EPISODE 1711

[INTRODUCTION]

[00:00:00] ANNOUNCER: ChatGPT has been out for more than a year and has become the centerpiece of intense discussion and debate about AI. Christian Hubicki is a renowned robotics research scientist and assistant professor of mechanical engineering at Florida State University. In 2023, he was a guest on Software Engineering Daily where he discussed ChatGPT and its implications with Sean Falconer. Christian now joins Sean again to check in about the state of AI and its future directions. 

This episode of Software Engineering Daily is hosted by Sean Falconer. Check the show notes for more information on Sean's work and where to find him. 

[INTERVIEW]

[00:00:47] SF: Christian, welcome back to the show. 

[00:00:48] CH: It's great to be back, Sean. Good to see you again.

[00:00:52] SF: Yeah. You too. How have you been? It's been probably about a year and a half, I think. 

[00:00:55] CH: Yeah. Year and four months-ish since we - a little less than that since we've last talked. Things are great. AI has taken over. I'm serving my overlord. I don't know what your particular fiefdom, how that's going. But things are going well here.

[00:01:08] SF: Yeah. Well, actually I was thinking - because I listened back to our prior chat where we - I think February of last year, we talked about ChatGPT, GPT, sort of general thoughts on generative AI. And I was thinking that, luckily, nothing has changed in the last year in AI. And we can just essentially just chitchat about other things. Like, why aren't you on season three of the Traitors? These types of things. And skip the whole AI conversation.

[00:01:34] CH: Yeah. Exactly. And, of course, in all seriousness, man, it is a whirlwind to keep up with. I feel like every day there's a new buzzword I have to be at least somewhat aware of in order to have conversations on these topics. It's been a busy year and a half. Thankfully, now I am now a tenured professor. That means I can spend all my time just keeping up with the lingo of the day.

[00:01:56] SF: Yeah. Absolutely. Now you have a lot more time for these types of conversations. Big congratulations on that. I know that's a big deal. 

[00:02:00] CH: Thank you. 

[00:02:01] SF: That was a path one point I was aspiring to do and then I kind of went in a completely different direction. And now I'm just out here podcasting and talking about AI with various experts. 

[00:02:10] CH: It seems like you made a fine choice, Sean, just from my perspective. That's what I'll say. Well done.

[00:02:15] SF: Yeah. When I was listening to the episode, I would say at the time our excitement about ChatGPT and some of this other stuff was somewhat tempered. It was kind of interesting. Potentially useful. But had some problems that we kind of talked about some of these things. And you even called ChatGPT a bullshitter at that time. 

And I was definitely not using ChatGPT and other LLM technology back then the same way I am today. Now, probably a single day doesn't go by where I'm not using some form of this technology in some capacity to help me do something. And given how fast things are moving, a number of models have come out. Everything else is happening in this space. Has that changed your perspective on any of these over the last year? And, also, how you might be leveraging some of these tools? 

[00:03:00] CH: Well, certainly, some things have changed and I think some things has remained the same for me. In terms of my usage, I would say that my usage has not explicitly gone up. Meaning, I don't seek out LLMs to do a lot of things explicitly. It's pretty rare where I go to ChatGPT and I use it for an explicit task, like, "Hey, help me generate some ideas for this." Maybe I'll do that once a week at most. 

And so, it's just like when I'm stuck, I'll try a thing. That's my explicit use. But that does not capture the way it's being starting to be embedded implicitly into the tools we use. And I think that that has gone up. I feel like, unintentionally, I am using it more just by what it's being embedded into. Whether it's now Google search will now give me an AI summary which we can now talk about. And I have mixed feelings about these things. 

One, it definitely hasn't gone away. It has only expanded. And people are finding uses for it. But I am still at a point where I am curious about the staying power of this particular tool. I do want to say that, short-term, I'm still skeptical. Long-term, I'm a long-term optimist on these various AI methods. Whatever will come next. But it is amazing to me all of the things that people have managed to cram language models into for good or for ill and in creative ways.

[00:04:26] SF: Yeah. I mean, I think it's like a natural cycle with any sort of innovation that happens in technology. I always liken it to like the pets.com era of the dot-com boom when did it make sense to create an online retail business for every business in the world at that time? Probably not. But people were just kind of throwing things at the wall. And some of them stuck and some of them didn't. And I think we're kind of in a similar place. 

I brought this up on a couple of shows previously when I've talked to people who are working in the world of LLMs and a big question around does it make sense to essentially stuff a Copilot into every single surface? And there's a lot of advantages when it comes to user experience to having like a more guided experience. Versus like not every interaction should necessarily just be in a free-form text box. Because if I don't know what to ask, it's not very helpful. Because I've essentially have this barrier to entry now where I have too many options to do anything. And I think we're kind of - my big concern is are we going too much in that direction?

[00:05:25] CH: I can only imagine that we're going a little too far in this direction, specifically in this particular use of prompts as our interface to things. I mean, it makes sense. It's like this exciting thing that is now been embedded in Internet culture and people are running to it. What can I use this for? And in a way - and it's not a great analogy, but it feels somewhat a little bit like blockchain to me. People are like, "Oh, my goodness. Blockchain." It's like everywhere. Bitcoin is worth a lot of money. What can I put blockchain into? And it might make some modicum of sense in a lot of different applications. But how many of them actually stick around? 

And it's only natural people will overshoot. I'm not mad at that. I think that people - and we can talk - and for those who don't know, hello, I do robotics. And we'll get to the point where I imagine we'll talk about the impact on robotics, which has been interesting. I think there have been some interesting innovations by using these various transformer models in robotics. 

However, I think that - I don't mind people trying to interface like this to different things. It's what are people again relying upon language models to do? And this was extremely relevant when Google recently rolled out their AI summaries on their searches. That was a big to-do that it was giving out unreliable results. 

I think that overshooting the market - look, I'm not a market guy. I'm not going to pretend - I don't have a stock portfolio. I put things in a savings account. I'm a dinosaur like that. Don't ask me for those kinds of questions, audience. I can't comment on the business side of what makes sense. But it feels like they would certainly overshoot in that direction. That's not too bad. It's when you compromise important reliability of services that we had that I think that that's when it's negative. And we got to be careful about that.

[00:07:05] SF: Yeah. Absolutely. I mean, I think the other analogy besides the blockchain one I think is when Facebook and these other social networks were blowing up. And, suddenly, every application needed some sort of social component to it, whether it made sense or not. There's some kind of like button, comment thing on every single surface. Eventually, that tempered down and went away and even got consolidated. Where if you were going to do that, you're essentially just like embedding one of these large social networks into whatever existing flow you have. There's going to be a certain amount of consolidation that happens over time as people start to figure out what works, what doesn't work and so on. 

[00:07:41] CH: Yeah. And I think that from our podcast, which I went back and relistened to, I think that some things hold up pretty well in the sense that the things that it really could be good for application are things where cost of failure is low and there's a lot of time saved via a speed up. And I think that to some degree, if you have a customer service sort of situation and it's not a high - you're just asking what's the status of my order return on my non-critical products? Not medication. You interface with a chatbot for that. I mean, it literally was developed as a chatbot. And I think that makes a lot of sense. 

