超知能には物理世界が必要だ:Arda共同創業者Bob McGrewが語るAIと製造業

Superintelligence Needs the Physical World with CEO and Co-Founder of Arda Bob McGrew

超知能には物理世界が必要だ:Arda共同創業者Bob McGrewが語るAIと製造業

超知能には物理世界が必要だ:Arda共同創業者Bob McGrewが語るAIと製造業

No Priors: Artificial Intelligence | Technology | Startups

0:0059:57

要約

元OpenAIのChief Research OfficerでPalantir初期幹部のBob McGrewが、ハードテイクオフ論は誤りだと主張する。研究の自動化は一度きりの加速であり、真のボトルネックはGPUや実験設備など物理面にあるという。彼が創業したArdaはAndurilと組み、製造エンジニア向けのAIプラットフォームを開発。フォワードデプロイドエンジニアリングの進め方や、自己複製する工場、フェルミのパラドックスへの見解にも話が及ぶ。

この回のポイント

  • ●McGrewは「研究者の作業を自動化すれば暴走的な自己改善が起きる」という議論を誤りだとする。実験速度を縛るのは研究者数ではなくGPUなどの物理的な計算資源だという。
  • ●研究の自動化は推論モデル登場のような「一回限りの大きな加速(S字カーブ)」であり、自己改善の連鎖にはならないと述べている。
  • ●知能が安価になると、ウェットラボや工場といった物理層が希少になる。デジタル側の高速化は吸収され、再び物理がボトルネックになるという見方。
  • ●Ardaが解こうとする課題は、頻繁に設計変更される製品(スマホ、防衛、データセンター関連)の自動化ラインを、人の再訓練並みに素早く組み替えられるようにすること。
  • ●Andurilとの協業では、アクチュエータの仮想ビルドが数日から90分に短縮された。Fury戦闘機など、ドローンからミサイルまで展開が広がった。
  • ●ロボット分野ではフロンティアモデルをオーケストレーターとして使い、専用アルゴリズムやツールを組み合わせる。独自の巨大モデルは訓練しない方針だ。

章立て

  1. なぜ物理世界に取り組むのか

    AI革命の最中にもかかわらず、実世界の原子レベルの変化は乏しい。McGrewは超知能には物理世界の変化が必要だと述べる。

  2. ハードテイクオフ論への反論

    研究者の自動化は一度きりの加速にすぎず、ボトルネックはGPUなどの計算資源だと論じる。

  3. ラボかスタートアップか、物理層の希少性

    知能が安価になると、ウェットラボなど物理層が希少になる。シミュレーションが進んでも最終的に実験は必要という議論。

  4. Ardaの課題:設計はするが作らない米国

    人件費の問題で米国は製造できない。柔軟な自動化によって、ライン切り替えを7日で行う構想を語る。

  5. ロボットの形態とクローズドワールド

    ヒューマノイドは不要で、組立ラインは閉じた世界。ルーブ・ゴールドバーグ式の自動化をAIで設計・調整する。

  6. Andurilとの協業とCursor for manufacturing

    製造エンジニア向けのIDE的基盤を作り、その上にエージェントを載せる。仮想ビルドが90分に短縮された。

  7. PalantirとOpenAIの経験、ロボティクスモデル

    課題起点とテクノロジー起点の違い、OpenAIでのロボットハンド、汎用ロボットモデルが機能し始めた理由。

  8. FDEの実際のやり方

    少数のパワーユーザーと組んで成果を積み上げ、その知見を製品に還元するのがFDEの本質だと説明する。

  9. 自己複製する工場とフェルミのパラドックス

    機械を作る機械、ノイマン・プローブ、フェルミのパラドックスに対するMcGrewの見解。

解説記事

元OpenAIのChief Research Officerで、Palantir初期の幹部でもあったBob McGrewが、No Priorsに出演した。現在は米国製造業の競争力をAIで高めるArdaのCEOを務める。「超知能は物理世界の変化なしには実現しない」という持論を軸に、ハードテイクオフ論への反論、Andurilとの協業、フォワードデプロイドエンジニアリング(FDE)、機械が機械を作る未来までを語った回だ。

ハードテイクオフ論は「一回限りの加速」にすぎない

McGrewは、研究者の仕事を自動化すれば暴走的な自己改善が起きるという議論を「端的に誤り」と断じる。自動化率が50%、90%、99%と上がれば速度も2倍、10倍、100倍になる、というのがその論法だ。しかし研究の速度を決めるのは研究者数ではなく、実験を回す資源だと彼は言う。2024年にOpenAIで研究を率いていた当時も、制約はGPUの量と、計算資源が適切な場所に集まりネットワークで結ばれているかどうかだった。

彼のモデルでは、研究の自動化はGPUの利用効率を「最適の35%」から「90%」に近づけるような効果をもたらす。これは推論モデルの登場に匹敵する大きな改善だが、90%から99%への差は小さい。つまり自己改善の連鎖ではなく、もう一つのS字カーブにすぎないという。

知能が安くなると、物理層が希少になる

知能が安価になれば、仮説の生成や実験の選択は容易になる。その分、ウェットラボやGPUの設置といった物理的な補完要素が希少になる、というのがMcGrewの整理だ。住宅建設の例では、デジタル作業を速くしても許認可や建設作業が残る。データセンター建設でブルドーザーが足りないとき、AIがブルドーザー工場を建てようとしても、整地にはブルドーザーが要るという循環に陥る。

聞き手のElad Gilは、タンパク質折りたたみのようにシミュレーションが実験を代替する領域もあるのではないかと問うた。これに対しMcGrewは、予測精度が上がれば必要な実験数は減るが、デジタル側が速い分その加速はすぐ使い切られ、再び物理が制約になると答えた。Ilya Sutskeverも、アイデアの良し悪しは最終的に実験で確かめるしかないと語っていたという逸話を紹介している。

Arda:頻繁に変わる製品のための柔軟な自動化

McGrewは、米国は製品設計に強いが製造は他国に任せていると指摘する。要因は人件費だ。iPhoneは毎年約75%の部品が変わるため、自動化できるのは約25%にとどまり、残りは手作業になる。従来の自動化は特定の問題向けの専用機械で、世代をまたいで使い回せないからだ。

Ardaの目標は、製品設計をクラウドにアップロードすれば、自動化ラインを7日で設計・再構成できる仕組みだ。想定する顧客は防衛、民生電子機器、データセンター関連のサプライチェーンなど、変更が頻繁な分野である。

ロボットの形態については、人間が立つ位置を置き換えるだけなので人型は不要で、アームとカメラがあれば足りることが多いと言う。ライン組立は掃除のような「オープンワールド」ではなく、環境を制御できる「クローズドワールド」の問題で、知覚と強化学習を組み合わせたエージェント型ロボティクスで扱えるという立場だ。

Andurilとの協業と「製造エンジニア版Cursor」

Ardaは設立前日にAndurilの課題を聞き、1月にシステムへのアクセスを得て、4月に最初のユーザーを得た。アクチュエータの仮想ビルドは通常数日かかるところ、90分で終わったと述べている。現在はドローンからミサイルまで全製品部門に展開し、Fury協調戦闘機の製造プロセスのAIによる可読化にも取り組んでいる。

