How A PhD in Machine Learning Revolutionized Codebases with Scott Dietzen, CEO at Augment Code

Today, we're talking to Scott Dietzen, CEO at Augment Code, about the future of AI and coding. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:14) I was reading your background. You've been running tech companies for thirty-plus years, and you've been the CEO at Augment since 2023. How did you end up there?

(Scott Dietzen at 00:00:25) So I actually did a PhD in machine learning a long time ago, different era. It was back when we were pursuing symbolic reasoning, but I was fascinated with AI, and I always thought we would make more progress. You know, it turned out to be a lot harder than I think many of us thought back then. But, you know, I decided I wanted to be an entrepreneur, and I spent that career in system software, right? So building really large, complicated codebases for things like serving applications on the internet, ultimately storage. Those businesses were successful, but the software engineering required to build a great product was really complicated and painful. You know, very hard work, sweat equity from great engineers. And so with the emergence of large language models, I saw the chance to tie these two threads of my career together and actually help build a solution that would target software engineering, not just simple programming, and remove a bunch of the toil and frustration associated with looking after these tens of millions of line codebases.

(Joel Beasley at 00:01:36) And did you see this opportunity before Cursor came out, or did you see Cursor come out and it lacks stuff and you're like, "Hey, we need to create something better"?

(Scott Dietzen at 00:01:44) So we got started two and a half years ago. Copilot was just emerging in the market. And, you know, Copilot was interesting, but didn't have any knowledge of your software, your codebase. And, you know, that's what matters in software engineering, right? An AI that is a novice is not nearly so helpful as an AI that's an expert in your codebase, like a co-programmer that knows as much as your best developer so that can be the aggregate, the synthesis of all of the knowledge that's gone into your product, is so much better. So, you know, we did get to watch the launch of Cursor. I do think, you know, they took a step forward from where Copilot was, but they're still lacking in that ability to tackle complex repositories, tens of billions of lines of code, codebases. And, you know, sustaining the work required to build higher software quality in those situations is something that's unique to Augment today.

(Joel Beasley at 00:02:46) Yeah, I thought it was interesting. I watched the YouTube video one of your engineers had put showing the comparisons of Cursor versus Augment. And at first, I was like, "Okay, let's see what's going to go," because my background is software engineering for seventeen years. And so I'm like, "Let's see how this goes." Surprisingly refreshing what Augment was able to do with understanding the codebase in full context all the time. I thought that was brilliant and the results it was giving, because I use Cursor. I just found out about Augment recently. And it just gets you where you need to go faster, especially with the next steps follow-up. And I think I forget what you called the feature, but the ability to, like, I'm going to make this change in my test file and then now I need to make this change in the model or I need to make this change over here. It was really interesting how it kind of—we're getting so close to just doing it for you.

(Scott Dietzen at 00:03:38) Yeah, we call the feature NextEdit.

(Joel Beasley at 00:03:41) NextEdit.

(Scott Dietzen at 00:03:42) It's an agent, although it's supervised, right, in the sense that the developer is still in the loop. But there aren't really simple changes in 10 million line repositories, right? There's just so much work that has to go into qualifying a change and making sure that you've got all the details right. And if you, you know, the simple example we often use is just adding a field to a data structure. But then if you are storing that in a database or you're passing it around through APIs or command line interfaces, you know, there's just all of these other touch points, all of the tests in the software, the documentation. And so having an AI that understands, you know, the codebase well enough to anticipate all of the work you need to do to pull that change through is just a great time saver. You know, so often we get interrupted in the middle of a complex workflow, and then you're trying to remember, "Hey, where was I? What part did I do? What part is remaining to be done?" And the fact that the AI has the context and can just pick you up and take you through the rest of it is just a great relief.

(Joel Beasley at 00:04:52) That's awesome. And I'm always looking for the edge of what technology can do. Can it, will it tell you when you're ready to actually run the test? Like, you know, you've made all the edits and all the different files. Now you're ready to see if it actually is passing?

(Scott Dietzen at 00:05:05) Yes. In fact, you know, what we'll see this year is AIs doing a lot more work where they actually write code and then run tests, and you can even iterate without humans in the loop in order to get, you know, say, test coverage right, like look at the tests and make sure that everything is working. Or if a bug is turned up, the AIs will be able to fix it. I have this dream, you know, instead of getting woken up in the middle of the night that there's a software bug, the AI has woken up ahead of me and looked at, you know, the root cause analysis of what was just reported and already has an idea of what can happen. Because so often, you know, you're asked to weigh into parts of the codebase that you're less familiar with and you don't have the full context. And so, you know, the fact that the AI can bring that for you and give you that kind of orientation really immediately into an issue that you're very unfamiliar with is very empowering.