And I think that, however, as we start to integrate more important things in our lives, is this thing going to touch our bank accounts? Is this thing going to manage my medications for like my schedule? Don't forget to take this medication at 8pm. Those are the things that we have to be careful about. And while so short-term, I'd be really queasy about that sort of thing. But longer term, as we come up with new networks and new approaches that are more safety-first or safety-critical, I think that's exciting. And that's where I would run as a researcher.

[00:08:48] SF: Yeah. I mean, I think that you mentioned the Google giving an answer directly in search now. And I think for a while, Google was the inventor of the Transformer in 2017. They had a huge head start. But they didn't invest in the technology. And there's a variety of different reasons why they might not have done that. But one of the perceived reasons could be that if they had to came out with something like ChatGPT, it actually goes against sort of their existing business model. Because their existing - most of their money comes from essentially serving up blue links that people click on that you have to go and dig for an answer or something like that rather than serving you an answer directly. 

And now, to, of course, stay competitive with ChatGPT, they have to do some of this work. It's potentially disrupting their business model there. But then the other thing that's happened with Google recently, which also ties into some of the things that we're talking about in terms of whether this is a mission-critical thing. But, also, the ethics and bias. There's all the stuff that happened when they released Gemini recently around ethics and bias. They basically erased white people from history. And they're putting sort of those guard rails in place to make it - in order to help deal with the fact that they're training these things on gobs of internet data. And gobs of internet data are going to have some of our own human biases within them. 

But this is I think also one of the big challenges when it comes to building these foundation models, is how do you balance sort of trying to create something that is maybe a better version of ourselves in terms of how it views the world? But, also, doesn't make these kind of mistakes that are even more maybe offensive like showing black Nazis or things like this. Then, I guess, my rambling question is around like what are your thoughts on how do we essentially move forward in terms of trying to balance those ethics? It's a very, very complicated problem. But until we can do that, it seems like it's pretty hard to make some of these mission-critical products based on this type of technology.

[00:10:45] CH: I think that, sadly, some of this is just foundational to how these particular methods work. They are imitators. They're imitating data. And you can clean the data going in. But, ultimately, it is learning patterns from that data. And you can try to put guardrails in. And when we say guardrails, we're not talking like I'm assuming guardrails that are going to make this thing become a Frankenstein's monster and try to take us over. We're talking guard rails of saying, "Hey, just don't do this. Don't do that. This could create safety concerns. It could create ethical concerns." It's hard to individually build those guardrails. I'm sorry. 

I can only imagine all the ways that these things can go wrong. And you can patch them up individually. But, I mean, these methods are generative. And it has weird interactions that finds ways to bypass these things. And I don't know. I'm skeptical that you're going to be able to patch out these kinds of issues. I mean, I hope they do. 

I think that this is an outgrowth to me of people chasing, I'm going to call it a shiny object. But that is underselling what these generative models are doing. They're very interesting. They're very cool. But they're chasing this thing. It's super successful at the moment. And trying to make it into the thing they need it to be. It might foundationally not be that thing. 

Now, somewhat appropriately or ironically based upon your perspective, these things are called foundation models. Pretty much anything that takes in gobs of data and then takes the inputs and get you an output that imitates those gobs of data. They call it a foundation model. Your ChatGPT is just one example. That's now has been the phrase for a while. But this is more popularized phrase these days as to what these are. But those are the foundation of what they're building on top of. 

And either they can kind of patch that up or figure some cool weird way to handle the hallucination problems, right? Or you kind of got to start with a little different foundation. One that might be harder at immediately generalizing all kinds of things, but is built with your values in mind. And that actually might be the hardest problem. That's maybe more a Sci-Fi question. It's like what do we really care about? What is it at its core? If we had an algorithm that we're going to build on top of, what principle should it be? 

And right now, the principle is predicting an output that is basically imitating a bunch of input/output relationships. There's stuff on top of that that's supposed to make it safer. And it definitely does make it safer. But what is that thing? And that's a bigger question. I'm rambling now, Sean.

[00:13:13] SF: Yeah. Well, I think that's one of the concerns and also criticisms of sort of big tech owning these models and making decisions about what is the ethics or the biases of those models. Because they represent a small part of the population. And it might not necessarily be politically aligned with the majority of people in the United States or beyond the United States, the entire world. How do you essentially make those decisions about something that potentially could be used by anybody in the world? There are things that are going to be seem like in isolation maybe a good decision that could be end up being like wildly offensive to someone else from a different background. 

[00:13:49] CH: Yeah. And I think that the traceability back to a principle and the whys. The transparent AI is I think definitely goes in that direction. Because as much I am annoyed - but let's go for maybe a smaller-scale example with Google's AI summary, right? I was very annoyed when I first saw that popping up on my search feed, to be quite frank. I saw it pop up. I was like, "I don't like this. I want to turn it off." And I couldn't turn it off. There is no way to turn it off without completely the kind of bypassing the search entirely. There are weird key phrases you can use to get around it. But I was annoyed that it was serving it to me. And realized it's serving it to everyone. 

And unless I missed a memo, I do not have trust and reliability in the reliability of these hallucinating machines. And that is not a knock against the people working on them. They're incredibly smart. They're doing really good work. But that doesn't mean it's ready to be serving up information to the public. Okay? 

Sorry. I'm going on a tangent here. But one thing I kind of at least liked about it is that it at least started to trace back to the articles that it got the answer from. At least it gave me that. I know they've been working on this problem for a while for how you trace back sources. At least I could click on the sources. But I realized, if I'm just clicking on the AI box and I go down to the sources, am I just doing a Google search at that point? And it feels to me - this is just my feeling. Okay. I'm not a business guy. But it feels like to me everyone does not want to be left behind in the AI revolution. And it creates - the fear of missing out is shaping all these products. And it compromises what is such a vital tool of the information age. Google. I mean, it's synonymous with searching. And that work worries me. That's the kind of thing that goes from me being annoyed to me being a bit worried as to where these companies are pushing these things, if that makes sense.

[00:15:39] SF: Yeah. The sourcing thing is something that we touched on last time. Where, that time, there was really no attribution. Because the advantage of doing a Google search - and now there's a lot of AI that's going on in order to serve up those links in an intelligent way. But you at least know as somebody doing the search and clicking on the link, you have some level of agency, because you can look at where's that coming from. Is this a credible source? You can kind of use your own decision-making to decide. Is this something valuable to me? 

And then the disadvantage of just spitting out an answer is like, "I don't really know. Is that hallucination? Is it real? Where did it come from? Is it credible?" And now, Google and I think others are trying to sort of bridge this gap. But then you run into that problem of like is that just Google search in the first place? 

[00:16:23] CH: Yeah. And I think my larger - and we're kind of going all over place. This is going to be a fun conversation, hopefully, for people listen - but I wish it would sometimes say, "I don't know." I really wish that you ask it a question, it would be able to collate the information in a way. It's like, "You know what? This is actually not very clear." At least I generally don't get that response. I don't know if you get that response when you ask things where it just says, "Honestly, I don't know based upon the sources." It might say something like, "Well, there are multiple schools of thought maybe. But I don't know." If it says, "I don't know ever." 

