Bob McGrew, CEO and co-founder of Arda and former Chief Research Officer at OpenAI, joins hosts Sarah Guo and Elad Gil to argue that superintelligence cannot be achieved through digital advances alone—it requires transforming the physical world. McGrew explains why he left frontier AI labs to tackle American manufacturing, how Arda is building an AI platform to make automated production lines as easy to reconfigure as retraining a human worker, and why the "hard takeoff" theory of recursive self-improvement is wrong.
> "I don't think you are actually going to be able to get to superintelligence without changes in the physical world." — Bob McGrew
> "If you look around outside, how different is anything from 10 years ago before AI happened? Not very different at all." — Bob McGrew
> "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." — Bob McGrew
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 [email protected]
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
Sarah Guo
If you look around outside, how different is anything from 10 years ago before AI happened? Not very different at all.
Bob McGrew
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.
Elad Gil
We need to break out of this pure focus on the digital world and go to the physical world.
Bob McGrew
I don't think you are actually going to be able to get to superintelligence without changes in the physical world.
Sarah Guo
Hi, listeners. Welcome back to No Priors. Today, Elad and I are here with Bob McGrew, the CEO and co-founder of Arda. Bob was previously an early key 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, and why the future is going to be machines that can make machines and explore the stars. Welcome, Bob.
Bob McGrew
Thanks so much for joining us today. No priors. Great to be here.
Elad Gil
So you've had a ridiculous career. You were the chief research officer at OpenAI. So you drove all those research directions for the company, or at least set 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 and will be displaced by code and AI and AI building itself. Why do anything in the physical world right now as this digital revolution is happening?
Bob McGrew
Well, fundamentally, I don't think you are actually going to be able to get to superintelligence without changes in the physical world. We've seen a huge amount of improvement from ChatGPT, from agents. All of this is happening in our digital lives, and very little of that has actually affected the atoms of our lives. If you look around outside, how different is anything from 10 years ago before AI happened? Not very different at all. Even when you're doing something that is genuinely very different, maybe you're in a self-driving car. You're in a Waymo going around San Francisco. The first time you were in your Waymo, you thought to yourself, this is amazing. There's no one driving it. The car is driving itself. And probably the third time you're in your Waymo, you're like, traffic. Man, this traffic sucks. Right? Even a Waymo is like halfway towards the physical world. But you didn't change the car. The car doesn't fly. You're still bounded by all of the physical barriers that made car trips around San Francisco pretty slow, even in the 1990s. So fundamentally, we need to break out of this pure focus on the digital world and go to the physical world.
Elad Gil
When you say you can't have superintelligence without changing the physical world, what does that actually mean? What does that substantiate as?
Bob McGrew
A good place to start is there's this common perception or argument that I've been hearing since like 2018, which says that if you can just automate the work of the researcher, you're going to be able to unleash a runaway acceleration, runaway self-improvement loop, and we're going to have a singularity.
Elad Gil
This is the hard takeoff.
Bob McGrew
Hard takeoff. Exactly. And I think this is wrong. I think this argument is just straightforwardly wrong. To recapitulate the argument, it says, okay, we can automate 10 percent of a researcher's time. That's going to give some improvement. We automate 50 percent of a researcher's time. Now you have a 2x improvement. So now you're going 2x faster. Then you automate 90 percent of a researcher's time, you're going 10x faster. 99 percent automated, you're going 100x faster. So by automating more and more of this job and eliminating this labor of the researcher, you're able to just zoom up and have runaway acceleration.
But this is straightforwardly wrong because researchers are not the bottleneck. The amount of time it takes to discover something is not primarily bounded by the number of researchers you have. Fundamentally, those researchers are running experiments. Even two years ago in 2024, when I was chief research officer at OpenAI, 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. Was the compute accumulated in the right places? Was it connected together with the right networks?
You can completely automate researchers. This is a big effect. I'm not saying this is not a big deal. Maybe the ability to do automated research is going to be as impactful as the ability to go from GPT-4 without reasoning to O3 with reasoning and tools. That was a huge one-time acceleration, but it's not RSI. The right way to think about research is that there's a bunch of experiments. You have to run these experiments. Right now there's a lot of waste. Researchers have to do a lot of engineering. They have to pick the right experiments. They have to go physically schedule them on the GPUs. All of these lead to the GPUs being less effective at delivering new discoveries.
