Bullish on Automation and Robotics, but not Humanoid Robots | Shahin Farshchi

David Hoffman:
[0:03] Hey, Bankless Nation. In this episode, I talked to Shaheen Farshi. He is a general partner at Lux Capital, which is a hard tech venture firm. They invest in frontier technologies, AI, automation, biotech, robotics, just to name a few. Not in crypto. This is not a crypto episode. I wanted to go learn a little bit more about robotics. And Shaheen is a veteran of automation and robotics. Automation is a very important word. There's a lot of hype and excitement about humanoid robotics. You know, the Tesla Optimus Prime, the figure robot. You know, and Andrew Kang really came into the crypto industry really talking about like robotics and humanoid robotics. And I wanted to learn, I want to learn about that. I want to learn about robotics and the investing space around there. Shaheen has been investing in robotics for a decade plus. And so he's seen a thing or two and he's been around a few hype cycles in robotics. And so he's just probably the foremost expert on this industry. And so I'm pretty honored to be able to host him on this conversation and just learn about a sector outside of crypto that's in frontier technology that I think is going to be very, very impactful from somebody who just knows all the ins or outs. And so with that preamble out of the way, Shaheen, welcome to Bankless.

Shahin Farshchi:
[1:21] Thank you for having me, David.

David Hoffman:
[1:22] Shaheen, there has been a ton of hype around robotics. And I think a lot of people are learning about the robotics sector for the very first time. And so this is why I reached out to you and Lux, because I want to get a veteran on the show to ask a veteran what they think about the robotics industry. There's just been a crescendoing of hype And while hype is exciting and it's fun and it's an opportunity to learn, it can also be dangerous. And so I think the first question I want to throw at you is, is the hype around robotics justified? What do you think?

Shahin Farshchi:
[1:58] Yes, it's absolutely justified.

David Hoffman:
[2:01] Okay. Okay, that's a pretty simple answer. So the hype around robotics, I think, comes downstream of like a lot of products or companies coming live with humanoid robots. Is that what gets you excited about robotics? the humanoid element, the figure in Tesla, Optimus Prime robots? Is that the same? Because that's my sector. That's what are getting people excited in my neck

David Hoffman:
[2:23] of the woods. Is that the same for you?

Shahin Farshchi:
[2:24] No. So I'm not particularly excited about the humanoids. I'm excited about

Shahin Farshchi:
[2:30] The most recent wave of innovations in AI and cheap manufacturing and the ability to make things that are extremely complicated systems that are very complicated at a very low price so you combine the intelligence with the cheap manufacturing together And you have products otherwise wouldn't have been possible for certain use cases. Robotics is not new. Automation isn't new. We've had automation for decades, if not a century, with the advent of the assembly line, with the creation of Ford. And so my view is that over the past, I would say, 15 years, with AI in the form of convolutional neural nets in the early 2010s,

Shahin Farshchi:
[3:17] And the continued slope of cost reduction with robotic arms, we've seen a whole new wave of new robotics applications. And I think the humanoid robotics are a manifestation of that, but my personal excitement in robotics and automation isn't rooted in the humanoids, is rooted in this more macro tailwind around the software that drives the robots and the commoditization of the hardware that the software runs on. If you rewind back to, say, the 1970s, there was a similar opportunity, a similar moment around the integration of microprocessors, which were the equivalent of AI of our time,

Shahin Farshchi:
[4:05] Into robotics. And that's how you saw robotics enter into, for example, automotive manufacturing. You had these robotic arms that had a level of intelligence in them that allowed them to be programmed to very fine specifications, do very specific tasks like welding, riveting, gluing, painting, various types of inspections. And that's what began this rapid peripheralization of robotics into the broader manufacturing setting. Whereas today, when you walk into most automotive manufacturing facilities,

Shahin Farshchi:
[4:44] A lot of the basic steps in automotive manufacturing are completely automated. You don't see any people involved at all until the later steps where various parts of the interior, the wiring harnesses, the glass are starting to be installed in the vehicle. And what we're seeing today, the point that you brought up, is a lot of the companies that sell into these larger companies also having the benefit of automation. So when you walk into a Ford factory, you see a ton of automation. But if you walk into the factory that sells the components that are sold into the Ford factory, you may not see as much automation, because they don't have the budgets. They don't have the investment to be able to invest the way Ford can invest in its automation. And now what we're seeing with AI is just like how it is now easy for anybody to generate code, it's becoming as easy for anyone to program a robot. And these robots are getting very cheap.

