Latest / Minds, Bodies, and Terawatts / Tesla Put Optimus Into Production on a Line Built for a Million Robots a Year. Musk Says the Hands Aren't Finished, the Brain Isn't Ready, and Customers Wait Until Late 2027

Transcript

Tesla Put Optimus Into Production on a Line Built for a Million Robots a Year. Musk Says the Hands Aren't Finished, the Brain Isn't Ready, and Customers Wait Until Late 2027 · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, the show where the bodies went into production before the minds were ready. The Verge reported on Thursday, October twenty-second, that on Tesla's earnings call the night before, Elon Musk announced his humanoid robot, Optimus, has started early production on the high-volume line in Fremont, California.
• 0:17
  [Guest] And then he spent the rest of the segment listing what doesn't work yet. The hands, in his words, are not perfected. The robot, as he put it, can't wear out after a month, and the AI still needs quite a lot of work before the robot is useful for anything open-ended.
• 0:35
  [Host] We covered this in September, when The Information reported Tesla was already turning out hundreds of these a week, with every hand put together by a person. So what actually changed on Wednesday?
• 0:46
  [Guest] The company's own label. Tesla's shareholder letter calls these the first production-quality robots it has ever built, which means every Optimus before this quarter wasn't. The line is now official, and the robot is still a draft.
• 0:59
  [Host] What does early production even mean when the hands aren't finished?
• 1:03
  [Guest] It means the factory is further along than the product. Musk said most of the equipment for a line built to make a million robots a year is already sitting in Fremont. Reaching that rate, about twenty thousand a week, will take at least a year, maybe longer, because parts of the design are still changing.
• 1:22
  [Host] How big is that next to what exists today?
• 1:25
  [Guest] The research firm IDC counted about twenty-five thousand humanoids shipped in the entire world in the first half of this year. Tesla is describing one line doing close to six months of global output every week.
• 1:38
  [Host] So who gets one, and when?
• 1:40
  [Guest] Nobody outside Tesla until around the end of next year. Musk said the first customers will be industrial, for work that's dangerous, repetitive, or has to run around the clock. If you work nights at a plant, he named your shift on an investor call.
• 1:55
  [Host] Then it's my shift he's coming for. How worried should I be about the end of next year?
• 2:01
  [Guest] Less than the headline suggests, because of everything else he said. Tesla didn't disclose how many robots it has built, what they can do on their own, or how often they finish a task without a person stepping in. Ken Goldberg, a Berkeley robotics professor who has spent his career getting robots to handle objects, put it flatly: you can produce them, but the software is lacking.
• 2:24
  [Host] Hold on, a month ago the bottleneck was the hands. Now it's the software too?
• 2:29
  [Guest] It's both, and this time Musk said the second part himself. A Tesla car learns from every Tesla on the road, and Optimus has no fleet, so the plan is to build ten to twenty thousand robots and have them practice tasks on themselves, in what the company calls an Optimus Academy. He compared it to a toddler learning to get around.
• 2:50
  [Host] The plan used to be people in recording suits showing it the moves. What happened to that?
• 2:56
  [Guest] That's the piece that moved. Recorded humans were supposed to be the teacher, and Musk is now saying the robot has to teach itself. The body is in production, and the mind is on back-order.
• 3:08
  [Host] Does a robot practicing on itself actually work?
• 3:11
  [Guest] Goldberg calls it a hypothesis. Practicing against yourself works in software, where a mistake costs nothing, and he told The Verge that in hardware it's not so clear. He also said task success rates in robotics labs have stalled around seventy to eighty percent, which no factory would accept.
• 3:28
  [Host] Musk promised a video soon of the robot doing tasks it wasn't programmed for. What should I be watching for?
• 3:35
  [Guest] Goldberg has a name for the genre, video emptor, buyer beware of the clip. A video can show the one take that worked, hide a person at the controls, or depend on a table arranged just so. The question is how many takes, and who's off camera.
• 3:51
  [Host] Musk also told investors nobody is even close to Tesla on this. Is that true?
• 3:56
  [Guest] Not according to Electrek, the EV news site. Hyundai's Boston Dynamics has been building its Atlas robot since the start of the year, says this year's deployments are all committed, and has it learning to line up car parts at a Hyundai plant in Georgia. And it gave Atlas sturdy grippers instead of human-like hands, which looks worse in a demo and gets around the exact problem Musk says is holding him back.
• 4:21
  [Host] So where does that leave the worry about who owns what the robot learned from people?
• 4:26
  [Guest] It got stranger. The robot didn't learn enough from people, so the next year of lessons comes from robots practicing on each other, inside Tesla, on a line that can make them faster than anyone can teach them. The factory finished first.
• 4:41
  [Host] So the threshold that matters was never the factory door. In Unscarcity, Patrick Deglon sets it at the moment a machine becomes the cheapest reliable way to do a job, his Substitution Threshold, and reliable is the one word Tesla's own CEO wouldn't claim this week.
• 4:57
  [Guest] For more on what happens to a night shift once a robot clears that bar, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.