
Figure AI はなぜ中国・Tesla に勝てるのか:Brett Adcock とロボティクス徹底討論
How Figure AI Is Outcompeting China and Tesla in Robotics | Robotic Deep-Dive w/ Brett Adcock

Figure AI はなぜ中国・Tesla に勝てるのか:Brett Adcock とロボティクス徹底討論
Moonshots with Peter Diamandis
要約
Figure AI の創業者 Brett Adcock が、ヒューマノイド競争の現状と同社の戦略を語った回。家庭向け出荷は「Apple 級の品質」に達するまで行わない方針で、Helix 3 や Figure 4 の手の進化、計算資源の急拡大、垂直統合の必要性を説明した。中国勢の動画の多くは実用性が低いとの見方を示し、姉妹会社 Hark や安全性・規制・半導体不足にも話が及んだ。
この回のポイント
- ●Adcock は家庭向けヒューマノイドの出荷条件を「遠隔操作なしで一日十数時間、毎日働ける高品質な製品」とし、最大の制約は知能だと述べている。
- ●競争を4段階(ハードウェア、AIレシピ、知能のスケール、ロボットのスケール)で整理し、現在は第3段階のデータと計算資源の拡大に注力している。
- ●中国勢のダンス動画はモーションキャプチャを使ったRL制御による開環の再生が中心で、汎用ロボティクスとは別物だと批判した。
- ●Helix 2.5 でデータ量と性能の対数線形関係(ヒューマノイド版スケーリング則)を確認し、未知の家30軒でゼロショット作業ができたと説明した。
- ●計算資源は年間9桁ドル規模で、NScale と3億ドル超...ではなく3十億ドル超・10万GPUの契約を結び、2027年後半に稼働予定。世界的に計算資源とメモリが逼迫していると語った。
- ●Figure 4 は最大の世代間ジャンプで、ロボット自身による『ライトアウト製造』を見据えて設計され、手は人と見分けがつかない水準になると主張した。
章立て
キャンパスとロボット1000台超
キャンパスにはロボットが従業員数を上回り、来訪者対応ボットが案内や音声対話を行う。全ロボットが音声のリアルタイム対話機能を備える。
家庭向け出荷の時期と条件
商用と家庭の二本立てで開発し、家庭は遠隔操作なしの高品質が出荷条件。未知の家でのゼロショット作業が当面の目標。
4段階の戦略とHelixの歴史
ハードウェア、AIレシピ、知能、ロボットの規模拡大という4段階を説明。Keurig のカップ事例からニューラルネット移行の経緯を語る。
中国勢と他社モデルへの見方
中国のダンス動画や大規模言語モデルによるロボット操作動画は、実用的なロボティクスとは異なると批判した。
計算資源とオンボード推論
学習用計算資源の拡大計画、推論は全てロボット上で行う理由、クラウドとの役割分担を説明した。
価格、Figure 4、自社開発
月額リース構想、Figure 4 の手の進化、アクチュエータを自社で作るに至った経緯を語った。
安全性・ライトアウト製造・規制
安全設計やサイバーセキュリティ、ロボットが自らロボットを作る未来、需要規模、規制の動向に触れた。
チップ不足と Hark、起業論
計算資源とメモリの枯渇、姉妹会社 Hark の成長、AI デバイス構想、シリコンバレーの優位性を語った。
解説記事
ヒューマノイドロボット企業 Figure AI の創業者兼CEO、Brett Adcock が Peter Diamandis らのインタビューに応じ、家庭向け出荷の時期、中国・Tesla との競争、計算資源やチップ不足、姉妹会社 Hark について語った。以下は、発言に基づく要点の整理である。
家庭向けは「Apple 級の品質」が出荷条件
Adcock は事業を商用と家庭の二本立てで進めていると説明した。商用は仕事内容が明確な反面、速度と性能の要求が厳しい。家庭は速度要求が低い代わりに、家ごとに環境もタスクも異なる。
家庭向けの出荷条件は、遠隔操作に頼らず、ロボットが迷って止まることもなく、品質の高い製品として入れることだという。Humane や Rabbit といった AI デバイスの失敗を引き合いに、不十分な状態で出せば機能しないと述べた。最大の制約は知能で、統合や安全性も課題としつつ、現在はカリフォルニアで20〜100軒規模の新規の家でテストを続けている。見通しとして、今後6カ月未満で任意の家に置き、数時間に及ぶ長いタスクをこなせるようになることを期待すると語った(約束はできないと断っている)。2027年末までには全体像が見えるはずだとも述べた。
4段階の戦略と「ニューラルネット化」
Adcock は競争を4段階に整理する。(1)実用的なハードウェア、(2)長時間の全身作業をこなせる AI レシピ、(3)データと計算資源による知能のスケール、(4)ロボットの量産、である。現在は第3段階にあたり、「データ、データ、データ」だと説明した。
第2段階の転機として、Figure 1 が Keurig のカップをコーヒーメーカーに入れるデモを挙げた。上半身を単一のニューラルネットで動かすもので、これが成功すれば約5億ドルを調達できると付箋に書いたところ、7億ドル弱を調達できたという。その後約1年かけ、上半身のHelix 1から全身ニューラルネットのHelix 2へ進んだ。足の置き方を含め、関節全体の軌道を同時に学習させる必要があると分かったと述べている。
中国勢・他社モデルへの批判
中国のロボット動画について、Adcock はモーションキャプチャの動きを強化学習(RL)で再現する手法が多く、環境に反応しない開環制御だと指摘する。見栄えはするが汎用ロボティクスとは「180度逆」の作業で、実用性は低く、大学4年生でも1台あればできるレベルだと述べた。最も近い競合は「集合体としての中国」だと認めつつ、ハードウェアやモデルはまだ良くないが年々進歩は速いとの見方を示した。
また、汎用の言語モデルでロボットを動かす動画は、固定カメラで簡単な作業を遅く行い、動画を5〜20倍速にしていると主張した。Figure は自前の AI(Helix)を一から作ったとする。
計算資源、オンボード推論、垂直統合
同社の計算支出は年間9桁ドル(低い9桁)で、Helix 3 の事前学習に半分以上が使われているという。NScale と10万GPU・30億ドル超の契約を結び、2027年後半に Vera Rubin を導入予定。現状比で100倍超になると語った。来夏には Helix で10億ドル規模の学習を行う見込みも示した。
推論は全てロボット上で行い、クラウドは音声対話と多数のロボットを束ねるエージェント管理に使う。5G の往復でも約50ミリ秒かかり、ミリ秒単位の制御には使えないためだ。さらに、市販アクチュエータは大きすぎ、重く、高価だったため、自社開発に踏み切った経緯を説明した。Adcock は「全てを垂直統合する必要がある」と強調する。
Figure 4 と将来像
現状を「ヒューマノイドのフィーチャーフォン時代」と形容し、Figure 4 を iPhone 初代に例えた。手は約2年・7世代の改良を重ね、手袋をした Figure 4 と人間の手はほぼ見分けがつかないと主張した。世代間では最大のジャンプになるとし、ロボット自身が組み立てる「ライトアウト製造」を想定した設計だという。5年以内には実現するが、時期は半年〜3年の幅で不確かだと述べた。量産は現工場で年4万台超、うまくいけば将来は年100万台規模の可能性を挙げた。家庭用は月300〜500ドルのリースを想定している。
安全面では、物理的安全と、ろうそくを倒さないといった意味的安全の両方が必要とし、サイバーセキュリティの専用チームがあると述べた。防衛分野は手がけない方針を維持している。
Hark とチップ不足
Adcock は Hark を、個人向けインテリジェンスを作るデジタル側の会社と位置づけた。Hark Pro は発売から約1カ月で120万件超の登録があり、売上は既に Figure を上回っている、3カ月〜6カ月以内に Hark のほうが大きくなる可能性があると述べた。
一方で、世界は計算資源とメモリが枯渇していると語る。入札に応募が集まらず、2026年9月に「世界が計算資源を使い切った」と感じたという。チップの自社開発には関心があるが、設計に3年以上かかる点がリスクだとした。
まとめ
この回は Adcock 自身の主張が中心で、競合への評価や時期の見通しは本人の見方である点に注意したい。それでも、家庭向けは知能の問題だとする認識、データとスケーリング則への賭け、垂直統合と計算資源の確保への執念は、ヒューマノイド開発の競争軸を示している。日本の読者にとっては、ハードウェアの精密さだけでなく、データ収集、計算資源、サプライチェーンを含む総合力が勝負になるという示唆が重要だろう。
文字起こし(英語・自動生成)
As you think about competition, is it Tesla and China? Is that the way you see it in the world right now? I personally haven't seen it much out of Tesla in two or three years. I've seen China do a lot of dancing. We care about something very different than what we think China cares about. We just gotta go beat them. It's a real race here. The work isn't really practical. The technology stack needed to get there isn't practical. And the AI stack to do it can be done by a senior in college. My views are there's like roughly four stages for this humanoid race. My chapter one of this book is you have to build, like, really good functional hardware. Two is you have to be able to build an AI recipe. Chapter three and four for me are, like, scaling intelligence and scaling robots. What might we see with figure four? What areas are you focusing on making improvements? We're in the flip phone era for humanoids. Figure four is probably the iPhone one moment we're going to have here soon. We need the robot to get cheaper. We're trying to make the robot more capable. A little sneak peek at that is we've been working on the hands. A figure four with a glove and a human with a glove is almost indistinguishable. That's like an iPhone moment. Figure 3 is probably the best humanoid in the market by far, but it'll make Figure 3 look like a caveman.
Now that's a moonshot, ladies and gentlemen. This episode is brought to you by the Abundant Summit and Link Ventures. So, Brett, good to be here for our annual catch-up. It's been about a year since Dave and I were here to see Figure, and quite the difference. So walk in, just park in the parking lot over there, walk in, and there's like 100 robots out to greet you. So my question is, what do you call a group of robots? I'm calling them a flock of figures. Yeah. Those are our visitor bots. They walk around all campus. They can open doors, walk on any stairs, use the elevators, and they're there to greet people, talk. And then we're going to have, like, we're right now working on trying to get them to, like, get coffee for folks. Nice. We have kind of a stretch goal. It would be great if you could come to campus and never talk to a human as you get set up. Get fully checked in, walk to the conference room, do a tour, go to the bathroom, all done through a humanoid. Let me show you where I was built. Let me show you what I'm doing.
Yeah, I know. Exactly. We do that all the time. So why not? We should have a humanoid. And you said something a minute ago that's pretty profound. You have more figure robots than humans now. Yeah, we have more figure robots on campus. When did that happen? probably like four or five months ago we uh crossed the number of i think we have like seven eight hundred people here so we crossed the number of uh yeah we have more robots now we have many more now than humans and you're up to how many produced at this point i don't know like uh 1500 or so on campus something like that in order of like uh we have well over a thousand on campus and we have robots out in other places in the world as well oh nice so we're seeing figure three around here. Last time we were here, you were just transitioning to figure three, I think, at that point. How long before voice, how long before I'm talking to the figure? You can talk to any robot here. Yeah? Yeah. You can stop it and just say, ask a questions. All right. I'll have to do that. It's got real-time speech to speech. Yeah. Yeah. So every robot here is equipped with speaker mic. We actually use Hark for our voice AI. And every robot here has got like
full real-time speech-to-speech. Incredible. Yeah. Incredible. So, you know, a lot of people are asking me, so when am I going to be able to get my figure? Let's jump into some timelines and applications, kick things off. I think that's the most interesting. When can I order a figure for the home? I think most people want to know that question. Yeah. I mean, at a high level, we have two tracks. We want to ship robots commercially and we'll ship robots in the home. Yeah. They're really different spaces. Like the commercial markets, in some way easier where it's like, you know, very well specified tasks. You can almost like write down a piece of paper exactly what the robot's doing every day. Yeah. It doesn't really deviate from it. But it's really hard in the sense of like we have a lot of speed and performance goals to go hit. And if you mess up, you get fired. No great. Not great. So really tough. And the home is a little bit of the opposite. We don't need to do it at the speeds we would need on the commercial market. but then there's not really a specified task. We, you know, every home is different. I bet if we ask, like, what kind of appliances three of us have,
they're all different in all of our homes. The setup's different in the homes. The things we'll want the robot to do will be, like, will change quite dynamically. So we're working on both in parallel. We think they're both an intelligence problem, really. Like, how do we get, like, the robot to – I mean, the robot figure three is really capable of doing most things in the home. So we're working really hard on trying to build, like, larger pre-trained models and doing a lot more work in post-training and RL to really try to crack this problem. What are we going to have in the homes? Like, Figure, like, so I think you'll get a little bit different of opinion if you talk to other maybe companies in the space, but, like, our view here is that Figure will ship a robot when it's, like, when it ships, like, a really high-quality product in the home. Like, it needs to be able to walk into the home, talk to you, be able to do real work. Like, we don't want to be teleoperating the home. We don't want the robot to be having any issues and failures or stopping because the robot's unsure what to do. The robot needs to, like, be, like, a high-quality product when it enters the home. Like, an Apple-style quality product when it enters the home. You've got to take the kids and the dog to worry about. Yeah, exactly.
But it needs to work. Like, it needs to be able to, like, go in there and be, like, this magical experience of having, like, another person in the home. So that's our, like, launch criteria. We're not going to launch before then. I think you'll see others do that, and it will just, like, in my mind, won't work. It'll be, like, what you've seen, like, AI devices. You've seen this now with Humane and Rabbit and these other groups in the AI device space. It's not going to work until you have, like, this really magical product. So for us, we're trying to crack that. The, like, key constraint here is, like, intelligence. Like, the figure-through robots, like, they have the ability to fold laundry, do dishes, like, unload groceries, like, open the, you know, open the refrigerator, open the oven, like, do real work in the home, like, clean the house up. we've shown some of that, but we need to be able to do that end to end for like for, you know, a dozen plus hours per day, every single day, seven days a week for years. So we're pushing on that. I would say the biggest constraint is like an intelligence constraint. Like how do you drop a robot into any home we would pick? Just talking to it, have it do useful work. There are other things. There's like an integration problem. There's a safety problem. There's other things
that are like really hard, but by and large, it's an intelligence problem we're trying to crack. But, again, without a promise, and I won't tell anybody, 27, 28? I mean, we're doing testing now, and at any given time now we'll be in, like, in order of anywhere from 20 to 100 net new homes we haven't been in in California doing testing as of right now. Okay. And we've been doing this since the launch of, like, of Helix 2.5. So we're doing that mostly for evaluations and evals for our models and, like, a ton of hardcore testing. My forecast, I think the first thing we need to prove to ourselves is we can do end-to-end work in a zero-shot environment, meaning a home we've never been in. We haven't trained in it. We haven't seen any object before. Just through speech, we can take commands, like, say, a prompt, and we can basically do end-to-end work. we're doing that really well today on like the micro level.