The correct mental model is to look at the GPUs and say, right now the GPUs are 35 percent of the way to optimal in order to be able to deliver discoveries. If we can automate more and more, maybe we can get them to 90 percent of the way to optimal. But the difference between going from 90 to 99 percent of optimal is actually not a big difference. 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 get closer to optimum. And then we have to go do something else.
Elad Gil
I feel like a lot of AI-driven research is basically stacking S-curves of different types, or AI development 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 the primary limiter on progress. You have to run the biology experiment. You have to build the motors for the 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 this physical world limitation that you have to overcome or at least optimize against in parallel? Will robotics and biology aggregate to the Anthropics and OpenAIs and Googles of the world and their subsidiaries like Isomorphic? Or do you think there'll be a bunch of startups that drive this sort of progress?
Bob McGrew
Before we talk about market structure, let's think about what the actual economic structure of this is. The labs, if we think about what they're doing, they're really producing cheap intelligence. What are the complements to intelligence? If we think about biology, or running these experiments for superintelligence research, 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 or laying out the GPUs. When hypotheses become cheap, then it's actually the physical part that becomes scarce. 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. So go back to the market structure. I think it could be either way. But do Anthropic or OpenAI have a comparative advantage in building out wet labs? I think they probably don't.
Elad Gil
Did you consider working in other domains before you decided to go work on manufacturing and robotics?
Bob McGrew
I did. When I left OpenAI, the one thing I knew was I didn't want to start another frontier lab. Maybe I wanted to be the only person that's ever left OpenAI and not started a new frontier lab. I'd done that and I wanted to do something new. I was captivated by this question that the models are so powerful. 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. Why is that? People love to talk about abundance as a concept in Silicon Valley. I think it's a very interesting one because it's a very physical word.
When you think about abundance, you're like, I have an abundance of food or housing or family. The way we experience the world around us is very physical. I don't think anybody is like, I necessarily have an abundance of digital work. In order to actually get something done, in economics it's called an O-ring production function, but you have to do all the things. If we're able to speed up the purely digital process of building housing, that turns out not to be very important. Actually, the hard part of building housing is getting the permits, which no amount of agents are going to help you do.
But also you have to do the construction work. I have a mental model that I think about for what the impact of superintelligence looks like. People say, if we can just have sufficiently smart agents, then we're going to have Dyson Spheres in just a few years. Here's a thought experiment. Let's say we're blocked on building data centers. We want to build a lot of data centers. We need to go build these data centers. They were slow. What is it? We have a shortage of bulldozers. No problem. The AI knows what needs to be done. It's going to go build a bulldozer factory. So now we got to build a lot more bulldozers.
Great. What's the first step in building a bulldozer factory? You got to clear the ground. What do you need? You need more bulldozers.
Elad Gil
I totally get your point on construction. I wonder for certain domains like biology, chemistry, materials, 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. I wonder if you 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 the manufacturing capability to build the next manufacturing capability. 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 in the materials shelf, or in the physics lab. I just wonder 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.
Bob McGrew
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. Let's go to wet labs. Let's say that we build amazing digital simulations of proteins and of how they interact in a cell. And we're able to predict 10 times better what's going to happen with our experiments. So now in order to get to a discovery, we need to run one tenth as many experiments as we needed to run. 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 in moving in the digital world. So we're going to eat up that acceleration. It's similar to the other thing. It's going to be a one-time acceleration, but then we're going to be blocked on the physical world again. And now we're 10 times better, but we still have to go run those experiments. There's an analogy that I remember from OpenAI. I was the chief research officer, Ilya was the chief scientist. 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, they have this idea and it works this way.
And Ilya would listen patiently for me to explain the paper to him. And at the end he would say, so what are the results? Because even Ilya, 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. Deep learning, which has no physical component, it's a purely mental, purely mathematical science. Actually, it's an experimental science. Even Ilya, he can tell you, yeah, this is a bad idea. Don't even try it. But there's a huge number of ideas that are actually very good. And just applying a lot more intelligence to figure out, is it a good idea or is it not a good idea?
