Ep 4: Fixie.ai CEO Matt Welsh on How LLMs Will Change the Way We Work

Jordan and Erica chat with Fixie.ai co-founder and former Harvard CS professor on the journey from researcher to Google eng leader and then to startup life prior to founding Fixie, a new platform for building LLM-based apps (a Redpoint portfolio company). We also talk about having Mark Zuckerberg in his CS class, ChatGPT Plugins and the AI ecosystem, and what the future might look like for our kids with AGI. You can find Matt on Twitter (@mdwelsh) and learn more about Fixie at https://www.fixie.ai/

• 0:04Speaker 00

  This is Unsupervised Learning, a podcast about the latest and the rapidly developing AI landscape and what it means for businesses in the world. I'm Jordan Siegel and I'm joined by Erica Brescia and we're excited to have Matt Welch, the CEO and co-founder of Fixie.ai, one of our newest investments at Redpoint. Fixie is a platform for allowing enterprises to easily connect LLMs to their organization's data, systems, and workflows. And Matt has an incredible background spanning startups, large companies, and academia, beginning with his career as a professor of computer science at Harvard, before becoming a principal engineer and director of engineering at Google, where he led several teams on Chrome. Matt then joined Exnor as head of platform, which was acquired by Apple, where he led several AI efforts. And then, most recently,

• 0:48Speaker 00

  prior to Fixie, was an early employee and SVP of Engineering at OctoML, a startup with more than $130 million in funding that makes deploying models to any hardware easy and optimized. Matt, we're so excited to have you here. Thank you so much for joining us today.

• 1:04Speaker 01

  Great. Thanks for having me.

• 1:05Speaker 02

  Matt, why don't you tell us a little bit about your, you know, earlier days, moving from, you know, being PhD, of course, at Berkeley and then, you know, going into research and then Harvard and then, like, what made you kind of, having gone through that experience, decide to go join Google?

• 1:25Speaker 01

  Yeah, so, you know, I did my PhD at Berkeley in the late 90s and graduated in, I think it was 2002, right at the height of the dot-com boom and then the dot-com bust, right? These days were really interesting, a little crazy. There was a period of time at Berkeley when I would be sitting in my office and the phone would ring and it would be, you know, "Hi, I'm CEO of, you know, such-and-such company" and they're just cold-calling Berkeley grad students to try to hire them. It was just crazy. It was it was absolutely wild and many of my friends and colleagues at Berkeley at the time did go into, you know, go into industry. And when I finished my PhD, I interviewed

• 2:10Speaker 01

  with this little startup company down in the valley called Google. And they, you know, they interviewed me, they actually made me an offer. I would have been employee number, I don't know, 400 or something like that at Google, which would have been crazy.

• 2:25Speaker 02

  Yeah.

• 2:26Speaker 01

  And I didn't want to do it.

• 2:27Speaker 02

  I'm kind of glad that you weren't, though, because we probably couldn't be investing in Fixie now. So everything happens for a reason.

• 2:34Speaker 01

  I would I would be living on the moon or something right now. Um, yeah, and it was just crazy times and it was great. I didn't want to join Google really because the idea of joining a startup back in those days, just it wasn't like a really established career path, especially for people with a research background. And of course, that's completely different now. But in those days, when you got your PhD, you know, the really great PhD students went on to be professors. And the not as great or not as academically inclined would go to work at a place like IBM Research or Sun Labs, and even before Microsoft Research existed. So, you know, it was the academic life for me. All of my

• 3:16Speaker 01

  heroes at Berkeley were also professors, of course, so it's a little bit of an echo chamber. But ended up going to Harvard as faculty, and that was a really interesting experience. You know, fantastic group of people, the faculty there are just wonderful colleagues to work with, obviously full of super smart people. And, you know, I had that opportunity to like, work with some like incredible students there. One of my students was Greg Brockman, who is now, you know, one of the co-founders of OpenAI. I had students that have gone on to be, you know, CEOs of very successful companies. And my most famous student was Mark Zuckerberg. He was in my operating systems class.