Shahin Farshchi:
[5:54] And so now you're seeing a lot of these companies otherwise didn't have access to robots. These industries otherwise didn't have access to robots, now getting access to them in a way we hadn't seen before. So that's what catalyzes my excitement around robots, which may be slightly different than what's catalyzing yours or perhaps others' excitement around robotics. And that would just perhaps be in the minority of folks that probably, you know, I'm optimistic, but I'm not as excited about what we're seeing right now in humanoid robots yet, as perhaps others are.

David Hoffman:
[6:30] I do want to talk to you about humanoid robotics, but I think maybe it's worth putting a pin in that and just saving that for later and really diving into what you just discussed right now. Maybe if I could just try and summarize your excitement, it's really about the integration of intelligence into automation. And inside of automation, in addition to that, intelligence is allowing for, you know, cost reduction, deflation for what it means to automate things. And so it's about the collision of automation and intelligence and also the accessibility of automation to be applied to more and more industries. And that's what you really get excited about as an investor and as somebody who is excited about like growing the GDP of the United States. This is a fair summary of your excitement.

Shahin Farshchi:
[7:20] As an investor, David, Many VCs are excited about, you know, giant markets and unfair advantages and monopolistic businesses in those markets. I'm particularly excited about opportunities to create markets where none exist. And what we're seeing with robotics is that markets for robotics are being created where markets for robots previously didn't exist for the reasons that you mentioned. The AI, the software, the commoditization of the hardware, which is now creating a market where one didn't exist.

Shahin Farshchi:
[7:59] And there's companies out there that are in a position to now dominate those nascent markets.

David Hoffman:
[8:04] Can we try and define some terms a little bit? We're using robotics and we're also using automation. Are those the same terms? Are they different? And how should we think about these things when we're talking about this industry?

Shahin Farshchi:
[8:15] Good question. So I look at automation as a solution to a problem. So there are many aspects associated with automation. There's a financing aspect associated to it. There's an engineering aspect associated with it, which has nothing to do with the robot. How do you engineer your factory floor and your workflows to be able to benefit from robotic participation? It is the physical robot itself. It's the software that runs on the robot. It's the systems that are put in place to maintain the robots and make sure that they're up 99.999% of the time. And so I view all of those resources culminating in the end goal of the automation of a task to fall under that envelope of automation with the physical robots and the software that is running on that robot being a component of that? But that's a great question.

David Hoffman:
[9:20] Automation is the broad category and we need a bunch of robots to create the process of automation, but automation and robots are not the same thing.

Shahin Farshchi:
[9:30] Correct. Robots obviously automate, but

David Hoffman:
[9:34] Automation is not limited.

Shahin Farshchi:
[9:38] To robots.

David Hoffman:
[9:39] Right. And then there's a spectrum of generalizability, I feel. I think people probably have an image in their head of the forward floor and there's a fixed arm that can actuate and move and it does a lot of the actual work. But it feels very precise. It feels very locked in. It's meant to do that one job. Maybe people can also think of an Amazon floor of little robots running around moving packages. and they're not fixed and maybe they're slightly more generalized than the car building arm. Is there a spectrum to discuss about the generalizability of robots? And are you particularly bullish on a region of generalizability or is this not really a subject that is discussed much when it comes to robotics and automation?

Shahin Farshchi:
[10:28] I feel like the automation solutions are as broad as the problem set. So as you stated, the problem associated with order fulfillment in an Amazon warehouse is very different from the problem associated with assembling the components of a vehicle. They're two very different problems and they demand very different automation solution. Now, I can see from the perspective of a founder, and or from the perspective of an investor in a company to be motivated to champion their investments or their company as the singular solution to all problems. There is an economic motivation in presenting that picture that, hey, I'm building a widget that's going to solve every problem associated with automation. And it has a human form factor because humans do many tasks, if not most tasks.

Shahin Farshchi:
[11:36] I would argue that that is more of an economically driven argument, more so than a practical one.