Like if you ask if you're like wanting like make the bed or like unload dishes. We literally just did that actually. Totally open-ended prompt. No preconceived what I'm going to ask you to do. And it takes a really good shot at any task right out of pure prompt. Yeah. So, okay, maybe like, sorry, a little verbose. But, like, we need to be able to, like, go off and do that, like, fully end-to-end now over long horizons. We're training now larger and larger models. What you saw there in Black Box today was Helix 3. Yeah. It's, like, probably the only people in the world that have seen this outside this office. And for me, I don't know what you guys are feeling, but, like, it feels to me like a really, like, a baby AGI. Yeah. I would characterize it like an adult body. It's a robot, but an adult human, like a really dumb adult human, like a two-year-old in an adult body.
And you can basically ask it whatever you want, and it can do a lot of things and some things it just can't do well at. But it was significantly under-trained. It was our first model checkpoint coming out of Helix 3. We now have much larger models that are being trained that were evaluated even this week. So my bet is you asked, like, what date? What year? My bet is within the next, I hope to see, I obviously can't promise, but I hope to see in the next sub-six months, perhaps even sooner, we can put a robot into any home and do fairly large, long-horizon tasks end-to-end. Right. Like over hours and hours. And then the question of when you would actually sell them to consumers. Yeah. Then we've got to figure out how to get them to consumers, make sure they're really safe. The reliability is really high. there's like no issues and faults. You know. Do you think you're going to keep robot maintenance or repair as a figure in-house capability
or are you going to sort of approve outside vendors to do that all over the world? For now, we're going to do it ourselves. Like we really want to make sure the robot is like the highest quality experience in the home and we know the robot's better than anybody else and we built him. So we're going to do that ourselves. You can just crawl into a box and ship itself back. Yeah, totally. there's like there's a label on it's open there's like the hand comes out slaps it on and close the box there's like many ways to solve the the the this problem like the the end state is for sure like us not doing everything but in the beginning as we like ramp the product up like you really want this incredible experience like in the home like it's basically just like another human is in your house uh helping you out and um we're only gonna get that with like the highest level of care and attention and we're putting everything our blood sweat and tears like all into this to make sure it comes out really rocking. The flip side of it comes out and doesn't work really well, you're kind of dead. So we really want to make sure it's a really high-companied job span as a kid. Of course. It's great. The parallels are unbelievable.
I was with Mike Rowe and Joe Polish at an event recently, and Mike Rowe was the guy who did the Dirty Jobs show. I told him I was going to be coming here, and he asked, when are we You're going to see figure robots doing welding and roof repairs and sort of all the blue-collar work. There's this interesting conversation that AI is coming for white-collar jobs right now, and it is from finance and legal and diagnostics. But the question of when will these robots start coming into vendors doing traditional bricklaying and all the manual labor out there. I mean, my thesis here, let me lay out my thesis for robots. I feel like it may be helpful to kind of give you guys my sense. My view is there's a high level, roughly four stages for this humanoid race. My chapter one of this book is you have to build really good functional hardware.
The robot hardware is going to be incredible. You can't build AGI on a robot without really good hardware. And we're doing that, you know, obviously now with figure one, two, and three. We have four coming out at some point. Like, we think we're like, you know. But anyway, the hardware needs to be able to do the kind of the speed, torques, like, necessary that, say, a human could with the same range of motions. Roughly at a high level. It could be affordable and you've got to make a lot of them. So that's one. Two is you have to be able to build an AI recipe that can run end-to-end longer highs and work. Full loco manipulation. I mean, like walking the body and moving the hands. And you really haven't, besides figure, we really haven't seen any of that in the world. You've seen, like, groups do, like, tabletop work with, like, all the shelf robots and other stuff that are, like, bolted to a table that are overfit to a certain environment. You haven't seen robots be able to, like, go off autonomously, be able to walk around and do stuff. And we started doing that with Helix, too. Like, full, like, full, like, full, like, room-wide scale. You know, I have a memory of being here when your very first robot,
the big deal was putting a curd cup into the coffee maker. That was, like, forever ago. I would be, like, maybe one of the most important turning points we've ever had. Getting coffee. Yeah, that's a turning point for everyone. I mean, honestly, it sounds funny. It sounds ridiculous. But, like, back then, that was, like, two, two and a half years ago. Like, everything was coded. Everything, all robots was, this is why the second chapter of, like, getting the ARSV is so important. Everything was, like, C++ code. Yeah. And we were using, like, kind of an autonomy, like, 1.0 stack for slam and optic detection and everything else. There was, like, really no path, like, making these really work at scale. Like, the robots just wouldn't – there was really no, like, how do we build a billion units and make them all across – build the iRobot movie. And that demo where we had figure one put a cake up into a Keurig, I was sitting there with the team, and I was like, listen, if we're able to pull off, like, you know, full upper body, like, dexterous manipulation with a neural network like this. It's A, the first time in human robotic history, I believe, that anybody's ever done this.
But it pays a path for all this being done through a neural network, like no code. And this is what we really need. It's just like the entropy is too high in the space to be able to go off in the world and do everything that we want to talk about, like going to everyone else's homes and all throughout the world and be able to do stuff. You think about the blue-collar stuff you just said, too. That stuff is really forceful interactions you're having with the environment, like grabbing objects, you're making impact. It's really quite difficult work. Anyways, I told the team, I was like, I think if you do this, it'll be like so historic that we'll probably raise like $500 million. And I put $500 million on a post-it note and put it on a coffee cup. And I said, if you're able to like actually do this demonstration, you know, show us internally that this like neural networks really work, we'll probably raise half a billion dollars. And we did it. And we raised like a little under $700 million, like shortly thereafter. And it was a huge turning point. It showed that like deep learning works for robotics. And that kind of entered for me like chapter two, which is like you've got to have an AI recipe that really works well, or you have like no hope to make a story work. People don't realize you jumped into robotics here before neural nets were really a thing.
Those were the same driving cars, too. They were 90% C++, 10% neural net. Yeah. And now it's completely inverted. Now, the cars, you've got to keep the C++. The brake, you know, is connected fly-by-wire. So you've got to keep some deterministic code in there. But basically all the decision-making is neural net. So for the Keurig Cup, was it literally 100% conversion to neural net, not a single line of handwritten code? We were standing in a controller that was in C++. The whole upper body, like all the hands and all this, were all a single neural net. So we were basically maintaining balance, but everything above that was a neural network. And then, you know, to be frank, it took us about a year to go from a tabletop like that to all neural nets all the way down the stack moving around. A year. That was Helix. Helix one was like upper body neural nets, and Helix two was like full body neural nets. And that took us a year. I was in a stand-up like right behind us in the lab here for a year trying to figure out how to move the robot off the table to do work.
And we always thought it was really trivial. Like I should be able to like joystick it. You know, in your head, like I could just joystick a robot somewhere and then do work again because you're like usually pretty static a lot of times when you're doing work. You're not like walking around, you know, making coffee. You're just, like, standing there. So we realized the hard truth was that this was a full-body experience, like where you put your feet. Almost like your whole body is there just to get your hands and your head to a certain place in the world and to stay balanced. Yeah. And what really mattered was, like, the body actually mattered more than the hands and the head. And what we really learned is that all of those joint trajectories needed to be done a single, almost like unified experience. You need to understand what the knee and hip and feet are doing as it relates to fingers and hands. The same exact time stamp. So we need to make all train together. I think a lot of people don't appreciate, like if you watch YouTube all day, you think that the Chinese have these insanely dexterous robots that are but under the covers, it's teleoperated or you build these
macros, like 10 movements that are pre-programmed and then when you stitch together the video you think, wow, China's miles ahead of us in robotics. There's a tech stack underneath it is totally contrived. It's not a real neural net that's learning on the fly. Let me just be clear about what we see out of China. We see robots dancing, and these are like deep learning RL controllers. Like one of the best forms of it is training through Deep Mimic. And these are humans wearing a mo-cap suit, like a Hollywood film suit. They're running around doing ninja moves, and then you're basically training RL controller to do this and replay it back on the robot. They're all done open loop, meaning that you can just, like, push a button. They don't react to the environment at all. A lot of them have little LIDARs in their head, so they localize when they're in parades together, so they don't run into each other. All of this work is completely orthogonal and different than what you need to solve general robotics. So while it's impressive to see, like, maybe for the hardware demonstrations that you're
doing, none of it's, like, practical for, like, what we need robots to do at scale. So you're seeing like, okay, the work isn't really practical. The technology stack needed to get there isn't practical. And the AI stack to do it can be done by a senior in college with one robot. This is like very easy stuff. I mean, we've done this work before in our past. So we understand it. So the coolest things about the investment community in Silicon Valley is that they get the difference. Because it's very subtle to the outside viewer, the difference between the two things. But if you watched, you know, ChatGPT become Astra while expert systems just kind of did nothing, and you look at early on, you know, it's giving, I can't spell strawberry, I can't add four and four, but you knew it was like on this trajectory where it was going to punch through. And the way my friend John, or mutual friend John, let me describe, it's like building a submarine underwater and you don't see anything. And then suddenly a periscope pops up. You're like, wait, that Keurig cup is the periscope. That means there's an entire submarine that's about to emerge.
This Kurekscope is like, we think, starting in Chapter 2, which is like A-Recipe work, we ended Chapter 2 with doing kind of like this 200-hour live stream with logistics work. You're seeing it right now. That guy, those robots on that line haven't stopped since we did that four months ago. They run practically every day. And then somewhere in the AI. They go Kurekscope, too. Could you please let me do something else? But you're doing an important part. Chapter three and four for me are scaling intelligence and scaling robots. And what we really want to see is what you're mentioning here with the submarine. We want to see Azure putting more input in. You're getting better output. And for us, our input is data and compute, and our output is getting robots to do more useful work. So stage three is data, data, data. It's data and compute. It's data and compute. It's AI scaling laws. This episode is sponsored by Google for Startups. Think about this for a second. You now have access to the same generative AI models that cost hundreds of millions of dollars to train. Google's Startup Technical Guide for Generative Media gives you a complete blueprint for deploying Google DeepMind's models in production.
Images, video, audio, all of it. Real architecture, real results. Find the link in the show notes below. When you stood up Helix, you needed to. the models that existed back then were not interfacing with the physical systems. And then recently you saw, you know, GPT-Astra actually be able to operate robots. So talk to us about that. How, you know, how far ahead is Helix from a general model that can operate a physical system? I would say, like, one thing I'll say for these, like, other, like, kind of type of language models and even multimodal models we're seeing, command robots, they're all being done for very basic tasks on a fixed camera frame very slowly. If you slow robotics down enough, none of it becomes a really difficult autonomy robotics problem in the limit. So these videos are shot with like 5, 10x, 20x speeds on a very slow –
People realize they're sped up. They're sped up because they slowed everything down because latency and these become like almost like these become like you're basically building like closed-loop control but doing it really slow. You need that feedback, but the models are not good enough to output this really quickly, and they're very like low-dimensional spaces on very basic tasks. We just have a different problem to solve. We need to move very different dynamics and frequencies than what we're seeing out in the world in these videos. We need a robot to walk around. It needs to be able to react to the environment really quickly. It needs to be able to move a package over there. at like sub three seconds of package. It can't really mess up very much, only a few percentage points, basically over like a whole shift. We need robots to go into home, do a very delicate, like fine-grained manipulation, grabbing cups and dishware and put them away and open doors and cabinets and navigating through tight spaces like a bedroom or like kitchen. And then we have to do that on like maybe one of the most complex mechanical machinery like ever built and not mess up. And you have to balance on like basically a quarter-sized point in your foot
above the center of mass in your torso. So it becomes like a very different problem when you attack it like this and what we'll see on like tabletop college setups. And you'll see that now with like Astra and the rest of the groups. They're not real models for robotics. If they were, we could run them. Sure. So I think – Well, there's another subtlety in there is your entire AI stack is totally owned by either Figure or Hark, which when you do a university setup, tabletop, whatever, It's on top of at least one layer of third-party neural net LLM. It's just not real. You can't put them into really high-dynamic robots like this and make them work. Yeah. And you can't scale them. There's no path where you can do, like, the Keurig K-Cup, go from there to full local manipulation, go to, like, there, like, to what we did with, like, Helix 2.5 recently. There is no – like, the architecture needs to dramatically change. These are setups that have, like, basically unlimited compute. They're plugged in with, like, really low-latency comms through a network. They have fixed cameras. They're overfitting the tables generally for data collection. And, like, there are not any scenarios that are realistic what we'll see.
So we had to go build Helix from the ground up to build, how do we build the right brain for the robot? And so right now I would say arguably we understand, like, how the brain works really well now. And what the brain is starved of today is data and compute. We're just, like, we don't have enough, like, world understanding of what forces feel like in joints and observations. to be able to get into every home in the world. But we do see a path now. I mean, with Helix 2.5, we went into 30 rental homes that we rented that day or day before. And we were able to do work, like make the bed and try to do the living room. Yeah, the world saw those videos. Without ever seeing the house or objects before. It was like fully zero shot. They call it zero shot in the air community. None of this was in training. That was the first time we were about seeing this. And that was just like a bunch of stair steps that we've been taking now over the last three years that are all building towards how do we get that now to go from there to drop into any home for like 10 hours and it can do work.
And that's the trajectory we're on now. So that really gets me to the heart of your question is like when we're going to see that. We've arguably seen some of that this year in 2026. I think you'll see more from us here even this year. And then I'd be very shocked if we exit 2027 and we haven't seen this whole picture. How much of the compute is running on the figure robot versus on Wi-Fi? So training versus inference. How much is training? How much is inference? Yeah, most of our compute today is in training, like the dollars going towards compute. We just doubled our – we're doubling our compute right now internally next week. And then we did a big deal with N-Scale for 2027. That deal, 100,000 GPUs, three-plus billion dollars. When does that come online? Those are VR Rubin. Yeah, those are like second half of 2027. Those are VR 2200s, the best computer in the world. Yeah. What's the ratio of training compute you'll have on that day versus, say, a month ago?