• 3:58Speaker 02

  That's so wild.

• 3:59Speaker 01

  At Harvard, it was wild. That class was in The Social Network, right? Yeah, there was a scene in The Social Network movie that featured an actor playing me.

• 4:09Speaker 02

  That is so cool.

• 4:10Speaker 01

  As the professor.

• 4:12Speaker 02

  What, do you remember Zuck from that period?

• 4:15Speaker 01

  Yeah, yeah, yeah, totally. And like, because the classes at Harvard were super small. I mean, this was like a 12 students in the class. Like, it was so tight knit that at the end of every semester of this class, this was the operating systems class, which was well known to be like this hardcore class that only the most dedicated students would take and they would, you know, pull so many all nighters to kind of finish the assignments on time. It was such an ordeal that the TAs for the class always made a T-shirt commemorating the class for everyone that had made it through. And so, you know, I've got one of these T-shirts and it's got the names of all the, you know, the students on the back. You can attribute Facebook's success to your class. It's 100%. It is completely me.

• 4:59Speaker 01

  That was every aspect of it. No, no, no, absolutely not, actually not, because I'm not sure Mark was in the class often enough for that to matter. But also, you know, he has so many talents and skills beyond what, you know, I could be helpful with.

• 5:14Speaker 02

  So, yeah.

• 5:16Speaker 01

  Yeah, so it was a great run. And then, um, you know, the thing that I've learned about myself over the years has just been that desire to try new things. It's like, if you've played a video game over and over again and you've gotten really good at it, you know, I'm the kind of person that wants to say, okay, I won that game, let me play a different game now, instead of just, just getting better at the, the same game. And so, I started a sabbatical at Google, really with the intention of being there only for one year and going back to Harvard to teach. But after getting to Google, after just a couple of months on the sabbatical, I just realized I was having way too much fun. And at heart, I'm a builder. Uh, I like to write code.

• 5:58Speaker 01

  I'm happiest when I'm building things and in particular, building things that people use. You know, the thrill at Google for me was, I was for the first time building software that someone other than like my grad students would use. And being able to build things that billions of people would use was just such an incredible opportunity and something I had not experienced in my career up until that point. And besides, just I'm a systems guy, I'm into systems and networks and mobile, and these are all things that, you know, if you're going to work in that, those areas, Google's a fantastic place to work on those things. Yeah. So that was that, that was kind of like how I made that transition back into, you know, from academia, a long period in academia, and then into industry.

• 6:44Speaker 02

  Yeah. And then you made this jump into the startup world with, you know, Exnur and then OctoML. So what drove that?

• 6:52Speaker 01

  Being in the startup scene, you know, like that feeling that I had going from academia to Google, going from Google to startup was like that same kind of like refreshment of my career in the sense that all of a sudden I was back to building again. I was back to doing things like really hands-on. It was like being back in the research lab a little bit. And just being scrappy and, and everyone in the company was pulling together in the same direction. That's a very different energy.

• 7:19Speaker 00

  You know, obviously you went from transition from academia to big company to startup. And I know you wanted to do a startup earlier before COVID, but what made you decide now is the time I'm going to go do fixy?

• 7:30Speaker 01

  You know, again, like coming back to that, being terrified. It's like I had no idea what I was getting myself into trying to start my own company. And there were so many things that in retrospect now, I'm like, boy, I should have done this a long time ago. It wasn't as hard as I thought it was going to be. But I just had no idea all of the things that one had to do to start a company. I decided I just wanted to do my own thing. I think that there's a period of time when you're in someone else's company and there's only so much you as an individual can control, or you as an individual can influence. And the experience at OctoML was awesome, but you know, I just sort of knew that at some point in my career, I wanted to do my own thing. I wanted to start my own thing.

• 8:14Speaker 00

  Well, I'm sure our audience would love to learn more about what Fixi does. Can you share more about that?

