Alan Nichol—co-founder and CTO of Rasa—shares his journey from physics PhD to conversational AI pioneer and the story behind Rasa’s rise from early chatbot experiments to a leading open framework for enterprise dialogue systems.
In this episode of The Dialogue Architects, host Lauren Goerz welcomes Alan Nichol, co-founder and CTO of Rasa, to unpack his journey from physics PhD to becoming a pioneer in conversational AI.
Alan reflects on the early chatbot experiments that seeded Rasa’s creation, the lessons learned from the first wave of conversational interfaces, and why today’s AI assistants demand a more thoughtful approach—one grounded in user trust, governance, and real-world problem solving.
Their conversation explores:
Alan provides guidance for enterprises navigating rapid AI innovation—and a grounded vision of what it takes to build human-machine conversations that truly work. Clarity, responsibility, and building AI that solves real problems.
Episode Timestamps:
(00:00) Introduction to Dialogue Architects
(00:07) Meet Alan Nichol: From Physicist to CTO of Rasa
(00:42) The Origin Story of Rasa
(01:13) Early Challenges and Innovations in Conversational AI
(03:03) Building Bots and the Evolution of NLP
(07:40) The Rise of Rasa and Open Source Contributions
(12:36) Company Culture and Hiring Philosophy at Rasa
(15:42) Maintaining Integrity and Trust in AI Development
(19:18) Early Bets and the Evolution of NLP Classifiers
(20:02) Challenges with Classification Paradigms
(24:14) The Shift to LLMs and Honest AI
(25:20) The Overuse of Conversational Interfaces
(28:06) AI Assisted Coding and Abstractions
(31:19) The Future of Conversational AI
(37:58) Advice for Enterprises and Final Thoughts
Links & Resources:
Subscribe & Follow The Dialogue Architects on YouTube, LinkedIn, and Apple Podcasts
[00:00:00] Welcome to the Dialogue Architects, where we explore how enterprises can thoughtfully design, scale, and govern conversations between humans and machines. Today we're joined by Alan Nichol, who's famously written on his LinkedIn bio doing chatbots since before they were good. An engineering PhD physicist turned co-founder and CTO of Raza, and also my boss.
[00:00:19] Thanks for joining us today, Alan.
[00:00:35] Hey Alan, welcome to the Dialogue Architects. Looking forward to having you on the show. Ultimately, it's also kind of an interesting scenario, conversation with your boss. Excited to, decided to interview you here today. This time you're in the hot seat, Alan. Yeah. And um, I'd love in this, uh, section, maybe if we can kick it off.
[00:00:50] Um, just let everyone know a little bit of an intro about yourself, why you're on this podcast today. Yeah. And in general, we're gonna cover a lot about kind of the origin story of Raza, maybe Alan himself as well. Well, no, I'm, uh, I'm excited that we're kicking off this project and I, uh, I saw the list of people that you have, uh, uh, joining us on the pod, and I think it's gonna be, I think people are gonna like it.
[00:01:11] I think there's gonna be a lot of good stuff. The first question, um, I know you've been in the conversation AI space for many years. Mm. Did it really all begin in 2016 or did kind of your journey. To create raa start earlier, kind of what about your background really led you into this space? Yeah, yeah, it's a good question.
[00:01:27] Um, 'cause you know, I was tinkering around with chatbot, like things definitely before some, anything that started to look like raa maybe kind of going, going right the way back. So my background is in physics and I was working on a PhD doing physics, but machine learning applied to physics and. I think this happens to a lot of people.
[00:01:51] You kind of get, and you kind of fall in love with like the methods and then maybe worry less about the application that you're building and you get really excited about the kinds of things that you're, you're able to build. Um, so kind of fell down the rabbit hole of ML and all the things happening there and, uh, and natural language.
[00:02:08] Was working on a search engine. So, uh, was building a search like, I guess you would call it now, sort of a, uh, a prehistoric version of glean, you know, sort of enterprise search rag without the G right. Without the G. That's just rock. Just right. Yeah, exactly. And also this was before GDPR, so we, we could do all sorts of, like, you can make these things simpler in, in many different ways, but let's not go into all that.
[00:02:35] Just violate all the privacy, all of that. Yeah, exactly. Well, no, 'cause it, we, we built, yeah, we built like a personal search engine, right? So it would search all of your stuff and it was like a browser extension, and it would look like the little spotlight search that you have on the Mac and. You could search across like all of your stuff, like your Dropbox and your Google Drive and like the different apps you use for work and, you know, you could find anything across all these places.
[00:02:57] And uh, that I think was the sort of the starting point for getting interested in natural language and how do you understand what people are asking for and how do you go beyond just looking for keywords and how do you really try and understand and, and have a conversation with people. So I think that was the sort of nexus of it.
