Uncanny Valley, a show about the people, power, and influence of Silicon Valley. Today we're talking about how the field of robotics is shaping up to be the next frontier in the race towards artificial general intelligence, or AGI. Earlier this week, we learned that OpenAI is ramping up its efforts in robotics, specifically by hiring researchers who work on AI systems for humanoid robots. Humanoids, robots that were built to resemble us and perform daily tasks, were famous for their clumsiness just a few years ago. But as our colleague Will writes about, advancements in machine learning and hardware have renewed interest in humanoid technology, particularly within the AI industry. We'll dive into what companies like OpenAI are hoping to gain by investing in robots and what their decision means for how experts are thinking about what is really needed to achieve AGI, and what it might mean for us, the real humans, if that happens. I'm Michael Calore, director of consumer technology and culture.
Kylie Robison: I'm Kylie Robison, senior correspondent covering the business of AI.
Will Knight: And I'm Will Knight, senior writer at WIRED.
Michael Calore: Will, this is an obvious question, but we have to start here. What exactly are we talking about when we talk about humanoid robots? How are they different from other robots?
Will Knight: Yeah, I think it is a good question really. The obvious answer is that they're shaped like a human, so they have legs, arms, head like we do. But the bigger point really is, and the reason why the industry's so interested in this, is because they are designed to operate in the world that we live in which is designed for humans. Going upstairs, sitting inside vehicles, anything a human could do that those machines could do. When you're talking about trying to build something that has human-like intelligence, if that's how you define AI, then a big element of that is being able to do things in the real world. ChatGPT is getting very, very good at advanced math, but it can't make you a cup of coffee. It couldn't come into your home and make a cup. Or maybe it can, given what they're working on.
Michael Calore: Right.
Will Knight: But that's the idea.
Michael Calore: As you reported on, Will, and others at WIRED have talked about, humanoid robots do not have the most stellar reputation. They're seen as gimmicky. We've all seen videos of the Boston Dynamics robot that dances. That video on YouTube has over 40 million views. As far as we know, based on what the company's put out there, that's about all they're good at. At their worst, they're seen as unreliable at performing desired tasks. They fall over, they bump into each other. How is that perception changing right now?
Will Knight: Yeah. I remember going to the DARPA Robotics Challenge, back in I think it's 2012 when it first happened. This was a really interesting idea. After the Fukushima nuclear accident, they couldn't bring robots in to work in this really radioactive environment, so the idea was can we develop robots that could go into that human environment. But back then, these humanoids, they moved unbelievably slowly. They just kept falling over, often in really comedic ways. In the years since that, actually a lot of what's happened is really the hardware has gotten better and better. And some key things, like motors. In order to move like a human, we think of robots as being more capable than us, but you need to be able to move very, very quickly and very explosively in order to do things like balance. That's what Boston Dynamics pioneered. That kind of hardware has gotten a lot more available, which is why you're starting to see more humanoids come on the market.
Michael Calore: Now we know that OpenAI is very interested in humanoid robots. What did you recently learn from your reporting about the company?
Will Knight: I've been following a lot of different labs that are working on robots. Humanoids is a really big thing because it allows you to explore this human physical element of intelligence. I discovered firstly that they were hiring people to work on humanoids. Then also, if you look at the job listings they have and other signals, you can see this picture where they're clearly ramping up their robotics work. Which I think makes perfect sense, given that a lot of other labs are doing that. A lot of other companies working on AI seem to be looking towards this physical side of intelligence now.
Kylie Robison: Yeah. When I read the piece, my first thought was Yann LeCun's position where we're not going to reach AGI or AI that's smarter than all of us without that physical intelligence bit, without it learning in the real world, like any other carbon-based life form. It's interesting that OpenAI is picking up their robotics work once again, after they shuttered in back in 2021. You product that they were doing pretty solid work on robotics, doing things like developing an algorithm capable of solving a Rubik's cube using a human-like hand. Now it seems like OpenAI is not just returning to robotics, but bringing that AI focus to it, exactly what LeCun wanted. Back in February, OpenAI parted ways with Figure AI, a startup focused on creating humanoid robots. How do you think internal OpenAI initiatives can be competitive in comparison to companies that have been consistently developing humanoid robots? The whole time I'm thinking about the robot on meat hooks. I would be remiss if I did not mention the robot on meat hooks.