(Joel Beasley at 00:06:10) Yeah, I'm starting to see—it's so interesting how all this is going to play out over the next five years. But I'm starting to see marketing, because, you know, my title on LinkedIn is CTO, right? So I get all the same marketing all the other CTOs get. And I started to get marketing in the new year about, "Hey, how great would it be to," or "Imagine having a full-time software engineer that's working around the clock that's just an AI model." And so they're starting to sell that, you know? And I don't know how close we are to that. Have you seen that actually happening in the wild?

(Scott Dietzen at 00:06:46) No. And I don't think the AIs are close to being able to be full-time software engineers. You know, the test that people usually talk about are things like SWE-bench, where, you know, you're asked to implement a pull request. And, you know, there are definitely—Augment included—AIs that are able to take those incremental steps. The challenge with software engineering, though, is there's so much more to it, right? From requirements gathering to, you know, building out an architecture, choosing database, choosing a microservices model, what cloud platforms, you know, how do you want to evolve features. You know, that kind of long-term view of the software is still very much the domain of human software engineers. And so what we can get out of these AIs is doing a bunch of the drudgery associated with the incremental steps to improve the software, but we still very much need human intelligence to get us to the promised land, to, you know, to shape the goals that we're aspiring to for that software.

(Joel Beasley at 00:07:55) And what are you seeing? I'm sorry, I'm a super curious guy. What are you seeing in the marketplace, right? Like, as far as, are people very interested in this? Are engineering firms? Are you trying to convince them to even use any assistive code editing tools? Like, where is the market at today in your opinion?

(Scott Dietzen at 00:08:17) So it's still very early days. You know, the vast majority of engineers are not yet using AIs, but I think, you know, the early adopters are very much in. If you look across our hundreds of companies in our customer base, we were not the first AI into any of them. So most often, we, you know, we find a Copilot, sometimes now a Cursor. But oftentimes, the most senior developers lose faith with the product and have decided not to use it because they want an assistant that actually understands the software as well as they do. And so we've been able to reignite their love of AI in seeing how much more productive they can be with an AI that is their peer rather than a novice in the codebase. So, you know, I think that's how the adoption is going. I mean, engineers are excited and curious to see what these AIs are capable of. And, you know, there's so much busywork associated with building these large, complex codebases, right? When you add new capabilities in, just telling the AI, "Hey, crank out a unit test for me. Update the documentation." Or better yet, you know, when you make a change, the AI prompts you and says, "Hey, should I refresh the documentation?" Or we've actually had cases where the AI has proposed deleting code rather than adding code. And to me, that's the definition of an expert AI. Expert AI wants to remove code, not just keep adding ever more code into your codebase.

(Joel Beasley at 00:10:00) And the video is so cool. Josh can put the video, the YouTube video, in the show notes that people can take a look at it. But my next question was, I was like, "All right, this is cool." And I get excited about technology, so I want to tell people about it. So I was like, "Who could I call?" I was like, "All right, what would happen if I called Ed from Mastercard or Jeremy from Pinterest and was like, 'Hey, you guys got to check this out.' What questions would they have for me?" And I'm pretty sure the first thing they would want to know is, like, security, right? Like, where is this model living? Who controls it? How is it being trained across—can I self-host the model that's actually doing this intelligence and do it in a secure way? How do I keep my data my data?

(Scott Dietzen at 00:10:45) Yeah, it's a very frequent, common ask in the discussions around security with these models. You know, people should keep in mind that not all of the models guarantee hygiene around your intellectual property. So, you know, if you're using a wrapper on top of one of the frontier AI models or using those AI models directly, I think you want to make very sure that the models aren't allowed to fine-tune or keep your data in any way, because you could potentially leak intellectual property. So this is something we were really careful about at the outset. So we do host the models that we use in the cloud. It is a blend of open source and frontier models that we use. But we strictly honor, unless our customer is open source and gives us special permission to do so, that we will never train or look at customers' software, so that there's—and this is the magic of how we implement, you know, this deep knowledge of customer code, but we do it without ever training on customer software. So, you know, the results and the suggestions you get are informed by your software, but there's no way you can leak intellectual property. We are a cloud service, so these models are hosted in the cloud, and we do bring code snippets up to the cloud, and that's how we inform our models for understanding your codebase. But it's done in such a way that at any point, you can revoke access to that software. And we were the first AI to get SOC 2 Type 2 certification to ensure that we are adequately protecting our customers' intellectual property.

(Joel Beasley at 00:12:36) Do you think in the future, it'll get to the point where I could just, inside of the Augment app, just select the source model that's being used and have it be a model that's self-hosted?