[00:03:16] And then. I would say another big factor in the whole thing was Slack as this app that, you know, 'cause we thought that search was gonna be the thing that pulled everything together. Right? We thought that's gonna be the sort of universal, universal entry point for for work. And then Slack kind of came out and it was this very exciting thing and then like everything was gonna be connected to it and all your apps were gonna like publish information into it and you would like trigger things from Slack and all this stuff.
[00:03:41] It kind of felt like, oh, that's actually the place you wanna have as a sort of central command. And then you could build bots for Slack. And that was a very exciting thing. Right. And that was sort of. Like 2015, something like that. You could build these little apps and we built like silly ones like based on cartoon characters and we built like serious ones that were useful and interesting, but it was just this sort of very exciting thing, right?
[00:04:01] It was something in some ways like so much more lightweight than building a mobile app or building a website. In other ways so much harder. Right. But conversational interfaces 1 0 1. Yeah. The, the interaction model is so much more complicated. But, uh, but we were building a bunch of these things and, and one of the ones that I like to talk about was we had built a chat bot for marketing teams and they could talk to their marketing data.
[00:04:30] So you could, okay. Um, ask questions like, you know, how much are we spending on Facebook campaigns? And you would get, you know, a graph or a table and, you know, it was sort of English. SQL was what we were doing in the background and we had some teams that were paying for this, right? We had some folks working.
[00:04:51] You could ask these really sophisticated questions. You could say things like, you know, what was the ROI on all my campaigns by country for the last three weeks? And we would like, you know, like we've very quickly found out, like been through it. Nobody talks like that. Like nobody, nobody, nobody thinks like that.
[00:05:07] If people thought like that, they would just write sequel themselves. So people would talk to this thing and they would go, how's Facebook? Right? Or, you know, how's Germany looking? And then you realize you gotta have a sort of real conversation with people to understand what it is that they want and they would've follow up questions and they would wanna do deep dives and all this interesting stuff.
[00:05:25] So we kind of, um. Discovered this interesting set of problems of, well, how do you have like a true multi turn conversation with people? Um, and you know, I mean it's obviously it was new to us. It's not like it was a new problem. And that was also the great irony of it was when we started looking into it, we said, you know, what's there for, you know, what were the tools there for developers to build these things?
[00:05:48] And it was all like very rudimentary, right? There were a hundred APIs and a hundred frameworks all there for building Hello World. And there wasn't something that was. You've done Hello world, and now you wanna do it for real, right? Yeah. And this is, this is the framework you wanna use now that you know what you're doing.
[00:06:02] Um, we said like, surely we should build that. And surely there's some good ideas out there that, you know, people have thought about this problem. And the great irony is that. Like in the old department where I had done my PhD in the floor above where I'd been sitting was this like research group on dialogue systems that I never knew existed at the time.
[00:06:22] Never knew those people or the work that they were doing. But they had, you know, done decades of work on spoken dialogue systems, which was sort of, um, anyway, very different set of techniques to what we use today. Even very different set of techniques to, to how we do things at Rasa. But was that sort of initial, Hey, actually there's lots of ideas here about how to do this in a better way.
[00:06:42] But there's nothing packaged in a way that actually sort of like developers are gonna use it, right? It's all sort of code for other researchers. So it's interesting because search led me into chat, and now chat is search again because everyone does drag. So you know, you kind of, you never escape the same set of problems.
[00:07:00] You just keep doing the same thing over and over. I guess if you add me at the end, it'll just be rage. You know, you see how much, how much longer it keeps going back and forth, but nonetheless, um, I think, I think that that's also maybe I, I met Rosaly later on than, because I think you all kicked off in 2016 and then I met Rosaly in 2021.
[00:07:20] But I remember one of the things that kind of resonated with me at that moment was this concept of trying to create. A foundational or default framework for conversationally I, and I think that's something that Raza kind of went after and in a way that other folks hadn't done yet is kind of say like, Hey, we're all doing this in lots of different ways.
[00:07:36] How can we find common ground or like a default framework to make this all hang together? So I think that was, and now that was how, how the, the thing started as well was before we had any inkling of what rasa should be or like, that there was gonna be a product there, or that there was even gonna be a company there.
[00:07:53] We were running a meetup and we had a meetup for chat developers called S Berlin. And you know, there was just lots of enthusiasm about the topic. Mostly because Facebook Messenger had launched as a platform, so it was like Slack, but with a billion and a half consumers that you can build apps for and, and, and try and monetize and.
[00:08:14] You know, there was all this like hype and excitement and all these folks were saying, yeah, okay. Well right now we're using, and this is also a funny parallel, everyone said, right now we're using, you know, Google's N-L-P-A-P-I, or there was one that had been bought by Facebook and there was one that Microsoft was offering that they'd built in-house and sort of like, well, I am.