Oh, man. Let me do the calculator. There are different types of GPUs, chips than we have now. I figured the VRs are 2X, like raw. You're using B300s today or something like that? Yes. So we, let's see, we have, they're roughly on order of 4. 4X? Yeah. So, man, yeah, like, if, like, we took the whole deal with, and scale and size it all the way up, well over 100X, what we have now. Over 100X, yeah. Going back to the question of. Yeah, I mean, we spend, like, we also spend, like, nine figures a year now on compute, before this. We're spending a large amount of dollars on compute here. That's pretty intimidating to everybody else. Inference on the robot, on Wi-Fi? We do all inference on the robots. You're seeing it here on board today. So you've got what on board in terms of compute? Yeah, we have all of Helix. It's fully on board the robot. So everything from the planning and the reasoning to even the low-level,
like sending trajectories out and torques, all that's done. on the robot today. And what gets done on the server? The only thing we do today is we do real-time speech-to-speech through servers, and we use the servers to help think about – we think of the robots like agents, like physical agents. And, like, you know, on the – if you think about, like, if I had 100 robots building a home or building, like, in construction, how do we get the robots to – like, how do we command that hopefully, like, minute-to-minute, hour-to-hour? Because as the home's getting built, the job's changing. The front of your board. Yeah, exactly. So that's all done from our agent orchestration in the cloud today, which is the only way to do it. We have to, like, in this case, even on this conveyor here, the robots are subbing out, they're recharging. If a robot gets damaged or doesn't feel healthy, it walks the maintenance, and we have another robot that comes in and flags the fleet. So, like, all of this is being networked together. So that has to be done in the cloud.
So that will always be done in the cloud. things there. We're, we're experimenting with like, you know, we've always have like with onboard and offboard computers, really hard. I think the struggle we have is like, we can't get around the physics of having like very high dynamics on the robot. The robot needs to be able to, you know, um, like we need to be able to talk, uh, to all the joints and sensors. Um, and, uh, and, you know, and then communicate back what the new plans are in a really fast way. We don't have like seconds we can wait or even like half a seconds. Like we have, like, we need to get down in the millisecond level. So there's no getting around that as we do really fast dynamics in the world. So in the best case scenario, you have a big brain living in cloud. You have a small brain living on board. In Extreme Example, you have everything living on board, which we're doing now. So jury's still out. We're still doing tons of experimentation. But all the robots right now are all networked on board. I think it's kind of nice because we can lose internet and the robots still do work, which is nice.
If you're even in a home, you have bad 5G service or something. You don't want your robot stopping mid-laundry. You want to continue to do your work. So there's a lot of different engineering trade-offs we have to go through to really understand this better. So millisecond, is that even viable on the cloud? No. The home Wi-Fi is a bottleneck, but then getting milliseconds. Really the best 5G networks we see that you can do is you can kind of get down to maybe 20 to 30 milliseconds per leg. so you're looking at almost 50 milliseconds just for a full round trip. Then you're at 20 hertz just in the network itself, and then you have all the forward pass you need to do on inference, and there's a bunch of other stuff we need to do as it relates to the images and torques coming back. So it's not just as easy as pinging a cell tower in Kenny back. We have a lot of other work to go through. Do you think that every home is going to have some kind of a brain box in it in the basement where the cable box is? Our hope is that everything's done on board and we have the right architecture now. But if every home has like 40 of them, you know, do you really want to?
That's a good market projection, 40 per home. Everything's done on board. Like we do have it on board right now. Like the robots still need it. Chips are scarce, though, if you can share it, you know, because you might have like a roof, you know, gutter cleaning bot up there. I don't think we ever, like, listen, there'll be some neural nets that no matter what, we'll have to live on the robot. Like they're just running too fast. It's about training. Um, world models versus physical versus, you know, what you've been doing in terms of, uh, task video recording. What's your vision on, on training to get to the level of fidelity you need? Yeah, we, um, we took a pretty big bet on Helix 3, uh, here, which, um, we talked a little bit about on the all hands, but we haven't released anything publicly yet. Um, if you roll back before that, we released Helix 2.5 about a few months ago, where we We basically showed, like, zero-shot reasoning in, like, 30 unseen homes before. We also showed the first, like, what we think is the first humanoid scaling loss for data.
As we increased data, we saw, like, a very, like, log-linear relationship, meaning, like, the models were, like, model losses were coming down and the robots were getting better, like, actually executing better in these environments with more data than with less data, which is unbelievable to see. We've seen this in, like, LLMs. We haven't seen this in robotics before then. If this trend holds, then basically as we're adding more compute and data to the problem, the robots are literally going to get smarter. They're going to get better at the tasks every single day. We have a new work we've been doing on Helix 3. I think we gave you guys a little preview today. And I would, like, largely characterize this as, like, little baby AGI. Like, we, like, have, like, a dude in a suit in the office. We have a secret lab called Black Box you guys were in. It's like a little home. And you said we're the only ones that have ever, outsiders that have been in there? Yeah. Wow. It's a mind-blowing experience because when the LLMs were improving, they didn't really look anything like a child.
They were doing complex math, but they were also hallucinating. It's nothing like a kid, but what we just saw in the black box looks exactly like a kid learning how to do things. And it's completely, you know, nothing is pre-programmed. It's all neural net. You ask it to do a task it's never heard before, and it looks exactly like a kid who's learning how to do it and experimenting and trying and rotating and, you know, pick up the chair and, like, oh, it's not going to work. And you can just visualize how this is evolving just like a kid. It's so much cooler than the way the LLMs evolved. So your production rate right now, I mean, the demand for these robots in homes, you know, I'm asking people all the time if these things are costing you. Are you still tracking towards a $30,000 price tag, $300 a month type of lease rate? Yeah. In the home, we'll have to make the robot really affordable.
You'll need some sort of like $300 to $500 a month lease. Yeah, $10 a day. I mean, it's like $100 an hour. I mean, I ask people when I'm on stage, you know, how many of these would you own if it was that? And, you know, the average individual is like at least two. Yeah. Yeah, you've got to either lease it or buy it. There's like two ways we've got to get the robot out. Do you have a preference? Not really. I think through a lease, I think you can really bring the cost down quite a lot and help amortize that better for the customer. I don't really know. And I think the technology would be radically improving. Like my view is that we're like in the flip phone era for humanoids. Yeah. I think I would make them maybe call that figure four is probably like iPhone one moment we're going to have here soon. It's just such a significant step up from where we're at with, you know, figure three. But we're like really early. So you know in the iPhone even from like Gen 1 to Gen 8 10 it was like every year it was like man this hardware is getting better Cameras are getting like the sensors Now it a slightly better camera Now you can really tell I mean it like all it like you know it like atomically approaching
like some perfect like, you know, iPhone vision for them. But like, you know, for us, we're really early. So I think the robots will continue to get like much cheaper, more dexterous, like just like overall safer to be around. I think a lot of the way you interact with the humanoid too will really improve. So I suspect if you do a lease model, you can really, like, get the newest version every few years, which will be good. It feels like you've been – your next gen, it's been like an 18-month center on next generations. Yeah, we, like – we really want to do our next gen at significantly high volumes. So we spend a lot of time on it. And we'll spend a little – we spend about a year between programs right now. We'll spend more than that on figure four. What might we see with figure four? What areas are you focusing on making improvements? Oh, man. That's a, yeah. I think the big drivers here at a high level, at the requirement level, are like, we need the robot to get cheaper.
Like, that's super important. We'll induce more demand. You know, so we're trying to, like, reduce costs as much as we possibly can. We're trying to make the robot more capable. Like, you want it to be able to do more things in the world than it has before. you know like a little sneak peek of that we've been working on the hands for like around almost like two years now and we're like probably the only group that's built like tendon driven hands and non-tendon driven hands we've done now seven iterations of our hand I think we really understand hands really well and one of the biggest step ups we've ever made in the hands is coming and I would make the argument that like a figure four with a glove and a human with a glove is almost indistinguishable really? are you serious? Yeah. That's like an iPhone moment. The whole thing I've seen so far is like they're functional. Yeah. Even the way they grip things, you know, when it was picking up the cups, it's sort of reaching into the cups rather than grabbing them on the side. I hate to say this, but, like, because I think figure three is probably the best humanoid on the market by far, but it'll make figure three look like a caveman.
Really? Yeah. So where are you going to use that, though? Yeah. It's funny because the iPhone analogy there is really interesting because, you know, the iPhone 3 was the first really big hit, but it had no camera. The camera came later, and that was a step function change in that trajectory. I'll bet the hand is the equivalent of the camera. You throw that hand on there, and it has human-level dexterity. I just think, like, you know, we've been – if you look at, like, all three robots, they've been – you know, they look different for us. We've obviously changed a lot across the three. We've changed, like, you know, from one to two. We changed, like, the architecture of the structures from, like, you know, it went fully from endoskeleton to exoskeleton. Like, all the loads on figure two, like, all the exterior take loads, almost like your aircraft does. Right. on how we build aircraft at Archer. So it's a very different type of architecture we went as we designed the robots out. But you'll line all these up and figure forward, it doesn't even look like the rest. It's just like, what happened here? It's like a full deviation from our lineup kind of like it's maturing, and there's just a four-foot fork. And that's what I mean by it's so significant of a robot.
We had a couple of hunches of where things should be headed, and we ended up doing a bunch of work on those, and those ended up painting out really well. So we took a couple of pretty aggressive bets on complexity to get there. And it's, yeah, anyway, fair to say, I think figure four would be like the largest step up we've ever made between generations. That's pretty extraordinary because two to three was a huge step up. You've never seen it like this before. It's like the largest step up is in figure four. Maybe you'll sneak us a peek later after the pod. On the hand, touch is super critical, right? being able to get that data. You know, a friend of mine has a company called TouchLab that is basically built out the most advanced touch systems for gathering data. Are you doing that all in-house right now? We do all in-house, yeah. We, like, the problem with, like, a lot of these external groups is we have to be, like, production ready. They have to work at a really high reliability. It's going to work perfect for our system.
It's going to be at low cost. We need to, like, rev it and fix it if we understand it. We need to, like, understand the firmware and embed a software around it. Like, we need to understand thermals around it. So, like, it's not – you might seem like we can just go off and, like, buy something and toss it in the robot. It makes this a magical experience. It's never the case. Like, never. Same thing from Elon concluded. You need to vertically integrate every particle. It, like, never fits. Like, you can't even make it fit in the CAD, let alone, like, the thermals don't work or it's not the right bus comms or, like, the, you know, the power loads are too high or the cost is unfeasible or we can't – It doesn't work a pass like five or six times and it breaks. Like none of this stuff is like, you know, it's like it's – and then we have some of the best engineers in the world here that want to move really fast. So then we're working with another group that like isn't like beholden to us to like be able to do that work. And it becomes really tough. And that's what happened with figure one. We went out hoping to go get the stuff and pull it all in on the supply chain side. And it just like all broke down really fast. We bought the first actuator. We brought it in. And then we still have it here. It's like, okay, we're going to use this for our elbow. This is an elbow actuator. It had all the right stuff, and it had a torque cell, and it had the right encoders, and the right gearboxes,
and the speed and torque profiles of the actuator were in the right place. And we got it, and it was just like this big. It was literally like this giant block. It was for your elbow. From China, probably. And it was like, we were like, what do we do with this? It's just like the best one on the market. And it's literally three times too big and three times overweight, and it was really expensive. We can't have it hundreds of dollars. It needs to be like tens of dollars at the most. And anyway, so, like, we had to go build everything ourselves. Somebody in that room says, okay, Brett, we're going to wind our own actuator. That's not a small step. I mean, no, we, like, we sat there. We said, this really sucks. We have to go do it ourselves. Is there any way we're wrong about this because we don't want to have to go do this? And we sat there for, like, a while trying to figure out every other possible thing to do besides, like, not build it. Like, if you build it yourself, it becomes really, really sucks. You have to have a firmware person. You have to have an electrical person. You have to have a thermals person, probably some work in embedded software. You have to work in mechanical structures. You have to work in the motor, which is like the rotor and stator and gearbox.
And you probably need a gear designer at that point. And then you need somebody to then put this all together in a really small package because our design team wants it to look really good. And then it needs to be able to move around, which means like the wire and the cabling and all the harnessing and connectors all have to be really dynamic and not ever mess up. And then you've got to figure out how to get the heat out. the system. And then you got to figure out how to, where to put them and how to daisy chain them all now together to build like a real, like a, you know, a functional like actuator and instead of shoulder or leg or hip or whatever you're going to do. And then it becomes like, man, this isn't really like a go get it and you put it together plan. This is like a, it's got to be done from scratch. And every time you make something better on the engineering side, it gets like, it could get bigger. And if it gets bigger, like you can't have your elbow, like it's too big or your knee. And sometimes it clashes with the rest of the robot. And then it doesn't look great. You really want, like, a really high aesthetic product. So it becomes, like, a very difficult conversation to have across, like, 10 people. Do you think, like, a year from now, if you think about the supply chain coming into this, it's just raw molten metal and bare wire that you're starting from?
Like, if you think about winding your own actuators, like, you're literally winding coils. Like, okay, but at the end of the day, you're not going to make your own CNC milling machine, and you're not going to make your own die casting. Or maybe you do your own die casting. I don't even know. But what does it look like? You're referring to the conversation at the Gigafactory where raw aluminum was coming at one end and cyber cabs were coming at the other end. I mean, it's buckets of molten metal just pouring, and you're like, you're kidding me. It's right out of a sci-fi movie. I mean, to get costs down and some of those times, yeah, some of that has to really happen. I can just imagine you're here to your team saying, more robots. We want more robots walking around. It's so much fun. It's so much of a robot space. though, right? It's amazing. It's amazing. You're living in the future. Even the weekends are all running. It's a bummer when they're not running. You just come in here. How many do you have at home? How are your kids going with it all? Are they enjoying? You have two or three young kids?
Three kids. Three kids. They're really young, though. They just want to run. They run over and they try to jump on the robot. I'm like, you can't really jump on these robots. We're not there yet. You can't jump on them. They come to the office every few weeks. They love being here with all the robots and stuff. It's great. Also, when you walk in now, there's 100 of them outside, just outside the door, and you don't even feel vaguely like you can't walk through them. You just walk right up to them, and they're moving smoothly enough that you don't feel like it's going to slap you in the face. I'm going to have some conversations with them on the way out. It's sort of the figure face, because that's an interesting set of choices, the uncanny valley of making something look too humanoid. Right now, you've got a black screen. What's been the trade in your internal conversations on what should the face of figure look like? Yeah. I'm sure there's been a lot of conversations on that. We still have conversations on this. The face is really tough because we're really missing a lot of the facial movements of a human.