(Scott Dietzen at 00:12:46) Yeah. So I think we're going to stay focused on being a cloud service. You know, the challenge with on-premises deployments—I mean, we make dozens of changes to our service per week, probably sometimes per day. We don't relish the packaging of software and then shipping it to customer data centers where we have inconsistent GPU footprints. You know, it's very hard to honor an SLA. I think it's a lot more feasible to do a VPC kind of deployment where we're co-resident with other services that the customer has in the cloud. You know, one of the things we point out is if you're using ChatGPT with code or Claude today directly, you're already sending code samples to the cloud, and, you know, we can do so in a way that's really secure. Or if a business is hosted at GitHub or GitLab, you know, they've already got their software in the cloud. So an AI cloud-based service, provided it's one that has very strong security, doesn't represent additional risk. In fact, we can make the case that it's actually more secure in many cases than customers' own data centers.

(Joel Beasley at 00:13:59) Interesting. Especially for the, like, government market too, right? The people that are building missile systems and all of that. Like, they need their stuff segmented, and I'm curious who's going to approach that side of the market. But I think we're just so early right now. We're just trying to get even the agencies that are just building consumer apps to get on board with this newer technology. Is that where we're at?

(Scott Dietzen at 00:14:21) Yeah, I think so. One thing I would add for, you know, the government customers, because they operate at scale, they often bring cloud footprint into their own data centers. And that is also a deployment model. It's very similar to the VPC model I mentioned earlier that you can deploy into someone's data center as long as you have that consistent cloud software stack for management and update. Smaller companies, I think that would be more of a challenge. And there are vendors that are specifically targeting on-prem. I just think there's so much easier ability to evolve and improve a cloud service quickly that it just doesn't make sense to tie ourselves to on-premises deployments.

(Joel Beasley at 00:15:05) Yeah. Well, you know, I think we've reached my level of understanding, my boundary of these deployments and stuff. I don't know enough. The only thing I know is that if I was working on a financial app or a banking app or something, I would want to either go on Amazon and deploy the model and just grab the endpoint and plug that into whatever IDE I'm using so that it can use that model. That way I know that everything's contained in buckets that I control.

(Scott Dietzen at 00:15:39) Yep. So you can do that today, of course, with open source models. There's a lot of excitement around DeepSeek, for example, and Llama. We've done extensive testing of those and other open source models. And there are five or six different open source models typically in use under Augment when you, you know, come into the service with a bunch of our post-training on top of them to make them better at what they do. We do also make use of frontier models, but they're hosted in the same cloud and under the same security model and security architecture. So I think we can deliver exactly the sort of security that you're talking about without you having to put all of the pieces of a service together yourself.

(Joel Beasley at 00:16:25) What's a frontier model?

(Scott Dietzen at 00:16:27) Sorry. By frontier, I just mean large closed source models like Claude or OpenAI's models.

(Joel Beasley at 00:16:34) Okay. I'm learning, man. I'm taking notes, Scott.

(Scott Dietzen at 00:16:37) Yeah. Technically, I think you could argue the open source models are frontier as well because they're pushing the envelope in what the models are capable of. And that's been a really exciting development, you know, just over the past couple of months to see the gap closing between open source and closed source models in terms of their capability. And, you know, remarkably good code models have come out over the past year. It's allowed, you know, ourselves as well as our competitors to all make really rapid progress in delivering great AIs for code.

(Joel Beasley at 00:17:12) Yeah. That's the engineer in me that, like, wants control, I want it. Yeah. That's my human fault. As far as your vision for this.

(Joel Beasley at 00:17:22) Right? Like, for the future for the next year or two, I won't go out five years because that's gonna be like chaos. You might as well have Elon Musk Neuralink chips in our brains. But for the next, like, year or two, what's your vision for where Augment's going?

(Scott Dietzen at 00:17:37) So we're building up ever higher levels of abstraction in terms of what the AI is capable of. So, you know, two different usage models are dominant today with Augment. One is, you know, we're looking and watching you type, and so we can help complete things. We can help suggest reuse of code. And we mentioned the next edit paradigm. But a bunch of our customers move to more of what I would call meta programming, where they talk in chat to how they want the program to behave or how they want to change its behavior.

(Scott Dietzen at 00:18:12) Let the AI make the change, and then they look at it and, you know, see the results and then may update that. And the nice thing about a product like Augment is you can mix and match. Right? You can tell the AI to generate code, but then you can jump into it and tweak it and modify it to your heart's content and then step back out. What we'll continue to do is just continue to raise the bar.

(Scott Dietzen at 00:18:36) Right? So where, you know, we can get to the point, I think, easily in the next two years where you can give the AI significant tasks. I mentioned earlier, you know, crank out a unit test. But I think you could say, hey, I want to move this application to a microservices architecture. And, you know, these are the, this is the rough breakdown, and then the AI can just go off and do that work.

(Scott Dietzen at 00:18:56) Or let's say I want to port this application that's running at AWS. I want to have it also running at GCP. I want you to migrate it and do the, you know, the database work to move from, say, Redshift to BigQuery, and the AI will be able to handle that. We've already got customers that are doing things like moving from C++ or Java to Rust. And the AI is not just translating the code, but it's actually mapping it to the new language's style, so that, you know, you get a natural Rust program, the same sort of thing that someone would design if they were writing it from scratch in Rust.