[00:08:36] Entirely dependent on this third party service for like every interaction with my app, right? It like, I don't know anything about what the users are doing until this thing, this N-L-P-A-P-I that I'm, that I'm calling is telling me what the users, uh. Are saying and doing. And that's a sort of precarious situation to be in because I'm trying to build a business here, and I don't know how long these things are gonna be free or when they're gonna start charging me for it.
[00:08:58] And I'm sending all this data to other people and I'm just like, complete. It's a complete bottleneck and I'd love to be independent of that. And so everyone kept telling us, oh, we're gonna do our own NLP Mm. And like, we don't have time for it now, but like next month we'll totally get around to it kind of thing.
[00:09:14] And that kept happening. Um, and it was all for pretty consistent reasons of sort of, you know, owning such a critical part of your stack. And we said, okay, well, you know, we've been building some stuff. We have a decent NLP engine or NLU engine, I guess we called it. Um, like, why don't we just open source our, and people can contribute to that rather than everybody, you know, trying and failing to, to roll their own perfectly naive thought, but it was just like, it was just the right product at the right time.
[00:09:41] You know, it was sort of that. You know, it had been long enough that everyone was fed up with these, depending on these APIs, and it was just an open source package. It was like, here's a drop-in replacement for whatever you're using today, and you could run it yourself and you could migrate within like a couple of hours.
[00:10:01] We'd really made it easy to, you know, ingest the data and the API was compatible and everything else. But it's also funny that. Now everyone's building all these apps that are totally dependent on like a third party API for, for like core function. And no one even questions that anymore, right? It's just sort of, well, yeah, that's it, right?
[00:10:19] I mean, aw, WS went down last week. Um, you know, what's it like when, when AI goes down or aro goes down or something, or both at the same time? Goodness. Yeah. I guess maybe then. I'd love to look in there. Well, first I think with, with the NLP pipeline that you were kind of talking about there, that might well be the longest lived part of the Raza platform.
[00:10:42] Is that possible? I think 'cause, you know, in, in some way or another, it's still evident today. Um, and we'll talk more about that later, kind of the future of intense manifestations, things that you've called out that came to, came to light, what may come to light in the future, but right. Nonetheless. I think that's, that's probably cool.
[00:10:58] You know, if, if I were to look at the entire Raza stack, I feel like that that's the longest live piece of it. It's definitely the longest live piece of it. It was really the sort of that initial piece and that initial prototype, and I just can't even remember, I can't even describe to you how janky it was, but it, it was just, it was so very, very simple.
[00:11:16] It was like, Hey, we can do, you know, text classification in a very, very simple way. And of course this was. Pre Bert, I mean, this was certainly pre transformer. Uh, and this was, you know, the big exciting thing we were all hyped about was called word to vec. And it was like, Hey, you can do maths with no with words now.
[00:11:36] Right? And you can represent words in this interesting way, and you can do, yeah, you can, you can build a simple classifier on top of that. And, um, but yeah, I think that's a, it's a, it's a very, very consistent part of the product in terms of, you know, how you feed data into it. And what the API looks like has been, uh, definitely the longest lived part.
[00:11:55] I mean, the. Components that go into it. Uh, plug and play, you know, the Lego bricks people have, uh, you know, had different choices over the years and we've introduced new models and things like that. But yeah, that core concept has been, yeah, really pretty remarkably stable. And then I remember you were mentioning earlier in our conversation that, you know, you were, when you were first starting, you were kicking around with Slack bots and kind of trying different things.
[00:12:19] I'm curious, just because I miss this kind of like golden early startup age, the Berlin office is thriving. It's there. Yeah. What was that like? Is there any good stories you have? Um, for, for some of us older, let's say, or later s not necessarily older, um, you know, of, of, I don't know, interesting stuff that you built.
[00:12:36] Stuff that worked, didn't work. Funny stuff. Um, I mean, you were, because I feel like especially in early startups, you have like a bunch of really clever but also very creative. You know, 20 somethings, 30 somethings, building things, which I think always is exciting. Definitely. And I mean, we were, we were very, very scrappy as you would want to be.
[00:12:56] Right? And like very, very frugal. And uh, you know, the nice thing about that. Ethos, especially if you're an early stage company that's open source, you can really punch above your weight and you can really be a thorn in the side of so much bigger companies. Right. And I think that was sort of something that drew us all into it and, and made us all very excited about it.
[00:13:15] It was like when we would like dislodge one of these enormous players from some large enterprise account because they were just, you know, all in on it. And something that I, I think was really special from the early days of Rasa was we, we bet on some people. Who on paper you would not hire to do a specific job, but just because, you know, we were never the kind to think about like years of experiences, the thing that we cared about.