• 8:20Speaker 01

  Yeah, so the idea behind Fixi is enabling companies to connect this new technology of large language models to any software tool, any data source, any API, and build applications that interface language models to external software. So the real premise here is we're seeing that language models are kind of a revolution in the way people build software. And in particular, the language models are so incredibly good at doing things that we conventionally think of as needing to write some special purpose code to do.

• 9:00Speaker 01

  So just to take an example, if somebody wanted to build a piece of software that could take an English description of, "Please set up a meeting with Erica and Jordan for sometime next week." Well, what does the API look like for that? It's going to have to be some kind of, "Well, I need to know who Erica is. I need to know who Jordan is. I need to know their email addresses. I need to know what next week means. I need to know what's on your calendar and my calendar." And that's a very complex piece of software. So if somebody wanted to build that, it's a fair bit of engineering exercise to do so, right? It might take a good programmer a few weeks to make something basically work. With language models, though, the amazing thing is you can use the language model to

• 9:45Speaker 01

  take in that English expression, "Set up a meeting with these two people for next week." And actually, you can teach the language model how to call out to APIs to look up the calendars, look up the email addresses, and make the appointment. And that process is so easy because you can teach the language model by example. You can just give it a few shot examples, and it can generalize from a few examples and then figure out how to call out to, say, Google Calendar to do that fairly complex thing. So what ends up happening here is you're building software tools, but you don't need

• 10:25Speaker 01

  to write a huge amount of code to interface between the software tools and the end user. And what we want to do at Fixi is take that idea and enable people to build software systems with that concept at the core.

• 10:40Speaker 02

  I love it how you referred to it the other day as like the ultimate no code. It's such a good tagline, but it's it's so true when you can use natural language now to build things, right?

• 10:50Speaker 01

  Right. Like, like, who needs, I don't need no stinking code. I mean, it's really, you know, we talk about no code as though that's like a, like a benefit, but like, why did code come into the equation in the first place?

• 11:05Speaker 01

  It's not necessary.

• 11:06Speaker 02

  The former CS professor.

• 11:08Speaker 01

  Well, it's like, it's like talking about a bicycle is no gas. This doesn't make sense. It's not the same thing.

• 11:16Speaker 00

  Totally. I mean, it's a really powerful concept, right? And I feel like the fact that you can sort of interface with the model in a very abstract manner and it decides and just knows and can reason what to do is really interesting. How are you sort of thinking through, you know, potential use cases in terms of, is there a ecosystem of integrations? Or how does the agent process work? I'm sure the audience would love to learn more about that.

• 11:39Speaker 01

  Yeah, the core idea in Fixie is that we make it really easy to build agents where an agent basically takes a large language model and has some small amount of code that's connected to that language model that then knows how to wrap some existing piece of software, some piece of data in a language abstraction. And so our GitHub agent goes from English to GitHub API calls. Or the calendar agent goes from English to Google Calendar API calls. And by making it really easy to build those agents, you end up with an ecosystem of many of these agents, each one of which understands how to do some very specific, you know, domain specific set of things. What's interesting then is the language models are a really effective way at composing these

• 12:24Speaker 01

  because every agent just speaks English and it's the lingua franca between any software tool to any other software tool. If you think about something like Zapier, right, this is like a natural language powered Zapier on steroids. You can effectively connect anything to anything. And here's the other kind of amazing magical bit. Which is, an agent can ask the other agents in English for help. And I did a demo this morning of Fixi for someone. And I wasn't expecting this to actually happen. I was trying a new idea. And I said to Fixi, make a bar chart showing the top contributors to this particular GitHub repo by the number of commits.