Will Knight: What's the robot on meat hooks?
Kylie Robison: The humanoid robot on meat hooks? I'm so shocked that you haven't seen this. Have you seen this, Michael?
Michael Calore: No.
Kylie Robison: Oh, my God. Do I just have the most cursed algorithm on planet Earth? I really hope that the listeners know what I'm talking about, and/or should go search it up because it's the most horrifying thing. It almost feels like a gimmick. Will, do you know what this is? Did you just look it up?
Will Knight: They'll often have a humanoid on a chain so it doesn't fall over and it goes berserk.
Kylie Robison: Yeah, yeah. And it loses it.
Will Knight: Is that the one?
Kylie Robison: Yes.
Will Knight: OK, yeah.
Kylie Robison: And then it breaks because there was some problem in the algorithm. Yeah, it goes berserk. It's horrifying, but probably a good marketing stunt. But I imagine OpenAI's not going to have humanoids on meat hooks.
Will Knight: If they're developing the hardware, they probably will have them on something.
Kylie Robison: Incredible.
Will Knight: So they don't break themselves if they fall over. They are going to be competing with a lot of companies that are already putting videos of robots out there and stuff. The advantages that they have is in the algorithms already. Large language models, one of the remarkable things about them, is they have a surprising amount of understanding of the physical world. But to get to the next step, whether they're developing hardware or just doing the algorithms, they're going to have to work on these systems that are better able to really understand how to operate in the real world. So how to move limbs and do things like manipulation, which is one of the really big unsolved challenges.
Michael Calore: Yeah. Some of the big competitors, we should talk about. Kylie, you just mentioned Figure. There's also Agility Robotics, Apptronik, some household names. Who are the ones that we should think about?
Will Knight: Yeah, there are a slew of companies doing humanoids now. You'll see these videos in your social media. Agility is one of the better known ones, Apptronik. Boston Dynamics of course has pioneered this. Figure AI is very prominent with some of its demonstrations or videos. Of course, Tesla. Be remiss not to mention that. I think Tesla is going to be, not just because of Elon, but because they are actually very steeped in developing AI for the physical world through their cars, and also through the manufacturing that they do, clearly Elon seems to be really focused on that. Now I think they will be a really prominent competitor. We should also mention Unitree, which is this Chinese company which is now the biggest manufacturer of humanoid robots. I think they just filed for their IPO. I got to see a bunch of their robots. They're really low cost. They're not super sophisticated I think, but they're very low cost and very able to do things like dancing. I went to a conference in China where they had them doing boxing matches, punching each other and doing kung-fu, which was quite impressive. But yeah, there is a growing number of companies. I think going is also doing some stuff with Apptronik, one of the US firms.
Michael Calore: It feels like the end game for the AI companies is to create humanoid robots that can function in place of a human in factory environments, in shipping and receiving, and around the home doing helpful tasks. The robotics companies have the same goals as well. They want to be able to sell a robot that can do this. But specifically with OpenAI, what are their primary motivations here outside of just the race for AGI?
Will Knight: Well, I think one of the big things to mention is GPT-5, the latest model which was meant to be a huge leap forward, kind of flopped. I think they, like others, are looking for the next big thing, the next direction in AI. Developing a better understanding of the physical world would mean building world models, so-called world models, is going to be a very important thing for AI. I think that this could be very important for consumer hardware, like smart glasses or whatever hardware they're working on, because robots, by navigating the physical world, especially if they have digits like us, will interact with the same world that we do. You do have the potential to build models that are going to have a much better of the understanding of the world you're walking through. I think there could be a bunch of commercial applications there as well.