[00:16:50] SF: Yeah. I think that's the challenge with the foundation models. I do think that when you start to get into more domain-specific - and we'll talk about some of the stuff later, some of the techniques like RAG. You can get to a place where it can say I don't know. 

For example, there's a project that we worked on at my company to build like an internal content Copilot where we did use RAG. And I did this recently in a talk where I did a demo where I asked ChatGPT who is the CTO of Skyflow. And it came back and said the CTO of Skyflow is Sean Falconer, which I am not. And then I asked our internal content Copilot and basically it says that it doesn't have public information on this. That's a little bit better. That's not perfect. But it was definitely better than just making up an answer in that case. 

And when you start to essentially think about this for research, and that's one of the things I want to talk about, is you might not be using things like ChatGPT as much to do relying on it for your work. But I'm sure your students are relying on this in some capacity. And I'm curious how that has changed things in the world of academics. 

Even recently, I was reviewing talks for a conference as one of the reviewers. And I could tell that probably 40% or 50% of them were using some sort of AI generation in order to write the talk. And I don't care if people are leveraging those tools to make the writing better. But I don't need to see delve in every 50% of those talks. There are certain signals that you can tell that they're written by an AI. You need to do a little bit more work. And I'm curious, what's happening in the world of university students at this point? 

[00:18:24] CH: Yeah. I mean, it's a great question. The way that - I was talking with one of my grad students just the other day. And he's probably one of the biggest adopters of these tools. But mostly for coding. And I'll just give you what he was telling me. And he was like he will use the Copilot autocomplete for a lot of things in his coding. And it's definitely accelerated his coding. It occasionally throws him a secret error that he has to then go chase down later and doesn't know where it's at. But, generally, that's been helpful. And he's like, "Yeah. But I don't believe in typing into ChatGPT. Write me an algorithm to do this. I don't believe in that." And that's sort of his line. 

For coding, I think that, overall, it sped up his productivity. He's definitely done a lot of work in the last year on new coding tools he hasn't used before. In terms of writing an English language, I definitely get - all of a sudden, the reports I get to read from my classroom students or the students I'm on, their thesis committees, it went from being English that wasn't particularly always there, to English is great. And I'm actually not sure though sometimes that the content makes a lot of sense. 

What that means for me is - I mean, before I could sort of skim and get the - if I had to read a lot of stuff quickly, skim and get the gist of what they're talking about. Now sometimes if I'm skimming. And if it's just a little too clean, I'm like, "Okay, I need to take maybe a little bit more critical eye to what they're talking about here if it's a whole bunch of text and really ask and figure out did this person just asked them to stitch a bunch of things together? 

Now, again, if they just take a bunch of lang - if they're using it as a tool to improve their English, that's great. I mean, I don't have a problem with that. What's important is that they understand the concepts and they're articulating this thing. They know how to articulate an idea. It's when they're generating entire content and the concepts are being completely generated by this thing that they don't understand, that's where a problem arises for me.

[00:20:11] SF: And how do you determine that? I mean, with the thesis, it's going to end up being a thesis defense where you can kind of sus that stuff out. But this is an undergrad class and someone's turning in a report, you might not really be able to do that. Unless you devise new ways of essentially evaluating students.

[00:20:26] CH: And, basically, the short answer is I really can't in terms of what the report has been generated. It actually forces me to make sure I'm evaluating the technical content. Like, "Okay, is this equation right?" Because at least I have the benefit of teaching a very quantitative course. I can look at the equations. Does that make sense? Does that not make sense? And at least I don't get the impression that they use it to generate the entire research paper, whole cloth, including the equations at least so far. I mean, anyone could be fooled. 

But what I can tell is that you can definitely see characteristic mistakes of a student in those equations that are not consistent with characteristic mistakes of maybe an algorithm generating those equations. Basically, I say, "Okay, you did this approach. You made this assumption. Does that assumption make sense?" I'm checking the content. Maybe that's actually where the rating should be. 

In a way, I'm not terribly bothered. And if they really need to learn how to write and they're not learning how to write, I guess, in a way, they're cheating themselves. But I don't have a way to say, "You wrote this with ChatGPT." I would not dare to use some of these checkers online. I don't think they're reliable either to accuse a student of cheating based upon doing that. Instead, I just focus on the technical content. That's what I do. 

[00:21:39] SF: Yeah. That makes sense. I mean, maybe I'd be curious to see what people who are teaching English literature or something like that or writing classes and stuff like that, how they're navigating all this sort of stuff and figuring out how do we tell whether - actually, you know what? This - and I think that also sort of, over time, is going to blur the lines, too, between what did I generate? Versus if I do the work to essentially make an AI do the thing, let's say it's actually good, then am I the creator or is the AI the creator? And we've seen some of that stuff in the world of art even where people won various art competitions and they're like, "Well, I spent, I don't know, multiple months, multiple hours prompting Dall-E or whatever to get the right artistic image that I want. I put in work." But that's a different type of work than I painted this, or I drew this, or whatever. And I think that's a big area that is also really hard for people to navigate. And everything's kind of moving so fast, it's going to take a long time for us to like kind of figure this stuff out.

[00:22:41] CH: I think so. But I also think that in that case - and I'm not an expert on the AI art of competition sort of thing. But to me, you make a new category. It's another medium. In a way, the AI in itself, the model is a new medium. Just like you wouldn't put a sculpture competition, I guess, a painting. You wouldn't put a painting in a sculpture competition. 

As long as you're honest as to what's being generated, yeah, I can't say that's not going to be art. There are definitely things that I would not consider art that come out of these, out of the stable diffusion algorithms. But you could be an artist that uses them. I don't doubt that. I think that that's the way that evolves. 

I mean, then you're talking about art. And maybe we talk about back to a student learning a technical concept. My first AI, let's say we'll call it AI with generous quotes. When I was learning - I got to figure this thing out for one of my algebra classes and I had a TI-83 calculator. And we had to do division of functions or something. I forget what it was. Divide functions by other functions. 

I said, "You know what I'm going to do? I'm going to create a calculator program. I wrote it that I can put in a function divided by another function and it will actually show me all the steps to do it by hand so that I can just write it out." And the way I - I think I showed it to someone. It's like, "Christian, that's cheating." I'm like, "What are you talking about? I had to understand functions so well so that in order to code this thing for myself to do it faster." And so, maybe people think I cheated in seventh grade, or eighth grade, or whatever it was. But that was work that I did. And I would have been offended if someone told me that what I was doing was not actually math. 

[00:24:23] SF: Yeah. Yeah, that's fair. One of the other things I wanted to touch on before getting into some of the technical stuff was around sort of jobs and some of your perspective on that. A lot of the things we're talking about is are these AIs ready essentially for sort of mission-critical stuff? 

And I was listening to an interview that Marc Andreessen from Andreessen Horowitz had done recently, and one of the questions that they asked was like what is his sort of prediction for how AI is going to change the world in the next five years or so? And one of the things - and, basically, what he said was he felt like it was going to democratize access to certain types of specialists. 