構想は、製造エンジニア向けにGitに相当するデータ構造と開発環境を作り、その上にエージェントを載せる「Cursor for manufacturing engineers」だ。設計から製造、自動化ステーションの設計まで左右に広げられる。独自の大規模モデルは訓練せず、フロンティアモデルを、モーションプランニングやCAD検査などの専用ツールを束ねるオーケストレーターとして使う。小さな専用モデルは、フロンティアラボでなければ現実的に作れるとも述べている。

PalantirとOpenAI、FDEの本質

McGrewはPalantirを課題起点、OpenAIを技術起点と対比した。GPT-3の頃、Ilyaは「一人の顧客から学ぶより、訓練データを10倍にしたほうが多くを学べる」と答えたという。ArdaはPalantir流の課題起点を取る。

FDEで重要なのは、少数の強力なパワーユーザーを見つけて共に苦労し、その知見を製品に還元することだと強調する。決裁者(スポンサー)だけでなく実際に使うユーザーが要る。大企業も小さなチームの集合なので、問題ごと、チームごとに進めればよい。FDEは単なるコンサルではなく、製品とユーザーの隙間を埋め、その知見を製品化して初めてレバレッジが効く、という。

また汎用ロボットモデルが機能し始めた理由について、ロボットデータの増加よりも、モデルが物理世界のデータの細部まで予測できるようになり、良い表現を獲得したためだと推測している。ただし微調整が難しい点は他のロボット企業の機会だとも述べた。

機械が機械を作る未来

McGrewは、データセンターに閉じた知能は「知覚」と「行為」という補完要素を欠くと考える。今後10年で多様な形のロボットが増え、あらゆる物理的なものが知覚を備えクラウドにつながると見る。工場が工場を作る自己複製は、機械工具の延長として近い将来に進み、長期的にはノイマン・プローブのような星間展開もほぼ確実だと語る。人類がその主導権を保つべきだとも述べた。

フェルミのパラドックスについては、文明は光速に近い波として広がるため、到達直前まで観測できないと論じる。インタビュアーは、到達済みでないのは希薄さを示すとの疑問も出したが、McGrewは自分たちが先頭の波になることを望むと締めた。

まとめ

本エピソードの主張は、AIの進歩を「頭脳の問題」だけで見ず、計算資源や工場といった物理的制約から捉え直すものだ。製造業に強い日本の読者にとっては、設計変更に追従できる柔軟な自動化、製造プロセスのデータ化、FDE的な顧客密着の進め方が参考になる。ただしハードテイクオフ否定や自己複製工場は、McGrewの見解として聞く必要がある。

文字起こし(英語・自動生成)

If you look around outside, how different is anything from 10 years ago before AI happened? Not very different at all. The first time you were in your Waymo, you thought to yourself, this is amazing. Probably the third time you're in your Waymo, you're like, man, this traffic sucks. We need to break out of this pure focus on the digital world and go to the physical world. I don't think you're actually going to be able to get to superintelligence without changes in the physical world. Hi, listeners. Welcome back to KnowPriors. Today, Alad and I are here with Bob McGrew, the CEO and co-founder of ARTA. Bob was previously an early chief exec at Palantir and the chief research officer at OpenAI. We talk about making American manufacturing competitive with intelligence, why physical intelligence is necessary for superintelligence, intelligence and why the future is going to be machines that can make machines and explore the stars. Welcome, Bob. Bob, thanks so much for joining us today. No priors. Great to be here. So you've had a ridiculous career. You were the chief research officer at OpenAI. So you drove

all those sort of research directions for the company, or at least had a lot of the agenda. You were at Palantir quite early and one of the main people there driving product and engineering. So you've had this amazing career spanning some of the most seminal moments of the digital world. And yet now you're working in the physical world. And it's right in the era when people are talking about RSI, which is recursive self-improvement, about how we will no longer have researchers, don't be displaced by code and AI and, you know, AI building itself. Why do anything in the physical world right now as this digital revolution is happening? Well, you know, fundamentally, I don't think you are actually going to be able to get to superintelligence without changes in the physical world. So, you know, we've seen a huge amount of improvement from ChatGVT, from agents. You know, all of this is happening in our digital lives. And very little of that has actually affected the atoms of our lives. Like, if you look around outside, how different is anything from 10 years ago before AI happened? You know, not very different at all. You know, even when you're doing something that is genuinely very different.

Maybe you're in a self-driving car, right? You're, you know, in a self-driving car in your Waymo going around San Francisco. And the first time you were in your Waymo, you thought to yourself, this is amazing. There's no one driving it. The, like, car is driving itself. And, like, probably the third time you're in your Waymo, you're like, traffic. Man, this traffic sucks. Yeah. Right? Because just, you know, you have something. Even, you know, a Waymo is, like, halfway towards the physical world. But, like, you didn't change the car. Like, the car doesn't fly. You're still bounded by all of the physical barriers that made the car trips around San Francisco pretty slow, even in the 1990s. And so fundamentally, we need to break out of this pure focus on the digital world and go to the physical world. So you can't have super intelligence without changing the physical world. What does that actually mean? What does that substantiate as? So maybe a good place to start here is I think there's this common perception and or there's a common argument. And I've been hearing it since like 2018, which says that if you can just automate the work of the researcher,

that you're going to be able to unleash a runaway acceleration, runaway self-improvement loop. And we're going to have a singularity. And this is a hard takeoff. Yeah, hard takeoff. Exactly. And I think this is wrong. I think this argument is just straightforwardly wrong. So just to recapitulate the argument, it says, okay, so we can automate 10% of a research time. That's going to give some improvement. We automate 50% of a researcher's time. Now you have a 2x improvement. So now you're going 2x faster. And then you automate 90% of a researcher's time, you're going 10x faster. 99% automated, you're going 100x faster. And so by automating more and more of this job and eliminating this labor of the researcher, you're able to just, you know, zoom up and have runaway acceleration. But this is, this is, this is straightforwardly wrong because researchers are not the bottleneck, right? Like the amount of time it takes in order for you to discover something is not primarily bounded by the number of researchers you have. Fundamentally, those researchers are running experiments.

And so, you know, if you, if you actually think about the, the, you know, even today, right, even two years ago in 2024, when I was chief research officer at OpenAI, like the actual limiting factor was not the number of researchers we could hire. It was the amount of GPUs that we had on hand. It was the physical amount of compute that we were able to put out there. And, you know, was the compute accumulated in the right places? Was it connected together with the right networks? And so you can, you know, completely automate researchers. This is a big effect, right? I'm not saying this is not a big deal. You know, maybe this is even maybe the ability to do automated research is going to be as impactful as the ability to go from, you know, GPT-4-0 without reasoning to O3 with reasoning and tools. That was a huge one-time acceleration, but it's not RSI. So I think the right way to think about research is that, you know, fundamentally, there's a bunch of experiments. You have to run these experiments.

And right now, there's a lot of waste. You have to, you know, researchers have to do a lot of engineering. They have to pick the right experiments. They have to just, you know, go physically schedule them on the GPUs. And all of these lead to the GPUs being less effective at delivering new discoveries. And so, you know, the correct mental model is to look at the GPUs and to say, okay, well, you know, right now, the GPUs are 35% of the way to optimal in order to be able to deliver discoveries. and if we can automate more and more, maybe we can get them to 90% of the way to optimal. But the difference between going from 90% to 99% of optimal, it's actually not a big difference. And so we can automate researchers. It's a big one-time benefit, but it doesn't kick in a self-improvement cycle. Instead, it's one more S-curve, one more big one-time improvement similar to reasoning that allows us to, you know, get closer to optimum. And then we have to go do something else.