• 13:06Speaker 01

  And what Fixi did was quite remarkable. It said, okay, the first thing I need to do is go off and find out who are the contributors to this repo and how many commits do they each make? I don't know how to do that. Let me ask if there's another agent in the system that can help. And it found the GitHub agent. The GitHub agent took that query in and said, sure, I can do that piece. I could go look up the number of contributors and the number of commits they made and I'll give you back a list. Then Fixi said, okay, now I've got a list. I need to make a bar chart out of it. Who can help with that? There's an agent called the chart agent that takes data and graphs it in various ways. And what was kind of amazing to see that, you know, the system itself basically figure

• 13:47Speaker 01

  out what it needed to do without my being explicit about, well, first you need to go to the GitHub thing and then you need to go to the bar chart thing. Like, I didn't do that explicitly. It was derived from that query that I gave it in English.

• 14:03Speaker 02

  It is, it's, it's really, truly remarkable. And I think starts to give you an inkling of what the future might look like. But kind of like zooming up a level and looking forward to pull out your crystal ball here. Like, how does work change in the next, let's say, five years, right? We can look at this on different time horizons and have very different answers. But like, how do you think work is going to change in the next several years?

• 14:30Speaker 01

  I think it depends a lot on the nature of the work today and the manner in which we think about integrating these new technologies into our work processes, right? So, you know, in the software development space, you know, we've, we've seen a lot of activity going on around things like Copilot automating the process of just typing out the code for you as you're in your IDE, which by the way, I could not live without.

• 14:59Speaker 01

  I was writing some code the other day and it was calling into an unfamiliar library. And, and I was like, okay, I could probably spend five or 10 minutes digging around the documentation or looking at Stack Overflow to figure out what are the, you know, eight lines that I need to write here to get the library to do the thing, but I'm too damn lazy. So I just wrote a comment and I said, do the following thing, and then started typing in Copilot, just filled out the rest. And I'm like, boom, that's it. And what it really meant was that I could stay focused on what I was doing. I didn't have to get distracted because if I go off and look at Stack Overflow, you know, 30 minutes later, I find myself on like Reddit or something, doing something random.

• 15:39Speaker 02

  Yeah, I was, I was talking to another founder the other day and he's like, literally my team now can do, in what used to take us like months, we can do it in days. Like it's just compressed time so much. And you know this, but for listeners who haven't seen the stat, you know, in, in repos where Copilot is enabled, like 40% of the code now that's being committed is from Copilot. It's just, I don't know if we've ever seen adoption and like productivity gains like this in such a short period of time. It's really, it's really quite remarkable.

• 16:12Speaker 01

  Yeah, it's quite remarkable. And, you know, I kind of liken this to, you know, like computer graphics. If computer graphics had gone like in one month time span to from, you know, like Pong to Red Dead Redemption 2 or something like, people should be freaking out because they never saw anything like this before. It wasn't like there was a crappy AI in my IDE before and it gradually got better over years and years and years. It was like, there was no AI and then there was this amazing AI.

• 16:41Speaker 00

  That's awesome. I'm a big gamer and I love Red Dead Redemption. So I love the analogy there. I mean, I think that's a good segue to, you know, talking a little bit about what's happened the last few weeks in the open source community with these LLMs. You know, we've seen the release of Alpaca and Lama and Dolly. And the cost of training and fine tuning these have dropped to as low as $600. And so I'm curious, when you think about sort of the future of this ecosystem, do you think it's going most of the usage to? A few providers like OpenAI? Or do you see hundreds of models being out there that people are picking and pulling from? And how does that sort of relate to Fixie down the road as well and the usage of that in the company?

• 17:20Speaker 01

  Yeah, that makes a lot of sense. My, my personal take is that it's just inevitable that we're going to have this Cambrian explosion of models. And they're going to be completely democratized and commoditized. We're already seeing that starting to happen, as you pointed out, and faster than we expected. Right? When Llama came out and then people started, you know, tuning it and doing other things with it with Alpaca. And I saw another one called Vicunia the other day. And I guess all of the Camelids of South America are going to be accounted for. It's just great to see that because it means that we're not going to necessarily be limited to just one model, one model provider. At the same time, there's a question about, you know, how much data you need to train these things and to make them work really well.