Shahin Farshchi:
[11:43] The way I view the world and the way I view the opportunity is that specific applications are best suited to certain types of robotics and automation solutions. And so my expectation is that there will be an opportunity for a humanoid form factor machine that does some subset of tasks that are performed by humans today, but they will be one of many automation solutions that will take on many shapes and forms. So if you look at the robot that is used in the Ford factory today, that is an arm that has, for example, a welder and defector, it probably has zero resemblance to the robotic cart in an Amazon warehouse that shuttles boxes from one location of the warehouse to another for either stocking or order fulfillment. These robots have practically nothing to do with each other, except perhaps both being made from metal and plastic. And I think that's having some computer chips in them and cameras on them, but perhaps that's where the resemblance ends, yet both bring significant value to their end use cases. And so it's my expectation that

Shahin Farshchi:
[13:12] Those specific form factors will continue to proliferate alongside the humanoids, which will have their own place in the market. But I personally, as an end of one, don't see a single solution dominating all problems because I just don't think a single solution will be able to be best served to most problems. problems.

David Hoffman:
[13:38] So would you say that there is the humanoid end of the spectrum? And maybe one of the reasons why that's fun and sexy to talk about is that the form factor is going to be consistent across companies. And I mean, they look like us and so we can get excited about them. But I think what I just heard you say is that that's kind of going to be the only consistent form factor. And when we talk about the rest of robotics that it takes to create automation, the form factors are going to be highly heterogeneous and non-overlapping and specific and tailored to their needs. And one of the reasons why we need to make, one of the inputs that we need to make that world work is the deflation and cost of creating this whole industry in the first place, which we talked about earlier. Would you agree with that assessment?

Shahin Farshchi:
[14:25] I would say yes, broadly. There are components that go into these robots that will be consistent across use cases. The cameras, the force sensors, the reduction gear sets, the encoders, the power management systems, the actual arms.

David Hoffman:
[14:46] The building blocks. The building blocks will be consistent.

Shahin Farshchi:
[14:49] For example, if you look at a smartphone, it may have little resemblance with, for example, a laptop, a small, inexpensive laptop. The use case for the cell phone may be very different from the use case, let's say, for example, for an iPad. Yet, a lot of the technology that goes into the phone, the ARM-based processor, the memory, the Wi-Fi interface, the video driver, even the LCD screen itself. There are different form factors. They have different specifications, but they share the same basic technological architecture. Even their operating systems may share many, many thousands of lines of code, but they're used for very different use cases with a similar technology basis. So when you see a robotic arm that is doing welding, it may have the same motors, encoders, force sensors, cameras as the robot that's doing the riveting. But the robot that's doing the riveting may have a larger, for example, range of motion. It has a different end effector attached to it.

Shahin Farshchi:
[16:02] And so it is engineered for that use case, yet that use case is very broad, but not as broad as doing everything for everybody. So I generally take issue with the notion of we're building a robot that's going to perform all tasks because I think it'll be very difficult for a robot that performs all tasks to be better than a robot that's specialized to a certain task. And why would someone not use a robot that specializes in that certain task?

David Hoffman:
[16:34] And so on the specialization end of the spectrum, the specialization to generalization spectrum, I guess the bottleneck is if we indeed do have all of the raw ingredients that it takes to make a robot, we have the power, we have the actuators, ball bearings, really the constraint is the ability to put them in the right package according to the actual task at hand. Is that the current constraint? Is great, we have the pieces, but we need to put them in the right shapes. And that part can be more difficult because there's more possible shapes to order all the pieces in than could, then, then, you know, there's so many possible shapes to order all the pieces that it's a little bit hard to bear.

David Hoffman:
[17:20] Is that a fair assessment?

Shahin Farshchi:
[17:21] And that's why you've only seen robotics proliferate in these larger industrial settings. Ford has tens of millions of dollars that it can deploy to companies who specialize in this, companies like Honeywell, Dematic, Symbotic, Winright, these companies who specialize in engineering robotic solutions for a particular use case. In Ford's case, it would be an assembly line. So the robots that, for example, put a sheet of aluminum into a stamp, the robots that take those sheets of aluminum, the stamped aluminum out of the stamp, put into another stamp. And then the robot that has a camera at the end that inspects the stamped part to make sure there's no defects. And then the robot that puts two pieces of stamped metal together and the other robot, a third robot, doing the weld between those pieces of sheet metal, those stamped pieces of sheet metal. And then robots doing the riveting and the gluing and then lowering it into a bath of acid and the robots that are doing the painting. So there are many tasks that need to be automated, that need to be engineered by the integrator, which is extremely expensive. And there are few companies today that can afford to do that with the legacy technology that was available to us until today. And my enthusiasm is rooted to the earlier part of the conversation and these robots becoming less expensive and