(Scott Dietzen at 00:19:32) They come up with the same thing. You know, migrations and integration are a huge pain point. One of our customers, Codem, they've been able to cut migration times in half with Augment today. And I think, you know, in the future, they'll be able to cut it by two thirds and three quarters. And then, you know, ultimately, the AIs will be able to do more of that busy work, freeing up human software engineers to think aspirationally about what their software dreams are.

(Joel Beasley at 00:20:02) Who is this Codem?

(Scott Dietzen at 00:20:05) It's one of our customers. So they do, they have an ecommerce business, and they support a bunch of customers on different ecommerce applications. And they're often needing to upgrade them from legacy architectures to something more modern. And they, you know, they come in, and Augment has accelerated their ability to deliver those solutions in half the time that they used to take. And I think it's very generalizable.

(Scott Dietzen at 00:20:32) You know, if you're an ISV that has a database business or, you know, anything that has an API where customers have to move their applications into your environment, we can cut that workload by order 50% today.

(Joel Beasley at 00:20:50) That's a huge win, Scott. You've got, they took, they cut their work in half. That's dollars, dude.

(Scott Dietzen at 00:20:57) We have a Fortune 500 customer I can't name, but they did a benchmark versus Copilot. They were a Copilot shop. They've moved over to Augment, and they cut their workload by 40%. So PRs were done 40% more quickly, and they had to be revisited 40% less. So, you know, just clear ROI of 40% relative to Copilot.

(Scott Dietzen at 00:21:24) Now this is a, you know, a big complex code base, which is in our sweet spot, you know, esoteric C++ and Rust code that, you know, required an AI that was able to dig in and understand the architecture, you know, of a tens of millions of line repository.

(Joel Beasley at 00:21:44) That's huge. Those are huge savings. So you must feel really good at work then. You must be really excited about what you're doing.

(Scott Dietzen at 00:21:51) Absolutely. Tremendously excited, you know, having felt this pain of software engineering, looking after teams of really talented people and, you know, the grind that was required to make commercial grade software is something we really hope to alleviate. You know, programming is fun. I used to love it. In college, I spent so much time programming, and I was really daunted at how much harder it was to produce commercial software.

(Scott Dietzen at 00:22:21) And almost every piece of software you encounter disappoints. Right? There's a long wish list of features. There's a long wish list of tech debt that people would like to eliminate. Imagine if you could burn all that tech debt down and deliver all those missing features.

(Scott Dietzen at 00:22:38) How much wealth could we create? How much human productivity could we unleash if we had, if all of the software met our aspirations? Just in the U.S. market, $2.5 trillion lost last year to software failures. What if we could eliminate that? You know, software didn't crash and didn't break because it was, you know, engineered by the combination of human and machine intelligence working together.

(Joel Beasley at 00:23:05) Yeah. I did an interview a couple, like, two years ago with these guys from Gremlin. They were on the Amazon team, and then they spun out their own product for reliability. And they were sharing with me, like, how much it costs per second when, like, services at Amazon's website are down. I was like, it was mind boggling.

(Joel Beasley at 00:23:24) You're talking, like, millions, like, tens of millions of dollars per second. And I was like, that's crazy.

(Scott Dietzen at 00:23:30) Well, I think that's just the cost to the Amazon business. But now imagine the cost of their customers, right, that are having outages. They're not able to serve their customers. And, you know, it's almost all software. Right? It's, you know, the hardware failures we've done a much better job of isolating and repairing from, you know, especially in cloud deployments.

(Scott Dietzen at 00:23:42) You mask those very easily. But the software failures, we don't have an answer for yet. And I think AI gets us there.

(Joel Beasley at 00:24:00) Where do you think the biggest gap is today in these types of tooling? Like, the class of tooling that you're in.

(Scott Dietzen at 00:24:06) Well, I mean, if you look across our competitors, they really fall short in being able to take on large code bases. And so, you know, as a result, they're fine tools for programming for weekend projects. You know, if you're starting from scratch to build a small application, you won't notice these limitations until, you know, a large team tries to look after tens of millions of line legacy code base. Right? All software aspires to be legacy.

(Scott Dietzen at 00:24:35) That's the best outcome. You know, in these production business critical systems, there's so much knowledge that's been put into them over so many years. Having an AI that can bring that knowledge to bear in helping individual programmers, I, you know, I think really moves the needle. One other thing I would hit, you know, some of the solutions in the space have taken to forking the IDEs, you know, producing their own proprietary version of VS Code, for example. We believe that's a mistake.