[00:13:42] It was purely, you know, what's our judgment? Can this person do this job? And I think we, we took a bet on some really bright people. Who did a fantastic job and also had, you know, tremendous career growth and like went on after Rasa and did other, uh, incredible things. Uh, but you know, we hired a lot of people into roles that had like, never done that role before, or some of us had never even had a job before, right?
[00:14:04] We're just at Fresh Outta University or something like that. But, um, I think we hired some really, some really great people and did some great things. Um, and it was always through that, um, that lens of. Everyone else is telling you that this is easy. And, and, and Ross is like actually here for when you've done that and you've realized that it's not so easy and maybe you wanna do it properly.
[00:14:30] Right? And uh, and I think there's just a lot of appetite and a lot of, especially at the time, like a lot of over promise of what conversation I could do. And, um, and that sort of clashing with reality of, of what, you know, people were able to achieve realistically. And a thing that. I think is really special about those early days of Rasa was that I, I used to describe it as quite a porous company because the big rasa champions at our biggest customers and our, you know, people who were looking at rasa and maybe potentially going to become customers.
[00:15:05] You know, they, they knew the code base, they knew the product. They were contributing to the code base. They were. As expert in Rasa as, as almost anyone who worked at Rasa. And at the same time, our, you know, customer facing teams, our support teams were so deeply embedded and, and so knowledgeable about the projects and, um, so sort of on, you know, intimate first name terms with all these people that they were building together with, that it wasn't always, it didn't always feel like there was even a clear line of like, who's actually.
[00:15:34] Officially in the company and who's out, right? It always felt like yeah. Kind of this continuum. And we had such a like large community that we were building that also felt like part of the community, right? And some of those folks end up joining Rasa and, um, and, uh, you know, we're, as some of those folks are still at Rasa, right?
[00:15:52] Yeah, I think that's, that was actually one of my questions I had for later, but at the end of the day, I think at Raza we talk a lot about building and designing for kind of high trust conversations. But I think it's, it's also, um, trickles down in the sense because working at Raza, I feel like also a lot of my colleagues are in incredibly, you know, have a lot of integrity and are, are high trust in that sense as well.
[00:16:14] Yeah. So I think it's part of the company culture that I, I really. Like about working at Raza. And what do you think, how does it come from that alone? Or how did you manage to keep that going over so many years? Because I, I, you know, I come from the Bay Area, I've worked with Bay Area Tech. It's a much different vibe than it is in, you know, this just incredibly transparent and, um, humble, I would say, kind of culture that we have here.
[00:16:41] Yeah. Um, it's a, it's a good observation. I think it's maybe one of those things where. It's there at the seed and then it becomes self-reinforcing. Um, yeah, and I mean, of course, you know, starting as, as an open source product. Right. And we also had a rule, you know, no brains on the website. Right. That was really our never any brains on the website.
[00:17:04] Um, and no, we, and we had that, that conversation, um, when we were kind of, yeah, thinking about our identity as a company, as an early company, and. Something that loomed large in my mind at the time was this deep mind acquisition, which was like, wow, you know, huge acquisition, incredible. Like what a company.
[00:17:24] We thought like, should we try and be like the deep mind of conversational AI and just be a deep research company and you know, product will come later and we'll just do lots of r and d and you know, who knows? So you kind of have these sort of forks in the road where you can choose an identity, right?
[00:17:39] And I think ultimately you kind of, you choose what comes naturally. To you. Um, and for us that was, it just felt right because there was so much fluff. And I mean, there's a lot of AI fluff now, but like, at least AI works a lot better. So Yeah, that's true. You can get away with that. Right. And you had the gap between the fluff and the claim is, is, is a lot less than it was.
[00:18:03] Right. And, and, and, you know, people would say, oh, this is like, you know. And I mean, that was, I think the, for me, a big motivation for starting the company in the first place was there was all this talk about conversational AI was gonna revolutionize X, Y, and Z. And then it was like, all right, how are you building these things?
[00:18:19] Well, first we have a classifier and then we have a bunch of if statements saying like, what to do? And I go, well, that doesn't sound like transformative technology to me, that sounds like, you know, the same, same way you would've built this 20, 30 years ago. So what's really new here? Yeah. Um, so. Yeah, I think it kind of came out of that early ethos of, um, building something that was loved.
[00:18:46] By the people working on it. Right. And, um, building something that people could identify with. And that was certainly an early goal, was that people who work with rasa aren't just users, but they think of themselves like, I'm a rasa developer, I'm a rasa user. Right. That the people would put that on the CB and they would think of themselves that way because that's, you know, obviously an important piece of, um, I guess like loyalty or affinity or just being a true champion of a product.
[00:19:14] I even notice that sometimes on LinkedIn when I go and search for, you know, people who work at Raza, I'll notice there's a bunch of people from other company who have like Raza even and sometimes like on their resume, on their LinkedIn, like they work there. You know, Raza developer. I was, I mean, yes, technically yes you are.