Shahin Farshchi:
[18:50] The engineering that goes into making them useful becoming far more accessible than,

Shahin Farshchi:
[18:58] Just like how anybody can now speak to a computer and generate code, I see a future where someone can speak to an interface powered by AI that will then program the robot to do a specific task rather than you having to hire an engineer to do that for you. And we'll see the more rapid adoption of automation and more and more settings that until now didn't have that access. So you're right. It is taking the individual components and building all the software and the tools that you need to actually make that system that you build for your use case useful.

David Hoffman:
[19:35] It feels like kind of building a developer platform of sorts in that we have all the basic ingredients and we need to be able to allow people to tinker with, play with, combine all of the pieces in order to suit their needs And making this more and more accessible is going to find, allow the market to place automation in deeper, more niche, more specific parts of the market, so long as we can just figure out how to open up robotics, and be more accessible. That's kind of what it feels like.

Shahin Farshchi:
[20:08] And that's what companies are doing today, which fuels my enthusiasm. I'll give you an example.

David Hoffman:
[20:12] Yeah, tell me about them.

Shahin Farshchi:
[20:13] So if you wanted to develop a robot that does some kind of unstructured task, so when you're talking about welding two pieces of sheet metal together, it is a very structured task. You are within the, you know, thousands of a millimeter, you know where this robotic end effector is going and you know exactly what it's doing. When you're talking about the less structured tasks, it becomes extremely challenging to develop automation for. If you fast forward to today, you have companies like Physical Intelligence that we invested in that are building models that specialize in various unstructured tasks. And they're putting many of those models on open source repositories for roboticists to use and fine tune for their individual use cases. This is extremely powerful and a huge enabler for the community that didn't exist until very recently. So what would have required many PhDs and millions of dollars of funding and many years now can be achieved by an undergraduate student, maybe even a high school student, downloading and installing one of these models and fine-tuning the robot to perform a specific task. Yeah.

Shahin Farshchi:
[21:33] Your question, follow-on question could be, well, why don't we have robots everywhere today? It's that we still have some work to do to make these robots more reliable and to make them faster and to make them competitive with labor. And we can get into that, if you'd like, on the role that labor has in the proliferation of robotics.

David Hoffman:
[21:58] I do want to get into that. Let me ask you this one last question. And it seems, I was going to ask you, are we on the cusp of something big in robotics? But it sounds like, and what I mean by that is that, you know, we have AI, we have, you know, decreasing costs of automation, feels like we have the raw ingredients. And so we're close, you know, we're close to a cusp. It kind of sounds like we're actually, the cusp is behind us. And actually the ball is materially rolling towards this direction. To your point, we don't, as me, as a consumer, you know, somebody who walks around the streets of New York is not impacting my life yet. But in the factories, the frontier of this intelligence is progressing and progress is being made. We don't quite yet have the very high reliability that we need. But in terms of, you know, a step function change in progress in the world of automation and robotics, it sounds like that ball is already rolling and it's not, it's no longer an if.

David Hoffman:
[22:58] If it's already like happening now. Is that correct?

Shahin Farshchi:
[23:01] That's right. And the analogy that I like to use and the evidence that I have that points to the theme that you're sharing is compute, right? If you look at the 70s and 80s, people regarded computers as pieces of hardware. When you thought about a computer, you thought about what the processor speed was, how much RAM did it have, what was the graphics interface. That was what defined personal computing in the 80s and into the early 90s.

Shahin Farshchi:
[23:35] The rapid proliferation occurred, I guess that tipping point occurred with the internet. And compute was no longer regarded as a piece of, let's say, capital equipment, something physical or a machine. It was regarded as a capability. You went on the internet, you consumed content, you stored your files, and the physical hardware was completely abstracted away. I feel like right now we're in the equivalence of the early 90s of compute with robotics, where when you and I have a conversation, we're still talking about the physical robots. We're talking about the machine. We're talking about the plastic and the bearings and the software. We're still not talking about the output of the machine. The emphasis is not on the output of the machine, the way the emphasis today is on the output of compute, which is evidence of this mass proliferation. So I feel like we are on that trend towards abstracting away the plastic and the metal. And the cool demos, and focusing on what these robots are going to do for us. And I feel like that is a direction that we're headed towards, and that's what makes me very excited.