(Scott Dietzen at 00:25:10) So, you know, the challenge anytime you fork an open source project, you end up with two different code bases. Right? And, you know, in the case of VS Code, Microsoft has a long roadmap of future enhancements that I think customers are interested in. And Microsoft has a whole ecosystem of plugins and capabilities that work with VS Code that don't work once you go onto a fork. So we've been really careful to, you know, to work within the constraints of how VS Code is designed so that our customers get full access to all of the Microsoft ecosystem capabilities, which, you know, is not necessarily the case if you choose Cursor or choose Windsurf, where they've elected to diverge away from the roadmap for VS Code.

(Joel Beasley at 00:25:58) Oh, so you're sticking with it so you get those updates competitively?

(Scott Dietzen at 00:26:02) Yes. It's a bit harder, the engineering work that we need to do to work within the constraints, but we think that value to customers that they don't get walled off from future enhancements to VS Code and they don't get walled off from Microsoft ecosystem capabilities really matters, you know, for these, especially in the large complex code bases. And then, of course, I should highlight, you know, we don't want to tell people they have to switch their editor away from VS Code, but also JetBrains and Vim and so on. Right?

(Scott Dietzen at 00:26:36) It's, the AI should come meet you wherever it is that you're working. You shouldn't have to switch your IDE out and learn a new environment in order to take advantage of AI. The same way we want to integrate in with anywhere else that developers work. So we have a Slack plugin, for example. And so if you ask a question on our Slack about how something works, an engineer can just invite Augment into the thread, and Augment can render a verdict, assign a PR, propose the fix.

(Scott Dietzen at 00:27:09) And so, you know, you have an AI collaborator that you can plug into whatever other tools your engineers are using, not just the IDE.

(Joel Beasley at 00:27:18) That is so cool. So will your technology be available in Vim and JetBrains and others in the future?

(Scott Dietzen at 00:27:24) No. It's available right now in JetBrains and Vim today.

(Joel Beasley at 00:27:29) Oh, is it really?

(Scott Dietzen at 00:27:29) Yes. So across the, across the full JetBrains family, and we have a plugin now for Neovim that brings all of the AI capabilities into VI.

(Joel Beasley at 00:27:42) Not even, I've misunderstood then because when I saw it, I saw, like, the VS Code customized. I thought, oh, this is just like Cursor. This, they just forked some. But no. Your technology is going to go meet everybody in their current IDEs.

(Scott Dietzen at 00:27:58) Exactly.

(Joel Beasley at 00:28:00) Oh, that is so cool. Yeah.

(Scott Dietzen at 00:28:01) Yeah. I mean, it's especially

(Joel Beasley at 00:28:03) I mean

(Scott Dietzen at 00:28:04) It took thirty four minutes

(Joel Beasley at 00:28:05) for me to figure that out. So

(Scott Dietzen at 00:28:07) Well, you know, there's so many vehemently dedicated JetBrains users. Right? Because, you know, the suite of their products are very popular for a good reason. No engineer should be forced to change that to get access to state of the art AI. So we want to be state of the art, but, you know, compatible with all of the existing tools in the developer ecosystem.

(Joel Beasley at 00:28:32) That is so cool. That is so cool. Now you said something to me that you kind of glossed over it, but I found it to be true and profound in an articulation of something I've known to be true, but I've never actually spoken. And I haven't heard it in a thousand interviews, like, without hyperbole. You said all software aspires to be legacy.

(Joel Beasley at 00:28:55) Where did you get that from? Is that you?

(Scott Dietzen at 00:28:58) That's a good question.

(Joel Beasley at 00:28:59) Is that Martin Fowler or something?

(Scott Dietzen at 00:29:00) I do think it's mine.

(Joel Beasley at 00:29:04) Okay. We'll give it to you. I haven't heard it. I've been doing this ten years, thousand episodes. I haven't heard that.

(Scott Dietzen at 00:29:10) Well, I mean, it's either legacy or obsolescence. Right? Those are the two, those are the two futures.

(Joel Beasley at 00:29:15) You don't aspire to obsolescence. I aspire to legacy. You know? Exactly. Did it well. Yeah.

(Scott Dietzen at 00:29:21) And there's so much knowledge. You know, there's this naive view that we can just toss all of our existing software, and then somebody's gonna write a natural language description of, you know, what the problem we're trying to solve, and then AI is gonna produce magic software that solves all the needs. It's such a pipe dream. There's so much knowledge, including, ultimately, the behavior you want to specify is in the running code.

(Scott Dietzen at 00:29:49) Right? So we, the information that is there in these large systems is so extremely valuable to these companies that we want to be able to mine all those assets and then be able to improve them much more easily. You know, so many of these systems get ossified because people get scared. The cost of making a change in, you know, one of these large complicated systems, people are worried they're gonna break something.

(Scott Dietzen at 00:30:17) You know, AI gives you the confidence that you understand the ramifications of a change that you want to make. And so I think it unlocks a lot of value in systems that today aren't getting as much attention as they could be.

(Joel Beasley at 00:30:31) I agree.