Yeah, I feel like a lot of AI-driven research is basically stacking us curves of different types, or AI development, I should say, in general. That also translates into other domains, though. There's robotics, there's biology, there's a few other places where the physical world is really, you know, the primary limiter on progress, right? You have to run the biology experiment. You have to build the motors for the sort of mechanical hand or whatever you're building. How do you think about the progress that the labs will drive versus startups will drive in the context of these other big markets where there's a strong research overlay or angle, but then there's sort of this physical world limitation that you have to overcome or at least optimize against in parallel? Like, will robotics and biology aggregate to the anthropics and opening eyes and Googles of the world and their subsidiaries like Esomorphic? Or do you think there'll be like a bunch of startups that draw this sort of progress? Yeah. Before we talk about market structure, I think like let's think about what the actual, maybe the economic structure of this is. You know, the labs, if we think about what the labs are doing, they're really producing cheap intelligence.

So what are the complements to intelligence? If we think about biology or, again, we think about, you know, running these experiments for superintelligence research, you know, the intelligence in terms of choosing which experiments to run, generating hypotheses, it's really a complement to the physical layer of having the wet labs, of, you know, laying out the GPUs. And so, you know, when hypotheses become cheap, then it's actually the physical part that becomes scarce. And so, you know, if you get better and better at generating hypotheses, you're just going to bang harder and harder against the limited number of wet labs that you have. And so then, okay, so go back to the market structure. Well, you know, I think it could be either way, right? But then do Anthropic, do OpenAI, do they have a comparative advantage in building out wet labs? I think they probably don't. Did you consider working in other domains before you decided to go work on manufacturing and robotics?

I did. Yeah, yeah. So when I left OpenAI, I think the one thing I knew was I didn't want to start another Frontier Lab. Like, in fact, maybe I wanted to be the only person that's ever left OpenAI and not started a new Frontier Lab. And, you know, I'd done that. And I wanted to do something new. And I was basically, I was captivated by this question that the models are so powerful, right? And they're really powerful now, but they were really powerful then too. And yet they don't seem to be making that big a difference in the world. You know, people love to talk about abundance as a concept in Silicon Valley. And I think it's a very interesting one because it's a very physical word, right? When you think about abundance, you're like, I have an abundance of food or housing or family or something like that. Like the way we experience the world around us is very physical. And I don't think anybody is like, I necessarily have an abundance of digital work. Yeah. And in order to actually get something done, you know, in economics, I think it's called like an O-ring production function, but like you have to do all the things.

And so if we're able to speed up the purely digital process of building housing, well, that turns out not to be very important, right? Actually, the hard part of building housing is getting the permits, which no amount of agents are going to help you do, right? But also you have to, you know, do the construction work. You know, I have a mental model that I think about for what it looks like to, you know, the impact of superintelligence. Where people say, okay, well, if we can just have sufficiently smart agents, then, you know, we're going to have Dyson Spheres in just a few years. And so, like, here's the thought experiment. Is, you know, actually, let's say we're blocked on building data centers. Like, we want to build a lot of data centers. Like, we need to go build these data centers. were slow. What is it? Well, we have a shortage of bulldozers. No problem, right? Well, the AI knows what needs to be done. It's going to go like build a bulldozer factory. So now we've got to build a lot more bulldozers. Well, great. What's the first step in building a bulldozer factory? Well, you've got to clear the ground. What do you need? Well, you need more bulldozers. Well, I think, yeah, that's interesting. I guess the comment I was going to make really quick on

all this is I totally get your point on construction. I wonder for certain domains like biology, chemistry, materials, et cetera, if you're able to build a sufficiently strong internal model that can actually simulate a lot of these things well enough. Protein folding is a great example where you really don't have to go and solve all these crystal structures anymore. You don't actually have to do the experiment anymore because the in silico prediction is so good. And so I wonder if we actually fragment what you can do in the physical world into two camps. There's a camp where you have to literally go and build the bulldozer and figure out, or build the manufacturing capability to build the next manufacturing capability or whatever it is. and then there's the ones where you can actually simulate the experiment to the point where you don't need the wet lab anymore or you minimize the number of things you need to do in the wet lab or you need to do in the materials shelf or you need to do in the sort of physics lab and so i just wonder if it's a little bit if if all these things aren't exactly the same there's two different types of problems that we're dealing with in the physical world yeah that's really interesting i think they're going to feel the same and the reason they're going to feel the same is because of the tremendous speed with which we're going to be able to do anything digital so let's

go to wet labs, for example, let's say that we build amazing, you know, digital simulations of proteins and of how they interact in a cell. And we're able to do this. And suddenly, we are able to predict 10 times better, how good what's going to happen with our experiments. And so now in order to get to a discovery, we need to run one 10th as many experiments as we needed to run. Well, we're going to be that period of time in which we are blocked on the digital world instead of the physical world, we're very fast and moving in the digital world. So we're going to eat up that acceleration. It's similar to the other thing is going to be a one time acceleration, but then we're going to be blocked on the physical world again. And then now, now we're 10 times better, but we still have to go run those experiments. I think there's an analogy that I remember from, from open AI where, you know, I was the chief research officer. Ilya was the chief scientist. And, you know, especially early on, I would read these papers and I'd be like, oh, these methods, they're so cool. And I'd go to Ilya and I'd be

like, oh, they have this idea and it works this way. And Ilya would like listen patiently for me to explain the paper to him. And at the end, he would say, so what are the results? Because, you know, even Ilya, right? It's like, yeah, sure. It's a cool idea. It could work. It could not work. But at the end of the day, you have to run the experiments. You know, I think if deep learning which is has no physical you know it's like a purely mental it's a purely in some sense a mathematical science actually it's an experimental science right and even ilia you know he can tell you yeah this is a bad idea don't even try it but there's this huge number of ideas that are that are actually very good and just applying a lot more intelligence to figure out like is it a good idea is it not a good idea is it going to help you you have to go run the experiment and then maybe you do, we're able to find proxies for this, which is great. As we find those proxies, then, you know, then we scale up to accomplish that. And then again, we're, we're banging hard on the physical world. Tell us more about what Arda is doing then specifically. Like how is Arda

approaching the physical world? What are some of the products or areas that you're focused on? Yeah. So, so let me start by, by talking about the problem we're trying to solve. So after I left open AI, um, you know, I, I sort of noticed that we are very good in America at designing products, right? So the iPhone, you know, famously designed in California, manufactured in China, it says that on the back of your iPhone. You know, we're amazing at designing defense materials, but we don't manufacture very much of it here. And so I went around, spent a year talking to manufacturing leaders and asking them, like, what's going on here? And, you know, fundamentally, it comes down to, fundamentally, it comes down to labor costs. And so if you want to build something in America, one possibility is that you're in defense, in which case you can charge whatever you want. And if you're if it works, the government will buy it. So that's great. But if you're not in defense, then you need to build an automated line. If you want to be able to build something cost competitively in the US, we're building out.