[00:19:30] I mean, yeah, no, pretty cool. Great. I'm very encouraging of it, right? It's, uh. I think part of this also has come from you. You've had a lot of, early on, big bets that kind of have also panned out in reality. And one of which is I, I'll, I'll, I think, um, the article you wrote was something intense, our House of Cards.
[00:19:46] Um, things like that. I, I'd love to ex, you know, dive in a little bit there because while the classifier, you know, the Raza classifier, NLP classifier is something that still exists in our product today, is like the cornerstone of early research. Yeah. At the same time, even from the beginning you were like, Hey, I think this is only, this is only short term.
[00:20:02] There's gonna be something that comes after. Um, what made you so have such conviction? A direction that we pursued, which was a kind of a dead end and we pursued it for quite a few years. So, so the literature was pointing in that direction, was kind of live in this world where you accept that the way you understand users is just like this classification thing, right?
[00:20:26] You have, you predefined these categories, right? You have the person greeting you or they're saying thank you or they're asking for, you know, an account upgrade or they wanna cancel or whatever, right? You these predefined categories and the job of like understanding a user is just to pick one of those.
[00:20:41] The idea is, okay, accept that as a fact and then try and build like a very clever dialogue manager that can, you know, handle the fact that actually meant, you know, that's not how conversations really work. People don't think of a category and then come up with a way to express it. They just, you know, say what they want and, um.
[00:21:00] I remember, you know, working on that problem for many years and, and, you know, we've published some papers on it and all that kind of stuff. How do you build like a more clever dialogue manager and kind of coming to the realization over time that actually what people wanted was a very simple dialogue manager and like a, like an oracle, like a perfect like understanding component.
[00:21:20] Like it just make sense of any of the complexity of anything that like people are asking for. And that's obviously a a lot closer to what we have now. Mm-hmm. But. When you're working on a project and you are trying to iterate and improve how something works over time, um, you know, in this sort of classification paradigm, you're building a supervised model that is classifying what users are saying into some buckets.
[00:21:44] And the, the thing you have at your disposal to improve that is to like improve the training data that's fed into that classifier. And so you are looking at conversations that people are having and you're saying, okay, what went wrong here? Could I, could I fix this problem by improving my classifier? And you realize that actually the, the fraction of conversations that didn't go well, that you could fix by having like a smarter classifier, it's actually very, very small.
[00:22:08] And what you see when you're looking at it is that there's just this, often, this mismatch between the way that you organize the world as a large company, building a support agent, and the way that your customers think about the world. We used to think of this as sort of disambiguation, right? You know, a person comes in and says, you know, you messed up my order.
[00:22:28] And you go, okay, well you've gotta do some disambiguation, right? Like, we don't know what is up with the order and how we're gonna help, but it's not like, it's not ambiguous what the person said. It's perfectly clear. It's just in their reality and it doesn't obey your, you know, your divisions and all this stuff.
[00:22:42] So I feel as though that idea of like, intense and, and, and forcing everything into a classification. Was just that very concrete manifestation of that mismatch and it just became very obvious very quickly that so long as we live with that assumption that we always have to classify messages into these predefined buckets, we're never gonna make meaningful progress because that's really the sort of the bottleneck, right?
[00:23:08] And you try and fix something. Because the, the thing that happens after that classification is all this, all these if statements, right? Then if you introduce a new category and they're like, okay, well what else is gonna go in that category? And what if people say that in the middle of the conversation, what they say that like, then, and then, you know, you have this, uh, you have this plate of spaghetti that, that, uh, represents your business logic on the back end of that classification.
[00:23:31] And yeah, uh, that just felt so brittle and so obviously the wrong way to do it, and just such a mismatch for how. Language understanding works. I mean, I used to joke, I said it's fine that we call it natural language understanding. So long as you remember that it's a lie and the computer doesn't understand anything, but like it's just assigning things that users said to some pre-define buckets.
[00:23:53] Yeah. I think what and assign today is a lot closer to to understanding True. I, I think that that always was hard for me to explain to people that were new and conversationally on projects was like, what is the confidence ranking? And, and, you know, oh, 80%, that must be really good. And I, I sit there and think, I, I guess it depends, you know?
[00:24:12] Yeah. Because it, it was so arbitrary between different vendors, you know, Roz as well. You, it was all relative to whatever you were doing in your little bubble, which I think is, is nice today. It's not. And, and, and, and that's a, that's a nice shift that we have now where. Okay. We used to always work in the supervised learning paradigm where you would take some data, you would annotate it, and then you would train a model to like do that classification or typically classification.
[00:24:37] And you would have these sort of confidence values and everything, and you would get all these sort of misconceptions. People would think that if now I, you know, retrain my model and it's now reporting higher confidence values, that means I have a better model, right? Like all, all sorts of issues with how people actually.
