Neil Gershenfeld: Self-Replicating Robots and the Future of Fabrication | Lex Fridman Podcast #380

The ribosome, who I mentioned a little while back, can make an elephant one molecule at a time. Ribosomes are slow. They run at about one molecule a second. But ribosomes make ribosomes, so you have trillions of them, and that makes an elephant. In the same way, these little assembly robots I'm describing can make giant structures at heart, because the robot can make the robot. So more recently to my students, Amira and Miana had a Nature communication paper showing how this robot can be made out of the parts it's making. So the robots can make the robots. So you build up the capacity of robotic assembly.

The following is a conversation with Neil Gershenfeld, the director of MIT's center for Bits and Atoms, an amazing laboratory that is breaking down boundaries between the digital and physical worlds, fabricating out objects and machines at all scales of reality, including robots and automata that can build copies of themselves and self assemble into complex structures. His work inspires millions across the world as part of the maker movement to build cool stuff, to create the very act that makes life so beautiful and fun. This is the Lex Friedman podcast to support it, please check out our sponsors in the description. And now, dear friends, here's Neil Gershenfeld. You have spent your life working at the boundary between bits and atoms. So the digital and the physical, what have you learned about engineering and about nature reality from working at this divide? Trying to bridge this divide, I learned.

Why von Neumann and Turing made fundamental mistakes. I learned the secret of life. Yeah. I learned how to solve many of the world's most important problems, which all sound presumptuous, but all of those are things I learned at that boundary.

Okay, so Turing and von Neumann, let's start there. Some of the most impactful, important humans who have ever lived in computing. Why were they wrong?

So I worked with Andy Gleason, who was Turing's counterpart. So, just for background, if anybody doesn't know, Turing is credited with the modern architecture of computing, among many other things. Andy Gleason was his U.S. counterpart. And you might not have heard of Andy Gleason, but you might have heard of the Hilbert problems. And Andy Gleason solved the fifth one. So he was a really notable mathematician during the war. He was Turing's counterpart then. Von Neumann is credited with the modern architecture of comp, and one of his students was Marvin Minsky. So I could ask Marvin what Johnny was thinking, and I could ask Andy what Alan was thinking. And what came out from that, what I came to appreciate as background. I never understood the difference between computer science and physical science. But Turing's machine, that's the foundation of modern computing, has a simple physics mistake, which is the head is distinct from the tape. So in the Turing machine, there's a head that programmatically moves and reads and writes a tape. The head is distinct from the tape, which means persistence of information is separate from interaction with information. Then von Neumann wrote deeply and beautifully about many things, but not computing. He wrote a horrible memo called the first draft of a report in the edvac, which is how you program a very early computer. In it, he essentially roughly took Turing's architecture and built it into a machine. So the legacy of that is the computer somebody's using to watch this is spending much of its effort moving information from storage transistors to processing transistors, even though they have the same computational complexity. So in computer science, when you learn about computing, there's a ridiculous taxonomy of about 100 different models of computation, but they're all fictions. In physics, a patch of space occupies space, it stores state, it takes time to transit, and you can interact. That is the only model of computation that's physical. Everything else is a fiction. So I really came to appreciate that a few years back when I did a keynote for the annual meeting of the supercomputer industry and then went into the halls and spent time with the supercomputer builders and came to appreciate, oh, let's see, if you're familiar with the movie the Metropolis. People would frolic upstairs in the gardens and down in the basement, people would move levers. And that's how computing exists today, that we pretend software is not physical, it's separate from hardware. And the whole canon of computer science is based on this fiction that bits aren't constrained by atoms, but all sorts of scaling issues in computing come from that boundary, but all sorts of opportunities come from that boundary. And so you can trace it all the way back to Turing's machine making this mistake between the head and the tape. Von Neumann, he never called it von Neumann's architecture. He wrote about it in this dreadful memo. And then he wrote beautifully about other things we'll talk about now. To end a long answer. Turing and von Neumann both knew this, so all of the canon of computer scientists credits them for what was never meant to be a computer architecture. Both Turing and von Neumann ended their life studying exactly how software becomes hardware. So von Neumann studied self reproducing automata, how a machine communicates its own construction. Turing studied morphogenesis, how genes give rise to Form. They ended their life studying the embodiment of computation, something that's been forgotten by the canon of computing, but developed sort of off to the sides by a really interesting lineage.