David Hoffman:
[24:57] Let's go into what you were talking about a second ago with a labor input into robotics. I'm actually not sure where this conversation leads, so I think I need you to actually take the reins here.

David Hoffman:
[25:07] What is significant about this part of this conversation?

Shahin Farshchi:
[25:10] Yeah. So there's always this ongoing debate about the question of how robots affect labor.

David Hoffman:
[25:21] Human labor, human jobs. Are we going to be automated out of a job like this question?

Shahin Farshchi:
[25:25] There's this general concern, which is a warranted concern, it's justified, that if you increase automation, then you are reducing opportunities for labor. And the enemy of labor, the enemy of work is automation. And if you observe historical trends, you'll see that economies who benefit from more automation, who adopt more automation, tend to also benefit from less unemployment. And economies who do not automate tend to suffer from more unemployment. So that's been the general trend. And so I would say that the enemy here of both

Shahin Farshchi:
[26:19] Labor and jobs, as well as automation, is cheap labor. So cheap labor here is the common threat or the common, let's say, enemy here. When you have cheap labor somewhere else, then you are now threatening the jobs or the employment in your market. And by the same token, when you have cheap labor available, then you're increasing the hurdle rate that automation needs to cross in order to be able to proliferate. And the observation that I've had is that when you automate, yes, you may be displacing

Shahin Farshchi:
[27:03] Jobs, but you are disproportionately creating higher quality jobs, where you have less churn, where you have greater worker satisfaction. We're investors in a company called Formic. And what they're doing is offering robots as a service. What they sell is not a physical machine. What they sell is the output of a machine. They abstract away the machine and they make it as easy for you as the owner or operator of a factory or a warehouse to automate as it is for you to hire labor. And these customers are not deploying automation because they want to reduce their headcount or save money. They're trying to do this because they...

Shahin Farshchi:
[27:51] Can't find workers. And these are jobs that have extremely high churn rates. And what they've actually experienced is creating more and more jobs that are higher quality that do not suffer from the worker dissatisfaction and the churn that they experienced prior to adopting automation. And again the enemy of all of this is not the enemy of jobs and automation again is not the robots it is the it is the cheap labor if that makes sense

David Hoffman:
[28:26] Yeah yeah i think it does and i think that takes us to where i want to go next which is just the the impact on the economy of a successful automation industry i think you talked about it just a little bit now we get more higher quality jobs. Yes, there may be some short-term dislocation of jobs, jobs from A to B. Perhaps there is short-term economic pain in some industries, but in the long term, the trends point towards more higher-paying jobs. What about just the rest of the economy, the cost of goods, the cost of food, the cost of building a physical non-software, non-SaaS-based startup? Yeah. Say the automation industry does everything that it hopes to do in the next decade. What would that mean for just the average consumer's life in terms of prices

David Hoffman:
[29:21] and just what the impact on the economy would be?

Shahin Farshchi:
[29:24] I mean, I'll give you the broad answer, and that would be more selection, more competition,

Shahin Farshchi:
[29:30] And ultimately a benefit to consumers. So if a factory, so factories, for example, that our company Formic is selling automation into, they have more productivity, their workers get paid more, they generate from better unit economics, and they ultimately are able to sell better products at better prices. And the benefit trickles down to the end customer. So it's my expectation, my belief, that if you're able to make automation simple enough to adopt and make it reliable enough, then everybody from the people that sell products to these automation customers to the people that purchase the products from them will ultimately benefit. And I think the key point here is... Educating our workforce. If we make sure that our workforce is educated and has the opportunity to grow their skillset, then they will be very well positioned for this future of more higher paying and higher quality jobs. And so I think that's something that we as a society have to take upon ourselves to keep our workforce, to make sure our workforce is prepared for this new generation of work that's going to come about as a result of automation, as opposed to looking at it as, oh, no, jobs are being eliminated. So what do we do with our workforce? We should be more proactive about it.

David Hoffman:
[31:00] Is there a particular industry that you get really excited about when it comes to the potential of said industry with automation? Like agriculture, I could imagine, gets scaled, but I'm sure there's a handful of industries that is possible to talk about. Does one stick out to you as particularly exciting?