• 18:04Speaker 01

  You know, we'll kind of see how that goes. But my own take on that is the cost and the time to train these things is, is going to continue to come down. So with that in mind, I think we're going to see kind of two things. I think we're going to see a proliferation of general purpose foundation models that are all going to compete on the kind of general use cases. But I think more importantly, we're going to see a massive explosion in fine tune models for very specific use cases. And with Fixie, we're kind of banking on that because each agent in the Fixie ecosystem can use a different model from a different provider. And you can fine tune a model specific to your agent. So the calendar agent or the GitHub agent that I mentioned earlier,

• 18:47Speaker 01

  these can use a heavily tuned model just for that domain, right? The GitHub agent model does not need to know how to answer questions about Shakespearean literature or, you know, Renaissance period in art history. It just needs to know how to speak the GitHub API. And so i- if you can find tune a model for that purpose, the quality will be better. But also you can run it much, much, much less expensively. And you can take smaller, cheaper models and fine tune them up to the quality level that you need for that particular domain task.

• 19:21Speaker 00

  No, I think that makes a lot of sense. And I think even just the amount of open source models that we've seen released recently, it's really crazy to think what's going to happen over the next year and just how many will be out there. And I think that that makes a lot of sense for Fixie too, and sort of the differentiation with something like ChatGPT and then the plugin system there, where you guys are agnostic to various models and providers, and whatever is the use case that you need, Fixie can accommodate all of those.

• 19:48Speaker 01

  Yeah, we want to be independent of any one model provider, and our customers are demanding that for a number of reasons. And I think one of the most important just being privacy, that people are going to need to be able to run these models in-house on their own hardware and not send out sensitive customer data to a third party. The other side is, as I said earlier about fine tuning, but I think that there is going to be value for companies to be fine tuning and training their own models on their own data that differentiate from the general purpose models that are out there.

• 20:20Speaker 00

  We talked a lot about sort of how Fixie can automate workflows and augment people and things like that. To get a little bit more theoretical, going back to the professor's side of you, you know, there's obviously this race toward AGI occurring. And I'm curious to hear your thoughts, you know, how long do you think it will take to get there? And what does the space look like once that happens?

• 20:40Speaker 01

  Gosh, you know, whenever I think about this happening, I tend to believe that, first of all, I don't think we're going to know exactly when we've reached AGI. I think it will be really debatable for a long period of time, whether a given system that we've built seems to have the right level of intelligence. And we're already having that debate right now, today with the existing models like GPT-4 and Bard. And so it's a very interesting problem, because I don't think that we can just declare, hey, next Tuesday, AGI is solved, right? It will end up being this quite fuzzy thing. But assuming that we all kind of agree that gradually we are going to be in a world where machine intelligence is at the level

• 21:25Speaker 01

  of problem solving ability and reasoning ability and precision that

• 21:30Speaker 01

  Rivals or even exceeds that of humans. I don't think we're very long from that point. And the only reason that I believe that we're closer to that than we thought we were even a couple of years ago is just how rapidly this field has progressed. Really, the only difference between, say, the early transformer models and GPT-4 has been more computers, you know, more GPUs and more data thrown at them. There hasn't been any major, heavy lifting necessary to bring the models up to that level of skill. It's just been a matter of throwing more resources at the problem. So assuming that that extrapolation continues, that, you know, if it's just a matter

• 22:13Speaker 01

  of throwing a couple of hundred million dollars more hardware at the problem, I don't see that there being a place where we can't get there and soon. Of course, that's extremely frightening in a way. And, you know, I like to, you know, plug the Nick Bostrom book called Superintelligence that deals with many of the philosophical issues around what happens when we have a world in which there are intelligent machines that are more intelligent than humans. And there's a lot of factors to be considered there. It's a very, very complicated space. But one way to think about it perhaps might be a little bit like what happened in the Industrial Revolution is that, you know, we made a societal shift from being a primarily, you know, agrarian-based economy into

• 22:58Speaker 01

  one in which, you know, people were now working in factories and moving to cities and, you know, so many things were getting automated. And certainly, that was a massive societal shift. But it didn't, I don't believe, lead broadly to a situation where, you know, people were just completely put out of work, right? Or starving because, you know, the machines were doing all the jobs. What happened instead was that societal shift actually led to a place where there was far greater opportunity for many kinds of jobs to be done that didn't exist before, right? Manufacturing automobiles was not an industry before.