We're designing all these innovative products. One part of designing innovative product is that the product itself changes. And so famously, you know, each year the iPhone has 75% of the parts in it change. Only about 25% stay the same. So that 25% is a highly automated line. The other 75% is assembled by hand. And the reason for that is that if you were to try to build automation that can go assemble this that automation is very specific The way we build automation today it like a bespoke machine that solves one particular physical problem And that machine doesn generalize from one phone to another. And it doesn't even generalize from one generation of the iPhone to another. So going back to, you know, what we can build in America, well, if you're building a car, right? Great, build a highly automated facility, spend, you know, a couple years and $200 million to make it happen and then produce that same car

with minor cosmetic updates for seven years. The economics work great. But if you want to do that for an iPhone, first off, you have to go from about 25% automated to about 90% automated for it to pencil. And then how are you going to be able to replace all of that automation? Like that's just a huge, incredible cost that you're not going to be able to amortize unless the automation becomes more flexible. And then there's a second piece of it, which is that who is actually going to be designing this automation? There are about a million people assembling iPhones or components in Asia. There are, alongside that, about 10,000 manufacturing engineers who are figuring out what is the manufacturing process that we need to do. How do we stand up these assembly lines? How do we spin down these assembly lines when it comes time to spin up the next product? And so that's, you know, fundamentally the problem we're trying to solve is how can you make it as easy to turn over an automated line as it is to change and retrain workers in a human plant?

So you can retrain a human worker in a day. Practically, manufacturers think they're extremely agile if they can bring a design change into production, even on a manual line in seven days. We would like to be able to build a system where we are building a system that allows you to upload your product design into our cloud. and then design and ultimately reconfigure an automated line in seven days. Putting aside the fact that everybody would like to be able to make whatever they want on demand as quickly as possible with that sort of pace of innovation, like what are the types of products and customers where this matters the most from a business perspective, where that high mix and speed to bring up the line is really important to them? Because you kind of said like for a car, it's kind of okay to have the same car for seven years. It's basically anywhere that you are doing highly innovative, frequently changing manufacturing or where you have a lot of variance. Defense is, I think, a very interesting example. It's very close to my heart as a Palantir veteran.

And, you know, in Ukraine, for example, you are now going you're now spending only a few weeks from changes in the battlefield into physical changes in the design that then need to be pushed out back onto the battlefield. That is a very short amount of time, and it's driven by these competitive pressures between Ukraine and Russia. Because if your adversary can turn that around faster than you can. Exactly. Well, there are also competitive pressures, although less violent ones, between Google and Apple, or the Android ecosystem and the iOS ecosystem. And so as you are building these larger and larger markets and these more and more innovative products, What you want to be able to do is just crank up the pace of innovation. And so consumer electronics is a great example here. A lot of things that go into data centers need to be changed very frequently because chips change very frequently. I think those are the kinds of industries, defense, consumer electronics, the data centers

and data center supply chain, where you see a lot of need to be able to turn things over quickly and where as a result, almost everything is built by hand. And when you talk about automating these lines, is that human-led robotics? Is it just existing manufacturing technology repositioned? How do you think about the form that's going to take? Yeah, I mean, the true answer is that it's all of the above. But practically, what does it look like? You know, and I think the most obvious thing is that you want to be able to have reconfigurable robotic work cells that you can immediately flow a design change into an actual change that's happening in the factory. That's a hard problem. Obviously, there's a lot of people working on different variants of that. The reason I ask about it is I feel like there's two theories of the world, and I'm not saying they're incompatible. One theory of the world is everything is designed around people right now in terms of manufacturing lines, et cetera, and so therefore the most natural thing to do is just copy a person, which means have a humanoid robot. But the other viewpoint is, while there's lots of these sort of very common manufacturing processes that you need to do,

a lot of them are done by hand because, to your point, it's a faster turn if you need to turn something around quickly. And those things could just be made into sort of stepwise sequential things like what we've done in automotive for welding arms or other things like that. And so I'm a little bit curious, like, where on that spectrum you think we are in the near term versus the long term. Yeah, so I would come in with two ideas. So first off, let's talk about the form factor. So for the form factor, given what I just told you, we are trying to replace humans, right? We're not trying to replace additional forms of automated lines. We don't need to work in automotive where they have these, like, extremely large Baroque robotics. I think, you know, two arms and a camera is enough. In a lot of cases, actually, one arm and a camera is enough. And maybe you need some more, some better sensing, you know. So from a form factor perspective, these lines, you know, you have a human standing there all the time. You don't need a humanoid robot. Does the capability of the arms have to change dramatically? Because people always talk about the sensitivity and fidelity of existing robotic arms versus humans and all the rest. Like how much progress actually has to be made in terms of the componentry for the arm versus, hey, it's actually ready to go and it's just all about the software.

Existing automated lines are able to build all of the products that we're talking about, or at least like 90% of them. You know, yes, wire harnessing is hard. But, you know, if you go look at something like the iPhone and you ask what is the barrier to doing the automation, It is not that robot arms can't sufficiently screw in every screw and fiddle with everything. That kind of sensitivity already exists. It's the economics because those arms and those, they call them CAPEX, capital expenditures, those machines have to be redesigned and rebuilt every year. They can do it. It's just too expensive. They can do it. It's just too expensive. Yeah. And so I think there you want to actually switch from thinking about the embodiment to thinking about the intelligence layer. So for the intelligence layer, I would distinguish robotics as, you know, some robotics problems are open world and some robotics problems are closed world. So if you were trying to clean a bedroom, that is an open world robotics problem. You have never seen the items here, you know, in a lot of ways that we think about this.

When you're deployed there, you've never been in this bedroom before. You're never going to go back into this bedroom. It's this, like, very long horizon planning problem. That's very hard. and it's very suited to, you know, a pure, you want to think about this as, you know, the intelligence you need is very similar to human intelligence. A closed world is one in which you actually control everything and you're doing something very repetitively and there's no novelty that's being introduced. Assembly lines are a closed world. That makes sense because we see that we're actually able to use sort of traditional automation approaches in order to automate them. And I think it's worth talking a little bit about what those look like. So, you know, suppose that I am trying to solve some particular problem with automation. Again, traditional robotic automation. What they end up doing, I think the best analogy for this is a Rube Goldberg machine. So I had to build one of these with my son who's 13 for school, you know, and you're like, okay, well, you know, the ball rolls down the hill and it hits the pencil and the pencil like pops the balloon.

and, you know, instead of that, it's actually you have a chute and, you know, the nut goes down the chute and, you know, there's a cup and the cup is vibrating and because it's vibrating, the nut drops out and it lands on a fixture. Yeah. And the fixture locates it precisely. So now there's no, you know, uncertainty about where it is and a robot arm without any perception whatsoever can come and pick up the nut and then they can go take it to, you know, the bolt that you're screwing the nut on, which is also on a fixture, is also precisely located. And then it can screw that on. It's basically deterministic. It's all deterministic. It's from this coordinate to this coordinate. Yeah. So what's hard about this, right? Like, why is this interesting? Like, why does this take six months in order to turn around? Well, there's two things you have to do. One is you have made a machine. Yeah. Right? So someone needs to go design that machine. That is not an easy process. That is a mechanical engineering task. It is a manufacturing engineering task. Um, well, we now have AI that can design machines.