(Scott Dietzen at 00:30:33) Do you

(Joel Beasley at 00:30:33) spend any time on X, on Twitter? I do. You do? Have you seen the guys doing the one shot challenges with the apps? The prompts?

(Scott Dietzen at 00:30:42) Yes. And, you know, I think it's interesting, but I think it's very different than, you know, what...

(Joel Beasley at 00:30:49) It's candy. Exactly. It's not, but it is interesting to see when they started doing those one shot challenges, like, two years ago, garbage. Now this dude did one that went viral about he, like, one shotted, and one shotting is where you, for people that are listening, so you take the prompt, and, like, in one prompt, it delivers a fully functional application code base. And this dude did it with a Spotify clone.

(Joel Beasley at 00:31:13) Did you see that one?

(Scott Dietzen at 00:31:14) Oh, I have not seen that one.

(Joel Beasley at 00:31:16) Oh, it was so good. Like, obviously, you know, I'm an engineer. Like, you're just not gonna throw it up on production. Right? And it's just gonna handle that.

(Joel Beasley at 00:31:25) But as far as from a single prompt, I was thoroughly impressed with how far things have come in two years. If that even just maintains the progression standard, I think within three years, we will be there. Yeah.

(Scott Dietzen at 00:31:41) Yeah. The idea of just being able to take a Figma and, you know, pass it into an AI and get the code that supports that interface, I really like that idea. You know, user interface is so tedious to develop and get right, and there's always these nuances. Having an AI to accelerate that, I think is gonna be, you know, fabulous. But at the same time, people shouldn't lose sight of application architecture.

(Scott Dietzen at 00:32:09) You know, if you're building a distributed systems microservices architecture, you know, there's so many insights required, you know, from data modeling and how you're dividing the work up. And these insights aren't captured anywhere. You know, they're in our software engineers' heads. So it's not like we even have the data to train the models on, you know, other than looking at sort of long term history as repositories evolve over time. So I think that kind of work, the vision architecture, you know, it's still very much human centric.

(Scott Dietzen at 00:32:43) And, you know, what we're seeing with one shot is we're automating more of the individual steps in that trajectory. But the, you know, human creativity and insight is still essential for orchestrating those steps.

(Joel Beasley at 00:32:59) You said your kids are around like 17, 13-ish. What are you telling them? You can see the future, even if you're like unconsciously competent about it, right? You know what's going on.

(Joel Beasley at 00:33:10) You've got these kids. What directions are you pushing or sharing with them that they should be looking at? Because they'll be entering the workforce in the next couple years.

(Scott Dietzen at 00:33:18) Yeah, I find, you know, when I first became a dad, I thought I was gonna be in charge. And one of the vivid realizations is these little humans, they come in very willful into this world, and they want to make the world the way they want it. And I've done nothing but encourage that. You try to raise adults and give them a lot of autonomy to decide what they're interested.

(Scott Dietzen at 00:33:45) So computer science isn't at the top of their list. You know, they're into music and art and history and philosophy. And, you know, I've obviously supported that. You know, at the same time, if they had been interested, I'm still very bullish on a career in software engineering. I think that there is a lot of work that, as a software engineer, I'd want the AI to do for me, but there are a lot of things that I still want control and insight over the process. And if we can unleash much greater productivity, we can pay down all this software debt and make it easier to build much less fragile software. You know, we can build more machine learning into the software so that it's a lot more robust and resilient in the face of change, that I think we can just deliver—every business and users can deliver the software that they aspire to and do so predictably and quickly—that can lead to a renaissance in software development. You know, people are just so sick of throwing so much money on failed software projects.

(Scott Dietzen at 00:34:57) I'd love to see that era end.

(Joel Beasley at 00:35:00) Well, improve the quality of our lives too.

(Scott Dietzen at 00:35:02) 100%.

(Joel Beasley at 00:35:03) Yeah. I'm an optimist in general with all of this stuff. But whenever I do get a little nervous, I just remind myself that there's still like these mainframes from the 1980s running critical parts of our infrastructure as a society. Just because the new technology comes out doesn't mean it immediately gets adopted. There's the human function of just the delay and the confidence and comfort of actually implementing it.

(Scott Dietzen at 00:35:31) Yeah. I would add too, I think a bunch of those systems are frozen because people don't believe they can make changes reliably to them.

(Joel Beasley at 00:35:39) That's true.

(Scott Dietzen at 00:35:40) And, you know, I think as AIs get more proficient in things like COBOL and FORTRAN, we may be able to bring those systems forward and ultimately reengineer them to be more modern and do more modern things. Right? There's still people having to deal with old terminal interfaces because it's the only way those legacy systems can be accessed. And there's just an opportunity to do so much better with an AI that helps you manage the details so you don't make mistakes that are gonna cause problems.