Shahin Farshchi:
[31:17] Yeah, I mean, if you look at agriculture historically, agriculture is one of the first industries to be automated. And as a result of that, we went from, you know, food shortages to food being completely abundant as a result. You know, you look at industries like automotive, consumer electronics. These are industries that are highly automated. And as a result, you know, we have cars that are extremely safe, extremely efficient. And, you know, many of us have the benefit of either keeping our cars for 10 years, like myself, you know, that that are reliable and run for a long time, or we can go out and buy the latest and greatest car. And the newer the cars get, the cheaper they get, and the more more capable they get, and the more safe and efficient that they get. So these are all benefits that that come to us.

Shahin Farshchi:
[32:05] I feel like the zeitgeist today, David, going back to your earlier question about excitement and is around robots and domestic settings. So we're talking about companion robots. We're talking about butler robots, you know, robots and household settings.

Shahin Farshchi:
[32:26] I, a little cautious about those kinds of use cases for two reasons. One, because I feel like there's still a huge opportunity for automation in factory and warehousing settings. There is plenty of work to do there. There's plenty of opportunities to automate in those settings. And I feel like it may not be as sexy as a companion robot or a maid robot, but I feel like those applications are still very, very real. We're seeing the proliferation of autonomous cars. I mean, they are a fantastic example of a robot. There's a lot more room there. We still haven't. We were promised the big rigs and the trucks to become automated. They still haven't become fully automated. So we'll see hopefully in the near future, more autonomous vehicles, autonomous trucks. And I feel like the future of a humanoid robot tackling the challenges in a domestic setting, cleaning up after a child, helping an elderly person, doing basic tasks like emptying the dishwasher, I think we'll see that come. But I feel like that is further out. And I'm more excited about these nearer term opportunities and more, I would say, industrial and commercial settings.

David Hoffman:
[33:52] Because those are the opportunities that scale to all of society in a very efficient way, right? It's getting a humanoid robot into everyone's home. That's like a huge last mile problem. There's huge constraints with the actual design of the robot. But I think one of the reasons why I think you're excited is that there are a few places where you can put robots and automation into that have just massive benefits for society as a whole. That's right. And so it's A little bit of the scalability of the effect, I think, is something that gets you excited about the automation side.

Shahin Farshchi:
[34:25] I don't want to say easier because all these problems are challenging, but I feel like the problems are more constrained in these industrial and commercial settings. And the value proposition or the economics that are more well-defined than robots going into a domestic setting. When you buy, for example, a piece of furniture, you know, I may sleep on my couch, you know, every afternoon and take a nap. You may not even sit on that couch, you know, for months, depending on like, you know, how we look at couches. And so given the widely varying utility that consumers get from like these kinds of discretionary purchases, I think it would be challenging to introduce a piece of technology like that into the domestic setting. I think I wouldn't say that it's that it's that it's impossible. I'm just saying that. It'll take longer than a lot of people think. And listen, like I'm a huge robotics nerd. I was a big Star Trek fan. I, I, you know, I wish to be able to interact with a robot, like, you know, commander data from, from, from series from next generation. But like, I'm excited for that day, but I just feel like it's a little further out than

Shahin Farshchi:
[35:45] what a lot of people hope.

David Hoffman:
[35:47] What are the constraints that are trying to be solved right now? Of the robotics automation companies that you guys have invested in at Lux, is there a common denominator of problems trying to be tackled? Or what are the modern problems in the proliferation of automation and robotics?

Shahin Farshchi:
[36:05] It really comes down to the breadth of applications, the reliability, and being able to demonstrate the unit economics to customers. Because you're talking about a relatively nascent product. Like when a customer buys, for example, a conveyor belt or a package sorting machine or a oven for their restaurant, there's very clear economics attached to that. When you buy, if you have a coffee shop and you buy a coffee grinding machine for $5,000 or whatever it is, it's very clear as to what the return on that investment is. When you're talking about a new product, a new technology with a sample set of 10 customers who've used this before, it's much more challenging to be able to justify that return on investment. So I think being able to demonstrate value, being able to create an opportunity for a customer to be able to properly underwrite is the challenge that a lot of these new wave of automation companies are facing right now, perhaps more so than just technology. It's being able to quantify the value add for their customers.