So that's one way we can make it faster. The other thing, uh, which, you know, really hit home to me as I was doing this Rube Goldberg project is like, you know, half the work is in the design. The other half is in the calibration, right? So, you know, you got to make sure the ball hits the pencil exactly right. So it pops the balloon. Um, you know, similarly, you know, if the nut is, you know, a 10th of a millimeter off, It may not end up in the right place on the fixture. And then, you know, the arm may not be able to actually, you know, place it in the correct place on the device. That sounds like a relatively simple RL problem. No? Well, I think there's a lot of, you know, this is, I think, where something like agentic robotics becomes an interesting way to look at the world. Because, you know, we are no longer in this open world, right? We're now in this extremely closed world doing a set of very specific motions. And you can bring in perception. And now, you know, it's, you know, if the nut's a little bit off, like you don't need to calibrate it quite so precisely, you can go pick it up.

And you can bring in RL and manipulation. And now, you know, maybe the nut's in the wrong place on the hand, for example. You know, the device is always going to be in a fixture. But now you're using actual manipulation in order to screw this in. and maybe you can start by automating the simpler kinds of stations. And then over time, you get into the ones that are really wacky where you've got elastic things that need to be stretched over a gasket. And again, though, speaking of closed world and an open world, maybe you miss a couple things in the bedroom and that's fine. In a closed world, actually like 99 or 99.9% of the time, you've got to go construct the device that you are constructing. How tall is the hill to climb to creating a great manufacturing engineer? Like, how deep is that field? How far do you think you are? You know, the starting point for us, you know, we've been talking about designing automated lines. We've spent the last year partnered with Andrel on building a software and AI platform that

allows manufacturing engineers to design lines. Now, you know, Andrel's a defense manufacturer. they don't need to have automated lines and a lot of the products they have aren't high enough rate yet to support traditional automated lines. And so what we've done is basically, so we went in there a year ago. We actually talked to them the day before we incorporated the company and they told us about their challenges. They talked about the need for a single source of truth for manufacturing engineering or one place where you actually encapsulated the manufacturing process and the manufacturing plan, and a reconfigurable factory that would allow you to change the manufacturing process and then also spin up and spin down different lines based on demand. And so we adopted those as sort of our north stars for automating manufacturing engineering. And, you know, we got access to the systems in January. In April, we had our first user. So he was able to take a actuator that he was trying to build. And normally that would have

taken him a few days in order to do the virtual build for this. And he was able to do that in 90 minutes using our software. And that was, you know, the very first thing we did. And since, you know, that worked out, that saves them a lot of time and they're very excited. And so since then, we have been moving across all of their, we're now deployed across all of their product divisions. We've done work with them, you know, on everything from drones to missiles. You know, we're currently working with them on the Fury collaborative combat airplane. So we're taking this entire airplane and, you know, taking that manufacturing process and making that legible to AI. And then, you know, a lot of what we're doing is basically you encapsulate that manufacturing process. Now it's not scattered in 100 different documents in 100 different places. And so you can have an AI agent that actually is able to add value. So, you know, going back to what does it take to be, to your question,

to what does it take in order to truly, you know, build an automated manufacturing engineer, we are lucky that we can start by saying, okay, we want to be Cursor here, effectively. Noting that Cursor basically said, okay, we're going to take this existing IDE and we're going to add an agent. There's no IDE for manufacturing engineers. So first we had to build that so that there's a true data structure. There's an equivalent of effectively Git that you can work with. And then we can apply the agent on top of that. And then now that we own and well, now that we control in a in a specific way, the manufacturing process, we can, you know, move left and right of that. So now you can say, okay, well, this design, you know, is manufactured this way. If we wanted to manufacture it more cheaply, how would we have to change the design? So we can send design for manufacturing feedback back into the design engineering, which is, again, a role of the manufacturing engineers. But they're able to use our agent to do that much faster and more effectively. And similarly, we can then say, okay, we can move right and say if we want to have an automated station, how do we design an automated station?

that can produce one of these, you know, they can do one of these manufacturing steps. Yeah, this is definitely the most pragmatic approach to this world that I've seen. And I feel like there's been a lot of iteration over time. I remember like a decade ago funding things that were going to be like Git for hardware or other things that kind of impinge on this in terms of how do you store the knowledge and then how do you act on it? And it feels like the unlock is really happening with AI. One thing that I think is really pragmatic about what you've described to me in the past in terms of the approach that you're taking is you're also not going and trying to build massive models of your own. And instead you're saying, how do we actually use things that are a bit off the shelf from an AI perspective to be able to really make this streamlined and fast and efficient and cost effective Can you talk more about that choice of like building your own models versus trying to use a broader suite of things against them Yeah well I would start but just from the perspective of simplicity, right? So, you know, when you want to go solve a problem, you want to use the simplest tool for the job. And, you know, maybe even to step back. So,

So I had one career at Palantir, one career at OpenAI. At Palantir, we were relentlessly problem-focused. We wanted to build the enterprise platform for the intelligence community, which ultimately became the enterprise platform for all enterprises. And we didn't care what technology we needed. And we had to solve a lot of very hard technological problems. But we used the forward-to-flight engineering approach. We talked to our customers. We heard what they needed. We brought that back into the product. And we just relentlessly expanded that. but by focusing on the problem. And of an AI, we took the opposite approach. We had no idea what any of this would be used for. You know, we, I remember having a conversation with Ilya once where we had just built GPT-3 and, you know, we were like, this is pretty good, but it needs to be better. And Ilya said, well, Bob, how do you actually build a company? Like, what would we need to do to make this product? I said, well, you know, there's a standard approach that everybody in startups knows, which is you pick someone who, you know, has a real problem. You actually solve their problem and then you expand from there.

And then you said, ah, yeah, that's very interesting. Yeah. No, that's not going to work for us. And I said, what? Well, yeah, I mean, thanks for thanks for asking. But why won't this work? And he's like, well, look, we can learn something by talking to one person. But instead, we have the ability to just, you know, 10x our training data and then we'll learn so much more from everybody. And that's much more than what we can learn from one customer. So I think that's that's the justification for why you want to take a technology first. approach to your market because fundamentally this technology has its own way of finding out about the world well at arda we're taking a technology first we're taking a problem first approach and so why are we not training models i've trained models before we don't need to do it um for the uh the first step of understanding the the manufacturing process designing a line from that that's something that we see that frontier models can do uh when you start going into something like robotics, then frontier models can do some of it.

But what we see is that basically the frontier models may work best as an orchestrator. And in fact, that's actually a lot of what we're doing with manufacturing engineering is that we have built specialized algorithms. We use motion planning. We have different tools that you can use to inspect the CAD. The frontier model is not adding all the value. They're an orchestrator of all of these different tools. Yeah, you actually see this a lot in the SaaS world, too, where you use the SETA model as the orchestrator, and then you actually have fine-tuned models or specific things that are very specialized for the application area. So it's interesting to see that it's also happened here. Yeah, and I think it's just like a good general purpose mechanism for thinking about things. Ultimately, if you want to have the most intelligent model, there will be a small number of very capital-intensive, very smart models. Like, you know, I don't think you're going to have specialist models unless you're thinking about cost. But when you're not a frontier lab,

it's actually extremely approachable to make, you know, small special purpose models that are better than, you know, what the off-the-shelf frontier models can do, at least for now. And training one of those models versus, like, writing a software package versus writing a tool, like, those are all kind of the same action. How did you end up at OpenAI coming from Palantir? Because to your point, you came from such a pragmatic world. And I remember visiting OpenAI in 2018, 2019. And at the time, there was a handful of robotic arms that weren't doing much in the office. And there was a bunch of, a lot of the focus, I think, was on video games and sort of beating video games initially. And then you guys started on the Transformer and sort of GPT train. But for a while, it was kind of this very broad-based research approach. Let's try a couple things. At least that's how I saw it as an external viewer. Was that a big transition? Why even join an organization that was so radically different from what you were used to? Yeah. So I would actually start the answer to this 10 years prior. So I was doing my PhD at Stanford in the AI lab.