Kylie Robison: Yeah. You previously brought up Elon Musk. Something I'd heard in my reporting and I'd reported a portion of this in my newsletter is during a recent all-hands with X and xAI people, which are the companies he also owns, he said that Tesla and the humanoid robot from there, he said that's their biggest advantage at xAI is having those robotics in the background. Which is not something OpenAI has, they're building this from the ground up which is notoriously hard. But the hype is very real, just like with the AGI hype. Now we've moved on to some sort of humanoid robot hype. Sam Altman, the CEO of OpenAI, recently told Bloomberg that people would soon be walking down the street and see seven robots walk past you doing this, or whatever. It's going to feel very sci-fi. He said that moment isn't very far away. You reported in your piece that it's a $5 billion industry right now and could be worth five-trillion by 2050. What do you think of this hype? Is it warranted?
Will Knight: Yeah, I think it is really hyped. I think one of the problematic things with robots, you can compare to ChatGPT and large language models, is you can do these demo videos that are really stunning and people naturally think, "Oh, the robot can do incredible things," if you see it making a drink or loading the dishwasher. But the thing is that often in those situations, it might be teleoperated. It might only work once time out of 100. That is the reality right now. Getting these things to be super reliable in any environment I think is an enormous, enormous thing. Just look at self-driving cars. We're 20 years in and they're just operating in some situations now, some roads. We're talking about something that's supposed to go into completely unstructured human world. I think it's a lot further off.
Michael Calore: OK, that's a good place to take a break. We'll come right back. Welcome back to Uncanny Valley. Today we're talking about Silicon Valley's renewed interest in humanoid robots. We just broke down why some AI players, like OpenAI and its competitors, are betting that these robots could be the key to achieving AGI. But human robots have also been billed for years and years as being the future aids that would help us in our daily tasks, and in some cases take over our jobs for us. I want to know if that's still the promise. If we are creating humanoid robots that can do the tasks of humans, how much further do we have to go for that to become actual reality?
Will Knight: There are a ton of robots that do work that humans do. Through the history of robotics, it's been the case of expanding what they do from incredibly niche things. Manufacturing robots do these very, very precise things, but have zero flexibility. We started to see in recent decades more robots expand what they can do. You see mobile robots that can navigate very simple environments. In the last few years, we've started to see even some humanoids being tested. Amazon has invested in this company, Agility, which has a humanoid. They've been testing it in some of their facilities, having it move boxes around. For a company like Amazon, they have a really, really strong history of deploying robots in their ecommerce warehouses in really valuable and useful ways. What they're doing with these humanoids I think is exploring how they can slot in for humans in some situations. They have a very cyclical need for workers and sometimes code a humanoid, fit it and do just a simple thing like moving boxes from A to B. But always with robots, the case is how much unstructured, unfamiliar stuff can you get those to do? It's still limited, so you have to narrow it down.
Kylie Robison: Yeah. You already mentioned Amazon. Some of these companies say they want to address labor shortages in industries where more workers are needed. My first thought as we're talking about robots at Amazon, they don't have to pee in cups and unionize.
Michael Calore: Yeah.
Kylie Robison: That's very nice for them I suppose.
Michael Calore: Very good point.
Kylie Robison: They need more workers in manufacturing, shipping, and logistics. The hope is that these humanoids will take over the more dangerous, less desirable jobs, too. Later this year, Boston Dynamics plans to put its all-electric humanoid robot Atlas to work in a Hyundai factory. The company already has a dog-like robot, Spot, and a warehouse robot, Stretch. What a funny name.
Michael Calore: Nope.
Kylie Robison: Which is already deployed in industrial sites. The Hyundai pilot will be the first time their humanoid is used in a commercial manufacturing setting.
Michael Calore: Yeah. If you watch the videos and the promotional materials, go to the websites of these companies, all of them show robots on the factory floor.
Kylie Robison: Yeah.