So there's no distinction between the head and the tape, between the computer and the computation. It is all computation.

Right. So I never understood the difference between computer science and physical science. And working at that boundary helped lead to things like my lab was part of doing with a number of interesting collaborators. The first, faster than classical quantum computations. We were part of a collaboration creating the minimal synthetic organism where you design life in a computer. Those both involve domains where you just can't separate hardware from software. The embodiment of computation is embodied in these really profound ways.

So the first quantum computations, synthetic life. So in the space of biology, the space of physics at the lowest level, in the space of biology at the lowest level. So let's talk about cba, center of bits and atoms. What's the origin story of this MIT legendary MIT center that you're a part of creating?

In high school, I really wanted to go to vocational school, where you learned to weld and fix cars and build houses. And I was told, no, you're smart. You have to sit in a room. And nobody could explain to me why I couldn't go to vocational school. I then worked at Bell Labs, this wonderful place before deregulation, legendary place. And I would get union grievances because I would go into the workshop and try to make something, and they would say, no, you're smart. You have to tell somebody what to do. And it wasn't until mit and I'll explain how CBA started, but I could create CBA that I came to understand this is a mistake that dates back to the Renaissance. So in the Renaissance, the liberal arts emerged. And liberal doesn't mean politically liberal. This was the path to liberation. Birth of humanism. And so the liberal arts were the trivium, quadrivium, roughly language, natural science. And at that moment, what emerged was this sort of dreadful concept of the illiberal arts. So anything that wasn't the liberal arts was for commercial gain and was just making stuff and wasn't valid for serious study. And so that's why we're left with learning to weld wasn't a subject for serious study. But the means of expression have changed since the Renaissance. So micro machining or. Or embedded coding is every bit as expressive as painting a painting or writing a sonnet. So never understanding this difference between computer science and physical science, the path that Led me to create CBA with colleagues was I was what's called a junior fellow at Harvard. I was visiting MIT through Marvin because I was interested in the physics of musical instruments. I this will be another slight digression. In Cornell, I would study physics and then I would cross the street and go to the music department where I played the bassoon and I would trim reeds and play the reeds and they'd be beautiful, but then they'd get soggy. And then I discovered in the basement of the music department at Cornell was David Borden, who you might not have heard of, but is legendary in electronic music because he was really the first electronic musician. So Bob Moog, who invented Moog synthesizers, was a physics student at Cornell, like me crossing the street. And eventually he was kicked out and invented electronic music. David Borden was the first musician who created electronic music. So he's legendary for people like Phil Glass and Steve Reich. And so that got me thinking about I would behave as a scientist in the music department, but not in. In the physics department, but not in the music department. Got me thinking about what's the computational capacity of a musical instrument? And through Marvin, he introduced me to Todd Macover at the Media Lab, who was just about to start a project with Yo Yo Ma that led to a collaboration to instrument a cello to extract Yo Yo's data and bring it out into computational environments.

What is the computational capacity of a musical instrument? As we continue on this tangent and will we shall return to cba.

Yeah. So one part of that is to understand the computing and if you look at like the finest time scale and length scale, you need to model the physics. It's not heroic. A good GPU can do teraflops today. That used to be a national class supercomputer. Now it's just a gpu. And that's about. If you take the time scales and length scales relevant for the physics. That's about the scale of the physics computing. For yo yo, what was really driving it was he's completely unsentimental about the strad. It's not that it makes some magical wiggles in the sound wave. Its performance as a controller, how he can manipulate it as an interface device.

Interface between what and what exactly?

Him and sound.

Okay.