Sure. When you talk to the robot, it has lights and screen, and it has obviously the gestures it can make with neck and the whole body. there's just still something missing to give a human-like experience where we're like do you think it'll evolve the face to something more human-like? because right now it's a black face it's it feels friendly but it's a little bit ominous to look at a shiny black skull we're working on the face we have some big plans for figure three on the head and face and we're working through this. It's just a hard problem. It's a human psychology problem. Yeah, like you really want to, I think we're seeing more and more internally. We're talking to the robots more. We're spending time with them. The robots have memory now. So if we've done work now or I can go up to a robot and it's like, hey, Brad, good to see you again and remembers our last conversation. So you're getting now to the point where you might want to just,
almost like a companion, you might want to talk to it, especially in the home when it's around family. and you've had it for a long time. So, you mean this, we call it HMI, but again, it's human to machine interaction. Really, like, well dialed. It's somewhere, we have, like, a whole team on it. We have, an HMI lead. We have, like, a whole focus on it. Every week we meet and do planning on it and demonstrations. It's really hard. A lot of it comes down to, like, not just, like, the face, but, like, how the body reacts to. Like, we're just, like, pretty expressive with our, like, like, obviously facial face, like, arms, all this really helps in my mind building something that's more natural to interact with so we have like a lot of work we're doing there and then yeah the phase is a tough one I think the thing that we've like really stayed away from is like putting like googly eyes on the robot I think it's like a really how do I put this I think like this is like one of the most like complex
pieces of technology on the planet And if we're there putting, like, googly eyes on a robot and saying, oh, it's, like, friendly, and it's, like, oh, it's, like, here, it's, like, safe, is just, like, a huge disservice. It's actually pretty – it's, like, a con. It's, like, the robot isn't having googly eyes. It can, like, literally, you know, walk upstairs, open doors. Like, it's one of the most intelligent things we've ever built. So I think we're trying to kind of hold a certain standard for design that we show, like, show what the technology is really capable of. So we're working through this problem now. I think it's a really hard one. It's one of those things where we have to get in a room with a lot of people, and we have to really kind of discuss where the future is really headed, and then we have to do a lot of product iteration. It's something you could test countless iterations of and how kids and adults interact with it. But, I mean, the end goal is to have, on the consumer side, a home companion that you trust and is part of the family. Yes? 100%. Yeah. You want to be able to talk to a human.
and you would have, like, persistent memory. You might want to spend a more percentage of your time, like, hanging out with a robot than you would, like, having to do work. So it might be used, like, in the home it splits to, like, you really want a companion there in the home that knows everyone, can be really helpful, a kid comes home from school, like, sad, like, asking why, and then you want something to be able to do work. Like, we all do. Like, this morning I was, like, getting stuff out of the dishwasher. Like, I can't believe I don't have my robot doing that right now. You know what I mean? But hopefully by the end of the year, we're doing that pretty repeatedly. So we're working through all those things. But we want to be able to connect with real humans every day. So on persistent memory, if you have a pet, a dog, or something like that, it passes away, you've got a new dog, you've got a new dog. But if you have an iPhone, when you get a new iPhone, you just port everything over. Totally. And the personality has now moved to the new device. So here comes your robot. You've fallen in love with it. It's got persistent memory. You've done a ton of tuning it.
You know, I don't like my laundry folded that way, or you shouldn't be doing that. You should be doing this in the morning. All of that state is now in your bond with this device. So does that live with the device, or does it become like your iPhone where you port the personality to the – and the other one just sort of, you know, folds it up. It definitely does not live with the device. You'll be able to, like, have that pass through any device. And then what if you have two of them? Do you swap personalities midday? Yeah. We're doing some testing here where, like, what if you're, like, walking by, like, these hundreds of robots and they're all talking to you as you're walking by. All the way through. But you're walking by all the way through campus and you're having a singular conversation. That's always mind-blowing, actually, because when you see them all lined up and they start doing things in synchronous lockstep, that is just a mind blow for anyone who sees it. Because you think of them as people and then you see them coordinated in a way that people could never be coordinated. So the equivalent when you're walking through is if you're having a conversation and the conversation is moving from robot to robot, But as you're walking down the line, you're like, wait a minute.
And that could be good or bad, you know, because like that uncanny moment. The shared mind. Yeah. When I think of these robots, I would want to personalize them, both cognitively but also physically. You know, right now they're all wearing the same outfit and costume. I assume you'll have the ability to design your robots. It's like your iPhone case. Yeah, outfit. Yeah, you're going to want to design your robot. Yeah. Yeah, for sure. Yeah, so these robots can all be offered with different soft goods and apparel. We have some here. We can show you on the way out. But, yeah, you'll be able to fully accessorize your robot, give it whatever personality you want, change its voice. What about tools? Are there going to be figure-branded tools that are? What do you mean by tools? Like a hammer? Oh, yeah, like a hammer. You know, like a mixing machine, you know. I'm making cake, you know. It's like, okay, well. That would be pretty cool. Like appliances, yeah. I spend half my life in the longevity business. The favorite example I use is if I've got a bunch of sensors in my body measuring my protein levels, hydration levels, peptide levels, and so forth,
my robot in the kitchen should be cooking me a meal specifically for what my body needs at the moment. The thing about tools in general, though, is right now we think, oh, I've got all this human-made stuff. My humanoid hand controls it. And we're like, well, wait, now we have for the first time ever the ability to snap on an entire function, just snap it in, or I can mentally control it. Like I don't need to, if the device has your IQ from Hark or from Helix built into it, I can just mentally control the device. I don't actually have to walk across the room to mix the cake. I can just brain it over because I can. From the oven, I'll not even go into it. Yeah, exactly. It becomes totally integrated in your whole life. Let's talk about safety because, I mean, one of your goals, you said in the beginning, for getting this out is it's got to be safe. Yeah. Safety, insurance, how do you think about building in the right safeguards for these robots? Because they are very capable. They are powerful to do what they need to do.
Yeah. What's your philosophy there? What are you building into the system? yeah to make it really safe you have to work from like basically like all the way down to like um like the substrate level like almost like a non-volatile memory like really understanding what those rules look like and affordances for the robot all the way up to like the system like architecture design um like a system safety architecture um it's really tough because you have like you have like both like physical safety and like i would call like more like semantic safety like you have a like the intelligence of it needs to also be pretty high like you can't um you know you don't want a robot walking around and knocking over a candle or like you know spilling the hot water uh these are things that like you need to understand there's like risks in those areas and how to deal with them uh i think separately like you know we have like 40 plus actuators on board like uh you don't want any of those like failing you don't want any of the robots falling or um you know, especially with like, you know, say kids or pets and stuff in the house. So we work in like, some of like, like we spent so much time last, say two or three generations
now thinking about what is the right architecture to build like a real time system that doesn't really have like, that approaches like never having faults and issues. And I mean, you see this in the world, like we have like turbo fan engines that are rated So, you know, 1 times 10 to the minus 9 hours for a cash-out event. And, yeah, so you have, like, systems, very complex systems out in the world that are – we've gotten, like, really high safety standards. Self-driving cars need to approach this, like, things like this. So I think we see a clear path how we, like, design a system that gets to a point where it's, like, it's safe enough to scale into, like, the millions and billions. But beyond just, like, the system architecture, there needs to be a lot of, like, iteration design of, like, being able to build and then test, learn, and then rebuild and retest. And then you basically need to build a track record behind that. So, like, you know, these are running software. Software can get hacked.
Do you have a master off switch that can be, you know, physically actuated? You know, we've seen agents escaping the lab. You know, it's been all the conversation on the podcast. I was with Daniela Roos, who runs CSAIL at MIT yesterday. They have a whole bunch of Chinese robots that they bought because they're really cool in the lobby, and they're all transmitting data back to China. They're recording video and audio. I'm not worried about that for Brett. Well, it's actually an advantage for Brett. I'm thinking about having a robot hacked in the home by some outside person. So how do you put the cybersecurity or physical off-command if needed? Is that something you're thinking through? We have a whole cybersecurity team. Just all they do is deal with this. Both on the product and corporate security side. It's actually a pretty large team as it relates to the headcount here internally. Yeah, I mean, listen, self-driving cars have the same problem. You don't want all the cars on the street to be hacked and driven into pedestrians.
Same with us. Like, so we have, yes, we have like safeguards in place for this work. We're continuing to refine that and make that better. It's an evolving thing. It's not like we have like a silver bullet where like we push like one line of code and we're like, yep, that will solve everything because it won't. So we have, you know, basically a lot like tremendous amount of work happening from cybersecurity, system safety engineering, all the way down into, like, what are, like, the fundamental rules within non-volatile memory that we need to put into the robot so it behaves, like, properly in the world. Actually, when we met nine months ago, you mentioned you were well down the road to having no Chinese or third-party components from outside the country. Have you made a lot of progress on that since then? Yeah, we've made a lot of progress last year or two on this point. I mean, it happened because, like, at first, like, tariffs just got so expensive. Yeah. I think at one point, like, last year, we were paying, like, one week was like 270% tariffs. It was like, you know, during all that negotiation stuff, I was like, I think we were like,
this is going to end soon. That's an incentive. But it was like, yeah, this is not good. And then obviously like the geopolitical stuff like really flamed. And so, yeah, we've been rotating like fully out of like just different piece parts and stuff we have out of China. So at this point, it's like in a really good shape. Have you seen a robot insurance industry pop up yet? But I'm assuming there will need to be insurance models wrapped around robots. Yeah, it's interesting. We haven't seen any of that. Like both commercially and consumer will need this. And in terms of financing robots, are you going to provide financing? You have outside financers coming in. So instead of buying for $30,000, I have a lease package. Yeah. We currently have a lot of demand for folks that want to finance robots for us. Right now, a lot of it's like a focus on commercial contracts they want to finance. There's no reason why this won't happen also in the home. This is a multi-tillion dollar space of doing, like, you know, credit card receivables and lease financing for automotive. Like, and we're working with, like, the biggest companies in the world around this.
But, yeah, we'll definitely, you know, I don't know if we'll personally finance them, but we'll work with different, like, banks and funds and stuff to do this work. Yeah. Do you think about space at all? Because I know this seems like a stretch because we're in the kitchen right now. But there's only like three, maybe four people on the planet that have the wherewithal to train a $3 billion foundation model for physical world. And meanwhile, things are starting to go into space, not just on Elon's rockets, on other rockets too. But if you think about Eric Schmidt launching stuff into space that needs to be maintained or changed or whatever, he doesn't have that foundation model, that physical world foundation model. So there's probably, you know, these guys in space, but there's also like, know my model controlling something else. But that physical knowledge model is just like this insanely valuable centerpiece. It's funny you ask. You also mentioned like, you know, like this AI stuff like really happened really fast. When I started Figure, we called it Figure AI.
Our mission was to advance AI for robotics. And we wrote my whole master plan was like, you know, this is going to be convergence of AI and robotics. and be really powerful. And it just, you know, I thought that maybe would happen 10, 20 years later. It happened like in next year. Yeah. And deep learning started working. And in that, we also talk about a key focus of trying to get the space. So it's an area that we're like really interested in internally. Oh, cool. We actually had a couple conversations on this last six months. And one group told us it costs like tens of, like low tens of millions of dollars per hour to put a human in space to do work. And we're like, wow, that's crazy. Well, it's really bad for PR when things go wrong, too. Totally. You know, it's – I think a humanoid should be great at this. Exactly. And that's the whole objection to data centers in space and other things in space is maintenance. Absolutely. Throw a humanoid up there and do it. Yeah, exactly. And how cool is it, like, floating around, you know, in those space views? We had a team here, a space team here, and they're like, you know, how is, like, you know, are you IP-rated to this and this and this for space?
And I'm like, oh, my God, no. Like, you know what I mean? IP-rated joints. Yeah, exactly. So I've got a lot of work. Well, I'll be candid. We are not ready. Brett, you very famously said you're not going to do defense work with Figure. Is that still the case? Yeah, it is. I respect that. I think it's kind of hard. I think it would be difficult to do defense work by also trying to sell robots into your home, doing work in, like, everyday commercial, you know, places that we all know. I think it just becomes really tough for humanoids. I think we need to pick a route. We're picking the whole civilian side, which we think is hundreds of times, thousands of times bigger than military applications. And it's also an area where all of us here at the company really want to solve. We want to solve a robot in the home. This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code.
Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task. Blitzy delivers 80% or more of the development work autonomously, while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Ready to 5x your engineering velocity? Visit Blitzy.com to schedule a demo and start building with Blitzy today. I am curious, you know, when the bird scooters started showing up in my hometown of Santa Monica, people started sort of like trashing them, hanging them from trees and so forth.
So I'm curious, when we start to see robots walking around the streets, how do you keep the robots safe from the humans? Yeah. Yeah, I guess we don't want to be like slapping humans, right? Well, it's more than, you know. That's the question. No, no, it's, you know, there's a significant amount of civil fear, disgust, anger towards technology. And, you know, you saw the Waymos burning in L.A. whenever that was, you know, 18 months ago. So I can imagine the robots themselves need to be kept safe from humans. Yeah. I mean, we're certainly not going to train any robots to, like, defend themselves against humans. So to the extent that a human wants to trash our robot, I guess there's really not much we can really do. There's a lot of cameras on that robot. I know, but, like, yeah, we wouldn't want to, like, you know, we wouldn't want to try to defend ourselves or try to stop us in any way.
I think, you know, I think it just is what it is. I think you're going to have some of that, but I think net-net is probably a very small percentage of what really happens. You see this way most, too, but you know what I mean? It's not like – I think it's just something we're going to have to basically build through. We haven't had it happen since. We're a little worried, like, is there going to be a blue van, drive up and snag one of our guys outside? Yeah, but again, there's things that are apart. Yeah, it's funny. It's really hard to imagine anyone getting away with that more than once. You know, there's so many cameras. It's all connected to the internet. It's really funny if the robot's like in the back, like, listen, man, you don't want to do this. You don't want to do this on timey. Oh, that's interesting. He's like, I have a two kilowatt hour battery pack. I can light on fire and take this whole thing down. Don't test me back here. Brett's not going to, you know, pay a ransom fee. He won't pay. I've already asked them before. Don't do this. That's the policy. It's really funny. I just try and talk him out of it. Can we talk about exponentials for a minute here?
Peter's favorite topic, actually. Go for it. So, BOTQ, are we allowed to talk about BOTQ? Well, what do you want to talk about? I want to know the exponential self-improvement equivalent that's going on right now. Yeah, RSI in the physical world is the coolest thing ever. Yeah. You know, von Neumann probes. But it feels like they're very close to the point where, not where they build themselves like winding the actuator, but they're controlling the machine that's winding the actuator. And then it's a closed loop. And I've always described it as like the way this is my kind of takeaway from the Gigafactory is that it's not robots building robots. It's robots controlling machines that build robots that control machines. But it's the same effect. It's a closed loop, just exponential takeoff of just the volume. And you can see it. You know, when we were here before, there's like a couple dozen. And now it's like, I don't know if the cameras can catch it all, but there's just massive scale going on. There's robots everywhere. And you slowed it down, too. So, like, when does this hit? We still make it a lot. It's just, like, we have a lot of robots coming off. So, it's, yeah.