Michael Calore: And they show them doing things like putting the bumper cover onto the bumper.
Kylie Robison: Yeah.
Michael Calore: Or taking boxes off of a palette and organizing it onto shelf space based on what's inside the box.
Kylie Robison: Yeah.
Michael Calore: Because they can scan it with their robot eyes. Which sounds creepy, but that's also how the world works now.
Kylie Robison: Cool band name, Robot Eyes.
Michael Calore: Indeed. The thing that I think a lot of people think about when they think about humanoid robots is the robot butler. You do see that imagery in some of the promotional materials. Like Figure AI, who we talked about, they have a video of their robot folding the laundry, watering the plants, putting the dog food into the dog food bowl. I think about a future like that where we have robots in our home as a lot further off from the manufacturing and the warehouse environments. Will, I think you've hinted at this by talking about the unstructured reality of life that we live in.
Will Knight: That's right. In any of those demos, you have to ask how much programming, how much preparation did they have to do to have it fill the dog food bowl? The thing that humans do is you go into an unfamiliar home, and can you feed the dog? You have to figure out where the bowl is, and you have to figure out how to pick it up, which is something that robots just struggle with. Anything that's unstructured like that, it's going to be much more challenging. I think it is really important to bear in mind when you're seeing those demos that the implication is that it can do the sort of thing that a human does, but it's a lot more complicated to do that in the real world. This is why things like manufacturing and ecommerce are going to be much more the environment that you'll see them come into first.
Michael Calore: I think when you talk about how to program a robot, the old way of programming it would be you show it where the bowl is, you show it how to pick up the bowl, you show it where the dog food is, and you show it where the bowl goes after it's full. But in the world of AI robotic programming, then there all of those little micro-decisions being made about grip and sensing touch, and the angle that you're supposed to hold the box at so that the dog food can flow out of it. All of those things are variables that the robot is going to have to figure out on its own, which is probably where the machine intelligence aspect of it comes in. It seems to me, when these companies are talking about, "Yeah, we're hiring these people to build these things out and turn them into these machines that we can use in our homes," those are the sorts of targets that they're looking at, right?
Kylie Robison: I am thinking about just about a decade ago, AI was figuring out whether a dog was a dog and a cat was a cat, and now we're at this point where can it even feed these animals feels strange.
Will Knight: That's such a good point. The thing that is interesting right now is that, as difficult as doing a task like that is, we're starting to see these models which are more general. The idea is you're going to have something that is eventually something like ChatGPT for language with the physical world, it'll be so general. You are starting to see these models that in very limited situations can take what they've learned to do one thing, and given a new unfamiliar task, do it in a slight more reliable way. This is one of the things that's creating that excitement. That is still a long way from figuring out a completely unfamiliar kitchen. But we're starting to see the beginnings of that, which is really exciting I think.
Kylie Robison: Yeah. All technologies have their own challenges, especially the nascent ones like large language models and robots. For large language models, there's hallucinations, sycophancy. I think the optimistic people that I talk to would argue humans do that, too. But sticking with the technology, what are the challenges for humanoids? What are you keeping your eye on?
Michael Calore: Don't kill people.
Kylie Robison: Right. Number one.
Michael Calore: Do no harm.
Kylie Robison: We could do a whole segment on that.
Will Knight: Yeah. Well, speaking of that, I've written some stories about how, if you take something like hallucinations in language models and then extrapolate to a model operating in the physical world, there's a lot more potential for it to go wrong. I think these things, there are going to be a little bit of security issues, reliability issues. Generally, developing the models that are going to be more general is not a sure thing. They don't know if you can get enough data to do that in a general way because we don't have an internet's worth of all physical actions that humans do in the real world. That is going to be a real challenge. I think also, it's really important to remember that the hardware is fundamentally different. We certainly don't have hands that are capable of as fine control as humans. The idea that you'll be able to do everything ... Some physical things you'll be able to do way better and that's always been the case. But there are certain things that the hardware can't do as well. It is interesting to see companies try to innovate there. Tesla's doing some interesting stuff, as are others.