And so, so what it led to was I had started by thinking about ops per second, but Yo Yo's question was really resolution and bandwidth. It's how fast can you measure what he does and the bandwidth and the resolution of detecting his controls and then mapping them into Sounds. And what we found, what he found was if you instrument everything he does and connect it to almost anything, it sounds like yo yo that the magic is in the control, not in ineffable details in how the wood wiggles. And so with yo yo and Todd, that led to a piece. And towards the end, I asked yo yo what it would take for him to get rid of his Strat and use our stuff. And his answer was just logistics. It was at that time our stuff was like a rack of electronics and lots of cables and some grad students to make it work. Once the technology becomes as invisible as the Strad, then sure, absolutely, he would take it. And by the way, as a footnote on the footnote, an accident in the sensing of Yo Yo's cello led to $100 million a year auto safety business to control airbags in cars.

How did that work?

I had to instrument the bow without interfering with it. So I set up local electromagnetic fields where I would detect how those fields interact with the bow. He's playing. But we had a problem that his hand, whenever his hand got near these sensing fields, I would start sensing his hand rather than the materials on the bow. And I didn't quite understand what was going on with those, that, that interference. So my very first grad student ever, Josh Smith, did a thesis on tomography with electric fields, how to see in 3D with electric fields. Then through Todd and at that point, research scientist in my lab, Joe Paradiso. It led to a collaboration with Penn and Teller, who where we did a magic trick in Las Vegas to contact Houdini. And sort of these fields are sort of like, you know, contacting spirits. So we did a magic trick in Las Vegas. And then the crazy thing that happened after that was Phil Ritmuller came running into my lab. He worked with this became with Honda and nec. Airbags were killing infants in rear facing child seats. Cars need to distinguish a front facing adult where you'd save the life versus a bag of groceries where you don't need to fire the airbag versus a rear facing infant where you would kill it. And so the seat need to, in effect see in 3D to understand the occupants. And so we took the Penn and Teller magic trick derived from Josh's thesis from Yo Yo's cello to an auto show. And all the car companies said, great, we where, when can we buy it? And so that became Ellis and it was $100 million a year business making sensors. There wasn't a lot of publicity because it was in the car. So the car didn't kill you. So they didn't sort of advertise, we have nice sensors so the car doesn't kill you. But it became a leading auto safety sensor.

And that started from the cello and the question of the computational capacity. Musical instrument.

Right. So now to get back to mit, I was spending a lot of outside time at IBM, research that had gods of the foundations of computing. There's just amazing people there. And I'd always expected to go to IBM to take over a lab, but at the last minute pivoted and came to MIT to take a position in the Media Lab and start what became the predecessor to cba. Media Lab is well known for Nicholas Negroponte. What's less well known is the role of Jerry Wiesner. So Jerry was MIT's president before that. Kennedy science advisor, grand old man of science. At the end of his life, he was frustrated by how knowledge was segregated. And so he wanted to create a department of none of the Above, a department for work that didn't fit in departments. And the Media Lab, in a sense, was a cover story for him to hide a department. As MIT's president towards the end of his tenure, if he said, I'm going to make a department for things that don't fit in departments, the departments would have screamed. But everybody was sort of paying attention to Nicholas creating the Media Lab. And Jerry kind of hid in it. A department called Media Arts and Sciences. It's really the department of none of the Above. And Jerry explaining that and Nicholas then confirming it is really why I pivoted and went to mit. Because my students who helped create quantum computing or synthetic life get degrees from Media Arts and Sciences. This department of none of the Above. So that led to coming to MIT with Todd and Joe Paradiso and my colleague, we started a consortium called Things that Think. And this was around the birth of. Of Internet of Things and rfid. But then we started doing things like Work we Can Discuss that became the beginnings of quantum computing and cryptography and materials and logic and microfluidics. And those needed much more significant infrastructure and were much longer research arcs. So with a bigger team of about 20 people, we wrote a proposal to the NSF to assemble one of every tool to make anything of any size was roughly the proposal one of any.