Essentially, suddenly, everyone is going to be able to have access to a really great doctor. Or at least like a consultation of a doctor. A lawyer, a tutor, which is currently something that maybe only really wealthy people can have easy access to. And I think outside of what he said, if you look at sort of sci-fi, a lot of times when we think about AI and the replacement of jobs, it's a lot of times kind of more on the blue-collar side of like, "Hey, automatic car repairs," or whatever. Flying robots and stuff. 

And given that you're in robotics and, also, you have some thoughts on this in terms of the limitations AI, what are your thoughts in the sort of the near future? Is there a danger? Or not a danger. But do you think that it's really going to be these sort of consultations of like a doctor, a lawyer? Or do you think there's going to be somewhere else that we're feeling the sort of near-term impact to jobs? 

[00:25:53] CH: I don't buy that it would be an immediately good idea to replace high-skilled labor like a doctor with one of these things. Specifically, because there's a lot of context that goes into what a doctor does. Or a psychologist, let's say. There's a reason a psychologist isn't supposed to diagnose someone without meeting them and actually like consulting with them. And it also comes from a lot of experience that is not data that's directly available to the algorithm. That there is a level of the expertise that's beyond I think just a bunch of training data. 

Now, call me a luddite. I'm not saying that's impossible to do. But I think what it requires is some kind of embodied data gathering. I mean, we talked about embodied AI is sort of the buzzword for kind of robotics these days. But at least it requires some embodied data gathering for the real-world. 

The kind of near-term stuff is really - again, if you need fast, low-quality, or it doesn't necessarily need to be high-quality content to be created, that's where this shines. And, I mean, I don't like being a Debbie Downer, the first thing that comes to mind are scams. I mean, scam emails, almost by design, are actually supposed to be kind of low-quality, because it'll supposed to filter out people who are discerning audiences for people who are not discerning audiences for scam emails and that sort of thing. 

I think that, as a fast chatbot thing, scammers use them a lot. All right? Now that's a really negative connotation. But, again, very low-stakes customer service, good. It doesn't matter too much. It doesn't fail if people get a little annoyed, right? I think things where the quality thresholds are low, that's where immediate impact is. 

I know that lots of call centers in other countries are now trying to replace with generative models. I mean, that makes some sense, hopefully. We'll see where that turns out in the long run. But high-skilled, I think that's a next-level challenge. That's my take. 

[00:27:49] SF: What about in terms of being more like assistive technology? We talked a little bit about code and Copilots, which now like probably - I don't know if I'd be hesitant to say the majority. But, certainly, a lot of people that are engineers that are arguably high-skilled are leveraging code and Copilots to help them do their work do it faster and so forth. Essentially, it's a lot of work. And, also, we set kind of unbelievable expectations on our medical doctors to be able to come into a room, talk to you for a couple minutes about while you try to articulate what's going on. And then they're supposed to have sort of photographic memory of everything that exists in medicine to be able to map that to a particular problem and then give you a diagnosis in the moment all within like 10 minutes. That's like a pretty crazy thing to expect on them. But if you have perfect memory through something like an LLM and then you build like an assistant there, is that essentially the path forward? 

[00:28:46] CH: In the terms of assisting, I think that there's a lot of potential there. I mean, we talk about coding Copilot. It's even like error checking. I'm pretty sure some kind of generative model popped up in my coding this morning and said, "Did you forget a comma?" "You're right. I forgot a freaking comma." I was coding in front of students live. I was like, "Oh, thank you. I appreciate the hint there." I have a far less of an issue with the suggestion. I think it's a great idea. I mean, it's basically super autocomplete or suggested autocomplete for various things. I have long thought that that was a pretty good thing. 

In terms of like labor replacement, I think that the labor replacement is on things that have a low-quality threshold and just a high-quantity, low-quality threshold. But I think it's a great idea. I've heard doctors say it would be great if when they're with a patient and they're about to prescribe something, something pops up. It's like, "Hey, make sure you check that they have those low-levels of this before you do this, because there's this interaction." "Oh, thank you." Or if they were like, "I don't know what you're talking about." It's as a reminder. Yeah. 

I think assistance can be helpful as long as they are appropriately specialized, I feel. I feel like there's this need to super-generalize these large language models of various things. It can do everything. I'm like does it need to do everything? It's like showing off. Yeah, it can do everything. But I'd rather it do fewer things reliably to do everything pretty well a lot of the time. I can understand people pushing toward that. 

But at that point, if you're doing an assistant technology, you have this whole user interface issue. And you know much more about this than me. My user interface is I tell the student what to do, then they do it and they give it back to me. That's my user interface. But that's its own challenge. Does it become smooth and workable? All those UX designer things that I know nothing about.

[00:30:36] SF: Yeah. I think on sort of the assistant versus the full replacements, I guess it's somewhat analogous to self-driving vehicles versus autopilot. We've had autopilot in cars for a long time where, essentially, maybe it maintains your speed on the highway or something like that. But you're there to drive. Or even in a Tesla version, you're there and it basically forces you to hold on the wheel for certain time so you can take over the car and stuff. 

But a human in the driver seat ready to take over at any point. Versus what the full-autonomous vehicles now, there are Waymos driving around San Francisco picking people up and driving them around. But there's a very, very high bar in terms of what it takes to be successful there. Because any kind of problem could be drastically terrible for a person leading to you know injury or death in the worst case. That completely changes essentially the expectation. And it's much easier to sort of get the assistive technologies out there in the hands of people to help them than the full sort of autonomy.

[00:31:40] CH: I love that analogy. And I think we actually didn't talk much about self-driving cars in our last podcast together. But I think it's a really smart blueprint for what has been working and what has taken a lot longer than what people expected. I remember the early days of the DARPA Grand Challenge, when DARPA, the government organization, put out the big competition for self-driving vehicles. And lots of university and incorporate teams came out. Made their self-driving cars drive across the desert. Drive in cities over the course of like four-ish years. And everyone thought self-driving was right around the corner. How incredible? And how awesome? And Google bought one team. Every car manufacturer tech company bought one of these teams and created a whole division. And we're still waiting on it. 

But at the same time, these assistive technologies, which frankly could be a lot simpler. It could be really helpful. For instance, just alerting me if I'm tired. They have those little alerts. I was driving a rental car one time that had one of these things and it's just beeping at me. Why is it beeping at me? It's like, "Oh, is because I am a little drowsy. That's right. I should pull over. I should pull over." Go to a convenience store for a bit. Maybe grab a Red Bull. Relax for a little bit. It was right. And, again, the stakes were low. And it was helping me. It's a simpler technology and I found it actually useful. 

Or like if I'm drifting in a lane, I guess. It's hard to tell what exactly what it's detecting sometimes. But that's great. The assistive technology I feel like should be much more of a focus. Even there, it can be complicated. Because we can talk about how these assistive technologies are trying to integrate everything, I guess. But I think that that makes some more sense than trying to straight up labor replacement. I think that's the way these things have gone for a long time.

[00:33:21] SF: Yeah. I think that, to me, in the immediate future anyway, that's sort of the way forward. We're going to see a lot of these kinds of assistive technologies come out that's going to help people, maybe they operate 5 to 10x faster. Or, essentially, the quality of their output is more reliable, because you can get sort of these error checks automatically like you do from like a coding assistant essentially. Make sure that I don't prescribe a drug that conflicts with the drug that the patient's already on. These types of things. I feel like that's where we will see the immediate impact.