Is it going to help you? You have to go run the experiment. And then maybe we're able to find proxies for this, which is great. As we find those proxies, then we scale up to accomplish that. And then again, we're banging hard on this physical world.
Sarah Guo
Tell us more about what Arda is doing specifically. How is Arda approaching the physical world? What are some of the products or areas that you're focused on?
Bob McGrew
Let me start by talking about the problem we're trying to solve. After I left OpenAI, I noticed that we are very good in America at designing products. The iPhone, famously designed in California, manufactured in China. It says that on the back of your iPhone. We're amazing at designing defense materials, but we don't manufacture very much of it here. I spent a year talking to manufacturing leaders and asking them, what's going on here? Fundamentally, it comes down to labor costs. 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.
If it works, the government will buy it. 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 designing all these innovative products. One part of designing an innovative product is that the product itself changes. Famously, each year the iPhone has 75 percent of the parts in it change, only about 25 percent stay the same. That 25 percent is a highly automated line. The other 75 percent is assembled by hand. 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's a bespoke machine that solves one particular physical problem. That machine doesn't generalize from one phone to another, and it doesn't even generalize from one generation of the iPhone to another. Going back to what we can build in America, if you're building a car, great. Build a highly automated facility, spend 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 percent automated to about 90 percent automated for it to pencil.
And then how are you going to be able to replace all of that automation? That's a huge, incredible cost that you're not going to be able to amortize unless the automation becomes more flexible. There's a second piece of it. Who is actually going to be designing this automation? There are about a million people assembling iPhones or components in Asia. Alongside that, there are 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?
That's fundamentally the problem we're trying to solve. 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? 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 that allows you to upload your product design into our cloud and then design and ultimately reconfigure an automated line in seven days.
Sarah Guo
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, 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 for a car, it's kind of okay to have the same car for seven years.
Bob McGrew
Basically, it's anywhere that you are doing highly innovative, frequently changing manufacturing or where you have a lot of variants. Defense is a very interesting example. It's very close to my heart as a Palantir veteran. In Ukraine, for example, you are 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. It's driven by competitive pressures between Ukraine and Russia. If your adversary can turn that around faster than you can.
Elad Gil
Well, there are also competitive pressures, although less violent ones, between Google and Apple or the Android ecosystem and the iOS ecosystem. 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. Consumer electronics is a great example. 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.
Sarah Guo
When you talk about automating these lines, is that humanoid robotics? Is it just existing manufacturing technology repositioned? How do you think about the form that's going to take?
Bob McGrew
The true answer is that it's all of the above. But practically, what does it look like? 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.
Elad Gil
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 is everything is designed around people right now in terms of manufacturing lines. Therefore, the most natural thing to do is just copy a person, which means have a humanoid robot. The other viewpoint is, there's lots of these very common manufacturing processes that you need to do. A lot of them are done by hand because it's a faster turn if you need to turn something around quickly. Those things could just be made into stepwise sequential things like what we've done in automotive for welding arms or other things like that. I'm curious where on that spectrum you think we are in the near term versus the long term.
Bob McGrew
I would come in with two ideas. First off, let's talk about the form factor. For the form factor, given what I just told you, we are trying to replace humans. We're not trying to replace additional forms of automated lines. We don't need to work in automotive where they have extremely large Baroque robotics. I think 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 better sensing. From a form factor perspective, these lines, you have a human standing there all the time. You don't need a humanoid robot.
Elad Gil
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. How much progress actually has to be made in terms of the componentry for the arm versus it's actually ready to go and it's just all about the software?
Bob McGrew
Automated, existing automated lines are able to build all of the products that we're talking about, or at least like 90 percent of them. Yes, wire harnessing is hard. But 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, they call them CapEx, capital expenditures. Those machines have to be redesigned and rebuilt every year. I think they can do it. It's just too expensive.
I think there you want to actually switch from thinking about the embodiment to thinking about the intelligence layer. For the intelligence layer, I would distinguish robotics as some robotics problems are open world and some robotics problems are closed world. If you were trying to clean a bedroom, that is an open world robotics problem. You have never seen the items here. 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 a very long horizon planning problem. That's very hard and it's very suited to a pure intelligence that 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 traditional automation approaches in order to automate them. I think it's worth talking a little bit about what those look like. Suppose that I am trying to solve some particular problem with automation, traditional robotic automation. What they end up doing, I think the best analogy for this is a Rube Goldberg machine. I had to build one of these with my son who's 13 for school.