David Hoffman:
[37:16] I want to learn a little bit more about the intersection of AI and robotics. We have we have AI now. We are now having like people talk about robots coming into the home. And I think the naive simple thing to you to do in your imagination is like, oh, we've got LLMs. Let's put them into the robot. And now we have smart, general realized robots. While preparing for this interview, I've learned that that's not quite how it works. I wish it was that simple. Shaheen, can you explain VLMs and VLAs and all the

David Hoffman:
[37:46] other details that we need to know about how AI actually becomes imbued in robots?

Shahin Farshchi:
[37:53] So AI comes in many flavors. And if you look at the technology that's more commonplace today, when AI is being implemented in a robot on the field today. Most of those robots use some combination of sensing with their cameras and radar and LIDAR or whatever it is. And then they have a processing chain that tries to perceive their environments, which is segmenting the sky versus the ground, what the objects are, what object is movable, what object is not movable, what is a target objects. They call that in general perception from what is being sensed from their sensors.

Shahin Farshchi:
[38:37] And then there is planning. Okay, so how do I reach for that object? If you're a car, what path should I take to get to the other side of the intersection, for example, if there is construction going on the other side of the street? And then there's the actual actuation, which is, okay, now I'm going to activate this motor, activate that motor to execute on that plan that I've generated from my perception of the environment that I have collected from my sensors. So that is the, I would say, common AI-based workflow that exists in robots in the field today. I'm not intimate with the technology behind the Waymo vehicles, but my guess is that the way they operate today is somewhere along these lines. Now, what we're seeing with VLAs and VLMs is some flavor of

Shahin Farshchi:
[39:35] Of using language to interpret a scene and then generating language from that language should then take some kind of action.

Shahin Farshchi:
[39:53] I am vastly, grossly oversimplifying this, and there's many people out there that can explain this better than I can. But the core of it is using language to interpret a scene and then using language to come up with some plan of action and then executing on that. And as you know, with us as individuals and animals, we don't necessarily talk through what we're seeing and we don't talk through what we're going to do.

Shahin Farshchi:
[40:25] There's many other steps that come into play and many other sensations that come to play and many other pre-planned reflexes and heuristics that come into play. And many companies are trying to bake that into their models. The jury is still out as to whether you can simply solve this problem with scale. So just make these absolutely gigantic models that rely on language alone to perform these tasks versus the more, quote unquote, traditional approach, which is this sensing, perception, planning, action process, which was popularized, you know, pre-LLMs. I think the jury is still out as to what's going to come together, but it's my expectation that it's going to be some kind of hybrid of the two, if that makes sense. But I can suggest many people that you can bring onto your show that can give a pretty thorough lesson on how these VLAs and VLMs actually work.

David Hoffman:
[41:31] That would definitely challenge myself as an interviewer to go that far down the robotics and automation rabbit hole, but I do find it very interesting. I probably should have defined VLM and VLA. That's vision language model and vision language action model. Would you say, is it fair to say like we got LLMs on the cloud anthropic opening eye side and we have VLMs and VLAs on the automation robotic side? Or is it just not that clean?

Shahin Farshchi:
[41:55] LLMs are basically, you know, chatbots. And then, and then VLMs and VLAs are, are, are basically interpreting a scene with language and then taking action based on,

Shahin Farshchi:
[42:09] Uh, uh, uh, putting, you know, a set of, of observations and intentions through a model and generating a plan from that and then converting that plan into action. So for example, okay, you know, taking a picture of a scene from a sensor, okay, like here's all the objects in the scene by putting it through a vision language model and then using language like basically a chatbot to come up with some kind of plan of action based on the robot's goals and then turning that language into like you know motor actuations to actually perform some kind of task okay you know you open the refrigerator okay i want beer where is the beer oh there is a beer. Okay, now you have to pick up the beer, you go grab the beer. So it's been shown that these work, but they're still relatively nascent relative to the more traditional approach. And they may not be as quick and they may not be as reliable. And so, again, the jury is still out as to how you can make them more reliable. Do you just continue to refine them and make them larger? Or do you kind of hybridize them with these more traditional approaches? And I'm not a roboticist myself, but I think it would be a great idea for your next guest on the show to talk about that kind of stuff.