We're in the process right now of scaling back up as we go into, like, December, January. So, when do we hit RSI and robotics? When do we hit robots, building robots? What's your – what's the time frame for that? Yeah, it was the very first robot where no human labor went into it at all. It was purely – Yeah, like, full lights out manufacturing. Yeah. Man, it just feels like we're – man, it feels like we're so – the thing is like every quarter that goes by, the progress that we're seeing here is just like – it's just astronomical. And I have to be real with you guys. Like when I started to figure it was like I was so excited about being able to work on the human problem. I think it's probably one of the most important businesses I could be building. And I was just like I'm going to go do this for 30, 40, 50 years. like just the rest of my life. Like there wasn't like, okay, in 2020, you know, 2024, 2025, we'd see all these breakthroughs and it would just be like on a full exponential. And in that way, we got really lucky. Um, you know, like basic deep learning showed up a year later, uh, in full
force. Uh, the robot hardware just got really capable. It's now doing work with neural nets, like fully end to end, like we're seeing out there. Like, um, it's, it's nuts and it's just getting better and better we're now you know we just took a robot uh down to a client site that it's never been in before it's never been in the warehouse or anything else this is like three weeks ago and it like fully zero shot the whole thing yeah and like before like you know lightings light would change or like the conveyor would be different package mix is different the robot would just get confused it was just hard and it made sense like right he's never seen this stuff before uh we did larger scale pre-training with helix 2.5 and that we took that that model down to this client site and it just zero shot the whole thing and if you asked me a year ago if that was possible a year from then i would said like you know that would be incredible if we did it but like you know or something like five years so the same argument here it's like everything's
just improving so fast. And we're seeing things like the hand for figure four and even figure four itself. We're seeing the work we're doing with Helix 3. We're seeing the amount of data coming online. We're seeing the amount of compute coming online. We have several billion dollars of cash right now on the balance sheet that we're working to deploy. Congratulations on that. Yeah, thanks. It's like, you know, does it seem certainly within five years? For sure. For sure. It feels for sure within five years where I have a full lights out manufacturing. But it's like, how do I, you know, but it's foggy if it's going to happen in like six months, one year, two years, three years, but like it's right there. Like calling it lights out manufacturing because there's no humans. No humans there. No power and no lights needed. That's really cool. Like full lights out manufacturing we call it. So like we designed figure four to help with this problem. Like we, there was a bunch of stuff with figure three design that we felt were really tough to do like full closed loop lights out manufacturing that we designed into figure four. That's why we're seeing like iPhone one moment. It's like a part of this is like, we think this is the robot that will like build itself. Let's flip it around and ask the other way. When the new hand comes along, which is imminent,
is there anything you can think of in the current manufacturing process where, you know, here's a person controlling a milling machine or whatever component, anything you can think of that won't be able to be done by that new hand? There's, like, blue-collar job opportunity. What's that last thing? Not really, man. I mean, I would have said, like, maybe ladders and this, like, you know, some of this forceful interaction stuff would be hard. But, like, the robots are climbing a ladder next door. Like, it'll just climb over and over every day. Like, just going up and down ladders. And we're, like, doing, like, very forceful interaction work now and carrying big boxes around our 50 pounds with the neural network. We're unloading things in a truck. We doing things in a home like dishes and folding laundry Like it just like all the puzzle pieces are right there The ultimate functionality for me is when is a robot doing surgery right When is a humanoid robot the best surgeon on the planet Well, the timeline lights out improvement to that moment.
I mean, you know, so we've just seen, you know, AI become the best diagnostician. I mean, the numbers are blindingly obvious that, you know, if AI versus human, AI wins every time right now. And then, you know, the next question is when we have a robotic surgeon so that, you know, it's federated learning. So every robot knows every surgical case ever done. It doesn't have any jitters. It didn't have a fight with its boyfriend or girlfriend the night before. It's there and it's able to do surgery. And what's fascinating is it's, you know, a dollar an hour for the surgery, right, any place on the planet. But, I mean, I think that's one of those. I also feel like I've been, like, half the time I've ever visited a doctor in my life, I've been misdiagnosed. Yeah. I feel like just having a network of, like, the smartest thing in the world with you that can be there with you in a room, checking you out, would be, like, an incredible, I mean, like, the world's best doctor for every country in the world. Yeah. We're there now. And then the question is, you know, when are these robots the best surgeons in the world?
Where, you know, you go, you need something, you say, I do not want that doctor touching me. I want the robot. does it perfectly every time. I think I want to capture this, though. If that hand comes out imminently and we can demonstrate every single step of the manufacturing of the figure four being done with that hand, that's a big moment in human history. And I think we might want to just document that, you know, just like literally every step of this entire product. Here's something that arrives in a box, open, assemble, build. It's like, yeah, I mean, guys, here's what I see. We see basically a task we'll pick out, like we need to be able to do this or that. We go get data for it. We train Helix. The robot thing goes and does it. And we've done that now over and over now for the last couple years. And we're like, man, these robots are really capable. How do we get it to generalize is the real thing. We want it to be able to walk into Baku with a minimal set of instructions, like whatever training a human would get, probably some slides and some documents
and maybe shadowing somebody for a few days. We want the humanoid to be able to go through that same process. And that's really the chapter that we're in right now, which is like how do we scale intelligence? How do we get this thing to do it? From a hardware perspective, the hardware is going to get as dexterous as, you know, hands. It's going to keep a steady rate of improvement. So if it has a hard time handling a needle one year, the next year it will just get better and better at this. Like this is all now happening really fast. This is like single-digit years. I think that's actually a much higher bar, though. The moment in history that's true lights-out manufacturing needs to demonstrate. Like, here are all the steps of creating a figure four. Pick anyone at random. This robot with this hand can do that step. That's a moment in history that is actually sooner and easier than the, I gave you an open-ended task that you've never even thought of before, and you have to go and figure out how to do it. That's actually a much higher bar, but it's probably only, you know, six months later. I think it's like the one thing I've learned is like these neural networks like love.
Love data. They love to learn. They love to learn. They love the amount of variety of data that's coming in really well. Like I would have said, like, even on this logistics line that we're seeing behind us now is running, the packages are all like, you know, like moving. They're like, they're squishy. They're moving on the conveyor. Like the robots in a lot of cases have seen these type of packages before. And for whatever reason, it's just good at learning these representations and training. and I would have thought it had never been the case so we're seeing we're going out now and doing large scale data collection as we're bringing more diverse data, the robot's getting smarter at the things it's not even getting we're going out and getting data like some guys I saw has a hose and washing down his dog another one is like somebody's painting a house and then we're like we're taking data like this and it's getting better at folding laundry and it's crazy it's like this data diversity We see this with LLMs in pre-training. This data diversity is just so healthy for the neural net to learn these representations across our world.
I don't know if you can quantify this, but how much do you think is general purpose, physical world knowledge, and how much is like, okay, if I want to train you to do this particular type of surgery, then I need exact data from just that surgery. So right now when people automate these things, they think entirely down these narrow swim lanes. But we've never had the general purpose neural net before. Well, we've been on the narrow swim lane for two years. That's what we do, right? We've been doing that for all – we're narrowly, I would say, arguably probably the best in the world at post-training these narrow use cases. And what you really want from a humanoid is it can adapt to things really quickly, things it's never seen before. And in the case of, say, Baki or even surgery, you want very minimal demonstrations to the robot to be able to do this work. that means like the prior here like the robot coming into it without any like say surgery training should be like should have like a decent understanding of like what it just with a prompt like what it should be doing and how to do it and then with some more specifics like I need you to do this task then go here and do this and watch me like here I just need this like needs to happen this way
without that specifics in context for a robot I should then be able to do the work so it needs to have a good base understanding of how to do this stuff like a human was like a human walks into like getting prepared on how to, what their first day of Baki looks like. They're a really capable human. They do a little bit of like verbal prompting or like show them how to do things. They'll really understand and, you know, latch on to the concept. So for us, it's like, it's really not about these narrow peaks anymore. It's really about like this broad level intelligence that we have like in LLMs. And then from there, the robots can be fine-tuned to a spot where it's like the best in surgery are the best in, you know, like logistics, the best in, you know, some specialized task. But it should be generalizable enough, though, to do most things. Like the surgery bot coming in should be able to go grab your coffee really quick without any training. You know what I mean? But might need some specific training on that surgery. Coffee bot, my favorite. When we were last, you were just gearing up to building a manufacturing facility. What's the status on that? Yeah, we're working on that now.
We haven't announced anything publicly. but like you know the current Baku capacity in my estimates like won't really hold here for the future so we have like you know we can do you know in order like maybe over 40,000 or so units now at full rate out of Baku next door a year across like four lines and you know if we if we can crack this next few steps here I think we just see it feels like almost like a potential potential limited demand. I can imagine that. So do you think, do you think figure four is going to be the first, you know, high mass production robot? Yeah. And then, you know, any projections for 2030, 2035, 2040? What do you mean? I mean, the demand is near infinite. I buy that. Yeah. So what are the numbers that you dream about? we see it like we see i arguably see a path where if we can um it's a really hard path but if we can
solve some of these like steps that we're on now that we could scale a million units like a million units a year out of a facility fairly quickly but that's the pinnacle must be able to crack some of these chapter three problems we're doing now and i remember when you had this conversation and I said, you know, do you, do you agree with, you know, Elon's 2040 prediction of, you know, more robots on earth and humans, um, 2040, 10 billion robots. Uh, I think, I think you were, you agreed with that potential size of the marketplace. I'm fairly confident we're going to have a lot more robots than humans. Yeah. Yeah. You're like, even if you size it, like there's 3 billion, 3.2 billion humans in commercial workforce, that'll probably scale to five to eight, maybe ten. And then every human will want at least one, so you're at roughly eight there. So you're roughly at an order of 20 billion humanoids. But you're going to have them everywhere. They're going to do all kinds of things. They're going to be a dollar an hour,
as you said. You're just going to have them waiting if there's any more work to do. One thing we're seeing now, which is going to help with this exponential, is the companies are all set up based on the human workforce. So if we can go in and automate, say, a certain package line or a certain part of manufacturing, we're still bottlenecked by the other human processes that run on certain shifts. So like the truck doesn't come at a certain turn time. So it's not like we can keep working. There's like literally no packages to move anymore. So if you're working 24-7. So you're like left with like, there will be this period where you're like, you're coming in and doing real useful work. And there will be another period where you come in and just fully automate end to end an entire company and work 24-7. And that's just another step forward. And like, we haven't even seen that in the world. Just kidding. I think you could ship – I think the world could hold, like, 20 billion human rights just in the first phase of that. But when you get to the point where, like, entire, like, logistics and supply chain companies, healthcare, like, data centers can be run fully 24-7 with lights out manufacturing, you're, like, in a different – you're, like, you're basically building, like, entire economies that run 24-7.
So I think it's a very different future we'll live in there. And this goes to the point of, you know, the conversation you and I had with Elon about UHI when the robots are actually able to massively demonetize everything, like building your home, doing whatever you need, at that dollar an hour price tag, it creates abundance for everybody on the planet, which is a beautiful thing. Yeah, pretty deflationary. It's going to be great. It's going to raise wealth everywhere. We had a late-night conversation, and you said something that really was a powerful idea, and I want to come back to this. you said there's sort of an ethical moral issue where when AI is becoming, when we reach ASI, and ASI needs things to be done, you don't want ASI telling the humans to do these things. You want ASI telling the robots to do these things. Do you remember that conversation we had? Yeah. Do you want to expand on that? Because I thought that was a really important point.
Yeah, I think like I have like a view that robotics is super necessary to kind of like avoid like a dystopian future. I feel if we like there's a question right now, you know, I feel like we have this question like are we at, you know, AGI, ASI moment here? the digital AI is like so damn smart but it's like it's also like not very magical. It feels like almost like a search engine replacement and it doesn't feel like HCI or ASI in my mind. It doesn't feel like the Jetsons right now. It just feels like it's fine. It can replace my Google searches. But as this gets more and more intelligent, like the digital systems that are living in the server in order to like move things around to make money and help more in the economy, these digital systems, there's a future where they're, like, bossing around humans to do this work,
paying them money to go do the work for them. Yeah, almost like enslaving humanity. Well, we have a long history of enslaving each other. And it's like – And that is kind of an outcome. Like, there's certainly, like, there's already, like, some companies popping up that are, like, asking humans to go do work and paying them per hour. to let agents go out and get work done. In my mind, if there is no robotic solution here, it can lead them to a really bad place. So my view is that as we build into AGI and ASI, the humanoids should be doing this work and not humans. And I think it's a really important part here of the story. It all needs to come together. It needs to come together really fast. And I haven't heard anybody speak about that, and I think it's a really critical point. you made that, you know, to avoid that dystopian future. I mean, no human wants to be told by the machine what to do.
Yeah. I think there's another. Peter's happening. There's like websites about you can go. Yeah, of course. Meat puppets. Yeah. It's just like. The agent needs to hire somebody to go open a bank account or go do something for them. 100%. Or go do a task. Yeah. There's like many like task rabbit type services that are up now that are doing this. They're getting real traction. Yeah. There's another type of dystopia that I think we're on a path to avoid because of figure, which is, I don't know if you ever saw any of those Russian cars back in the Soviet era, but everybody in the country has to drive the exact same color piece of shit. It's like one form fits everybody. And when you look at robotics in manufacturing today, and you look at like a Honda Accord assembly line, it's like this robot is welding, it's welding, it's welding, it's welding. And what comes out the end of the factory is an identical Honda Accord for everybody, different color, maybe a couple configuration differences. But now the LLMs are capable of mechanical design for the first time, which means somebody with an idea for a new device, a physical device, a gutter cleaner, a pool cleaner,
like a companion, whatever, you can vibe design it, but you can't make it. So unless all the parts are 3D printed or something like that, it'll never come into the world. All of a sudden, figure robotics comes along capable of making itself, which means it's this far away from being able to make virtually anything. So then you're in a world with massively bigger variety of physical things because it's just a completely different supply chain. It's all, you know, anything that can be done can be done very cheaply, fully automated. 100%. Yeah. They're just going to work their butts off. All they're going to do is work. We literally have had that robot since the logistics live stream running just every day. It's just still running right now. It just hasn't stopped. Instead of a webcam in a fish tank, you've got webcams on robots doing packages there. Can I ask you about chips? Yeah, of course. So when you look at the – suppose there are 10 billion robots or more,
and suppose figure is a huge fraction of them, maybe a dominant fraction. The compute behind them wants to get smarter and smarter and smarter. So Elon is looking at all the supply chain components. He's like, TeraFab, this is the one dependency. I can eliminate all the other dependencies. I've got a roadmap. Oh, wait, the chips. So if we were talking nine months ago like we were, the idea that you would have the capital to, you know, chips is about a $16 billion problem. But now you've got billions of dollars of cash. and then suddenly it's viable? Are you going to go down that road? Is it something you're talking about? Yeah. How far down the rabbit hole are you going to vertically integrate? Yeah, I suspect at some point we'll – yeah, I mean, it's really a question is, like, can we control our destiny well enough with the current supply that will exist over the next five or ten years? memory, silicon
and then it even comes down to the cost per chip and power we're getting. Even the chips we have now run on robots, they're just still too costly and too high power. If you really want a really affordable robot, you're going to have to really spend a lot of time in there. Also, you remember you were talking about the actuator comes out of the box, it's this big. The people designing chips right now, they are sold out for decades. Yeah. And so they're not going to start thinking about innovating in the direction that you need to go for robotics. They're just going to try the cheapest, fastest fab they can put up and spew out, and you just have to live with it. We actually met with, like, C-suite of one of the largest memory OEMs in the world. And it's, like, a few months back. And they're like, we have, you know, trying to talk them through, like, you know, our future and roadmap and stuff. They're like, we have no capacity for this. Yeah. All of our capacity is going to, like, phones and servers and everything else. And, like, if a phone, like, doubles in memory, it takes the whole capacity of the whole market.