You're like, okay, the ball rolls down the hill and it hits the pencil and the pencil pops the balloon. Instead of that, you have a chute and the nut goes down the chute and there's a cup and the cup is vibrating and because it's vibrating the nut drops out and it lands on a fixture. The fixture locates it precisely. So now there's no uncertainty about where it is. A robot arm without any perception whatsoever can come and pick up the nut. Then they can go take it to 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 all deterministic. It's moved from this coordinate to this coordinate. What's hard about this? Why does this take six months in order to turn around? There's two things you have to do. One is you have made a machine. 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. We now have AI that can design machines. So that's one way we can make it faster. The other thing, which really hit home to me as I was doing this Rube Goldberg project, is half the work is in the design. The other half is in the calibration.
You got to make sure the ball hits the pencil exactly right so it pops the balloon. Similarly, if the nut is a tenth of a millimeter off, it may not end up in the right place on the fixture. And then the arm may not be able to actually place it in the correct place on the device.
Elad Gil
That sounds like a relatively simple RL problem.
Bob McGrew
I think this is where something like agentic robotics becomes an interesting way to look at the world. We are no longer in this open world. We're now in this extremely closed world doing a set of very specific motions. You can bring in perception. And now if the nut's a little bit off, you don't need to calibrate it quite so precisely. You can go pick it up. You can bring in RL and manipulation. And now maybe the nut's in the wrong place on the hand. The device is always going to be in a fixture. But now you're using actual manipulation in order to screw this in. 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. Speaking of closed world, in an open world, maybe you miss a couple of things in the bedroom and that's fine. In a closed world, 99.9 percent of the time, you got to go construct the device that you were constructing.
Elad Gil
How tall is the hill to climb to creating a great manufacturing engineer? How deep is that field? How far do you think you are?
Bob McGrew
The starting point for us, we've been talking about designing automated lines. We've spent the last year partnered with Anduril on building a software and AI platform that allows manufacturing engineers to design lines. Anduril is a defense manufacturer. They don't need to have automated lines. A lot of the products they have aren't high enough rate yet to support traditional automated lines. 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, 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 spin up and spin down different lines based on demand. We adopted those as our North stars for automating manufacturing engineering.
We got access to their systems in January. In April, we had our first user. He was able to take an actuator that he was trying to build. Normally that would have taken him a few days to do the virtual build. He was able to do that in 90 minutes using our software. That was the very first thing we did. Since then, we have been moving across all of their product divisions. We're now deployed across all of their product divisions. We've done work with them on everything from drones to missiles. We're currently working with them on the Fury collaborative combat airplane. We're taking this entire airplane and taking that manufacturing process and making it legible to AI.
A lot of what we're doing is encapsulating that manufacturing process. Now it's not scattered in 100 different documents in 100 different places. You can have an AI agent that actually is able to add value. Going back to what does it take to build an automated manufacturing engineer, we are lucky that we can start by saying we want to be Cursor here. Cursor basically said we're going to take this existing IDE and add an agent. There's no IDE for manufacturing engineers. So first we had to build that so there's a true data structure, an equivalent of Git that you can work with. Then we can apply the agent on top of that.
Now that we control the manufacturing process in a specific way, we can move left and right of that. We can say, this design is manufactured this way. If we wanted to manufacture it more cheaply, how would we have to change the design? We can send design for manufacturing feedback back into design engineering. That's a role of the manufacturing engineers, but they're able to use our agent to do that much faster and more effectively. Similarly, 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 manufacturing steps?
Elad Gil
This is definitely the most pragmatic approach to this world that I've seen. I feel like there's been a lot of iteration over time. I remember a decade ago funding things that were going to be Git for hardware or other things that impinge on this in terms of how do you store the knowledge and then how do you act on it. It feels like the unlock is really happening with AI. One thing that is really pragmatic about what you've described is you're also not going and trying to build massive models of your own. Instead you're saying, how do we use things that are a bit off the shelf from an AI perspective to make this streamlined and fast and efficient and cost effective? Could you talk more about that choice of building your own models versus trying to use a broader suite of things?