(Joel Beasley at 00:36:12) Maybe you can help some of the CTOs, tech leaders, VPs, engineers, those types that are listening. I'm curious about like pitching this to the executive team in the context of like ROI of AI development tools. Right? So let's say I've got some interest here. I'm a CTO.

(Joel Beasley at 00:36:30) I've played with Augment. I've played with these tools. I know I think we should be implementing them. How do I put that into business terms? Because you can't say, "Hey, guys, we're gonna lay off 80% of the staff because we're gonna be more"—you can't do that.

(Joel Beasley at 00:36:44) That's not allowed. Like, how have you seen technology leaders have a desire to implement Augment, sell it to their peers, and then deliver the results? Walk me through some of that.

(Scott Dietzen at 00:37:00) Yeah. So, you know, I will say, and I'll admit that the return on investment case is still something that's a work in progress for these AIs. I mentioned, you know, the 40% productivity improvement for a Fortune 500 and Codum's 50% reduction. One other one that I can cite is Webflow and Lemonade are customers. They've been able to use Augment to dramatically reduce the load on mentoring.

(Scott Dietzen at 00:37:29) So when you move somebody to a new portion of the code base or you bring a new hire on board, it's often six to nine months in a complex system before they're fully productive. We've been able to cut that ramp time down dramatically. And just as importantly, the senior leaders don't have to answer the questions because they now have an AI that's an expert in their code base. We've even had them test it. You know, CTOs and chief architects that have unparalleled knowledge of their internal systems will quiz Augment and be shocked that Augment is able to answer questions that most of the other senior engineers on their staff weren't answering, and they still had to come to them.

(Scott Dietzen at 00:38:09) And so, you know, if everyone can have an expert at their beck and call that knows the code base, it drives up productivity for everyone on the team from novice to expert. And you see that in really happy engineers. You know, that's probably the easiest way that ROI is quantified right now for customers is that the engineering team is more productive, and they're much happier, more satisfied with their job. You know, I think what we'll be able to do is establish ever better metrics for software quality until we can show that the AI is improving software quality as well as human productivity. And that, you know, I think that's still a work in progress, but I am convinced that that's happening today with Augment, that we are actually making our customers' software better, not just their engineers more productive.

(Scott Dietzen at 00:39:06) And so I think some leaders will find ways to cut costs. I'm much less excited by that than the engineering leaders that realize that they can now do things they were not able to do before. You know, the things that are on the roadmap for next year can be pulled in to this year, and you could start thinking thoughts about what else—what other frontiers do we want to go conquer with software?

(Joel Beasley at 00:39:31) That's how I feel about it too, because I think a lot of the job reallocation will happen in the form of when people naturally move on, not necessarily replacing them right away, versus cutting teams and staff like that. And that's because exactly that. Like, as a builder—

(Joel Beasley at 00:39:52) I'm excited about—I want the efficiencies that we can build more stuff faster. I don't want the efficiency so a number on the spreadsheet can change for department headcount or salary. I want to deliver more quality faster. And yeah.

(Scott Dietzen at 00:40:06) And the world is just short great software. Right? I mean, again, like, what piece of software do you use and say, "Well, that software is perfect"? It'll be so awesome if we could get a lot of—

(Joel Beasley at 00:40:18) Oh, man. Oh, snap. See what I did there? Yes. Oh, man.

(Joel Beasley at 00:40:24) Sorry for cutting you off. But yeah, you make a great point too. I didn't even think about that. When I saw it, I was thinking about how it helped me build faster and how it helped teams build faster in the context of in-engagement code. I didn't even consider the fact that the senior architect—you know, you're making a change, and you've got this whole team over here that's managing the Redis queuing system.

(Joel Beasley at 00:40:46) Right? And you're gonna interface with that system. So now you're gonna have to go talk with that team, ask a bunch of questions, figure out how the change you're gonna make, how it's gonna implement or affect them. But now you can just ask the question to the model in the context of the code base and get like 80% there, and then you could just go confirm what you've learned with the senior people in charge. And that's just beautiful.

(Scott Dietzen at 00:41:11) That exact use case comes up over and over, you know, just swapping a library out. You know, it's like, hey, your mate tells you there's a better library for this or that. You can just ask the AI, "Hey, sub in and change my code to use this library," and then test it and see how it works.

(Scott Dietzen at 00:41:32) And Augment is, for example, often able to do that complete substitution for you and generate a working application that then you can experiment with before you check it in.

(Joel Beasley at 00:41:46) That is so cool. That is so cool. This is an exciting time to be alive, Scott.

(Scott Dietzen at 00:41:53) I could not agree more. You know, I've—you mentioned earlier that I've been doing this for a long time. You know, I got to be part of the internet wave and mobile and, you know, got to see a lot of changes to application architecture and modern programming languages and clouds and microservices. So we've made programming better, but this is gonna be the biggest leap ever for software that we can—it takes such a step up in terms of improving quality and functionality and the life of software engineers.