And the phones are getting choked off. Think about, like, the amount of margin on a phone is insane, and it can't even get the supply. Ernest was complaining about that as earnings call said memory is destroying us. Memory is bad, yeah. We actually have done a few pretty cool, like, memory partnerships here to shore up memory over the next few years internally. So we have like a near term, we feel very confident in our ability to scale up robots. And then, you know, you're asking a lot of good questions about in the long term. So, yeah, we'll see. I think like, as you mentioned, it's a huge capital investment. The technologies, the models itself and the way we're building the models and running them is really evolving fast. Yeah. So I'd be, you know, one concern I have is like building, you know, these chip work takes like two, three plus years. So you have to look out three plus years minimum and say like, okay, what is the design going to look like? And if you told me three years ago what's the design of my inference compute here, it would have just been completely wrong.
Yeah. And then we have like some, you know, work we're doing now with Helix 3, which you guys just saw, that I would say is a radically different type of way we're running now than we're used to. So I think, yeah, it's a little bit hard to, like, lock us into a certain amount of constrained power and cost and, you know, chip architecture given how things are accelerating so quickly. So it's been nice to be able to, like, move to the right chip for us now. But, like, you know, if we look out over, like, a long enough period of time, we're going to need access to like unprecedented amount of silicon. Well, there's a pro tip, you know, there's a gap to, if you think three to four years out, that's the earliest you could get your own chips. But between here and there, Positron, Etched, Grock, are getting incredible headway on six nanometer or thereabout process nodes that are still available. But you have to race right now to book that capacity. But a six nanometer node that's directed toward your use case is dramatically better than the two and three nanometer nodes.
that are, you know, that's all the iPhones now and all the NVIDIA GPUs are built around the two and three node. They can't get any headway out of the sixes. Yeah. But that's the stepping stone that fills the gap. Yeah. But it's, you know, those conversations are moving really, really fast. And those companies are, you know, it's funny in the physical world, entrepreneurship world, there's only like two or maybe three of you. So, I mean, it's a huge bar, but someone has to fill that void. Otherwise, we're all going to suffer for lack of robots for some period. How old are you now, Brett? Forty. Forty. So congrats. I think. Made it to 40? What's that? Made it. No, no. No, I think the congrats. I think, no, because the congrats is, you know, you built serially increasing, more valuable companies, which is extraordinary. I think you were making the quote earlier, the stat earlier that, what was it? It's staggering. It's staggering. So all entrepreneurs all time, only Zuck is ahead of you at age 40.
But serial entrepreneurs. Zuck? God damn it. It gets better. You're not sure. Serial entrepreneurs, which is basically Jobs, Elon, a handful of others, are far more likely to change the world in the future, which totally makes sense to me having invested in hundreds of entrepreneurs over the year. If you can do it once or twice, you're going to do it three, four, five times. If you do it once, it might have been a lottery. It might have been just an outlier. It doesn't mean you can do it again. But you've already proven to do it like four times now, I guess. So the forward trajectory of that profile is so much better statistically, not to put too much pressure on you. But then among serial entrepreneurs at age 40, you're 20x ahead of where Elon was or Jobs was in terms of resources that you have available. I'm going to just phrase it that way. Now, part of that is just a sign of the times. There's just a lot more capital available today. Also, we were talking at breakfast this morning about if you – I think you were destined to be a physical world guy.
You know, it just feels like it's in your kind of Illinois farm DNA, you know, machinery. I like touching hardware. I do, too. But if you were born 20 years earlier, you would have run headlong into the hardware is hard, Silicon Valley is never going to invest in it, you know. The moment is so perfect right now. So you're just hitting it. It's just going to be so fun to watch where this goes. Let's talk regulation. So, I mean, so much conversation around regulation of AI, but I haven't seen that in the robotic space. I've seen sort of the U.S. government putting a regulatory block on importing Chinese robots. Yeah, that's easy. which I get, but are you starting to have conversations or hearing from regulators about, we've got to regulate robots, they shouldn't be on the street, we need to be careful about in the homes? Where is that space right now, and what are you expecting? What are you sort of gearing up for? I think the FCC ban on Chinese robots was a little bit of a shock.
yeah I think from my perspective on that like I'd love to see like a really level playing ground to compete I'm going to come back to regulation in a second but as you think about competition is it Tesla and China is that the way you see it in the world right now I personally haven't seen much of Tesla in two or three years I've seen some totally operated robots but I haven't seen any you haven't seen this right have you no okay um i've seen china do a lot of dancing and um but we talked about that before the the neural network architecture is completely it's literally 180 degree different work if you if you said brett we need you to out dance china they do they do a damn good job we'd have to pivot everybody's strategy we don't have a like uh we we are dancing roadmap and go dance like it is um it is literally 180 degree opposite at work. We need to get a bunch of humans
wearing Hollywood suits, dancing around. We need to then work on replaying that on robot. And, you know, we need to get out in the middle of the woods and go around and leap over rivers. Boston Dynamics was making the best robot dance videos. I mean, listen, I think if I answer your question more directly, I think if I had to pick the most the closest competitor to figure, I would say, is certainly China today. Like the collective China. There's just a lot of work happening there. They're moving quite fast. I happen to view they're moving a little bit different direction than we are. And there are a lot of them. There's a lot of them, yeah. There's like 150 humanoid robot companies. Yeah, it's just, it's kind of all crap right now. And the hardware's not great. The policies are running and models don't look that great. but the amount of progress they're making year to year is really great. So I think I would expect China to continue to do much better,
and it's a group we intend to beat. Sure. I mean, China needs robots desperately. Yeah, really. For manufacturing and for elder care and all of those. Yeah, of course. But are there any of the particular Chinese robot companies that you think are doing a better job than others? Not really. I think, like, we care about something very different than what we think China cares about. China's almost in this race to, like, show, like, how capable their robots are in gymnastics. We care about, like, how do we ship a robot in a home to do, like, zero-shot long-horizon reasoning. And you're not seeing that out of China. Like, you can't even, like, these groups, like, even, like, Unitry, they don't build good hardware when you, like, buy it. They're, like, super tiny. They come out of the box. They wave. With the controller, you're like, what else do I do with it? It's unclear. They have no hands. and they have compute that's kind of like a USB stick. So you basically can't run any real, like, models on them. Seriously, like, all the groups that are doing research put, like, compute backpacks on them. And they're not – the sensors aren't great. The hardware was, like, basically duct taped together.
It wasn't manufactured well. But, like, you know, will they build better robots over time? Certainly. They're going to build great – like, I think over time probably great robots. We just got to go beat them at this. It's a real race here. So we're spending all of our time and energy thinking about where do we want to head long term? And we're trying to move that left on the time curve as hard as we possibly can. And every morning I come in thinking about how to move faster because there's always a way to move faster. And, you know, some of the time. For entrepreneurs out there, hear this. There is always a way to move faster. And speed and agility is your. You can literally always move faster, which is like quite an unbelievable thing to have access to. You can even just, you know, working harder, like work double the amount of hours. You will probably make an order of like one and a half, maybe even two times more speed per year. So like all of that really matters as we compete against like the global collective of China. Welcome to the health section of Moonshots brought to you by Fountain Life.
The discussion in this section is for informational purposes only and is not medical advice. Please consult a qualified health care professional about your individual needs. I'm a co-founder of and client at Fountain Life. You know, AI is having an outsized impact on every aspect of our lives, how we teach our kids, how we run our companies. It also is having a huge impact on health, helping you prevent heart disease, one of the key things. I'm here with Dr. Dawn Musalem, our chief medical officer at Fountain. Heart disease has been personal for you as well, hasn't it? It really has, Peter. My daughter was five. My husband died of sudden cardiac death. And so this is a topic that is one that I am mission-driven to try to eradicate. Prevention first and early detection is absolutely critical. Fifty percent of people die of heart attacks with no warning signs. No shortness of breath, no pain, no nothing. No, silent killer. They just don't wake up in the morning. They don't wake up. And so, you know, AI, this is our mission to advance science, to try to help to one day democratize wellness. We know at Fountain Life, when we do this CT angiography with AI analytics, we are actually finding that 88% of people coming in have detectable coronary disease.
But, Peter, what's more alarming to me is 23% of those individuals had soft plaque. This is the plaque that would not traditionally be seen on CT looking at calcium scores alone. And this is the plaque that we must intervene with, with the multimodal testing we're doing, including diagnostic laboratory studies partnered with healthy lifestyle recommendations. So listen, make sure you understand what's going on inside your body, genetically, metabolically, and cardiovascularly. You can know, and it's your obligation to know. So check it out at fountainlife.com slash Peter to find out more, and really make sure that you're the CEO of your own health. All right, back to the episode. You have the advantage of being a founder-led company versus a large industrial company trying to get into the space. You're willing to take bigger risks than almost anybody else. You're looking at disruptive strategies and approaches and iterating rapidly. And also the talent. When you walk around the building here at the talent level, you'll never find it at a big industrial company.
I mean, yeah, there's a bunch of small little things here. I think the only way to really build really capable gender-purpose robots is in the Bay Area. We invented almost like robotics. The Shenzhen of, yeah. That's better than Shenzhen. We built robotics here. AI was literally invented here. You have silicon was built here. Like, you know, I have batteries and motors and embedded software and thermals and AI, deep learning, all of infra, all of it was built here. Like, you're talking about the best place in the world to go do this. Self-driving cars were, like, here before they were here in China. AI was here before it was in China. You know, like, the Internet. Like, these are all things that were, like, birth tier that, and the people, like, growing up here have worked in these areas and really understand the technology as well. And we've assembled, like, what I think is a top tier of them. And they're, like, you know, the whole team is a bunch of maniacs. They're just, like, they're really great at what they do. They love being here and mission-driven. And we're all here to kind of beat everybody in the space. What's your journey to Silicon Valley?
Because you grew up in Illinois in farm country, right? And then you went to college in Florida? Yeah, I went to college in Florida. I moved to New York City after that. And then I started a software company called Vetery. And I ended up selling that in 2017. And still in New York? Yeah, I was in New York. It was basically an online recruiting marketplace. And so we had basically a two-sided marketplace to bring on employers and job seekers to help with recruiting. and ended up selling at the world's largest recruiting company in 2017 for a little over $100 million. And then as soon as I got the term sheet for this, I was like, I had a few things I really wanted to do in the world, and a few of them were in the hardware space. And then I ended up starting Archer Aviation like that next year. Started it here or in New York? I started in New York with the intention to move. and it was first I was going to move to LA because like all the, like, you know, some of the best aerospace folks in the world. Yeah. And then I found out it wasn't really what I was working on with Archer, like building like electric, like, you know,
basically like electric, like electric, it's basically like robots. It's like electric motors and battery systems and control software and stuff. None of that was like stuff that aerospace was building. And it was all happening in the Bay Area. You had some groups out here building like eVTOL aircraft, like Kitty Hawk and Airbus and a few others. And you have like all this electrification community. Yeah. And robotics community. And so I moved here then, and I've been out here now ever since. And it's been great. I mean, it's like, you know, Archer is basically like a flying robot. Yeah. And it's got all the same, you know, it's got electric motors, batteries, control systems, embedded software, sensors. Like, it's a flying robot. And so, you know, doing robotics out here is arguably the best place in the world. Maybe second to, like, number one, it's got to be the Bay Area. Maybe number two is maybe Shenzhen. Those are the places that you'll see. If general bodice will be solved, you'll see it in one of these two places. I totally agree. It's amazing, actually, if you've stayed in New York or moved to L.A. Today, you know, with $9 trillion of new liquidity from Anthropics and IPO
and NVIDIA's run-up and everything, if you woke up tomorrow morning and said, you know, we were on this podcast yesterday, I think we need to start building our own ship fab. I need $20 billion. You can do that here. Cannot even come close to doing it in New York or L.A. This is the new capital in all those ways. You know, I was in New York for a long time. It's like, I love New York. New York is one of the best places I've ever lived. And, you know, I was like, man, what's this deal with SF? Like, I'm in New York City. You know, probably like in software-wise, probably the number two market in the U.S. besides SF for like doing software engineering. Yeah. I was like, I don't really get this like SF thing. And, you know, I ended up moving out here for Archer. And, man, no, do I get it now. It's like this most condensed place in the world of talent. that are all kind of in this right cultural phase of their life to come over here and really, really, really punch it. It's actually the easiest place to import. Like if you go back to Chicago or Boston, a huge mass of talent. Can't get them. But you can migrate anybody to here.
Yeah. If you're in like Austin or Chicago or whatever, trying to do robotics or Boston, you'll get some folks, you know, like one out of 50 or so, two out of 50 to move there from SF. But, like, unless you get 50 out of 50. Yeah. And in robotics, like, there's, like, an order of 12 different divisions of folks you need to build one robot. So it's like, okay, good luck getting those 12 people to come and those 12 groups to build out by hiring them all from SF. It's just not going to happen. It's just too difficult. When you moved out, everything you said was true when you moved out, except it wasn't the financial capital of the world yet. Now with all the new liquidity it all that plus it the financial it downhill road in S Speaking of that you took Archer Public what billion or thereabouts Peggy Johnson just took Agility Public When's the right time for a company like Peggy to go public? What do you need in your mind having gone through a public offering? Well, I think Peggy went public because she was – Yeah, she needed capital. Yeah, of course. No, I understand.
I've known Peggy for years, and I'm not going to speak about Agility. Like, what are the criteria, having learned what you did with Archer? Yeah. I mean, listen, a high-level Archer went public because we needed the money. Yeah. Like, we were raising our, like, Series B, and, you know, we just couldn't, we needed hundreds of, we needed, like, a few billion dollars to get through. I needed a few billion dollars. Hardware. Hardware to get through, like, F.A. certification, which, you know, is very difficult to go do. and I was just not seeing it in the private markets and ended up going public and raising a little under a billion dollars. It was the right move at the time, but we were really early. We were just a really young company. So we were too young at that stage to be public, but I needed the capital. So I took the company public. And as a choice, I would have done it over again, just given the constraints I had at the time. But if we had access to the private markets, Archer would still probably, to this day, would still be private. We'd been healthy for the business if we were even today. so on the figure side you know I think it's somewhat
inevitable that we'll be public at some point like you know access to the right public capital public trust as well for disclosures and such totally I think it's great for employees and liquidity and everything else for folks and investors but you know I think for the audience here I am a investor in figure yeah early investor too for real cool You know who I invested? It was you. It was specifically you and the team you were able to assemble. And, you know, again, just a note for the entrepreneurs listening, the signal was so strong based upon what you had done as a serial entrepreneur, but even stronger, you were putting your own capital into it, and the team that you had assembled was like, that's a rock star team. Team is lit. Still are. Still lit. But I couldn't have recreated it again if I tried. It was just right place, right time, and team just unbelievable.