Bob McGrew
From the perspective of simplicity, when you want to go solve a problem, you want to use the simplest tool for the job. Maybe to step back, 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. We didn't care what technology we needed. We used the forward-deployed engineering approach. We talked to our customers, heard what they needed, brought that back into the product, and relentlessly expanded that.
At OpenAI, we took the opposite approach. We had no idea what any of this would be used for. I remember having a conversation with Ilya once where we had just built GPT-3 and we were like, this is pretty good, but it needs to be better. Ilya said, "Bob, how do you actually build a company? What would we need to do to make this product?" I said, "There's a standard approach that everybody in startups knows. You pick someone who has a real problem, you actually solve their problem, and you expand from there." Ilya said, "Ah, yeah, that's very interesting. No, that's not going to work for us." I said, "Why won't this work?" He said, "We could learn something by talking to one person, but instead we have the ability to just 10X our training data. We'll learn so much more from everybody, and that's much more than what we can learn from one customer." That's the justification for why you want to take a technology-first approach to your market. The technology has its own way of finding out about the world.
At Arda, we're taking a problem-first approach. Why are we not training models? I've trained models before. We don't need to do it. For the first step of understanding the manufacturing process and designing a line from that, frontier models can do that. When you start going into something like robotics, frontier models can do some of it. What we see is that the frontier models may work best as an orchestrator. That's actually a lot of what we're doing with manufacturing engineering. 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 are an orchestrator of all of these different tools.
Elad Gil
You actually see this a lot in the SaaS world too, where you use the soda model as the orchestrator, and then you have fine-tuned models or specific things that are very specialized for the application area. It's interesting to see that it's also happened here.
Bob McGrew
I think it's 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. I don't think you're going to have specialist models unless you're thinking about cost. When you're not a frontier lab, it's actually extremely approachable to make small special purpose models that are better than what the off the shelf frontier models can do, at least for now. Training one of those models versus writing a software package versus writing a tool, those are all kind of the same action.
Elad Gil
How did you end up at OpenAI coming from Palantir? You came from such a pragmatic world. I remember visiting OpenAI in 2018, 2019. At the time there was a handful of robotic arms that weren't doing much in the office. A lot of the focus was on video games and beating video games initially. Then you guys started on the transformer and the GPT train. For a while, it was this very broad-based research approach. Was that a big transition? Why even join an organization that was so radically different from what you were used to?
Bob McGrew
I would start the answer to this 10 years prior. I was doing my PhD at Stanford in the AI lab from 2002 until I dropped out in 2005 to join Palantir. I remember sitting around with my research group, making fun of people who thought that neural networks would ever work because there was no proof that they would work, unlike SVM or regression which have clear provable guarantees. Even if they did work, you couldn't figure out why they worked. Why would you ever use these things when all these other things work? I left and went to Palantir because I thought AI wasn't really happening back in 2005.
I spent 10 years at Palantir running engineering and product management. 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. I'm thinking maybe it's time for me to move on. I thought to myself, maybe AI is really working now. Where is it going to be an application of all the vision models? You can go recognize people's faces in photos. If you want that, you should work at Facebook, not start a company. 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 deep learning from watching Andrej Karpathy's videos. I got to the point of teaching a robot to play checkers from vision. Having done that, I realized this is way too hard. This is not the right time for a startup.
I knew some people and got introduced to this nonprofit full of smart people that had the dramatic goal of building AGI. I thought, I'm not going to start a company. Why don't I go work for a nonprofit for a while? I started there. I didn't think I would stay for years. I expected I would be there for a couple of years, learn from people like Ilya or Wojciech Zaremba or Andrej Karpathy who were there at the time. I ended up joining the robotics team. The goal for the robotics team was to show cool advances in robotics in a way that demonstrated deep learning was actually a real thing that people should pay attention to, that it might lead to AGI.
We settled on this project of having a humanoid robot hand, a shadow hand, be able to solve Rubik's Cube. The story behind this is that Wojciech Zaremba, one of the co-founders of OpenAI and head of the robotics team, went around at a robotics conference and asked people, "What would be absolutely mind-blowing to you if we solved it?" People said 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?" "I don't know, solve Rubik's cube." We spent two years on that and it actually worked.
I am 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.
Elad Gil
If you fast forward 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 the end of people training models specifically in general robotics?