[00:24:54] Operationally worked with supervised learning, but the, the way we work now with LLMs where you just type in English, like what it is you would like to have happen, is actually a lot closer to, I think, the way that people wanna work with ML and with ai. Because what we saw was that people would just, they would have those heuristics in their head of what they wanted to happen and they would kind of launder it through some data generation process and then force the, force the model to learn what it was that they think the model should learn.
[00:25:26] So in some ways it's more honest, you know, the, the sort of prompt and pray thing that we do now is that. It's a bit more honest of like what we're actually trying to achieve. I think maybe I'll give my, like, one of the things that I sometimes am frustrated with in this space today is that everyone assumes that everything should be a con conversational interface.
[00:25:42] Mm. Like just everything. Like your tv, your fridge. Yeah. Your water, you know, kettle like, and then sometimes I struggle with this idea of like, okay, what makes a good application? What makes a good conversational interface? That's my soapbox. One day, maybe I'll get off of it, but nonetheless, I like it. I'd love to hear from, from your side, like what are some of the things, like the new sins of this era that you're seeing pop up?
[00:26:05] Uh, honestly, I think feeding the output of an LLM back into an LLM or back into another LLM is, is a mortal sin. Okay. And that's, that's a slight exaggeration, but it's not that much of an exaggeration. I think, you know, the idea that we're just gonna call. An LLM multiple times in series and just feed its output back into itself.
[00:26:29] Um, I mean, there are, there are times when it's appropriate and there are times when it works, but it's just such a bananas way to engineer a system, right? Um, you have all this un unpredictability that on the, out the output side of an LM, and now you're. Like multiplying it, right? You're plugging an amplifier into another amplifier, into another amplifier, and you have this like horrible feedback loop.
[00:26:53] But now, so at some point some noise gets introduced and you're like, oh, you know what, what, which, which noise caused which thing to sort of, you know, diverge in some crazy direction. So, um, yeah, I think that there, there's a time and a place for it, but generally speaking, I think it's something to be avoided at all costs unless there's really a good reason why, why you should be doing that.
[00:27:14] Um, so you. Can you, can you do things in one shot? I mean, the other is, and if you go on sort of, you know, Reddit and things like that, you check out sort of developer forums, um, and you look at like people asking questions like, how do I do X? How do I do I, how do I do z Things that are. You know, people are really scratching their heads, how do I solve this problem that are o they're only a difficult problem because you're starting with the assumption that everything goes through an LM and every, you know, and you can only work in this paradigm.
[00:27:41] This is the only way you, you can do things, right. And you're like, that's actually a very simple and a very well solved problem. Um, but you know, you're just, you're just coming at it in a silly way, sort of narrow things down a little bit. So, um, there, there are plenty of sins, um, but I think that would probably be.
[00:28:00] Be the main one. Well, I guess we've, we've had, we've covered maybe so far cardinal sins, so, uh, you know, continually using an LLM for things that shouldn't be used as brains on the website. Absolutely no go. Um, you know, but what are, what are some of the things you're kind of more excited about or more hopeful?
[00:28:20] So, I mean, I think that AI assisted coding is. I mean, this is a cold take. I mean, everybody thinks this, but it's just, it's just amazing. It's just truly, truly amazing. The, the thing in terms of someone who builds products, the thing that it really unlocks that I never thought was possible is that I can interact with Cloud Code and be building a project and describe what I want at any level of abstraction that I like, and that makes sense to me at any given point in time.
[00:28:54] I sometimes talk about Raza as a 10 year quest to find the right abstractions, right? Like what are the right Lego breaks? What are the ways, the things you really need to compose together to build good conversationally ai? And that's, you know, that's always something we're talking about internally, right?
[00:29:08] Is this the right abstraction? Can we move it up? Can we move it down? Can we reuse this? Can we merge these things? And. The, you know, the original sort of, um, coining of the term vibe coding was around sort of just talking at a high level, like looking at a UI that's being built for you and just describing the change you want, not even looking at the code that's being generated and working at that level of abstraction and all the way down to, you know, being very, very specific about the specific change that you wanna make.
[00:29:34] And so there's all the productivity gains and all this kind of stuff from Yeah. From AI assisted coding, but I think that ability to work on different levels of abstraction as someone who builds products to me is just kind of mind blowing. Um, so that's something I'm very bullish on and, and very excited about.
[00:29:55] I think for me too, in that direction, one of the things I'm excited about is. The element of validation in that, that you get when you're working with something new. Because I think of my own journey, you know, learning some code here and there. Yeah. One of the biggest issues you have always is that. You can't scan a project and be like, oh, that looks wrong.