(Joel Beasley at 00:42:29) Conversations like this make me feel great. Like, I'm gonna get off this call and tell my wife, "I feel like I had a great podcast," because you're at the point where it's advanced enough to be cool and exciting, but it also has immediate and direct business impact. And that's—you've been in this thirty years. Like, that is the Goldilocks zone. That's the sweet spot.

(Scott Dietzen at 00:42:50) Yeah. This market is moving faster than any I've ever seen. And, you know, when I would say the big upside surprise is how much better the models have gotten at software. You know, I think it was in part because the big players see software as a way to get to AGI in general, you know, that you need algorithms to solve problems. We were not counting on that, but it's allowed us to deliver so much better of a product, and especially when you combine it with the techniques that we've developed to bring all of the knowledge that's existing in your software to bear in the AI.

(Scott Dietzen at 00:43:27) That combination is so unbelievably powerful.

(Joel Beasley at 00:43:31) How can people get their hands on it? They want to see it. They want to touch it. They want to play with it.

(Scott Dietzen at 00:43:36) So, augmentcode.com. The plugins for the IDE are publicly available, so you can have them up and running in minutes to give them a try. We'll actually start looking at your code base. Generally, the first experience is we've examined this repository and it looks like it has these capabilities and these modules and this structure. And, I mean, in the case where you know the code base, you probably won't see any surprises, but that's a wonderful experience if you happen to try Augment out on a code base you're unfamiliar with.

(Scott Dietzen at 00:44:13) If you want to get oriented inside of code you don't know so well, the AIs make that really phenomenally easy. And then so you just start digging in and testing out and trying things, and hopefully, you'll fall in love like so many before you.

(Joel Beasley at 00:44:30) I do have a recommendation too. For me, my background—software engineering in Rails. Right? Like, I've never contributed to the Rails core. But I understand the terminology in Rails.

(Joel Beasley at 00:44:41) I understand it enough to know how to get around somewhat. But then putting that into Augment and being able to engage with it about Rails, I was like, "Woah." It was way better than—like, if I would have picked the model for the self-driving cars, I don't know those terms. I don't know about the vectors and the points and the lines and everything.

(Joel Beasley at 00:45:03) So I am just like, "Okay, I guess that's right," you know. But when you have experience, everyone has an open source tool. All software engineer leader type people, they have an open source tool that they're familiar with the nomenclature, right? So they could plug that open source project in and really get up and running quickly, gain confidence, and then plug their own code base in.

(Scott Dietzen at 00:45:25) It is a lot of fun, and we do get a bunch of customers that try that. And, you know, they pick an open source repository. Ideally, we encourage them to pick a bigger one just because if you want to see the difference between Augment and the other products on the market, the larger the repository, the more clear our differentiation is, you know. Context, it turns out, is quite expensive for these models.

(Scott Dietzen at 00:45:48) So, you know, some folks naively suggest, "Here, you just pass the code base along as context of the model," but the cost of that is the square of the length of the context. And so a system like Augment that is very judicious in its real-time selection of relevant context to deliver that knowledge gives you a much, much better experience. And it's fun, you know, to take an open source code base that you maybe know a bit but don't understand the internals of, and you can get familiar with it very quickly.

(Joel Beasley at 00:46:20) Yeah. And you have a competitive advantage too, because if I'm Cursor and I go to sell IBM, I'm like—they have to all download Cursor. Right? And all the engineers typically, companies will let engineers pick their tooling, right, in general. But if you have this broad support across them, across JetBrains, across, you know, all these different tool sets that you can provide this level of quality to.

(Joel Beasley at 00:46:50) I mean, do you see that? Am I getting that wrong? Or do you see that as a competitive differentiation?

(Scott Dietzen at 00:46:56) It so matters a great deal in the mid-market up to the enterprise. You know, I think in general, small companies, you know, they're a little more flexible with letting people bring their tools in, and they're not maybe sweating exactly where their code lands and what the security architecture is used. But, you know, if you start in the mid-market, you know, we end up doing a lot of due diligence with—you know, the sort of—our sweet spot in these more complicated code bases where you've got tens to hundreds of engineers, in some cases thousands working on the software. Then there's a level of security due diligence and, you know, due diligence around licensing. Right?

(Scott Dietzen at 00:47:41) You know, do I want to go live on a fork of an open source project that's gonna divorce me from future enhancements and cut me off from this part of this ecosystem, for example? And so, you know, we definitely find that we have that sweet spot sort of upmarket that I think it's gonna be harder for the folks that fork to penetrate.

(Joel Beasley at 00:48:04) Nice. Nice. Yeah. No. This is good. We gotta have you on like next year and see all the growth that's been happening.

(Scott Dietzen at 00:48:10) We would be thrilled to come back whenever you want us.

(Joel Beasley at 00:48:13) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn, or send me an email: [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.