Kylie Robison: Is it too spicy to talk about robots and warfare? In a recent podcast with Tucker Carlson, Sam Altman said, "I don't really understand how the military is using our models." Tucker really pressed. It's crazy. He's like, "Well, I don't have an army of humanoids." One time I was doing source reach-outs and someone told me, "I can't talk, but you're the first on my list if I find out Sam Altman has an army of humanoid robots at the bottom of Chase Center." Seriously, we are going toward this future, and we don't even understand how it's going to be used in the military. It feels far off, but it's concerning.
Michael Calore: Yeah.
Will Knight: I've done some reporting about AI pre-language models as well as current stuff being used in the defense world. I think one of the things is that actually, those models have to be used in very limited ways because they're so unreliable. The worry I guess is that people want to rush so much and you have this pressure, this hype, that you start to bend what you're willing to do. By and large, those systems are too unreliable to be used in a lot of contexts really.
Kylie Robison: Thank God.
Michael Calore: Yeah, thank God. All right. Well, thank you both for bringing your best human selves to this conversation. We're going to take another break and we'll come right back. Thank you both for a great conversation. It is now time to go to our very human recommendations. Kylie, you're our guest in the chair in the studio, so you get to go first.
Kylie Robison: Yay.
Michael Calore: What's your recommendation?
Kylie Robison: I am addicted to a podcast called Armchair Expert with Dax Shepherd. They interview celebrities and sometimes experts in fields. They did one on hoarding, which was really interesting.
Michael Calore: Whoa.
Kylie Robison: One on multidimensional; we live in a simulation.
Michael Calore: Nope.
Kylie Robison: Very, very creepy stuff. They don't talk about tech regularly. They are doing celebrity interviews. But recently they've been talking about AI a lot more. As a reporter who’s listened to this podcast for years, I'm just screaming at my phone. At one point, they were trying to remember the name of Yoshua Bengio. I'm washing my dishes going, "Yoshua Bengio! Yoshua Bengio!" Then this morning when I was listening as I was getting ready, they were talking about mechanistic interpretability, but they didn't say that. They were saying, "Didn't you hear that AI models have this own language that we can't decipher and they can talk to each other?" I'm once again screaming at my phone. I'm like, "Oh, my God, talk about mechanistic interpretability. Who would care about that but me?" I love that podcast and highly recommend listening, after you listen to Uncanny Valley of course.
Michael Calore: Even with all of its flaws, it's still a good listen?
Kylie Robison: It has many flaws, many flaws. I think sometimes you hate listen to it. I don't know how to describe you just violently disagree with some of Dax's points. Otherwise, I don't think I would hear those points. I'm in San Francisco. I'm in the AI bubble, and it's an interesting point of view. But yes, with all of its flaws, it'd still recommend.
Michael Calore: Great. Yeah, it sounds like you're learning a lot, too.
Kylie Robison: Yeah, that's true. I didn't know about the simulation theory, how deep it went and who subscribes to it. Sam Altman in Keach Hagey's book, he says that he sort of believes in it. I was like, "Oh, how nice, the guy creating the simulation believes he's God."
Michael Calore: Yeah. That's a whole other topic.
Kylie Robison: Yes, it is. Will, what do you recommend?
Will Knight: I'm going to recommend what I think is a genuinely useful home robot. I have a cat, Leono. He's very nice but will bring in various animals, dead or alive, into our house overnight. Often including large rabbits that then run around.
Kylie Robison: Oh my God.
Will Knight: This cat flap has a camera and computer vision, and it can tell if your cat has a rat or something in it's mouth. It will say, "Contraband detected," and then block them from coming in, which I think is a good thing.
Michael Calore: When it sends you a notification that contraband is detected, do you get a photo of what your cat has in its mouth?