Bob McGrew
The end of people training models probably depends on venture capital. That's more a question for you guys than for me. But why does this work? It's a bitter lesson. The challenge in robotics has always been that the amount of data you have to work with is very small. You would see companies saying we're going to solve robotics by teleoperation and build giant teleoperation labs. Then it's like, we can use these UMIs, little devices like chopsticks that record people's operations, and collect data 10x more cheaply. All of these are extremely scarce sources of data. If we were trying to train GPT-4 but we had to pay Scale AI to have humans write all of the data, we would never have been able to train GPT-4.
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. Then you can fine tune the models towards a particular problem. Where do those representations of the world come from? They come from predicting the last few significant bits of some sort of real world data that you're dealing with, which could be internet videos. I'm sure there is a tremendous amount of robot data that has been brought in, but I suspect it's very small compared to the fraction of overall data in a model like Astra or Fable. As the models progress, as the data gets better, as the algorithmic improvements change, the model becomes better at eking out those last few bits of randomness and being able to predict those. Suddenly it's able to predict events in the physical world. Once you have those really good representations, you can target them at robotics. Suddenly robotics begins to work.
For a long time, there was this theory of emergent capabilities. I remember watching on Twitter where someone said, "I watched GPT-4 and it can go debug, it can iterate with software, you can run this program, it can look at the output and go fix bugs. This crazy emergent capability." I'm like, Jerry spent like two years working on that. That's not as emergent as you think it is. There are robotics teams at OpenAI, there are robotics teams at Anthropic that are doing this. 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 finally able to get down to the last bits of the physical data that they already had.
Elad Gil
What do you think that means for the overall robotics ecosystem? What do you expect the player landscape to look like? When do you expect people to have robots in different use cases?
Bob McGrew
I think this really democratizes it. There's an interesting tension here that I don't know how to resolve. For digital tasks, few shot learning works very well. It's really easy to collect data that is similar in format to the data you're actually using. In robotics, because of my suspicion that a lot of the value is not coming from the robot data in these models, I've always believed you're probably going to want to be able to fine tune in order to work in any particular area. There's a tension here between something like Astra or Fable. It's very good out of the box. It's also slow, that can be overcome, but you can't fine tune it. It's very hard to fine tune effectively on a specific use case. That's an opportunity for other robotics players. If you look at Pi or you look at generalists, you can see them using this opportunity to take data and use it to fine tune against a specific use case. The bitter lesson is very strong. The idea of training on the test set is very strong. Both of those are very powerful forces. We're going to have to see where those come in.
Leaping out of that frame, agentic robotics is an interesting way in between these two things. You can think of these models as just one more tool that the agents have. If you're doing something like we are, working with the closed world, then you have a lot of different tools. 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. For people who are actually deploying these, there are many more tools that they can use. That's a very powerful mechanism.
Elad Gil
People also talk about public versus open world versus private, more closed loop systems. In a public open world, you have a robot in somebody's home or walking down the street. You can't have the lag of an internet connection or an API call because you're worried the robot may freeze, fall over, something could happen. You want the intelligence to be embodied on the device as much as possible, or at least certain control mechanisms to provide safety. That's the other aspect people talk about in terms of state of the art models versus specialized fine tuned models running on device. That's interesting from a safety perspective.
Bob McGrew
There's a safety perspective and there's also just speed. You want your robotics to operate on a very fast clock cycle that is challenging to do with these larger models. Historically, that is typically solved by distilling into smaller models. Whether that's something the end user is doing or whether you have your Astra and by the time that gets turned into a Luna model, it can still work on some of the most popular use cases, but it's able to operate very quickly and efficiently. Actual edge on device models are also likely to be very important here. They typically end up being solved by open source that then gets improved through data from the big labs.
Elad Gil
Amazing run at Palantir and OpenAI. You're used to doing things relatively fast, including with difficult ecosystems. Conventional wisdom is that working with large incumbents with physical world plans, trying to get them to change big processes, is hard. How do you think about the pace for Arda and what's possible here?
Bob McGrew
There is a huge difference between developing technology for yourself as a first party, building models for use in ChatGPT, and building something for use in a large enterprise. The way we're approaching this is using the forward deployed engineering model. We developed this at Palantir way back in 2005. 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. Building a product for people that have secrets they can't tell the world is actually really good training for building a product for any enterprise. Any enterprise that is working has secrets and they don't want to tell the world everything about their IP.