[00:33:51] CH: And I think that that's something where you get the immediate impact together. And you can do some real studies with those professionals on how it actually impacts their reliability as practitioners. Because you can come up with all kinds of ZoomState scenarios where doctors just don't even think about it anymore. They just rely upon the assistant or whatever, which we don't necessarily know what happens. But really study that problem from an outcomes perspective. What happens when you do some real trials on it? What happens when start integrating these things into care? There hopefully should be a nice stepping stone of the type of practitioners where it is more and more safety critical. But I think that's a good way to go. And really do some good academic study on the impacts of these technologies on professionals. 

[00:34:37] SF: I want to start to talk a little bit about how you're seeing some of the impact within your world of research in robotics. How is generative AI starting, if at all, to impact the that goes on in robotics? 

[00:34:50] CH: Oh, it absolutely is. I can say for certain, it is very much at our doorsteps in robotics and very much in the field. I just came back a few weeks ago from the International Conference of Robotics and Automation. It's the largest academic robotics conference in the world. And there were entire whole sessions of insert your type of robotics here and learning. 

And, specifically, a lot of generative models. And people have mixed feelings about it. Just for background sake, the idea how ChatGPT would be in implemented with a robot would be to handle the high-level reasoning. While you have some other robot controller actually moving the arms or moving the legs around, what you tell it to do would come from a large language model. 

And from relistening to our podcast, one of the things I threw out as a statement was that, well, we can't even have a robot make a turkey sandwich at this point. And that's actually changing. That's changing quickly. I'm not saying it would reliably make a turkey sandwich. But now nowadays, you could type into an equivalent of a ChatGPT, a large language model like it the instructions to make a turkey sandwich. 

And a lot of researchers have then said let's use that with a robot and take those individual instructions, turn them into individual tasks for the robot to do the sequence to actually do a task. That is a thing that happens in robotics. You see it both in academia and also in this completely crazy number of startups that have jumped out of basically nowhere it seems to create humanoid robots that decided to help you in the home and in factories. 

There are YouTube demos of one by Figure AI, one of the many human robotics companies, showing their robots saying, "What is that?" "That's a rack of dishes." "Well, please put the dish we it belongs." And it picks up the dish and does those. It's using a large language model to reason about those tasks. And there are whole slate of papers at ICRA, the conference I was at, which were some people said it's ChatGPT on a quadruped robot. A robot dog. It's just like jump to the right. Okay. And it jumps to the right. People are doing it all over the place. It is definitely in the field. 

[00:37:02] SF: In terms of being essentially influencing a robot or using it for reasoning, does that have to be like an edge deployment directly on the robot? Or is the reasoning essentially okay if you're making essentially the slowness of a network hop? Because I would think that when it comes to actual like movement. That has to be sort of decisions or powered directly on the robot so that it's essentially fast enough. It's just like with autonomous vehicles, you can't do back and forth to the server. It has to live essentially on the vehicle itself in order to make decisions. Otherwise, the reactive time is not fast enough. Can it exist essentially on the server when it comes to these robotic decisions versus what has to live on the robot itself? 

[00:37:41] CH: The thing that can live remotely and then require a server call would be the plan that it needs to get. If I need to make a sandwich, what they currently do is say, "Okay, it requires these steps. You go to the cabinet. You open it up. You find the dish." You get those commands, right? Once you have that sequence of commands, everything else needs to run on board the robot with feedback control as we call it, where things are a little different. The robot will kind of nudge it around. 

The highest-level sort of this sequential planning level is where you can rely upon something that's slower. You wouldn't rely upon that for feedback control or for something that's going to happen on the order of 100 Hertz. Typically, the standard for what a robot needs to be making quick decisions about is something like 500 Hertz to 1,000 Hertz, kilohertz. Okay? And that's pretty standard for what's controlling your actual motors on the robot. Clearly, you don't rely upon that. That uses lower-level controllers, which can be traditional robot controllers that we've been using for decades. Something you might learn in a senior level mechanical engineering class where you literally are just looking at joint angles. And you say, "Okay, I want this joint angle. I throw this much torque to the motor to make sure it gets there." Simple controllers. 

Or it can be more complicated controllers sort of in the middle, which can also be based upon generative methods. Stable diffusion is actually the thing that generates your Dall-E images has been adapted for what we call Behavior cloning in robotics. The high-level - the task might be, "Oh, reach for the door." Okay. Got it. Reach for the door. Then the mid-level control would be something that comes out of, say, stable diffusion. Also, transformers, the T in GPT, that actually comes up with the joint commands to go to the door. Put your hand toward the door. Okay? 

Then there's a lower-level controller down below which is a thing that you can just write in a single line of code, which is typical control theory stuff. But people have found a way to train these diffusion models or these transformer models. Those are sort of the two camps, if you will. I think Elon Musk said, "Oh, which are you? Team transformer or team diffusion?" And he some kind of joke about that. The Sharks and the Jets, if you will. Okay? 

Pick one of those two and you train it with like on the order of 50-ish demonstrations of a human using a joystick or a haptic interface device doing that task. And it can then do the task. Okay. That stuff will happen on the order of like 100 Hertz or faster. That it's taking data from the robot. Sending it through whatever network it has. Comes out with an output, which is probably a joint composition command at 100 Hertz that's doing that. You don't need a server call for that. That's much faster. Okay? That's the distribution for software and where you store the information. And how you get it embodied in a real space for real-time control.

[00:40:35] SF: What is some of, from what you've seen, the most interesting stuff that's happening that's combining essentially generative AI and Robotics? 

[00:40:42] CH: Well, first off, it is genuinely impressive to me how many tasks some researchers have gotten a robot to do using either the diffusion policy method. Again, diffusion, that's the Dall-E. Making the art images adapted for this. And then the other one is called ACT, or action chunking transformer is what it stands for. It is actually genuinely impressive the number of things that they've been able to demonstrate using human demonstrations. Again, it's not watching the person necessarily in these cases. It's a human controlling the robot arm with a joystick. And then the robot can do it. 

Now, how reliable? Sometimes it's reported by researchers. And that's great. Sometimes not. But it can do a lot of things. If you want to look at a great example, just search the Aloha Arm. It's a project out of Stanford University where they created a very low-cost robot arm. Or two robot arm. We call it bi-manual manipulation. Bi, as in two arms. And they have it doing things like putting on a part of a fitted sheet of a bed. Or squirting ketchup into a dish. A variety of manipulation tasks. It's an impressive variety. Things that even require some delicacy to as well. That's really impressive. 

Again, what they can do - and my mantra is it's always impressive what they can do. But what can they reliably do? But I am super impressed by what they can do to the point where a lot of research institutes, a lot of robotics companies are rushing toward these methods. Thinking it's the key to making their robot really useful in the real-world. Because if they could crack that, that would be huge. Again, I'm an academic. I got to be skeptical. But it is impressive, Sean. 

[00:42:28] SF: Has it changed your research at all? Are you leveraging any of this stuff? 