And, you know, this from 2002 until I dropped out in 2005 to join Palantir. And I remember vividly sitting around with my research group, making fun of people who thought that neural networks would ever work. because, you know, there was no, you know, proof that they would work, you know, unlike SVMs or regression, which have clear provable guarantees. And even if they did work, you couldn't figure out why they worked. And so, you know, like, why would you ever use these things when obviously, obviously all these other things work? And so I left Palantir, I left and went to Palantir because I thought that AI wasn't really happening back in 2005. So I spent 10 years at Palantir, run engineering and product management there. And, you know, around 2015, I'm looking around. It's been a few years since AlexNet happened. Vision models are beginning to really work. Palantir has grown tremendously, and I'm, you know, thinking maybe it's time for me to move on. And, you know, I thought to myself, like, maybe AI is really working now.

And where is it going to be an application of all the vision models of people living? well, sure, you can go recognize people's faces and photos. And if you want that, you should work at Facebook, not start a company. But robotics seemed like maybe this would be the place. I was very wrong about that. I spent a year and a half trying to get this to work. I taught myself, you know, deep learning from watching Andre Kapathi's videos, you know, got to the point of teaching a robot to play checkers from Vision. And having done that, I realized this is way too hard. This is not the right time for a startup. And so I knew some people and, you know, got introduced to this nonprofit full of smart people that had the, you know, somewhat dramatic goal of building AGI. And I thought, well, this isn't a I'm not going to start a company. Why don't I go work for a nonprofit for a while? And so I started there. And, you know, I didn't think I would stay for years. I expected I would be there for a couple of years. I would learn from people like Ilya or Wojciech Zaremba or Andre Karpathy, who were there at the time.

And I ended up joining the robotics team. And the goal for the robotics team was to show cool advances in robotics in a way that demonstrated that deep learning was actually a real thing that people should pay attention to, that it might, in fact, lead to AGI. And so we settled on this project of have a humanoid robot hand, a shadow hand, be able to solve Rubik's Cube. And the reason for that, the way the story behind this is that Voychek Saramba, who's one of the co-founders of OpenAI, but also the head of the robotics team, went around at a robotics conference and asked people, what would just be absolutely mind-blowing to you if we solved it? And people said, well, either pick and place that can pick and place arbitrary objects or doing anything with a robot hand. What would be the most impressive thing you could do with a robot hand? Solve Rubik's Cube. Yeah. And so we spent two years for that, and it actually worked. although I'm now permanently not a believer in robotic hands after the experience of trying to make one do anything.

I hear the engineering is better now, but I'm probably ruined for life. If you fast forward all the way to this point where Astra and Fable look like they're showing signs of life that excite or freak out the robotics community in terms of actually being able to generate policy to do things in the real world from a general model, what do you make of that? Is that like the end of people training models specifically in general robotics? Well, you know, the end of people training models probably depends on venture capital. So, you know, that's more a question for you guys than for me. But like, why does this work, right? Like, yes, it's a bitter lesson. But fundamentally, you know, the challenge in robotics has always been that the amount of data that you have to work with is very small. And so you would see companies that, you know, we're going to solve robotics by teleoperation and we're going to build giant teleoperation labs. And, you know, then it's like, oh, well, we can use these UMIs, you know, little devices like chopsticks that record, you know, people's operations.

And then I can collect data 10 to 10 X more cheaply. All of these are extremely scarce sources of data. I think about it like if we were trying to train GPT-4, but we were doing it on top of, you know, if we had to pay scale AI to go have humans write all of the data, like we would never have really been able to train GPT-4. But fundamentally, you don't need to train the model on exactly what you're deploying on. What you need to do is generate really good representations of the world, and then you can fine-tune the models towards a particular problem. And so, you know, where do those representations of the world come from? They come from predicting the last few significant bits of, you know, some sort of real world data that you're dealing with, which could be Internet videos. I'm sure there is, in fact, a tremendous amount of robot data that has been brought in. But I also suspect it's very small compared to the fraction of overall data in a model like Astra or Fable. And so as, you know, the models progress, as the data gets better, as the algorithmic improvements change,

the model becomes better at sort of eking out those last few bits of randomness and being able to predict those. And suddenly now it's able to predict events in the physical world. And once you have those really good representations, then you can go target them at robotics. And suddenly robotics begins to work. I suspect, you know, that for a long time there was this theory of, you know, emergent capabilities that, you know, I remember watching on Twitter sometime where someone said, oh, you know, I watched GPT-4 and it can go debug. It can, like, iterate with a software, you know, you can run this program. It can, you know, look at the output and it can go fix bugs. This crazy emergent capability. And I'm like, well, Jerry spent, like, two years working on that. Right? That's, like, not as emergent as you think it is. And so, yes, we know that there are robotics teams at OpenAI, there are robotics teams at Anthropic that are doing this. But my guess is that the actual motor of progress here is not that they collected a lot more robot data than the next people,

but that the models were able to finally get down to the last bits of the physical data that they already had. What do you think that means for the overall robotics ecosystem? Do you expect like maybe that's also a question for entrepreneurs and venture capitalists, but what do you expect the player landscape to look like? When do you expect people to have robots in different use cases? Yeah, I think this really democratizes it. I think there's an interesting tension here that I don't know how to resolve, which is, you know, for digital tasks, few shot learning works very well. and also, you know, it's really easy to collect data that you can actually, that's similar in format to the data you're actually using. In robotics, you know, again, because of my suspicion that a lot of the value is not coming from the robot data in these models, I've always believed that you're probably going to want to be able to fine-tune in order to be able to work in any particular area.

And so I think there's a tension here between something like Aster or Fable. you know it's very good out of the box um it's also slow that can be overcome but you can't fine-tune it so it's very hard to fine-tune effectively on a specific use case and so i think that's an opportunity for uh other robotics players and i think if you if you look at pi or you look at generalists like you can see them you know using this opportunity that they have to take data and use it to fine-tune against a specific use case and so you know you know the bitter lesson is very strong. And also the idea of training on the test set is very strong. Yeah, right. Both of those are very powerful forces. And I think we're going to have to see where those come in. I do think this does sort of leaping out of that frame, though, is that agentic robotics, I think, is an interesting way, like sort of in between these two things, where you can think of these models as just one more tool that the agents have. If you're doing something like we are,

where you're working with the closed world, then you have a lot of different tools. Now, if there's one thing that requires an incredible amount of dexterity, maybe you put Fable on that. But maybe you don't want to do that because Fable is going to be really slow. So you don't want to do that for everything. And so I think that if you think about these, now for people who are actually deploying these, there's many more tools that they can do. And I think that's a very powerful mechanism. Yeah, I think people also talk a little bit about, to your point, sort of public slash open world versus private, more closed loop systems. And in the sort of public open world, like you have a robot in somebody's home or walking down the street with them or something else, you can't have like the lag of an Internet connection or an API call because you're worried that the robot may freeze, fall over, something could happen. And so you want the intelligence to be embodied on the device as much as possible or at least certain control mechanisms to provide safety relative to the device. And that's the other aspect that people sometimes talk about in terms of state-of-the-art models versus specialized fine-tuned models or just running on device and things like that, which is kind of interesting from a more safety perspective, I guess.