• 23:35Speaker 01

  That. And so what does it mean when AI is now automating a lot of things that we do by hand, right? The programmers of today are kind of like the farmers of, you know, the 18th century, doing the stuff by hand when they really should have a tractor.

• 23:51Speaker 02

  I mean, not just the programmers, though. I mean, when I think about this, I think about people working in support roles, right? And I think, you know, when I was at GitHub, support rolled up to me a while. And, you know, we were already doing a lot of work to automate our own processes. Or if I think about things like an internal, like an HR help desk, right? Where traditionally, people would be submitting a ticket and waiting for a person to reply. There's just across the spectrum of how a company operates when I think about how reporting works, or, you know, if I'd go to a team and want some data related to their business, like there are opportunities with AI in particular, to go from having to learn

• 24:34Speaker 02

  to use this BI tool, or even have a license for it in the way that like seat licensing works for a lot of these products today, that it's going to change. Like, you know, I think I see a world, and I see a world with Fixie, where I can go in and, you know, ask a question that involves looking across multiple data sources. And to your point, you know, maybe using one of the, you know, gen models to create some text based report, interpreting the data, or all kinds of other things that you can do that before I would have asked somebody to do. So, you know, I'm, I'm always thinking about like, what will those jobs look like in the future? And, you know, to ask you a question around that, you and I both have kids around the same age, which is a lot of fun to talk about.

• 25:18Speaker 02

  But also, you know, as an investor, I'm, I'm so excited about everything that's happening in AI. At the same time, as a parent, I'm like, how do I think about preparing my kid for the world that he's going to have in 10, 20, 30 years? And right now, he just came home, he started to learn JavaScript. I told you he's learning Python, like he's super into coding. I'm like, I'm really glad you're learning this for now, but you're totally, you know, it's hopefully good brain development. You're actually not going to use these tools in the future. But like, how do you think about the world for your kids and what jobs are going to be there for them?

• 25:53Speaker 01

  It's a wonderful question. And, you know, like, well, first of all, we must have the conversation sometime. When is the right time to talk to your kids about JavaScript?

• 26:03Speaker 01

  Yeah, my son is also doing a ton of JavaScript programming and it kind of makes me cringe a little bit because he's asking me a lot of very good questions, but I'm just not a very good JavaScript programmer and I kind of don't like JavaScript that much personally. So, anyway, I think that's a very hard question to answer and, you know, what is the world going to look like, you know, when I was a kid and if I wanted to learn something about the world, I mean, I had to go to the library or, you know, pull the Funk and Wagnalls encyclopedia off the shelf in my living room and like there wasn't this idea that I could just pull up something on the internet and it was like immediate and I could learn anything about anything just right now. And so, you

• 26:43Speaker 01

  know, knowing how to extrapolate from what's happening in the technological space to what the world will look like for my kids in 10, 20, 30 years, that it's so hard to extrapolate, so hard to know what that's going to look like. I do think that, you know, one thing that we are gonna see is that the introduction of natural languages, the way of instructing computers how to do things, is going to have a massive impact on the world because it means that humans don't need to have this super specialized programming knowledge in order to make use of computers. In other words, it opens up computing, I think, to, you know, orders of magnitude more people than have access to it today because they

• 27:26Speaker 01

  don't need a specialized skill set like knowing JavaScript programming in order to instruct them how to do things. And so, what I think that means for the future is the idea of automation and computation being embedded into our daily lives. There's not this priesthood of people that are, you know, building our apps for us or building our cloud services for us, that instead, you know, imagine a six-year-old that can speak, maybe can't even write or type yet, being able to instruct their system at home, their computer, their, you know, assistant, whatever it happens to be, to make something happen for them that they wanted to automate and it didn't require, you know, waiting for the latest app to come out. And I don't even know that any of us can imagine what that world's gonna be like. I certainly can't.

