3271: Inside Poolside’s Mission to Reinvent Enterprise Software Engineering

[00:00:04] What does the future of software development look like when AI isn't just assisting developers, but working alongside them, like a colleague? This is exactly what I wanted to find out more about today at the AWS Summit in London. I was hiding in plain sight in a rather swanky podcast booth to catch up with Isocant. He's the co-founder and CTO of Poolside, one of the most talked about startups in the foundation model space.

[00:00:34] You might have heard their name after raising $500 million to build what they're calling the most capable AI for software engineering. But what does that really mean? Well, in our conversation today, my guest will cut through the noise around generative AI and productivity hype and talk about what makes Poolside different in a crowded AI marketplace and why it might not be as crowded as you might be led to believe, and how reinforcement learning from code execution is really changing the game,

[00:01:02] and why enterprises need to stop outsourcing their intelligence and start treating AI like a core internal asset. So we'll talk about the evolving language of AI adoption, where we're hearing a shift from tool to teammate, and how Poolside's deep partnership with AWS is making it easier for enterprises to deploy AI securely behind their own firewalls and most importantly on their own terms.

[00:01:29] So whether you are an engineering leader, CTO, business leader, or just someone curious about what comes after AI assisted development, this episode is aimed at giving you practical insights and maybe even a few bold predictions along the way. But enough from me. Time to beam your ears directly right into the heart of the show floor here at AWS Summit.

[00:01:54] Well, thank you for joining me here on the show floor at the AWS Summit here in London. An incredibly fancy podcast booth, I've got to say. But can you tell everyone listening and watching a little about who you are and what you do? Yeah, so pleasure to meet you, Neil. And thank you for having me. I'm Iso Kent, co-founder and CTO of Poolside. And Poolside is a foundation model company that focuses entirely on closing the gap between AI capabilities today

[00:02:20] and human software development capabilities. And so we're really a company that's in the race for making sure that AI is going to get incredibly more capable to help us assist in a day-to-day basis in the way we build software. And there's a lot of hype around AI and a big focus or an increasing focus on the ROI of every AI project. So for every person watching who might be unfamiliar with what you do, can you start by explaining what Poolside is

[00:02:48] and what makes it different from other AI companies in the software development space? Because it is increasingly crowded out there, isn't it? So I think if you look at who in the world is really building capable AI from the ground up, the space actually isn't that crowded. I think it's about, you know, six, seven companies in the West that we can find that are building true foundation models.

[00:03:09] So meaning that we are investing these extremely large amount of resources in compute and data to build increasingly more capable AI that can help us and assist us in lots of areas of knowledge work. What's unique about Poolside is that we focus entirely on our AI's capability to help you in building software. And so we do that through special ways that we train our models. And we can talk a little about this in a minute, but it's also not just about the model.

[00:03:35] When you talk about ROI, it's, you know, an AI model is one component, but actually when you're bringing this in the world's most, you know, high consequence environments, and that's really where we operate. So Poolside brings the full stack for AI enabled software engineering inside enterprises and the places where high consequence software gets built. So think here, the largest banks, defense, you know, retailers, essentially the software that runs our world.

[00:04:01] And when we do that, we bring a model behind your firewall that learns from your data, a context engine that ingests your code and other data sources so that the model has all of the context required to make the right, you know, feedback and do the right tasks. And then the application experience that show up for developers in their editors where they work on a day-to-day basis.

[00:04:23] So in a nutshell, Poolside is AI for software development, specifically built for enterprises that goes all the way from investing in building the models to the end user experience. And you said it is a much, it's not as crowded as you might imagine. I would imagine if you take out all the companies that say they're AI but have no AI, it gets even smaller, right? A hundred percent. A hundred percent. Look, if we kind of look at the landscape, like building capable foundation models today is, I would say it's Google.

[00:04:51] I would say it is Anthropic, OpenAI, XAI. There's a few others that you can throw in that list, but it's quite a small set of companies that are doing so. But while others are focused on general purpose, touching every area, we really focus on how do we make our AI the most capable in building software. Because software runs the world and if we can give superpowers to developers so that AI allows them to be, you know, 1x, 1.5x, 2x more productive.

[00:05:20] That's really, I think, our way that we want to push forward capabilities. And you mentioned the word there, capable. And before you join me on the podcast today, I was reading that you said that Pulside is building the world's first or most, should I say, capable AI for software development. And I love how you're turning developers into giving them superpowers. But what does this actually look like in practice for that developer that's listening?

[00:05:42] Yeah. So I think we live today in a world that is human led, developer led, AI assisted, right? Models today are useful to help us on a day to day basis with our tasks. Everything from looking for where we can improve things in our code to helping us write code, test, fix bugs, etc. But it is not yet at the level of capabilities that we ourselves have as human software developers. So in that world, we live in this developer led world that is AI assisted.