Yeah, I think that's an important thing. I mean, there's a safety perspective, and also there's just a speed, right, that you want your robotics typically to operate on a very fast clock cycle that is challenging to do with these larger models. You know, historically, that is typically solved by distilling into smaller models, whether that's something that the end user is doing or whether, you know, you have your Astra. And by the time that gets turned into a Luna model, you know, it can still work on some of the most popular use cases, but it's able to operate like very quickly and very efficiently. Actual edge on device models, I think, are also, you know, likely to be very important here. and they're going to rely, typically they end up being solved by open source that then gets improved through one way or another coming from data. So amazing run of Palantir and OpenAI. You're used to doing things relatively fast, including with difficult ecosystems. It conventional wisdom that working with large incumbents with physical world plans trying to get them to change big processes It hard right Like how do you think about the pace for Arden What possible here Yeah So there is just a huge difference between being you know developing technology for yourself as a first party

like, you know, building models for use in chat TVT and, you know, building something for use in a large enterprise. And the way that we're approaching this is that we're using the forward deployed engineering model. And so I think sort of classically, we developed this in Palantir way back in 2005. The way I like to think about it is back in 2005, I spent a lot of time in secure spaces in northern Virginia, working with people with secret workflows that couldn't leave the building and had to figure out how to build a product for them. And it turns out building a product for people that have secrets that they can't tell the world is actually really good training for building a product for any enterprise. Because any enterprise that works, like if they're working, they have secrets, and they don't want to tell the world everything about their IP. And so the outcome of that process was the forward-deployed engineering model, which I think mostly Sean Stanker, who's the CTO at Palantir, deserves the most credit for. But my job was to figure out how to build products for it.

And so what that does is you basically take your engineers and you put them, you know, embed them with the customer, I would not develop a feature unless I had a customer that was ready to use that feature. And so when I think about the product roadmap, a lot of the beginning point for the product roadmap is not, okay, I'm going to go build this feature. It's how do I find a customer that needs this feature so I can go build it? As a consequence, then you need to start embedding with, understanding, you know, seeing these enterprises. You still want to move fast, though. So how do you move fast? Well, typically, you want to start by finding a small number of power users. You know, classically, at the agency way back in the day, there was a woman named Sarah Adams that was, like, the only user of Palantir for about a year. And we learned I would go visit, and, you know, I would meet her, and, you know, she would be like, this stuff doesn't work. And I'd be like, okay, I've got to go fix this. And eventually it worked. And then suddenly lots more people used it, right?

But it's the power of that just very small number of very strong users that are willing to fight through the pain with you. You know, I told the story about our first user for Arda, our first user at Anderil of a sort of lead manufacturing engineer who is trying to do the virtual build for an actuator. Well, what I didn't say is that this was a tremendously frustrating process for him being the first user. and he was able to make, you know, tremendous improvements and be able to move very fast with that actuator. But it only works because our head of product manager was sitting right next to him, you know, and every time he did something, it didn't quite work the way he wanted it to. And, you know, our head of product manager was just sitting there like taking notes. And then two weeks later, all those things worked. Thank you, using number one. Exactly, exactly. And so, but like, it's not like this process ends because, you know, six months later, we're at Anderil, they're like, okay, go see if we can figure out how to build missiles faster. Well, this only works if we find that power user who's on the missiles team,

who's now interested in working on a product that has a hundred times more parts than the actuator, two orders and attitudes, more parts than the actuator that we worked with. Yeah. Once again, out of the box, the product doesn't quite work. You find that user, you, you, with that user, you fight through to make it work. And then they can introduce it to everybody else. And so you're, you know, classically, when you're doing enterprise sales, you want to find that visionary early on. You know, this is from Crossing the Chasm, the classic business book. Classically, that visionary is a sponsor, and you do need that sponsor. But what the book doesn't talk about is you actually also need a user, someone who's actually willing to use your software to go accomplish an outcome. And if you do both these things, then even very large organizations will work on their problems faster than you expect. Well, there's not really any such thing as a very large organization. Every, you know, people will be like, well, what does OpenAI want? It's like OpenAI doesn't want anything. Sam wants something.

Greg wants something. Like, you know, this person down here, this research scientist on this team wants something. And big organizations are just an assemblage of very small teams. And so, you know, you have to proceed problem by problem, team by team. you know, one of the really interesting ideas in the FDE approach is you're not just throwing the software over the wall to them. You are partnering with them to achieve an outcome. For example, you know, figure out how to build missiles faster. And you are not successful unless the customer is successful. And once you achieve that outcome, you can then look around at all the other outcomes, and you've earned the right to work on something bigger. In our case, maybe that's we go work on an airplane, we go work on the Fury with the Underworld, right? But also it could be we work on an entirely different problem that's bigger in terms of dollars for them. It moves the needle more for them. And maybe they wouldn't be willing to trust you, to trust your startup, to solve that problem unless you had demonstrated some form of excellence. And so, you know, really within a lot of this within an enterprise is snowballing, you know,

the success as you've made, snowballing the outcomes you can achieve while still baking that into your product. Because I think the thing that a lot of people who try the FDE approach now get wrong is that they see this as, okay, well, I'm just deploying some engineers at the site, and they're going to make the product work at the site. Somewhat famously, like, people didn't believe the FDE approach worked, right? Like, it was widely mocked in, you know, 2015 to 2020. Folks thought there would never be leverage. That's right, yeah. These are just consultants. Like, oh, Palantir just sends a bunch of consultants to do things. And then people started doing it when Palantir became a $300 billion company and started profit. And the difference is that they weren't just consultants, right? The FDE strategy is a product strategy that teaches you how to build products. The FDEs are the gap between the product and the user, or they're filling the gap between the product and the user. And if they have to keep filling the gap, you get no leverage. Instead, you have to productize what they're seeing and bake that back into what you're doing. And so that's what allows you to move quickly and discover new things and bake those into your products.

So if we take all this to the extreme and you have an army of FTEs over time who are deploying across multiple industries, you're learning very quickly, you're iterating, you're building a lot of products that then gets reused at scale. Where does this all go? Like, what is your vision for the endpoint of all this? Yeah, well, I would go back to think about superintelligence at the beginning. So we started by talking about, you know, what it looks like to have digital only superintelligence. And, you know, yes, the need for laying out lots of compute, that's a requirement for superintelligence. But superintelligence also has consequences. So, you know, I love Dario's phrase of, you know, a country of geniuses in a data center. But if all they can do is sit in that data center and think, and like you can send them emails and they can send you emails back, like they're not going to be able to do very much. They're kind of trapped in the data center. And so, okay, so intelligence has become cheap. What are the complements to intelligence? Well, perception on the one hand and action. as we have all these AGI's, you know, right now, humans are the eyes and hands of the intelligence

that's in the data center, right? And instead, what I think we're going to be seeing over the next 10 years is this tremendous profusion of, you know, various different kinds of perception, various different kinds of manipulation, you know, there'll be, you know, drones and sidewalk robots, Every physical thing that we own will become, you know, perception enabled and hooked up to the cloud. And, you know, whatever it does will become controlled by AI. And then all sorts of things that we want to do, we will be doing because there are robots that are out there doing these things for you. Some people think that they will just be an army of humanoid robots. I think instead we're going to see just this tremendous profusion of different forms. We think that, you know, the world is made for humans. But in many ways, you know, if you look at sort of the broader world, it's actually, you know, we'll probably see robots that begin to look like all different kinds of animals, right? There will be robots that fly. There will be robots on four legs. There will be robots that carry burdens.