The outcome of that process was the forward deployed engineering model. Sean Sankar, the CTO at Palantir, deserves the most credit for it. My job was to figure out how to build products for it. You basically take your engineers and embed them with the customer. I would not develop a feature unless I had a customer that was ready to use that feature. When I think about the product roadmap, a lot of the beginning point is not "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?" You need to start embedding with and understanding these enterprises.
You still want to move fast. How do you move fast? You start by finding a small number of power users. Classically at the agency way back in the day, there was a woman named Sarah Adams that was the only user of Palantir for about a year. I would go visit and meet her. She'd be like, "This stuff doesn't work." I'd be like, "OK, I've got to go fix this." Eventually it worked. Suddenly lots more people used it. It's the power of a very small number of very strong users that are willing to fight through the pain with you.
I told the story about our first user for Arda, our first user at Anduril, a lead manufacturing engineer who was trying to do the virtual build for an actuator. What I didn't say is that this was a tremendously frustrating process for him being the first user. He was able to make tremendous improvements and move very fast with that actuator. It only works because our head of product manager was sitting right next to him. Every time he did something, it didn't quite work the way he wanted. Our head of product manager was just sitting there taking notes. Two weeks later, all those things worked.
Elad Gil
Thank you, user number one.
Bob McGrew
Exactly. This process doesn't end. Six months later, we're at Anduril, they're like, "Go see if we can figure out how to build missiles faster." This only works if we find that power user who's on the missiles team, who's interested in working on a product that has a hundred times more parts than the actuator. Once again, out of the box, the product doesn't quite work. You find that user, with that user you fight through to make it work, and then they can introduce it to everybody else.
Classically, when you're doing enterprise sales, you want to find that visionary early on. This is from Crossing the Chasm. Classically, that visionary is a sponsor and you do need that sponsor. What the book doesn't talk about is you actually also need a user, someone who's willing to use your software to accomplish an outcome. If you do both these things, even very large organizations will work on their problems faster than you expect. There's not really any such thing as a very large organization. People be like, "What does OpenAI want?" OpenAI doesn't want anything. Sam wants something. Greg wants something. This person down here, this research scientist on this team wants something. Big organizations are just an assemblage of very small teams. You have to proceed problem by problem, team by team.
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, figure out how to build missiles faster. You are not successful unless the customer is successful. Once you achieve that outcome, you can look around at all the other outcomes and you've earned the right to work on something bigger. In our case, maybe that's working on an airplane, working on the Fury with Anduril. It could be working on an entirely different problem that's bigger in terms of dollars for them, that moves the needle more for them. They wouldn't be willing to trust your startup to solve that problem unless you had demonstrated some form of excellence.
Within a lot of this within an enterprise is snowballing the successes you've made, snowballing the outcomes you can achieve while still baking that into your product. The thing that a lot of people who try the FDE approach now get wrong is that they see this as just deploying some engineers at the site to make the product work at the site. Somewhat famously, people didn't believe the FDE approach worked. It was widely mocked in 2015 to 2020. Folks thought there would never be leverage. "These are just consultants. Palantir just sends a bunch of consultants to do things." Then people started doing it when Palantir became a $300 billion company and started profit. The difference is that they weren't just consultants. 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. If they have to keep filling the gap, you get no leverage. You have to productize what they're seeing and bake that back into what you're doing. That allows you to move quickly and discover new things and bake those into your products.
Elad Gil
If we take all those to the extreme and you have an army of FDEs over time who are deploying across multiple industries, you're learning very quickly, you're iterating, you're building a lot of product that then gets reused at scale. Where does this all go? What is your vision for the end point of all this?
Bob McGrew
I would go back to think about superintelligence at the beginning. We started by talking about what it looks like to have digital only superintelligence. The need for laying out lots of things, that's a requirement for superintelligence, but superintelligence also has consequences. I love Dario's phrase of "a country of geniuses in a data center." But if all they can do is sit in that data center and think, and you can send them emails and they can send you emails back, they're not going to be able to do very much. They're kind of trapped in the data center. Intelligence has become cheap. What are the compliments to intelligence? Perception on the one hand and action on the other.