Tool to make anything of any size?

Yeah, so there usually nanometers, micrometers, millimeters, meters are segregated. Input and output is segregated. We wanted to look just very literally how digital becomes physical and physical becomes digital. And fortunately, we Got NSF on a good day and they funded this facility of one of almost every tool to make anything. And so with a group of core colleagues that included Joe Jacobson, Ike Trang, Scott Manalis, we launched cba.

And so you're talking about nanoscale, microscale, nanostructures, microstructures, macrostructures, electron microscopes, and focused on beam probes for nanostructures, laser micro machining and X ray micro tomography for microstructures, multiaxis machining and 3D printing for macrostructures. Just some examples. What are we talking about in terms of scale? How can we build tiny things and big things all in one place?

Yes. So a well equipped research lab has the sort of tools we're talking about, but they're segregated in different places. They're typically also run by technicians where you then have an account and a project and you charge. All of these tools are essentially when you don't know what you're doing, not when you do know what you're doing in that they're when you need to work across length scales where we don't. Once projects are running in this facility, we don't charge for time. You don't make a formal proposal to schedule and the users really run the tools. And it's for work that's kind of inchoate, that needs to span these disciplines and length scales. And so you work in the project today. Work in CBA today ranges from developing zeptajoule electronics for the lowest power computing to micro machining diamond to take 10 million RPM bearings for molecular spectroscopy studies, up to exploring robots to build 100 meter structures in space. Space.

Okay, can we the three things you just mentioned. Let's start with the biggest. What are some of the biggest stuff you attempted to explore how to build in a lab?

Sure. So viewed from one direction, what we're talking about is a crazy random seeming of almost unrelated projects. But if you rotate 90 degrees, it's really just a core thought over and over again, just very literally how bits and atoms relate, how digital and just going from digital to physical in many different domains. But it's really just the same idea over and over again. So to understand the biggest things, let me go back to bring in now Shannon as well as von Neumann.

Claude Shannon.

Yeah. So what is digital? The casual obvious answer is digital in 1 and 0, but that's wrong. There's a much deeper answer, which is Claude Shannon at MIT wrote the best master's Thesis ever in his master's thesis, he invented our modern notion of digital logic. Where it came from was. Vannevar Bush was a grand old man at mit. He created the post war research establishment that led to the National Science Foundation. And he made an important mistake, which we can talk about, but he also made the differential analyzer, which was the last grade analog computer. So it was a room full of gears and pulleys, and the longer it ran, the worse the answer was. And Shannon worked on it as a student and he got so annoyed. In his master's thesis he invented digital logic. But he then went on to Bell Labs. And what he did there was communication was beginning to expand. There was more demand for phone lines. And so there's a question about how much, how many phone lines you could phone messages you could send down a wire and you could try to just make it better and better. He asked a question nobody had asked, which is, rather than make it better and better, what's the limit to how good it can be? And he proved a couple things, but one of the main things he proved was a threshold theorem for, for channel capacity. And so what he showed was my voice to you right now is coming as a wave through sound. And the further you get, the worse it sounds. But people watching this are getting it as, as packets of data in a network. When the computer, they're watching this gets the packet of information, it can detect and correct an error. And what Shannon showed is if the noise in the cable to the people watching this is above a threshold, they're doomed. But if the noise is below a threshold for a linear increase in the energy representing our conversation, the error rate goes down exponentially. Exponentials are fast. There's very few of them in engineering. And the exponential reduction of error below a threshold if you restore state is called a threshold theorem. That's what led to digital. That means unreliable things can work reliably. So Shannon did that for communication. Then von Neumann was inspired by that and applied it to computation. And he showed how an unreliable computer can operate reliably by using the same threshold property of restoring state. It was then forgotten many years we had to rediscover it in effect in the quantum computing era, when things are very unreliable again. But now to go back to how does this relate to the biggest things I've made? So in fabrication, MIT invented computer controlled manufacturing in 1952. Jet aircraft were just emerging. There was a limit to turning cranks on a machine, on a milling machine to make parts for jet aircraft. Now this Is a messy story. MIT actually stole computer controlled machining from an inventor who brought it to mit, wanted to do a joint project with the air Force, and MIT effectively stole it from him. So it's kind of a messy history, but that sounds like the birth of computer controlled machining. 1952. There are a number of inventors of 3D printing. One of the companies spun off from my lab by Max Lubovsky's Forum Labs, which is now a billion dollar 3D printing company. That's the modern version. But all of that's analog, meaning the information is in the control computer. There's no information in the materials. And so it goes back to Vannevar Bush's analog computer. If you make a mistake in printing or machining, just the mistake accumulates. The real birth of computerized Digital manufacturing is 4 billion years ago. That's the evolutionary age of the ribosome. So the way you're manufactured is there's a code that describes you, the genetic code. It goes to a micro machine, the ribosome, which is this molecular factory that builds the molecules that are you. The key thing to know about that is there are about 20amino acids that get assembled. And in that machinery, it does everything Shannon and von Neumann taught us. You detect and correct errors. So if you mix chemicals, the error rate is about a part in 100. When you make elongate a protein in the ribosome, it's about a part in 10 to the 4. When you replicate DNA, there's an extra level of error correction. It's a part in 10 to the 8. And so in the molecules that make U, you can detect and correct errors. And you don't need a ruler to make u. The geometry comes from your parts. So now compare a child playing with Lego and a state of the art 3D printer or computerized milling machine. The tower made by a child is more accurate than their motor control because the act of snapping the bricks together gives you a constraint on the joints. You can join bricks made out of dissimilar materials. You don't need a ruler for Lego because the geometry locally gives you the global parts. And there's no LEGO trash. The parts have enough information to disassemble them. Those are exactly the properties of a digital code.