There will be robots that, you know, plow fields. I think we already have those, right, with tractors. and so the part of the story of superintelligence is not just creating superintelligence and having it live in the data center and maybe scheme about how to take over the world you want it to actually be able to go out of the data center and help people and do things for us and so that I think is the call to action for manufacturing and why what we're doing manufacturing at large is doing in this world of superintelligence. You know, von Neumann, and I don't remember when he came up with this, he came up with von Neumann probes, right? These self-replicating probes that you send out of space, they land on an asteroid, they mine it, or they land on a planet, they mine it. They produce more versions of themselves and they keep spreading. Yeah. And so it's the way to effectively explore the universe at scale in a sort of distributed system way where you have things self-assemble and then keep going. And it's exponential, right?

You land, you make 10 copies, those 10 copies go off and make 10 more copies and you kind of spread through the universe. do you view that as an eventual manufacturing model on our planet? In other words, eventually do you have these self-replicating factories? Or what do you kind of view as the end state from an almost physical form perspective? Let's narrow in and think about this concretely, right? So maybe you have a problem that you need to go build more bulldozers, and so you need to build a bulldozer factory, right? Well, right there you have a recursive loop where the right way to do this is that you're deploying your intelligence to go manufacture more things. And so, you know, I think very, very quickly we will see, you know, yes, factories that build other factories. But, like, that's machine tools, right? Like, that's what we have, right? But the difference is doing this in an intelligent way where, you know, it is a tremendous amount of human work to be able to go build a factory. It is a tremendous amount of human work to be able to design machine tools for that factory.

All of that, I think, is going to become increasingly automated, and it will have to in order to be able to support the vast network of things that we need in order to... It's basically the digital singularity versus the physical singularity. Yeah, and not even versus, just alongside. Like, part of the digital singularity involves converting increasing fractions of the Earth's crust into data centers. Yeah. Right? And maybe those data centers are going to be in space, too. And, like, that is a physical process. process. And it's going to have to be directed by AI, just like increasing human civilization across the world was directed by human intelligence. So I think that's a very near-term outcome. You know, that's like what we're doing in the next 10 years as humanity. And then in the longer run, I think, well, once you have self, you know, machines can self-replicate and they are a better form factor for spreading between the stars than humanity is. And so it seems almost certain that it will happen.

And so, you know, then you want to make sure that this is a thing that's happening that is that is liberating humanity, that, you know, humanity is benefiting from that humanity remains as much as possible in control of rather than something that, you know, where humanity is a little sideline that just remains here on Earth and, you know, watches a lot of videos, which are probably will also be generated by AI. They're going out there to terraform for us so that we may also explore the stars. Exactly. I think we will also have a role to play if we choose to take it in getting to the stars. Why do you think, I mean, this is back to the Fermi paradox, right? To some extent, AI and superintelligence and AGI and these self-replicating factories, they all suggest we should have come across other intelligent life sooner, actually, right? In other words, it increases the likelihood of running across it than if you just had, you know, our biological forms trying to make it through space on our own.

Actually, I disagree. So I think this is the resolution of the Fermi paradox, which is, if you think about it, what is the most effective, you know, civilization spreading through space? It's one that, you know, goes to a star system, very quickly sets up a factory, builds a lot of self-replicating probes, sends those out. And so you have basically a wave of civilization spreading through space at the speed of light. And so why don't we see that? Because it's spreading through space at the speed of light. I see. And, you know, we will only see it, you know, a few minutes or perhaps a few years before it arrives. Isn't the argument that it should have arrived by now? In other words, there's arguments in terms of what's the density of potential civilizations throughout space. and the rate at which they arise, and then therefore if you send something out at the, you know, sub-speed of light, what is the rate at which it would eventually get to you, et cetera.

And so that would suggest sparseness in the universe, right, in terms of... Well, I think, I mean, I think this is a mystery, right, is like, why are we the first ones? You know, and there's some evidence that says if you look at sort of the uncertainty of all the different factors in the Drake equation, then you end up seeing that, okay, well, maybe actually it's quite unlikely. and so it is plausible that we're the first civilization yeah but i think or there's very far away yeah or that it's very far away right yeah it's like somewhere out there in the universe but like you know our light tones intersecting you know is going to take some time but i think you know to some extent evidence shows that this hasn't happened right and then you know the question of are we going to run across you know alien civilizations i think this argument says no we're not. We're only going to detect them, you know, a little bit before they arrive. Or perhaps we will be the wave. Hopefully we will be the wave of civilization spreading out through the galaxy at light speed. Amazing. Thank you for joining us today. Thanks for having me.

This was fun. Find us on Twitter at NoPriorsPod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-buyers.com.

番組の概要欄(原文)

Is the AI singularity actually coming? Bob McGrew, former Chief Research Officer at OpenAI, argues that superintelligence won’t arrive without changes in the physical world, and that the “hard takeoff” argument is wrong. Sarah Guo and Elad Gil sit down with Bob, an early key executive at Palantir who later led research at OpenAI. He is now CEO and co-founder of Arda, which is leveraging AI to make American manufacturing competitive. Bob explains why automating AI researchers is a one-time boost rather than a runaway loop, and why GPUs, compute, wet labs, and factories are the real bottlenecks.  They cover Arda’s work with Anduril on everything from actuators to missiles to the Fury combat aircraft, why Arda is building “Cursor for manufacturing engineers”, and how the forward deployed engineering model really works. Plus, the big-picture future view: machines that make machines, self-replicating factories, and Bob’s answer to the Fermi paradox. Sign up for new podcasts every week. Email feedback to show@no-priors.com Follow us on Twitter: @NoPriorsPod | @Saranormous | @EladGil | @bobmcgrewai | @Arda_Industries Chapters: 00:00 – Trailer 00:05 – Bob McGrew Introduction 01:05 – Why work on the physical world during the AI revolution? 02:27 – Why the hard takeoff argument is wrong 05:48 – Labs or startups – who will win in robotics and biology? 13:00 – Why America designs products but doesn’t build them 16:33 – Who needs flexible automation: defense, consumer electronics, data centers 18:19 – Humanoid robots vs. robotic work cells 21:07 – Open-world vs. closed-world robotics 25:20 – Building with Anduril: from days to 90 minutes 28:59 – Palantir’s problem-first approach vs. OpenAI’s technology-first approach 33:15 – From dropping out of Stanford to joining OpenAI 37:13 – Why general robotics models suddenly work 40:28 – Implications for the robotics ecosystem 42:46 – On-device models, speed, and safety 44:18 – How forward deployed engineering really works 50:57 – Superintelligence out of the data center 53:41 – Self-replicating factories and Von Neumann probes 56:46 – Bob’s answer to the Fermi paradox 59:33 – Conclusion

出典・この記事について

番組
No Priors: Artificial Intelligence | Technology | Startups(Conviction )
元の回
Superintelligence Needs the Physical World with CEO and Co-Founder of Arda Bob McGrew
配信日
要約作成日

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

X でシェアSpotify で聴くApple Podcasts で聴く

関連エピソード