[00:30:14] 'cause you don't have the knowledge or the skill yet. Right. But to have something actually tell you, oh, you messed up here, and go through that learning cycle, I think that's super useful. And expedites the process. Maybe for the worst. 'cause I didn't have to, I'm not sure yet. 'cause I didn't go through that process of like, you know, figuring out where I messed up.
[00:30:30] I just had someone serve it to me on a bladder. Yeah. Someone a thing. So I think, yeah. And anthropomorphizing is that, is that one of the deadly sins as well? That's a deadly sin up there with brains. Don't anthropomorphize machines. Yeah. Put that up there. Um, no, that, that, that makes, that makes loads of sense.
[00:30:49] So in general, just kind of vibe putting, being able to iterate, build quickly. Um, yeah. Yeah. Um, I mean, uh. I heard it described as like, I think it was Sam Alman, he said it's like the fast fashion era of software, which, you know, you have ethical qualms about fast fashion, which are totally valid. But, uh, from a software perspective, you know, sort of on demand, you know, software, um, and being able to interact with things just in a way that makes sense to you, to your point, whether you're, maybe you're knowledgeable about the technology, but you're totally new to a project.
[00:31:25] And you're just like onboarding to a new code base or a new project or something like that, and you, you can just be effective right away. Yeah. Although you were mentioning about abstractions, and I think maybe this is one of my, when I think about the future of conversational interfaces, one of my concerns is actually now the exponential.
[00:31:43] Let's say sprawl of extractions. I feel like when I think about all of the different ways that you can manage dialogue today, we actually, we have a session coming up here at RA and we introduced some new things and I'm like, goodness. I mean, even someone who's deep in this field, it's hard to look at all of your options and understand what will make sense.
[00:32:01] Yeah, no, you're you, you're right. And the, I would say 2018 through 2023. We're five years that I was very deeply frustrated at the lack of appetite for new ideas, conversational ai, especially in enterprises. And, you know, we were sort of pushing a lot of stuff and, and coming up with new ideas and different approaches and things, and everyone was pretty dead set on, well, no, this is how it works, right?
[00:32:27] Like we have our. And then we have a bunch of if statements, and that's just, that's just how things go. And, you know, sort of be careful what you wish for, because now there's just such an explosion of, of new ideas and we're really in an exploration phase. Right. And you see that with, I mean, every team that we, you know, that we talk to, that you and I talk to is.
[00:32:48] Working with the assumption that we're building something now, but it needs to be flexible enough that, you know, because we're assuming that in 12 months time we're gonna want to change how this thing works. 'cause some new underlying, uh, capability has been unlocked. Right? Or just the models have gotten smart or cheaper or whatever, or something else has changed about it.
[00:33:06] So. I think it's still very much a moving target and everyone's trying to, of course, plant a flag and say like, this is the thing and this is the right way to do it. But, um, it's, uh, yeah, it's moving, it's moving too quickly in too many different directions, I think. Yeah. And I guess, um, when you, when you think about like new skills or mindsets that people need to adopt, and I'm, I'm always like trying to.
[00:33:32] I, I think especially when you live through the past of conversation, it's hard to think beyond because you're so used to Yeah. What works skeptical about does what doesn't work. Yeah. But what, what are those skills and mindsets that people need to switch? Yeah. I, I would say currently I would say there are kind of, um, they're just two completely different worlds.
[00:33:53] There are people who. Worked on conversation I had before chat, GPT, and there's one who, who came to it afterwards and they just have very, very different understandings of the universe and, and just think about a very different set of things and ask different questions and use different vocabulary and just have different starting assumptions.
[00:34:11] And something I've been feeling in the last, um, three, four months or so, is that.
[00:34:24] It is not that hallucinations have been solved, but people got bored of talking about them. So it's less of a topic now, just the attention span of, of the hype cycle to focus on something, right? So it's, it's definitely some problem. It's definitely still a thing, but you just hear less about it because it's like, well, maybe it's just the cost of doing business or maybe it's just, but either way it's just not, you know, people kind of got bored of it and moved on to talk about different topics.
[00:34:55] So, um, you kind of have things moving at very different. Cadences, the, you have the hype cycle and then you have sort of consolidation of technology abstractions and things that are actually working and are actually being used. Right. And, um, yeah. Often the, the thing that's most talked about is still very, very premature and still very, um, uh, very experimental.
[00:35:19] Um, but I will say it's very exciting to have so many more brains thinking about this problem now than we did five, six years ago. Right. That's. That's pretty exciting. And like, all this stuff works so much better. Like, like if you had, if you had something like what we can build today in, in a few minutes time, and you could show that five years ago, uh, you know, it would've absolutely blown people's minds, right?
[00:35:42] So everything that, that people were promising in the past is, is, is hypothetically possible. The other thing I think about is, um, so many things change, but what, what doesn't change? Mm. And that there, I'm just, you know, shamelessly ripping off Jeff Bezos and that. But I think it's a really important question to ask yourself, and it's helpful for us at Rasa, which is, well people, you know, need infrastructure that evolves with them.