And, yeah, it's been great. You know, listen, I think one of the big focuses for us right now is we want to solve what we call general robotics, especially general purpose robots. And you saw a little glimpse of what that looks like today, a little future, and I think it will get much, much better. that will really kick off like significant scale manufacturing, even maybe the larger significance in inference and compute that we might need as we scale and just overall working capital needed to, in CapEx needed to grow the business. So I think we really want to hit that milestone privately. And we feel like we're approaching this. It feels like, man, every new model we train, we're like, is it the one? And then there's like, you know, bugs we want to work on and all this stuff. I think we have, you know, sitting in our servers today, 10 times more data that we can train a new model on than what you guys saw today. Wow. Really? I think it's a little – We have 100x to compute coming on. Sorry. I wish I should make it bigger. I think it's maybe way more, like much more than 10. Maybe a factor of like 15 or so more data sitting there.
I mean, we bring data in. Is this data captured from humans wearing cameras? Yeah. It's from a variety of sources. but basically we need we need really rich information about like physical world information that we can't get from online and scrape online. You just can't get it from like the internet and we need to know how a human grabbed this cup and how are they orientated and what are the forces they feel what are the hands doing the video is all over the internet but the feeling in that feedback can't get it so we basically have like an effort we call index here which is like large-scale data collection effort. And we, like, but we get data in. What I'm trying to say is, like, we have to go, like, audit that data. We have to make sure it's all the right distribution. We have to make sure we clean it, annotate it, get it ready for training. That's something that's time. And then when it's ready for training, we've got to be at the right place where we're ready for a new checkpoint for pre-training and go through. It's not like we can just instantly train on it. We have, like, you know, most of our compute. Over the last, like, month, more than half of our compute has been taken
from, like, one giant pre-training run. a large percentage of it. So, like... What's the price tag on a big, big run for you? I mean, we pay, like... We pay nine figures a year for compute right now. Yeah. Like, low nine figures. So, like, you know, that was, like, maybe 50 to 60, 70% of our compute will be used for pre-training. Okay. So, like, you know, a lot. Yeah. Nine figures a year. Yeah, yeah. Of, like, of is going into, like, just pre-training for Helix 3. So, quite a lot. And if we had 10 times more compute, would train on 10 times more compute well next summer you have 100 times more are you gonna next summer you're gonna do a billion dollar training run yeah i think like yeah i think hopefully i i expect that next summer we'll be doing billion dollar training runs on helix um all goes well i mean like it's like you know everything's gonna go well but it looks like it looks like you'll have the data we'll have the data we have the compute we have the capital like getting these like larger like clusters set up is is hard so we went out like six seven months ago for compute it's the current cluster we have now
I think it was like 7 months ago we had like 4 vendors that came back that had like compute like pretty decent like chips like new Blackwell series chips and we like we beat the piss out of them you know in pricing just like got it down like you know in GPU per hour it was like a great deal we signed it like great it's going on we went out like I think it was like 3 months ago 4 months ago like a couple months after that we need like double our compute capacity we just need more for Helix 3 we went out and we got like we went out and got like no bids back And I was like, we've got to push harder. We're not again. We got one bid back from a big cluster. And then they beat the piss out of us on the price. It was just like buyers and sellers. I was like, what just happened? And we just couldn't get any more compute. And we met the NVIDIA team. I was like, listen, everything's at capacity. The whole world is just there. And I made this prediction earlier in the year. I was like, you know, in 2027, you're not going to be able to compute anymore. And that happened in like September of 2026 for us. September 2026, the world ran out of compute. You just can't get it. I'm so glad you're saying that.
So many people are misinformed right now about, oh, you know, the say-high thing is overblown. The bubble's done. Dude, you can't. There's going to be chips flying everywhere. We can't get H100s. Yeah. We can't get the old chips. They're trading above par. We can't get those. We can't. I can get, like, somebody wrote me yesterday, like, I got compute. I'm like, great. I'm on just, like, eight nodes. Yeah. I'm like, dude, eight nodes. Like, I can't even stop and talk to your eight nodes. Like, I need, like, more than eight nodes, man. It's not going to work at all. So, yeah, the whole world is out of compute, and then compute is constrained by memory. The whole world is out of memory. It takes three or four years to build those fabs from scratch. So, yeah, there's like a really – and NVIDIA has been like one of the best – you can't train on anything else besides this. So it's really just interesting environment. Yeah, you know what else is you mentioned staying in the private markets longer, but a lot of the investment is coming from NVIDIA and other public companies. So it's actually coming from the public markets through large public companies. But they never used to invest in startups or early-stage companies or even, like, fast-growing companies like Figure. Now they're just wide open for business.
That's also a Silicon Valley kind of invention, which allows much bigger things to be built highly efficiently in the private market. Maybe a fab, you know, that would take it to the next level. Because you were saying a second ago, like, the new fab, the new HBM memory fabs won't be online for, like, three years. But if you look at how many are being built, it'll be overbought before they even come online. Like, it's not even – even if they come out on time, they won't. Like, none of these are forecasting humanoids. Yeah. They're not even forecasting. Not even – I mean, like, we literally met the CTO of one of the big groups, and we're like, we don't even have any of these in our projections. But, you know, everywhere you go, too, whether it's software for mechanical design or whether it's space, whatever, everyone's like, yeah, they haven't forecasted my demand yet. Yeah. Like, okay, we have enough of these conversations. and you stitch it all together, you're like, oh, my God. I mean, if you look at, like, the scaling laws for compute, too, like, literally just the same amount of data and same parameter models being trained on more compute has, like, log-linear relationship. You're seeing, like, real ROI there. Yeah. So I would say, like, man, it doesn't seem like this trend is anywhere close to stopping.
Yeah. So it just – I think we're really early in the – what's coming out of this, like, both digital systems, like say we're doing a Hark or even figure stuff it's just so magical it's going to be way it's going to be 100 times bigger than the internet this stuff is going to be so important so I think we're just like, I don't know my view is we're still really early and it's not time to panic yet so keep calm and keep building so I'm a Hark Pro user nice and I would love to ask you what's the relationship between Hark and figure in terms of where is Hark going to be powering part of the figure ecosystem? Yeah. Well, maybe I'll give you a quick overview of both just real quick and then what we're doing. I mean, Hark's on campus here, 100 feet away. It's nice. You have your children next to each other. It's nice. I don't have to travel and I can always be around and help wherever I can. Yeah, Hark's basically trying to build personalized intelligence, so almost like the digital version of what we're trying to do at Figure.
We do a lot with the computer use and speech and memory and vision and a few other things. And we're building our own, like, AI devices internally, which I'm really excited about. We launched Hark Pro, which is our kind of, like, our digital version of Hark, like our non-hardware version. We're about a month in now. And it's been, honestly, a crazy launch. We were, like, you know, we're really interested in hardware. We think hardware is going to 10X this whole thing. and I was even debating whether we want to put out anything before hardware. They're like, okay, let's just put some stuff out. We think the models are pretty good that we're developing now and I actually use them all the time. That's all I've been using for a while now. And I said, let's put it out and see how it goes. And it was just crazy. We've now had over like 1.2 million signups. We've done, you know, we do, you know, probably like, I think it's almost like 5 million browser sessions like since launch. like the usage is just like going through the roof so much that we had to curb some countries like last week just because we're, yeah.
Anyway, so that's good. So we're developing models and hardware. Right now we're very focused on one of the areas is in real-time speech-to-speech. So when you talk to the robots on the way out, they're using the HARC voice real-time speech-to-speech model. And we're doing a lot of improvements there. I think real-time speech-to-speech is going to be a really important part of the UI layer for robotics. There's no third-party components, just totally internally developed real-time speech-to-speech. Our speech-to-speech model is our own model here. There's a lot of people demoing speech-to-speech on top of something else. It's very similar to what we're talking about with the Chinese robots. The difference is night and day if you have a vertically integrated. Can I say something about speech-to-speech? Speech-to-speech today kind of sucks in the world. The hardware is not the right hardware to be using it on. We don't use it as a primary thing. It's like a thing we'll put on every once in a while in the car to talk to. It's not really that human-like. It doesn't do well with tool calling. like, go and get information. It's, like, it's not very good, like, low latency. You interrupt it, sometimes it doesn't hear you. It's just, like, not quite great. So we're, you know, we at both Figure and Hark strongly believe, like,
speech is, like, the right UI layer, like, the default layer into robotics and AI hardware. So we're, like, all in on speech. Like, we really want to make that work. So we're working on that at Hark. We have a collaboration with Figure on this. Figure was, like, the first beta user of Hark. So we've been using Hark now for five months. uh the so we use like uh we don't like force people to use it but if they want to use this uh they can use hark internally the numbers are just like like basically through the roof internally here which has been which has been great um we can get like help with productivity around the company um and then uh and then we're also working on um we're really big at hark on like larger scale multimodal pre-training we're like really we think it's really important for the space. Obviously, Figur has been working on how do we generate action predictions out of the robot. Anyway, we're doing a lot of... There are some areas of classic collaboration we're doing. We have a running bet inside the Moonshot Mate team here, especially with Alex.
So Elon very famously started XAI, then SpaceX bought it, and then X and so forth, consolidating the companies. So the bet is that Figure will buy Hark eventually. I mean, I would probably make a different... I would say, this is just my guess, so don't hold me to it, but I bet Hark is bigger than Figure in sub-six months. In sub-six months? Maybe sub-90 days. Are you serious? From a revenue standpoint? I mean, it's already bigger than Figure from a revenue standpoint. We've been growing like crazy. the Hark numbers are just through the roof. We have some, like, very big stuff coming on hardware and some very big commercial agreements. So, yeah, I would say, like, you know, given that it's digital, we can scale, like, very different. Like, we're in a different slope in scaling robotics. But I think, listen, this is my guess. We don't have anything going on right now. But I think, you know, I do see a path where, yeah, certainly.
I think I do see a path where in the next six months there could be a chance that Hark's bigger than figure. And then at some point, longer term, maybe figure gets bigger than Hark. I don't know. Yeah. So who requires you? Maybe a trillion here or a trillion there. I mean, the reason I wanted to work on Hark is, like, I was sitting here watching, like, doing this human-owned robotics thing. We're like, we're an AI business. We want to build the AI brain that fits in a – the hardware is just a distribution for the AI. It's like the vector it's going to head out into the world in. and I was like playing around with like chat GPT and the rest of the AI stuff and they were just so stupid and I was like I was like I'm just like talking to this like chatbot that was like a search engine replacement from Google right and I was like this isn't what I really want I want something that like can be like with me everywhere I go that like knows me really well has access to my accounts and can like do things for me like pay the property tax bill remind me of something before I forget it help book my travel for me like uh do something more complicated and that It might be long horizon, like, for me with, like, with my permission. And, like, almost like Jarvis would have. And nobody was doing it.
And, you know, OpenAI led my Series B at Figure, and I, you know, got the vibe of being close to those folks. Like, they had no real interest in making this a huge priority internally. And I also feel that on the consumer side, like, interacting with, you know, say, computers and phones with AI is just, like, a really broken experience today. So I sat here and said, okay, I'm kind of in my prime as a builder. I'm trying to understand AI more and more. I'm, like, fascinated by this digital version of AI. It's like, you know, even going to a computer and using it. It's like a little mini-humanoid. Like moving a mouse and, you know, keyboard and trying to get everything to work. I'm just fascinated by it, like, personally. So I spent a lot of time just, like, researching it, understanding how this deep, how basically deep learning works in this space. And so I got to a point where I'm like, I'm going to be like, I'm just going to sit here and watch all this stuff. Or I can go build and come off the sidelines. And I decided, like, it'd make me a better founder and CEO if I came in and helped build something I think was really missing for the world.
And I think it's going to be really special. There's going to be a point in time, you know, like all those little charts you watch, charts you see over time, like over a span of like your whole lifetime. Like, how can you see your kids and your parents and everything, like, kind of goes up and it goes down. Like, it goes down, down, down until you're, like, you know, the rest of your life. I think there will be a chart here of, like, both in robotics but certainly for, like, even for stuff like Hark where this line will be going up all the way to there. And it will be such an important – you have the smartest person in the world with you everywhere you go that has the same interests as you do in helping you. And it's trustworthy, though. The key is that that's actually common between Hark and Figures, that you have to trust this entity with your most personal information for it to be effective. We tell the team every day, it's like you have the person's, over time, diary, and we need a new level of security, a new level of privacy that's never been done before. Like, a thing that we do at Hark is we try to protect that at, like, every angle, and it makes it really hard on us. It's like we can't do what other companies do as it relates to, you know, advertising,
advertising other stuff for their products because we're trying to, like, protect the interests of the consumer as much as possible. So, anyway, I think – Well, congratulations on a meteoric growth. I mean, it's been – when did you announce Hark? It's not – what? We're, like, over a year old now. Is it a year? We launched Pro a month ago, so we're, like, a month – you know, we're, like, a month out of Pro. It's been great, man. I, like, I'm learning every day. There's, like – we're, like, in the early, early chapters of this AI. like almost like with like 1999 back at the internet, it just wasn't the right it wasn't quite the right internet yet but it was like getting seeded, it's all getting seeded now I thought for a minute the last like six months we were real late at Hark and then I've now realized we're pretty early and so we have a lot of, you know we're launching I think four or five pretty big product announcements this week we launched one this morning so there'll be a lot, you know, and then we think about getting into hardware this thing is good, we have hardware now at Hark in the lab and dude is it crazy so peripherals but when you say hardware uh is there going to be
you know sort of hark glasses or can i give you my like thesis here for like devices um so my view is that um there's like there's like two different like areas for ai devices one is like there's this like centerpiece for ai devices which i i think you know for today for consumer electronics that It's the phone and computer. Call it like a billion units a year devices. So it's right there in the middle. And then those are the main platforms. And then around there you have like accessory devices. This could be like AirPods or like watches or glasses or whatever else. I think AirPods is like 3% of Apple revenues. So like Apple's not sitting there getting super excited about the AirPods, but it's an important platform play. They're super interested in like the computer and phone. That's the main thing. what will happen is there will be AI devices that will kill those in the middle. They will do a billion units a year devices that will be way better than your phone and computer today. And then there will be just as similar as it is now, like accessory devices on the outside. Those devices will be single-digit revenue percentage, and they will only work with the main show.