As we have all these AGIs, right now, humans are the eyes and hands of the intelligence that's in the data center. Instead, what I think we're going to be seeing over the next 10 years is this tremendous profusion of various different kinds of perception, various different kinds of manipulation. There will be drones and sidewalk robots. Every physical thing that we own will become perception enabled and hooked up to the cloud. Whatever it does will become controlled by AI. 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. The world is made for humans. In many ways, if you look at the broader world, we'll probably see robots that begin to look like all different kinds of animals.
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 plow fields. I think we already have those with tractors. 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. It needs to actually be able to go out of the data center and help people and do things for us. That is the call to action for manufacturing and what manufacturing writ large is doing in this world of superintelligence.
Elad Gil
Von Neumann came up with Von Neumann probes, these self-replicating probes that you send out into space. They land on an asteroid or a planet, mine it, produce more versions of themselves, and keep spreading. It's a way to explore the universe at scale in a distributed system way where things self-assemble and keep going. It's exponential. You land, make 10 copies, those 10 copies go off and make 10 more copies, and you spread through the universe. Do you view that as an eventual manufacturing model on our planet? Eventually, do you have these self-replicating factories? What do you view as the end state from a physical form perspective?
Bob McGrew
Let's narrow in and think about this concretely. Maybe you have a problem where you need to go build more bulldozers. So you need to build a bulldozer factory. Right there, you have a recursive loop where you're deploying your intelligence to go manufacture more things. I think very quickly we will see factories that build other factories. But that's machine tools, that's what we already have. The difference is doing this in an intelligent way. It takes a tremendous amount of human work to go build a factory, to design machine tools for that factory. All of that is going to become increasingly automated, and it will have to in order to support the vast network of things we need.
It's the digital singularity versus the physical singularity. Or not even versus, just alongside. Part of the digital singularity involves converting increasing fractions of the earth's crust into data centers. Maybe those data centers will be in space too. That is a physical 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. I think that's a very near-term outcome. That's what we're doing in the next 10 years as humanity.
In the longer run, once machines can self-replicate and they are a better form factor for spreading between the stars than humanity is, it seems almost certain that it will happen. You want to make sure that this is happening in a way that is liberating humanity, that humanity is benefiting from, that humanity remains as much as possible in control of, rather than something where humanity is a little sideline that just remains here on earth and watches a lot of videos, which will probably also be generated by AI.
Elad Gil
They're going out there to terraform for us so that we may also explore the stars.
Bob McGrew
Exactly. I think we will also have a role to play if we choose to take it in getting to the stars.
Elad Gil
This is back to the Fermi paradox. AI, superintelligence, AGI, and these self-replicating factories all suggest we should have come across other intelligent life sooner. It increases the likelihood of running across it than if you just had our biological forms trying to make it through space on our own.
Bob McGrew
Actually, I disagree. I think this is the resolution to the Fermi paradox. If you think about what is the most effective civilization spreading through space, it's one that goes to a star system, very quickly sets up a factory, builds a lot of self-replicating probes, and sends those out. You have a wave of civilization spreading through space at the speed of light.
Elad Gil
So why don't we see that?
Bob McGrew
Because it's spreading through space at the speed of light. We will only see it a few minutes or perhaps a few years before it arrives.
Elad Gil
Isn't the argument that it should have arrived by now? There are arguments about the density of potential civilizations throughout space and the rate at which they arise. If you send something out at sub-speed of light, what is the rate at which it would eventually get to you? That would suggest sparseness in the universe.
Bob McGrew
I think this is a mystery. Why are we the first ones? If you look at the uncertainty of all the different factors in the Drake equation, you end up seeing that maybe actually it's quite unlikely. It is plausible that we're the first civilization.
Elad Gil
Or they're just very far away.
Bob McGrew
Yes, or very far away. Our light cones intersecting is going to take some time. To some extent, evidence shows that this hasn't happened. The question of whether we're going to run across alien civilizations, I think this argument says no, we're not. We're only going to detect them 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.
Sarah Guo
Amazing. Thank you for joining us today.
Bob McGrew
Thanks for having me. This was fun.
Sarah Guo
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. Sign up for emails or find transcripts for every episode at no-priors.com.