[00:36:09] And, you know, we've rebuilt rasa completely many times over the, the, the lifetime of the company and sort of reinvented the engine quite a few times. And, um. Some abstractions stay and stick around and some you get to deprecate and gets in, you know, invent new ones and, and build new things. Um, the teams that we support are.
[00:36:30] Heterogeneous and complex and build for a large, complex set of use cases and millions of end customers, whether that's on the phone or via chat or whatever it might be. So, and you can, you know, derive some, some constraints from that, just from the organization that's trying to deliver something. Right.
[00:36:47] So, um, while there is lots of change of like what's trendy, I think a lot of the core principles are actually remarkably solid. That's so true. I think even as you were saying earlier, like the, the promise finally fulfilled, I realized recently that I changed how I introduce what I do at like a barbecue.
[00:37:09] If someone were to ask me, you know, in the past I almost started with an apology. Like, ah, yeah, you know, I work at a company, we build those little things at the bottom of the screen that are annoying and pop up and you know, but they're getting better. We're working on it. We're trying to make it better.
[00:37:21] And now, and now it's like pH. Ai, like, you know, you just throw it in there. Well, okay, so, so we, we have a, an onboarding meeting that I run for anyone who joins Rasa. And we do a little round of introductions and we ask a couple questions about yourself. And one of the questions is, what's your favorite AI agent?
[00:37:41] Mm. And. For most of the age of the company, people really struggled to come up with something that they had used and they had enjoyed, and that they, and now of course, everybody has, you know, five or six that they've encountered recently that were quite good, that they're using personally for their own, uh, their own experience.
[00:37:57] Right. So, um, it's, uh, yeah, it's a, it's obviously a very remarkable shift. Really cool space to be in. Um, I think as, as we wrap up today, um, two final questions. Um, if you could give one piece of advice to kind of an enterprise that is, let's say, maybe just starting their journey today, what would it be?
[00:38:24] I mean, it's a good question. Are there any enterprises or just starting their journey today? I feel like yeah, definitely. They're, they're starting like, I guess maybe this is it. Here, I'll, I'll give you mine first. I would say like, don't think that rag alone is your only option. That would be my advice because I think when, when people like, and I'm even like small, large companies, when they start dabbling mm-hmm.
[00:38:43] I, I'm seeing like so many just q and a bots, and I guess I wanna say like, you can do so much more, you know, you can actually solve problems in chat. It doesn't have to be, and I know that's like kind of the first foot in the door, so it makes sense to maybe do first, but that would be my advice is like think beyond.
[00:38:58] Think beyond that. Yeah. Yeah. Fair point. What would you give to them? To someone who has been, who's been, let's say, a, you know, tried and tested battle scarred. Conversationally enterprise, which, you know, what's your advice today then? Um, I mean, I think the,
[00:39:17] the, the thrust of it's got to be to stay focused on actual user problems that you're trying to solve and not checking boxes that you integrated, like the latest three letter acronym or, you know, the latest model or whatever else it might be. Um, because there's a lot of fluff. And a lot of hype and a lot of things that aren't real.
[00:39:39] And the only thing that you'll really be judged on at the end of the day is how happy are you making customers and you know, whatever technology you're using, don't sleep on the value of just looking at some of the conversations that real end users are having with your system and developing your own intuition and judgment about.
[00:40:00] How good of an experience that is, where it could be better, what the opportunities are, what else, you know, could be because yeah, instant gratification. It's easier than ever to build a sort of semi vaporware kind of very cool demo. But it, you know, it's still very difficult to, to really move a business number.
[00:40:22] Right? Move a metric that people care about. I think that rounds out top three cardinal sins at Raza. Brains on the website, anthropomorphizing machines, and three letter acronyms standing for vaporware. That could be like the top three. Alright, well thank you so much Alan. Um, finally, how can listener, listeners really stay connected with your work and kind of keep in touch with Raza, um, give you feedback.
[00:40:51] If they ever give our, our, our, uh, product a try, I can definitely recommend anyone who's listening in, go and try out Hello Raza. We can maybe link it in the description and uh, give it a go. Learn how a Raza works. No, I mean, you beat me to it. But, um, we have a nice new experience for getting started with Rasa and that's feedback we've had consistently over many years is that rasa is very powerful, but it's maybe a little intimidating if you haven't used it before.
[00:41:18] And we built a very nice experience that doesn't require any upfront rasa knowledge that lets you build something, uh, very cool in a very short amount of time and immediately see how Raza agents are very explainable and understandable and composable. And help you build that confidence in how the system works and, and what you've built and how you can scale from there.
[00:41:41] Amazing. Thanks so much Alan. Tell the next time. Alright. Thanks for having me online.