Will Knight: Yes. I think you get a video feed of it trying to bring in. There's some on the website, which are great. There's one that has a very large frog in its mouth.
Kylie Robison: You got to put a wanted poster with these.
Will Knight: Yes.
Kylie Robison: "Wanted: cat with frog in mouth."
Will Knight: Yes. I should say it's called OnlyCat. That's the name of it. I'm actually genuinely curious how she brings these rabbits into the house.
Kylie Robison: Me, too.
Will Knight: I need to put a camera down there, because they're enormous. I feel like she has to go through, go back and get it, and then pull it. It must be.
Kylie Robison: Wow. Free rabbit.
Michael Calore: She's really working hard.
Kylie Robison: Paying her rent.
Will Knight: Yeah. Yeah. Over to you, Mike.
Michael Calore: OK. I would like to recommend merino wool T-shirts.
Kylie Robison: Why? I would imagine merino wool is quite sweaty, actually. I don't know why I think that.
Michael Calore: No. It's the opposite.
Kylie Robison: I see.
Michael Calore: We're in mid-September here in San Francisco, which is also known as actual summer.
Kylie Robison: Yeah.
Michael Calore: It's very hot, so I've been thinking a lot about thermal regulation. I'm also doing a lot of activities, like hiking and canoeing and doing things out in the wilderness and getting sweaty. The merino wool T-shirt has been really great. A bunch of different people make them. My colleague Scott Gilbertson, who works on the WIRED reviews team, has been testing a bunch and he's written about them. He has a buying guide that gets continuously updated once a season with all the best picks. His favorite is the Proof 72-hour merino wool T-shirt. Mine is the one that Smartwool makes. Merino wool is very, very good at wicking moisture away from your body so it actually keeps you cooler. It keeps you from really feeling sweaty, which is nice. Even if you're not particularly sweating hard, it'll just make you feel fresher. And also, it has natural antimicrobial properties.
Kylie Robison: Whoa.
Michael Calore: It doesn't stink after a couple of days, which is really important if you're camping.
Kylie Robison: Right.
Michael Calore: I've been wearing them on runs. I've been wearing synthetic blends a lot, and I'm over those, because those are starting to fall apart. I'm investing in merino wool. They're quite expensive. They can cost 60, 70, $80 for a good shirt. But you'll only need a couple-
Kylie Robison: Yeah.
Michael Calore: ... because you don't need to wash them as often. Yeah, they're great. It's an old-school innovation.
Kylie Robison: Yeah. It reminds me that merino wool socks are recommended for Burning Man. That makes a lot more sense now.
Michael Calore: Is this where you tell us that you went to Burning Man this year?
Kylie Robison: Oh, God. No, no, no, I never. I don't know what you're talking about.
Will Knight: I'm actually wearing a merino wool T-shirt right now.
Kylie Robison: Whoa.
Michael Calore: Are you?
Will Knight: I can endorse. I've got an Icebreaker one-
Kylie Robison: #notsponsored.
Will Knight: ... which is very nice.
Michael Calore: Yeah. Yeah, this is not spon con. We're living the real life here. All right. Well, thank you both for being here and talking about robots. We appreciate you very much from the bottoms of our cold mechanical hearts powered by sycophantic LLM models. Will, thank you for coming on the show.
Will Knight: Yeah. Thank you for having me. It's been great.
Michael Calore: Kylie, thanks for being here.
Kylie Robison: Of course, happy to.
Michael Calore: Thank you for listening to Uncanny Valley. If you liked what you heard today, make sure to follow our show and rate it on your podcast app of choice. If you'd like to get in touch with us with any questions, comments, or show suggestions, you can write to us at uncannyvalley@WIRED.com. Today's show was produced by Adriana Tapia and Mark Lyda. Amar Lal at Macro Sound mixed this episode. Mark Lyda is our San Francisco studio engineer. Kate Osborn is our executive producer. Katie Drummond is WIRED's global editorial director. Chris Bannon is Condé Nast's head of global audio.