For AI devices, we've only seen the accessories today, like the metaglasses, which are horrible, by the way. and other type of things on the periphery that are not the main show that don't work well today because they interface with things that are in your main platform. Your iPhone fundamentally will never interface well with an AI device. If you try the Meta glasses, it has to have your app opened at Meta on your phone, or it can't get a network. Even with that, it's a horrible experience. The products may be one of the worst products I've ever tried. How do you really feel about this? I mean, let's just get real. I feel like when I watch the spectacles, like it's just, I feel like I'm taking crazy pills. Like nobody's going to wear these. They keep getting bigger and uglier somehow in this space. But like, you know, so for us, you have to go build it well. What's the form factor look like for the ideal AI peripheral device? We're not going to tell you. Come on, we're telling you, buddy. We're going to tell you. I think my analogy, my best analogy. But you have a vision for what it becomes. Oh, we have the hardware in the lab right now.
Oh, wow, really? Yeah, we have it right now. It's in the lab. Right next door. Right next door. I'm using it every day. Does it have windows? Unbelievable. It's unbelievable, guys. Like, it is just, this stuff is going to be really incredible. Are you coordinating with Johnny Ive and OpenAI? Do you guys, like, talk to each other? No, we're not. We're not coordinating. Like, not at all. Not at all. We're not talking at all. I mean, like, we fired them, like, two years ago after the Series B work, and it just wasn't a good fit for us. How so? Is there a backstory there? I mean, so they came in. I met Sam, you know, and they came in and led our Series B at Figure, and they co-led it with Microsoft. And they ended up joining the board, and we did a lot of work together internally. We worked with them for about a year or so. And it was like, you know, the goal was to build, like, general intelligence in a robot. And I was in every weekly meeting. We had a whole team dedicated to it. They had a team dedicated to it.
And we're like, great, let's get the figure every day. on this rock. They're like, yep. And then they never showed up to the office. They were like, it's too far from SF. And, you know, it's like, it's just loud over there. And it's like, you know, we're like, we got grease on us. And it's like loud. And we're working on robots. They just would never show up. We couldn't get in there. I was like, it's like we couldn't. I was like, okay, let's do a big event. We just couldn't get into the office. And then like every week I was just watching the progress that we were making. And then we were sharing with them. And then they were making and sharing with us. And it was just completely lopsided. We were making 10 to 1, 100 to 1 progress versus what they were doing. and we were, like, sharing all the stuff we were doing. Meanwhile, they're sharing with Microsoft. The chain of... And then, you know, like, and then it became pretty obvious they really wanted, I think after seeing the results from us, really wanted to get into robotics again. And then it was just like, you know, it was just something I just didn't want to do. When did Johnny Ive come on board at OpenAI during this time? Was that after? I don't think Johnny's working on the humanoid at OpenAI. I don't know if he knows. I think he's working – my guess is he's working on, like, AI devices work.
Yeah. I don't know if he's working on – It's the iPhone killer, but now they got afraid of saying that. But it's some device that has AI embedded in it that's so compelling. But your point about there being a centerpiece, your personal information isn't going to be everywhere. You're only going to trust a limited number of centerpieces. And so I think they want that device to be the centerpiece, not the iPhone. You obviously have the same vision with Hark. Yeah. So I can see how there's going to be. Listen, it's certainly not going to be the iPhone. Like, the iPhone is going to go away. All right, we're calling it here the decline of Apple. I mean, like, let's just call it. Listen, if you take the extreme case, in 10 years we're not using a traditional phone. Yeah. You know what I mean? Yeah. So we know that. It's like, Ternus really has a challenge right now because they've been gliding forever on the iPhone, and it's going to get challenged from a dozen different places. Man, I've had this thing where I thought I'd wake up at some point every day and see Tim Cook or whatever, John, get on some emergency PR events and basically say, hey, we bought $100 billion of GPUs.
Yes, exactly. We're going to be on the world. We're going to go at pre-training hard. In the meantime, we're going to use other models. And this is going to be so core to the experience and interface with consumers long term that we really need to own it and understand it and embed it. and it's been crazy to see nothing. Yeah. And it's now crossed the point, probably about a year ago, to a point where it's, like, really too hard to turn around. You think? It's for sure. It's lost so much momentum. Yeah. You want to co-design the hardware and the models together. You really want to think about these AI devices as just a pure interface. It could be BCI in the brain. It could be language devices that we're working through now. Like, this is just the interface to this raw intelligence in our side. It's going to be human-like just for you. And they need to be co-designed together. And it's not something you just, like, flop in. Totally. Are you interested in BCI? Yeah. I think, like, I'm very interested in BCI. I think it's one of the most – I think it's a very important vector to be on for kind of, like, AI interfacing long term.
I'm invested in four of the companies right now. And I've just seen some – I just had some meetings yesterday. It's blowing my mind. We're on the precipice of being able to actually get to a mouse connectome. Yeah. And then a human connectome. Yeah. And then the world changes. Yeah. That's sick. We should debrief at some point. I would love to make some introductions. It's an amazing company. Yeah. I think we're in this phase of, like, kind of beyond where we're at now. It's like language device phase. I think there's a phase kind of after that of, like, BCI and alignment into the brain. I think it's going to be really important. I expect Hark to be working through that whole process all the way through. So I think we're, like, certainly right now there's, like, an upgrade cycle coming for traditional computers and phones. And what I can tell you by just using some of the products here internally is it's going to be phenomenal. It's going to be a much better experience. Yeah. Nice. We haven't changed in 20 years. It's literally 20 years. and now are just plopping AI into it.
And it's like the interfaces are just really reactionary. Like you have to kind of control the phone yourself. But, man, we've built this thing that can think now. And it can use computers, and it can speak to you, and it can see, and it can build persistent, almost perfect memory. And, you know, you want to be able to interface with that real time and have it with you everywhere you go. And we don't have anything near like that in the world today, but we will shortly. It's awesome. Yeah. After AI is the interface, you basically use the phone as a camera, but the cameras are like $3. It's like, okay, why did I pay $1,500? I got a $3 camera. I mean, phones are incredible. There is just a next upcoming. It's going to be great. Yeah. It's about time. Let's wrap with the conversation on regulation because I am curious what you're thinking here. You've got Bernie Sanders and et al saying, you know, we've got to stop, slow down, prevent RSI. And I have to imagine, I don't know if you're in those conversations at all, but I have to imagine there's going to be a similar kind of like slow down.
We don't need robots on our streets. It's not safe for the seasonary. How do you think about that? Are you having those conversations? And is there any, you know, is there real concern that the public should have in this regard? I can speak for our plans. Our plans are to launch a really high-quality product that works really well into the market. And it's safe. And it's safe. Like, we don't want to be out there, like, you know, tele-hopping the robot around and doing unsafe work, shipping too early. So we're making incremental steps now. We shipped first commercially last year. We're shipping even more this year. Have you seen any inkling from regulators to start speaking about humanoid robotics the same way they're speaking about superintelligence? We really haven't seen that, like public-facing. We've even rumblings of this. We've seen, like, there's been, like, you know, the only thing we've been aware of the last, like, 90 days is the FCC ban on Chinese humanoids.
Beyond that, we haven't seen anything, you know, come across our desk. I think, you know, also a little bit personal to me, like, I don't fly to D.C. every week and try to get my nose in this. Like I really want to be here. I'm here. Good for you. I'm here. I'm here trying to build shit and I want to build awesome robots. Regardless, like there's going to be robots coming and they're going to be great for the world. I think the same with AI, it's going to be great. If like humanity woke up one day and just didn't want AI, we can just shut down the internet. It's, you know, we can just close it. And we don't, but we don't want that. It's going to be like really helpful for the whole entire world. And you know, there will be like whatever Bernie Sanders, the world or whatever here. But like for us, we're just like really focused on the mission of trying to build and ship your robots. and we're just like we're here every day trying to go do that. I got a question just to inspire all the entrepreneurs out there. So, you know, I grew up a big Steve Jobs fan too and I think a lot about how much more he could have done in his life. But in that era, you could only do one thing at a time and nobody had super voting control and all that. And then when Google came along and then Zuckerberg came along,
it switched to this era where the founders control the companies and they could be much more aggressive with the business plan. And that works really well here in the Valley. But then Elon came along and reinvented it again where you can do multiple things at the same time. So this is working out incredibly well for Figure and Hark. Have you just said Hark nine months from today could be bigger than Figure? This is a mind-blowing statement. I think you said three months from today. Three months from today. I mean, Hark's just going like this, guys. It is going really well. So we'll see. You've also got this other thing about school safety, which is not going to be bigger than the other two companies, but it's important for the world. The power of an entrepreneur today is so much bigger than it was in the Steve Jobs era, let alone the Zuckerberg era. So I think it's really inspiring. Do you think a 20-year-old entrepreneur today can walk in your footsteps, or is AI going to make it all too late and the ship has sailed? Yeah, I mean, I think one thing I'll say is I'm, like, completely dying.
Like, three kids, like, three companies. Like, I'm just, like, I've been here every night for a whole year. I'm just, like, man, has it been hard. But it's, like, you know, it's the best time in my life I never want to do again, you know? It's, like, it's been going great. Yeah, listen, I like the future generation is going to be great. Everything is going to be AI native. You'll be able to build any type of hardware just through – Promptive product. Exactly, just through talking and robots. It's going to be so cool. Really where the AI is really missing is it has no imagination today, and we're going to be bound by this imagination. This is what entrepreneurs do best, right? They always have a little notepad. I still have a notepad of all the business ideas I can think of, and I still keep them to this day. They're a great idea generation, great about thinking about the future. Like, this is going to be an unbounded opportunity for every, like, new entrepreneur, like, behind me growing up.
And it just – maybe one of the – it'll be the – it'll mark the best time in the world for entrepreneurs in history. And now you have an ecosystem here, like, today for entrepreneurship. Like, when I was graduating school, like, it wasn't – like, it wasn't conceivable you'd start as an entrepreneur. Like, it was just like you don't do – nobody does that. Like, so I had a job for a few years before I started. I wish I started even sooner, starting companies. But, like, you know, now you can go do that. You can get capital for it. There's incubators. There's people who want to work at startups. It's fun. Like, you have this whole, like, community built around it, too, which is just, like, going to propel these future entrepreneurs further. I like to joke we've all been promised a three- and four-day work week. We've invented a 10-day work week. I mean, I like working, man. This is the most exciting time ever in your life. I like grinding out and working on this. Like, I just, we, I think everybody here with me likes it too. We just like, we have a bunch of groups that like to work hard and we've like, we've made it this like dirty word out here in the Valley. And so, you know, like a lot of folks come in here and be like, you guys are like the most unsilken Valley company I've ever seen.
Like we work hard. We come in on the weekends. Like we work nights. Like we just, we like it. And, um, there's not enough of that anymore. You know, it's just like almost like we like interview people like, Oh, what's your like day off strategy and like work from home. Like we don't have any of that shit. You come in here every fucking day and you like work with us and we're going to work with you. Like, you know, I bought you guys can't work from home. We got to get on the line every day. And same with us. We got to get on the line too. So, like, anyway, it's just. What you're building is so much cooler than virtually any day job, you know. I mean, it's got to be, even if you're working twice as hard, it's got to be so much more rewarding at the end of the day to look at what you've physically built. Also, I didn't anticipate how much the office environment is elevated by the walking around. So much better. It's bizarre. It just adds energy to the building somehow. So a year from now, we're going to come back and do our figure four recap. You come back, figure four, figure four gets you all set up, checks you in, gets you water, coffee, gets you around, gives you a tour, and then I'll see you for the interview.
Yeah, we'll include figure four in the interview. Maybe just like honestly figure four can speak on my behalf. That's so fun. You know what I mean? Buddy, thank you so much. Thanks for having me. Yeah. Thank you. And now when we cut cameras, we'll go see your secret device in the figure four. Okay. Perfect. Good deal.
番組の概要欄(原文)
The Mates sit down with Brett Adcock, founder and CEO of Figure AI, to discuss the global race in humanoid robotics, how Figure AI is competing with China and Tesla, and the latest developments shaping the timeline for bringing humanoid robots into everyday life. Get access to metatrends 10+ years before anyone else - https://qr.diamandis.com/metatrends Peter H. Diamandis, MD, is the Founder of XPRIZE, Singularity University, ZeroG, and A360 Dave Blundin is the founder & GP of Link Ventures Brett Adcock is an American technology entrepreneur and the founder and CEO of Figure AI, a robotics company developing AI-powered humanoid robots. He previously founded Archer Aviation, an electric aircraft company, and Vettery, a technology-driven recruiting platform. – This episode is brought to you by: Get the blueprint for generative media https://goo.gle/startupgenmedia Go to Blitzy to book a free demo and start building today: https://qr.diamandis.com/blitzy Your body is incredibly good at hiding disease. Schedule a call with Fountain Life to add healthy decades to your life, and to learn more about their Memberships: https://www.fountainlife.com/peter _ Connect with Peter: X Instagram Substack Website Xprize A360 Connect with Dave: Web X LinkedIn Instagram TikTok Connect with Brett: LinkedIn X Instagram Website Figure AI Listen to MOONSHOTS: Apple YouTube Follow MOONSHOTS: Instagram TikTok X Threads – *Recorded on November 9th, 2026 *The views expressed by me and all guests are personal opinions and do not constitute Financial, Medical, or Legal advice. Learn more about your ad choices. Visit megaphone.fm/adchoices
出典・この記事について
- 番組
- Moonshots with Peter Diamandis(PHD Ventures)
- 元の回
- How Figure AI Is Outcompeting China and Tesla in Robotics | Robotic Deep-Dive w/ Brett Adcock
- 配信日
- 要約作成日
- 公式サイト
- www.diamandis.com/podcast
この要約・解説記事は、元の音声を文字起こしした内容をもとに AI(Claude)が日本語で作成したものです。 固有名詞や数値に誤りを含む場合があるため、正確な内容は元の音声でご確認ください。制作方法について
関連エピソード

ジェンスン退任とNVIDIAの後継問題、AI終末論への反応、ハサビスのRSI警告|MOONSHOTS #311
Moonshots with Peter Diamandis

数学を襲った雪崩がAI開発にも来る――AI研究の自動化と再帰的自己改善の見通し
Dwarkesh Podcast

ベン・ブキャナンはAIで米国に勝たせたい。だがそれは「競争」か「罠」か
Your Undivided Attention

ジェンスン・ファンはCEOを続けられるか?NVIDIAの次の一手|MOONSHOTS #310
Moonshots with Peter Diamandis

政治とAIの大衝突:中間選挙、データセンター規制、再帰的自己改善(#309)
Moonshots with Peter Diamandis

Jeremie Harris が語る「本物のAI災害」後の世界と、中国との信頼構築が難しい理由(NonZero からのクロスポスト)
80,000 Hours Podcast