The unreliable is made reliable.

Yes, absolutely. So what the ribosome figured out 4 billion years ago is how to embody these problems, these digital properties, but not for communication or computation in effect, but for construction. So a number of projects in my lab have been studying the idea of Digital materials. And think of a digital material just as LEGO bricks. The precise meaning is a discrete set of parts reversibly joined with global geometry determined from local constraints. And so it's digitizing the materials. And so I'm coming back to what are the biggest things I've made. My lab was working with the aerospace industry. So Spirit Aero was Boeing's factories. They asked us for how to join composites. When you make a composite airplane, you make these giant wing and fuselage parts. And they asked us for a better way to stick them together because the joints were a place of failure. And what we discovered was instead of making a few big parts, if you make little loops of carbon fiber and you reversibly link them in joints and you do it in a special geometry that balances being under constrained and over constrained with just the right degrees of freedom. We set the world record for the highest modulus ultralight material just by, in effect, making carbon fiber Lego. So lightweight materials are crucial for energy efficiency. This let us make the lightest weight high modulus material. We then showed that with just a few part types, we can tune the material properties. And then you can create really wild robots that instead of having a tool the size of a jumbo jet to make a jumbo jet, you can make little robots that walk on these cellular structures to build the structures where they error correct their position on the structure and they navigate on the structure using all of that. With NASA, we made morphing airplanes. A former student, Kenny Chung and Ben Jeanette made a morphing airplane the size of NASA, Langley's biggest wind tunnel. With Toyota, we've made super efficiency race cars. We're right now looking at projects with NASA to build these for things like space telescopes and space habitats, where the ribosome, who I mentioned a little while back, can make an elephant one molecule at a time. Ribosomes are slow. They run at about one molecule a second. But ribosomes make ribosomes. So you have thousands of them, trillions of them, and that makes an elephant. In the same way, these little assembly robots I'm describing can make giant structures at heart, because the robot can make the robot. So more recently, two of my students, Amira and Miana, had a nature communication paper showing how this robot can be made out of the parts it's making. So the robots can make the robots. So you build up the capacity of robotic assembly.