• 28:11Speaker 02

  It makes me feel better that you can't, since I can't either. I'm like, I don't have the answer yet. I just don't have the clarity. Teach them to be good humans, right? Which is always what I come back to.

• 28:22Speaker 01

  Well, you know, and one of the things I learned being an academic was that, you know, as much as people think about academic researchers as their job being to like foresee the future and to kind of, you know, envision these amazing things, you know, as late as, you know, 1999, the research group that I was part of at Berkeley, which is, you know, arguably one of the top computer science programs in the world, was envisioning a future in which people would be walking around with PDAs using infrared point-to-point communication with kiosks as the new way of doing computation. Like even us in the bleeding

• 29:04Speaker 01

  edge on the very forefront of computing research could not have foreseen the iPhone or Wi-Fi. And that is, I think, a really interesting lesson because it says like, no matter how much you know about what's going on today, your ability to actually envision the future is unrelated to that.

• 29:24Speaker 00

  Well, we like to end these podcasts with some hot take questions, keep them the same each time. And so, you know, first one is, obviously, AI Twitter is more active than ever before, given what's been happening in the space. And so what's one thing that you think is overhyped and one thing that you think is underhyped?

• 29:41Speaker 01

  Yeah, well, I definitely think there's, there's too much overhype around generative AI in general. And that's mainly because I think that we're all finding new use cases every day. But the real trick is going to be to make this stuff not just a demo, but something that is deeply embedded into the way that we build products. Underhyped, you know, I think definitely the idea that building out the rest of the system support to build tools around AI. Right now, there's just not a lot of machinery to do that. It's really hard to build AI-powered systems. And I've seen, you know, snarky comments from people saying, oh, you're building this infrastructure, but it's just the same thing as we had, you know, three years ago for this other thing. No, I think that this is a different world

• 30:26Speaker 01

  and that the problems are different.

• 30:27Speaker 00

  I think that's great. I think a lot of the tools out there right now are very much sort of prototyping tools, for lack of a better word.

• 30:34Speaker 01

  Right.

• 30:34Speaker 00

  Things that can really advance it to the next stage where everyone's using it in enterprises or in their workflows. I think that makes a lot of sense. Second one, how do you see governments getting involved in AI going forward?

• 30:45Speaker 01

  Hopefully as consumers and not as regulators, I think, is the way I see that. We already know that AI plays an extremely important role in many government functions and I think should play a more important role. I think that there's a huge opportunity here for AI to improve the way that the government operates, about the way it manages itself, the way it, just like any other company. The regulation is something that I think would not be well-advised, in large part because we don't really understand what the future holds. And I think well-intentioned regulation could very much backfire. And fortunately, we've seen, at least in the United States, I think, a very much a hands-off

• 31:28Speaker 01

  approach to that to date, and I kind of hope that continues.

• 31:31Speaker 00

  Awesome. Well, this has been awesome, Matt. We really appreciate you taking the time here. I'm sure folks are going to want to learn more about yourself and learn more about Fixie and keep up to date. What's the best way for them to do that?

• 31:42Speaker 01

  Yeah, just check out our website. It's fixie, F-I-X-I-E dot AI. And you can sign up for an account, you can play with the platform, you can check out all of our tools and examples. And if you want to get in touch, our email address is on the site.

• 31:56Speaker 00

  Awesome. And do you and Fixie have a Twitter account?

• 31:59Speaker 01

  We are on Twitter at Fixie AI, and I am on Twitter at MD Welsh.

• 32:04Speaker 00

  Awesome. Love it. Well, this has been great. Again, I know we covered a lot of different things here, all very interesting. And so, thank you so much again. This was fantastic.

• 32:14Speaker 01

  Yeah, thank you. I had a good time.

• 32:15Speaker 02

  Thanks, Matt. Let's go Fixie.