David Hoffman:
[43:28] It seems like it's an important ingredient nonetheless to add to the generalizability and just the practicality of automation and robotics to fit into more spots in the world. Yeah. Because it seems like you can take this VLM or VLA and apply it to a robot in

David Hoffman:
[43:47] different settings and it just kind of works. Is that right?

Shahin Farshchi:
[43:51] So the whole, the thought process behind these types of like language models is that, yes, they're more generalizable. You can teach them to do things by simply just, you know, showing them the task the same way you would show a child a task. That's the thesis behind them. And I'm really excited about them. I'm optimistic that over time, we'll figure out how to make them faster, more reliable, easier to train. Because right now, you know, again, you need a team of hundreds of engineers to teach a Waymo, for example, how to go from A to B safely. Is there a future where you can do the same with a VLA? You know, perhaps. But the question becomes, how much does it need to be trained to get to what level of reliability? Like, for example, I'm trying to teach my four-year-old how to write numbers. And, you know, some numbers she can write, like, after I show her twice. Other numbers, for whatever reason, she has a hard time, like, stopping when she's doing a curve. Like, for example, at the number two, you have to curve and then stop and do a straight line. That's the challenge for her. So maybe that challenge is limited to human children, or maybe it's a limitation associated with neural networks. Who knows? So that will be figured out in the near future.

David Hoffman:
[45:13] I would also imagine the chatbots had this very incredible advantage in that they just had to train on all the data of the internet, which was accessible to them. I would imagine that robots and automation don't nearly have the same qualitative and quantitative amount of data to train how to move your arm to grab the thing with the right amount of force.

Shahin Farshchi:
[45:33] So we are investors in XDOF, which is specializing in generating this training data for robotics. Physical intelligence obviously also has a huge capability around amassing this data internally for training as robots. So, yes, you're hitting on a very good point, which is companies that are able to access and build these training sets for these particular applications will certainly be advantaged.

David Hoffman:
[45:59] Shaheen, this has been very exciting and very educational.

David Hoffman:
[46:03] How are you hoping that automation and robotics impacts, positively impacts your actual personal life? So your day-to-day changes and your house gets an upgrade, your car gets an upgrade. Is there anything that you're like trying to get your hands on as soon as possible to have like a material improvement in just your day-to-day life?

Shahin Farshchi:
[46:23] I'd like to have an autonomous car that is always available to me rather than having to call a rideshare. But the reality is that you kind of have that today with Waymo, and you sort of have that today with Tesla's FSD. But I wouldn't mind taking it a step further where the vehicle could figure out, you know, where to park and go off on its own. And for me not to have to deal with it, I think that for me would, and I'm just a big car enthusiast. I'm a car nerd. So having a car like that would be to own is something that I personally find fascinating and exciting.

David Hoffman:
[46:58] What about how you think like automation actually enters the home? Not the humanoid robots.

Shahin Farshchi:
[47:04] Yeah, yeah, yeah, yeah. So if it was just me, David... I would totally nerd out on having even a modestly capable robot, you know, in my home that could do simple things like, you know, turn the stove off, you know, or pour me a glass of water and bring it over, you know, just for the purpose of, you know, nerding out on something like this. I would find that just personally exciting, but, you know, being, you know, married and having two small kids, I just think it'll be, it's extremely unlikely that my wife would allow something like that in the house until it's proven to be, you know, safe and never trip over a child or anything like that. So, you know, just for myself, just having a robot around that I can physically

Shahin Farshchi:
[47:52] interact with would be a huge novelty and interesting.

David Hoffman:
[47:55] Of the companies that you guys have invested in at Lux, if listeners wanted to just go a little bit deeper about what we've been talking about today, are there any good companies that are doing something exciting that also provide like an educational opportunity to just learn more?

Shahin Farshchi:
[48:09] They should absolutely learn more about physical intelligence. They should check out the company. They should check out the models that they have available out there. They should also learn about Formic, which is deploying robots in real manufacturing logistics settings. They have hundreds of deployments across the country. Most of their customers had no automation before automating with Formic. And their goal is to be the largest employer of robots globally. We'd like to say the equivalent of U.S. Robotics from the movie of iRobot, but not evil. And so, you know, those two companies, they should absolutely check out.

David Hoffman:
[48:45] Shaheen, thanks so much for coming on the show today.

Shahin Farshchi:
[48:47] Loved it. Loved it, David. Thanks for having me.