[00:42:32] CH: Yes. The short answer is we're leveraging more of the deep learning approaches. Yes. We're trying to find the right places to include it. And I think I sort of talked about the two general approaches to control. There's one where you have a model. You've written the equations of the world somewhere into the robot's reasoning. And then there's the one without a model. Or model-free. Right? 

A lot of stuff these days is model-free that's really impressive. Again, if you talked to me 10 years ago, I was like, "Oh, model-based are the only people that are doing anything good at this point." At least anything impressive on a piece of hardware to be clear. Researchers from 10 years ago who did model-free, I did not mean to insult you there. You do great stuff. And look who's laughing now? You guys. Right? 

And right, now the model-free, that's your reinforcement learning. That's your stable diffusion. That's your transformers. All that stuff. That's really good right now. However, last time we were talking about how do we mix the chocolate and the peanut butter together to get your nice - just Reese's pieces? I don't know. I haven't eaten candy in a while. Get your beautifully-tasting mix, right? 

And a lot of people have been asking that question for a while. And, recently, there's been some really interesting approaches that do mix the chocolate and the peanut butter in interesting ways. And, certainly, we have our own approach to it. And one thing that is really interesting is people will use the neural network to handle all of the stuff that's really hard to model. And when I say model, I mean the physics. And that's things like your friction. That's things like the actuators in your motor. And, yes, you can write down models for these things. But in real life, it can be very hard to get the parameters right when you're running your robot out in the real world. 

One of the great examples of people who have been super successful with this is people at the ETH in Switzerland. A colleague by the name of Marco Hutter. His group was one of the leading groups to put deep reinforcement learning on a quadruped. And they're still just doing fantastic work. 

And the thing that that quadruped was able to do with these model-free, deep reinforcement learning methods that model based wasn't is you're able to handle lots of different contacts with the ground. It could kind of clamor around on obstacles to clamor over things. You give it a weird push, it will come up with have some weird strategy to put its leg down at the right places. That's impressive. And that was something that was very hard for us on the model-based side to replicate. Because we had to basically have a model for how things make and break contact. And that's actually not that easy in robotics to write down. 

People like our group are looking at ways to mix deep reinforcement learning where it makes sense in methods like that. But also, leverage things where, if you have a model, that's really handy. When you have a model, you have a physical understanding of the world, you can make predictions as to what's going to happen. When you can make predictions, you can come up with good plans for what to do. 

If I had no idea what tomorrow would bring, Sean, then it would be very hard to plan for tomorrow. Having a model for what will happen tomorrow is useful. The sun will likely come up, hopefully. And I can hopefully plan upon that. Right? And so, finding ways to plan ahead when your policy - sorry. Your controller. We'll call it your policy. Is this deep reinforcement learning thing that's kind of crazy and weird from a mathematics perspective. How do you use that and use that to plan into the future? That's a tricky problem. 

Recent papers this year have just come out that they come up with new ways to do that. That's a long way to say, yes, Sean. We're looking into this. But, certainly, we are far from the only ones. But it is still quite a spectrum. But one thing has definitely changed. There's so much more bleed over now than there was when we talked a year and a half ago on the field, at least in my specific subfield of robotics of legged locomotion.

[00:46:17] SF: And I'm curious beyond just sort of the general space of transformers or even sort of the Dall-E diffusion model type of stuff. There's a lot of stuff that's happening in the world that I know of outside of robotics to help build these AI-powered applications. There's things like retrieval augmented generation. There's all this growth in vector databases. Long context windows, fine-tuning stuff. I'm curious, out of all those things, how much of those are also touching the world of robotics? Is there applications for some of those types of techniques there as well? 

[00:46:48] CH: In terms of where these things are showing up in robotics yet, I can't say I've seen a paper that's done this. That doesn't say it doesn't exist. But the ideas are really - there's a lot of parallels to this RAG approach, right? The thing about this retrieval augmented method is that it's checking a source. It's going back and it's looking back to verify something. And verification is an entire field of robotics that predates any of this AI stuff. It's an important tool for showing that I can guarantee my robot will work. Guarantee subject to our model assumptions, performance not guaranteed. 

But from a mathematical standpoint, try to guarantee that it's going to work. And that is definitely a huge issue in robotics is verification. There are all kinds of cool mathematical approaches to it. There are people who try to look at a neural net and say, "How, with this messy, weird combination of functions, can I still verify in a traditional control sense that my robot would work?" 

This retrieval, this idea of a retrieval - and, again, this is me pitching a proposal to you if you'll fund our research here, Sean. If I were pitching this in a room with a program manager, it's like, "Look, we could use some of these tools. But our retrieval is not looking up a reference online. Maybe our retrieval is running a physics simulation of this thing our robot plans to do and see what happens." 

Imagine, we have one of these control workflows like this Figure AI that's doing a dish, right? You give it a command. You say, "Hey, put the dish away." All right. It calls ChatGPT on a server somewhere. Comes back. But before it does the policy, says, "Here's the thing I was planning to do." And uses the physics of the world to say, "Is that going to do anything like I wanted it to actually do?" And as a verification, in which case, you are retrieving a call to a model of reality. That's the thing you verify against as opposed to some reference that you're drawing. That's a cool idea in principle. 

[00:48:50] SF: That reminds me of actually a conversation that I had with Nvidia earlier this year, where they're doing a bunch of work with open USD and this idea of creating digital twins. Essentially, a model of some sort of real-world. It could be like a factory and creating a digital model of that, so that then you can run all these different simulations within the model that mimic, essentially the factory. And that way, they can figure out like, "Okay. Well, this type of machinery with this setup is potentially risky, because it's going to like cut someone's finger off or something like that." And I would think that that sort of thinking is also super, super applicable to the world of robotics. What you're talking about is like, basically, we can run essentially this plan in this sort of digitized space before we actually run the plan in the physical world.

[00:49:33] CH: Yeah. I totally agree. And Nvidia is definitely a big player and trying to come up with good simulations. And they call it digital twin. And manufacturing is called digital twin. And that language is starting to come into robotics more and more, where we just call it a simulation. That's one of the many trickled out effects of the AI boom is the more machine learning language is starting to now come into our phrasing in robotics where we already had paralleled phrases. We kind of got to learn to talk each other's language. 

And having physics simulations is super important. And it's one of the ways that that group in Zurich, the ETH does their good work in terms of robot control and many others to be clear. I can go on a great length at the various groups that do. And if they have a good simulation, at least. And they and they simulate the right things right. 

One of the key tricks that this and other groups figured out, that if you want to train how you're going to make a robot work or some kind of factory work - I'll talk about robots. This is what I know. Make a robot work from training it from a simulation, is that the physics you got to really get right is the physics of your motor and your box attached to it. 

One of the key things that they did was they basically ran the real robot through a ton of different test maneuvers. Flailing the legs around and stuff like that. And they took that data, used a much smaller neural network and then trained the neural network to come to model the motor and what we call the actuator. Right? And that became part of the simulation. And then, voila, as one of the major things that allowed the transfer from the simulation, the digital twin world, to the real world. 

Whatever field you're in, whether it's in operations research from a big factory, or you're trying to model how to make a fusion reactor work, a good digital twin is critical to the success of your method. 