It can self replicate. Can you linger on what that robot looks like? What is a robot that can walk along and do error correction? And what is A robot that can self replicate from the materials that is given. What does that look like?

What are we talking about?

This is fascinating.

Yeah. The answer is different at different length scales. So to explain that in biology, primary structure is the code in the messenger RNA that says what the ribosome should build. Secondary structure are geometrical motifs or things like helices or sheets. Tertiary structures are functional elements like electron donors or acceptors. Quaternary structure is things like molecular motors that are moving my mouth or making the synapses work in my brain. So there's that hierarchy of primary, secondary, tertiary, quaternary. Now, what's interesting is if you want to buy electronics today from a vendor, there are hundreds of thousands of types of resistors or capacitors or transistors. Huge inventory. All of biology is just made from this inventory of 20 parts, the Amino acids. And by composing them, you can create all of life. And so as part of this digitization of materials, we're in effect trying to create something like amino acids for engineering, creating all of technology from 20 parts. As another discretion, I helped start an office for science in Hollywood. And there was a fun thing for the movie the Martian, where I did a program with Bill Nye and a few others on how to actually build a civilization on Mars that they described in a way that I like as I was talking about how to go to Mars without luggage. And at heart, it's sort of how to create life in nonliving materials. So if you think about this primary, secondary, tertiary, quaternary structure, in my lab, we're doing that, but on different length scales for different purposes. So we're making micro robots out of like nano bricks. And to make the robots to build large scale structures in space, the elements of the robots now are centimeters rather than micrometers. And so the assembly robots for the bigger structures are. There are the cells that make up the structure, but then we have functional cells. And so cells that can process and actuate. Each cell can move one degree of freedom or attach or detach or process. Now, those elements I just described, we can make out of the still smaller parts. So eventually there's a hierarchy of the little parts, make little robots that make bigger parts of bigger robots that up through that hierarchy in that way you.

Can move up the length scale.

Right. Early on, I tried to go in a straight line from the bottom to the top, and that ended up being a bad idea. Instead, we're kind of doing all of these in parallel, and then they're growing together. And so to make the larger Scale structures. There's a lot of hype right now about 3D printing houses where you have a printer the size of the house. We're right now working on using swarms of these, you know, table scale robots that walk on the structures to place the parts much more efficiently.

That's amazing. But you're saying you can't for now go from the very small to the very large.

That'll come, that'll come in stages.

Can I just linger on this idea starting from von Neumann's self replicating automata that you mentioned? It's just a beautiful idea.

So that's at the heart of all of this in the stack I described. So one student, Will Langford, made these micro robots out of little parts that then we're using for Miana's bigger robots up through this hierarchy. And it's really realizing this idea of the self reproducing automata. So von Neumann, when I complained about the von Neumann architecture, it's not fair to von Neumann because he never claimed it as his architecture. He really wrote about it in this one fairly dreadful memo that led to all sorts of lawsuits and fights about the early days of computing. He did beautiful work on reliable computation and unreliable devices. And towards the end of his life, what he studied was how, and I have to say this, precisely how a computation communicates its own construction.

Yeah, so beautiful.

So a computation can store a description of how to build itself. But now there's a really hard problem, which is if you have that in your mind, how do you transfer it and wake up a thing that then can contain it? So how do you give birth to a thing that knows how to make itself? And so with Stan Ulam, he invented cellular automata as a way to simulate these. But that was theoretical. Now the work I'm describing in my lab is fundamentally how to realize it, how to realize self reproducing automata. And so, you know, this is something von Neumann thought very deeply and very beautifully about theoretically. And it's right at this intersection. It's not communication or computation or fabrication. It's right at this intersection where communication and computation meets fabrication. Now, the reason self reproducing automata intellectually is so important, because this is the foundation of life. This is really just understanding the essence of how to life. And in effect, we're trying to create life and nonliving material. The reason it's so important technologically is because that's how you scale capacity. That's how you can make an elephant from a ribosome. Because the assemblers make assemblers.