[00:51:31] SF: In terms of like some of the stuff that we were talking about with applying RAG in this space, some of the advantages there help sort of fix some of the problems that we talked about earlier of like not having source attribution, hallucinations. Having sort of more reliable results rather than just essentially being able to do something. Like, can you do that thing sort of reliably? Based on your knowledge or the way that you've been thinking about this, why is it that applying, essentially RAG, it helps solve or address some of those issues? 

[00:52:02] CH: Well, I mean, it's something, say, analogous to RAG in your case. Or what I'm just saying is having some kind of verification before you roll out your policy onto your system is important. And my concern is that if we put all this stuff on a robot, you call LLM to come up with a series of commands that's going to hallucinate. It's the same analogous problem that you're going to get for all these non-embodied AI approaches. 

And so, especially, all of these applications are going to have some kind of issue with the hallucination like that. And the applications that people want for, say, humanoid robots. Basically, my field. Recently, people have been pushing for industry. Elon Musk is trying to say he's going to have his robots rolled out into factories by the end of 2025 doing real task in factories. I was asked to comment on this. And I was like, "Well, you got to be specific about what you mean by that." Is it's just going to be like a demo showing, "Hey. Look, it does this one task. And is that really that important?" I bet you could do that by the end of next year. 

But in most factory tasks, you have to have a lot of nines of reliability. How many 99.9999 do you need for a factory to be reliable to give you a product? I mean, the standard for a lot of companies was Six Sigma. Six Sigma is like one in a million things that go wrong. You're telling me an LLM or one of these methods is a one-in-a-million reliability on top of the hardware reliability of these robots? 

Anyway, my point is, in industry, there's a reliability issue. Okay? But the good news in an industry is that you can keep your robot away from people. And you don't have to worry about tripping and falling on somebody, right? 

Now the other one is home robotics. Home robotics. Wouldn't it be great to have a robot in your house that's doing all kinds of cool tasks? Look, my garbage disposal, I have to go fix it. Right now, I have to go fix it. And I got to go look up how to do it. Wouldn't it be great if my robot could do that? That would be cool. And maybe an LLM-powered thing with a transformer with an ACT or diffusion policy could do that. 

But when it's in my house, I'm going to be near it by definition. Grandma's going to be near it by definition. The baby and the dog. Safety becomes incredibly important at that point. You have a reliability there. There are techniques people try to do to solve either problem. People might try to make household robotics really soft and squishy. If it did fall on you, what's the worst that could happen? All right? 

None of these solutions are super compelling as of yet for what I wanted to do. I forgot what your original question was, but it had something to do with - oh, yeah. And the verification. The role of verification. Both things have reliability issues that are critical. And verification, it could be a step to solving that problem. 

[00:54:47] SF: It's interesting you brought up the sort of like six nines of reliability or whatever you need essentially. ChatGPT was actually down earlier today. And I looked at the - they have a public page, so people can you know verify this. But it lists essentially what's their uptime for their APIs, ChatGPT. And so, ChatGPT's service is 99.42% is their uptime. 

Even that, essentially this thing that where it's okay - and maybe that's why it's the way it is in terms of its uptime. But it's kind of okay if it's down. But, suddenly, when you start to put AI in a robot that has essentially the strength to hurt somebody or even just fall over and hurt somebody and you put that near people, then, again, the bar for the precision and the reliability goes way, way, way up. And that's really been something that I think slowed down things like autonomous vehicles, where, there, it's a much - I mean, self-driving is very, very difficult. But it's a little bit more constrained than having sort of just like absolutely general-purpose robot in my home where I can tell it to perform any type of task. 

[00:55:52] CH: The self-driving car analogy has always been my go-to for when people ask me, "Oh, are robots going to be in our homes in like X number of years?" Well, we've been trying to get robots on the road for decades now. And there are at least rules of the road, at least in theory. What are the rules of getting around your house? And part of the challenge is that robots now need to learn to play by our rules, which is so different than so many other technologies that we've had to develop over generations now. 

We created computers. We realized these are really great tools. And we trained a generation how to use them. We all learned how to type. We all know how to interface with them. And then, over time, we figured out how to make user interfaces that work better with people to streamline the process. Now they literally live in our world. That is a challenge. And I think that we would be foolish to expect it to be solved quickly. As optimistic as we can try to be, I think we need to be clear-eyed about the complexities of real-life. 

[00:56:55] SF: Do you think that if we - a faster path. I mean, obviously, this would be a faster path. But like a more reasonable path would be if I just want to create the robot that can essentially fix my garbage disposal. And that's the only thing it could do. Now we can argue whether you would really want to put money into that or not. But that's the only thing that it could do, that's a much easier problem to solve than having this sort of general-purpose robot. And that robot that fixes your garbage disposal could still potentially make use of some of these leap forwards and AI technology around deep learning and transformers and stuff. And maybe that leads to a better garbage disposal fixing robot than what is previously possible.

[00:57:33] CH: What you're asking for is the killer app in a way. What is the killer app of a application that people are going to be like, "Yes, things are way easier with this robot than they were before." Because, certainly, robots have a place in our industry. I mean, if you need to pick in place a thing in your factory down in that, that's perfect robotic tool. And, certainly, we've gotten way better at control and AI since then. What is the next thing? 

And I can understand why people are going to try - they say, "Oh, what if you had just one robot that did a couple of things?" Right? Instead of a robot that just did one thing, it did a couple of things. Like, "Okay. All right. We'd trade a few policies on that." Now what are those couple of things? Are they in a factory? Are they in your home? Again, once they're in your home, different standard. 

I mean, what do we have in our homes right now? I have a Roomba. I felt obligated as a robotic professor to buy a Roomba. And it does an okay job. Right? It's kind of a novelty. Maybe home is tough. Home is tough. 

I can see why people are going for industry. That is a huge market. You get into the industry. I can see why all these robotics companies like humanoids in a factory, but then you're like, "Why does it need to walk? These factories are pretty flat. Can it roll around, and you can have -" 

What I'm reasoning through with you right now, Sean, is all the conversations that we have about what's the killer app for robots. It's like, "Well, does it need to walk around?" Well, maybe, maybe not. Well, maybe it could roll around, so maybe then it has arms. Maybe it gets us closer to this idea of these two arms on wheels rolling around, doing a couple things around factories, maybe doing less specific repetitive tasks, something with a little bit more reliability. 

Oh, here's an idea. We're spit balling ideas here, Sean. We're doing it, okay? Start up right here. You run to the patent office, all right? We are - when you have a factory full of all kinds of stuff spilling around, maybe some things fall off the assembly line. Maybe a robot that goes over, okay, able to pick up the thing and put it back on the assembly line. Okay, it could roll around and gets it. Maybe that's a little bit of loss prevention that's worth the time. Or it's a jam over there. I'll clear the jam, right? 

Something that might happen in the middle of the night frequently enough in enough places where that one robot can do those things. Does that need to be humanoid? I'm not necessarily sure. I would bet not. But those kinds of tasks where a couple of things you do pretty well are the most likely to be successful, barring some major leap forward in the capability of these algorithms on robots. 

[01:00:08] SF: Yes. Or I guess another place is where the sort of human equivalent is not that capable, essentially. Or it's too dangerous for a person to do it. Obviously, it'd be better to have a robot. Probably the bar for success to have a robot explore Mars is less than have a human explore Mars. Similar to, I don't know, exploring a volcano. 

But I wonder if even certain types of rescue like human rescue where the risk to the person trying to do the rescue is too high to justify the rescue attempt. But then you could have, essentially, a robot attempt that. But, of course, there is challenges there as well because it kind of gets back to some of the political challenges around autonomous vehicles, I guess, as well. 

[01:00:47] CH: Well, sure. I mean if you had talk to any legged robotics researcher, I think, prior to 2019, they would probably have given an answer like that. Well, what's the use for a legged robot? Like, "Well, search and rescue, exploring areas that are too dangerous for us right now." 

The DARPA Robotics Challenge which ran in 2013, 2014, 2015, technically started 2012-ish, the challenge was to make a robot that would go into a simulated nuclear power plant disaster and shut off the valve. That's what it was about. It needs to be a humanoid because it has to do all the things a human could do in this plant. It was this incredible challenge. I wasn't on one of the main teams. I was on one of the offshoot teams for this, but I got to see it up close. 

These people had to make a robot drive a little RV to a nuclear disaster site. The zone in Pomona, they ran this whole competition. You get out of the vehicle. You would open a door. You would walk in. You'd have to grab a power drill and then cut out a thing and go over cinder blocks, go upstairs. All those things, right? That was - a lot of great robotics work came out of that, pushed the field forward a lot, the challenge of that. 

It has the similar issues you're talking about, which is it does need to be reliable. But because the operations are so niche, you could have a human overseeing it remotely, which what they called semi-autonomous applications. That was okay. The issues we were dealing with there were reliability of the algorithms on a different level or just reliability of the robot. Sometimes, these things were such prototypes. A cable would come unplugged or something like that or silly things that from a product perspective, that would cause these things to fail. 

There are these teams, just amazing work on these algorithms. I remember hearing a talk by one of them, and they said like, "We were running the robot, and it had this great algorithm." Someone accidentally left a switch flipped on when it shouldn't have been on. Then the whole robot started vibrating in the middle of the competition, right? It's that kind of thing we were dealing with. 

We haven't been paying much attention to that since the last five years of booming robots back and factories again. It's a cool application. I still think it's a great one. Frankly more ready to go but a much smaller market than revolutionizing factories. 

[01:03:05] SF: Yes, absolutely. Well, as we start to sort of come up on time here, I want to leave on sort of a positive note. What is some of the areas that you're sort of most excited about in terms of either these impacting your day-to-day or impacting the things that you're looking at from a research perspective?

[01:03:23] CH: I am excited from the idea of these assistance robots, to be honest. I have not dealt personally much into the co-pilot things, but my students are starting to convince me I should. I think that's cool. Again, coding, use a lot of learning how to code and program. There is that bar of learning the syntax of a language, right? That was cool. Like, "Oh, I can kind of translate this code to a different language, even though it's not perfect. It gets me started." 

Then it is a useful tool for very specific things, and I think that that is really - I love the approach that they took, this whole transformer attention-based. It's so interesting. It's an approach I never thought would take us this far. I sort of encourage people to say, "Hey, who knows what the next giant breakthrough will be? Don't get too laser-focused on this one approach. Maybe there's something from a completely different angle that's going to take us to the next level." 

I think that the meta lesson of these breakthroughs coming from a chatbot of all things should be something we take seriously. You remember the Alice days I imagine. Do you ever read what's called the Loebner Prize? That was like the Terrain Test prize from years ago. I would read the transcripts. For those who don't know what I'm talking about, there used to be Terrain Test prizes. Can you create a chatbot that would trick people into thinking it was a person and not a chatbot? You would read this transcript, and it was hilariously bad. I was like, "Man, this is going nowhere fast. We had a long way to go." 

Eventually, we just leapt over that hurdle. I think certainly, if nothing else, ChatGPT has leapt over the Leibniz Prize hurdle long, several times over. Then it also - I would never expect that it would be helpful for coding. They're both languages. I didn't think of that. That's my big positive takeaways. Look, even if I'm skeptical and I somehow end up being right, hopefully, I'm not, that like, "Ah, LLMs, they're good for some things but not others." There's something else. 

I mean, clearly, we can do all this stuff. We just got to figure out that right architecture that works for synthetic computing, and we're in business. That's very - that's by part super exciting. 

[01:05:38] SF: Yes. I mean, I think the belief right now is that this was such a step forward that people think like if we keep making these models big enough, get enough data, then essentially it's going to get us to potentially human-level performance. Then there's other people that I think are in another camp where they think there's essentially a ceiling to how far you can push these methods. That is going to require some other type of innovation, which is kind of like what you're talking about, that is going to be that next leap forward. 

I'm really blown away by what people have come out with in the last year as well. I have found a lot of use. Originally, I was pretty resistant to the idea of code and co-pilot. I'm like, "Hey, I've been coding since I was a kid. I don't need a machine telling me what to code." But it's incredibly useful, especially where I don't do as much coding day-to-day now, and some of my skills have slipped. Having that sort of reference there does allow me to move much faster.

I think at the very least, they're incredibly useful, even if there is a sort of limit or bar to them becoming these replacements for certain jobs or in the robotic sense coming in and being able to be my personal assistant or to clean my house and do all these other things. We'll probably get there in some capacity with something, but it's going to, I think, take much longer than some people are predicting today. 

[01:06:58] CH: I understand why everyone's rushing toward this thing. Very exciting. I think if I were to underscore this whole lesson here, I mean, this whole thing all sounds very sci-fi. But at the same time, I can't think of sci-fi that was written that's like ChatGPT, where the way that people write artificial intelligence in these novels, it's like, "Oh, all of a sudden, this rogue program comes to life and takes over." Not people were making a chatbot, and it kind of sounds eerily like a person, but it's actually kind of dumb. They're making it incrementally smarter.

The Arc of what the way this is all happening as of right now doesn't resemble what we had imagined these things going about. I think people need to be really open-minded for the ways that this thing can go around because it certainly exceeds the limits of how our imaginations predict things will go. 

[01:07:48] SF: Yes, absolutely. As we wrap up here, first, Christian, thanks so much for coming back. Then last thing, how can people follow you? How can they find out about some of your research?

[01:07:57] CH: You can follow me on website formerly known as Twitter @chubicki, C-H-U-B-I-C-K-I, where I occasionally talk about robotics, occasionally AI. But robotics is really my wheelhouse, so you're going to see me very upset at people making fake robot videos and passing them off as real. You'll find me there. 

You can find my website, christianhubicki.com. You can get in touch with me there. Search me on Google Scholar and see what our lab is up to or optimalroboticslab.com which is the name of my laboratory here in Tallahassee. 

[01:08:24] SF: Awesome. Well, thank you so much. Hopefully, we can fast-forward another 15 months or whatever. We'll be talking about even more sort of innovations that have happened in the space, and maybe it'll change a few things, I'm sure, very, very quickly over the next year. 

[01:08:37] CH: Looking forward to it, Sean. 

[01:08:38] SF: Cheers.

[END]