[00:06:12] As models get increasingly more capable, we're going to see AI show up in a more agentic form and more in the form where it's becoming a co-worker for us, right? Adding to our team. But right now it shows up for developers in web assistants, in editor extensions, in CLI tools, all of the APIs that you can build around it, all of the places where you're doing your software development work on a day to day basis. And there seems to be a shift in terminology here.

[00:06:37] I'm hearing at more and more conferences, AI was considered a tool, but increasingly it's being called a co-pilot or a colleague, more importantly. Are you noticing that shift too? So I think we're looking in the world and particularly with marketing, marketing often likes to run ahead of where the capabilities are in the world. But we are definitely on a trajectory where, you know, if we look back several years ago, pre-chat GPT moment, AI for software development was code completion, saving us seconds. Then it became chat starting to save us minutes.

[00:07:08] Now it's becoming agentic, meaning we can give it longer tasks for us to do. These tasks today are on the order of minutes that they, you know, solving for us for 10, 20 minutes worth of work and several minutes worth of AI time. But as the AI and models in this case get more capable, that time horizon and the abstraction that you can give the objective on gets bigger and longer. And at that point, we're going to live in a world where we will start looking at it like colleagues and like co-workers, but we're not there yet.

[00:07:34] And I think it's important to kind of temper expectations of where models are right now, but not to underestimate the fact that we're on an incredibly fast trajectory now that in the coming two to three years, that gap is going to almost entirely close. Yeah, 100% with you. And Paul, you seem to have taken this unique approach by using reinforcement learning from code execution rather than relying solely on bigger models and more data. Why is that so important, that shift? So in our space, we talk about scale.

[00:08:02] And so you mentioned size of models, size of data, and we talk about reinforcement learning. So Paul said, when we started the company two years ago, the premise in the world was we're just going to scale up models in size with more next token prediction, and we will get to human level capabilities. And we said those are incredibly important access to scale models, apply more compute, and to make them more capable. But we think the most important scaling access for capabilities of models is going to be reinforcement learning.

[00:08:30] Now, reinforcement learning at its core is giving a model in an environment in which it can do tasks, and you can reward it when it's slightly more correct, and do the opposite when it's more wrong. Yeah. And this goes all the way back to the days of AlphaGo, when DeepMind kind of made this first breakthrough in 2016 to show that by letting a model play games against itself, it could learn from when it was winning and when it was losing. Now, we have taken that entire philosophy behind this and applied it to software development.

[00:08:59] And the reason that is because software development is something where you can say, is it more correct or is it less correct? Yeah. So much so that at this point, we have built the world's largest code execution environment. To give you a bit of the numbers, we're close to a million code bases, real public repositories that we have fully made executable with their whole test suite and everything associated with it.

[00:09:21] What that means is that during the training of our models, we can give them hundreds of millions and soon as we scale up billions of tasks in these environments where they're fixing bugs, creating features, building software, and learning when they're correct and when they're wrong. Where, you know, the first wave of model building was about getting models to learn from the existing data that's out there.

[00:09:44] This next wave of how we're improving models is about really experiential learning, given these environments in which they're becoming increasingly better. And that goes hand in hand with giving models time to think and time to reason. And you're starting to see that in the industry. And I think we've been very privileged that this was our starting point two years ago before the world kind of agreed. And now we're on a trajectory where we're really reaping the benefits of that.

[00:10:08] And for those business leaders that are tuning into our conversation today, what would you say are some of the practical ways that their enterprise could possibly use Poolside today? And how does AWS, where does that fit in and how do you help support those customers at scale? So let me take those as two questions. I think the AWS one is one I really want to dig in for a second.

[00:10:27] But for enterprises today, you are very likely an organization that is operating anywhere from 500 software engineers to maybe 100,000 software engineers, right? The global 2000 enterprises are thousands or tens of thousands of software engineers. So software is already at the core of your business and increasingly more so. Now we have a technology available that can come inside your organization.

[00:10:52] AI and a model that can become yours, that can learn from your data, that can start assisting your software developers and drive more productivity, but also frankly become a core part of the fabric of your company. And this is, I think, one of the bigger advices that we have for enterprise leaders is let the model come to your data. Let it become your model. Because what's going to happen in the coming years as AI gets more and more capable is that it is powering and empowering your workforce.

[00:11:22] And so start early. Now what's nice in the realm of software development, your comment earlier of like ROI and AI is that it's the one area where I think ROI is undisputed, right? It is measurable. You can see the impact. Your developers will tell you the impact. And hence it has an impact on the speed at which you can build software and the volume of which you can build it, right? This is not about cost saving. This is about allowing you to go into your competitive landscape and ship your products faster, ship your features faster.

[00:11:49] Now, to your second question related to AWS. At reInvent last year in November 2024, we announced something quite unique. We announced a first party partnership with AWS, something that has happened less than a handful of times in the history of AWS as far as I'm aware. And it means that if you're an enterprise that is running on AWS, you can bring poolside online entirely under your AWS terms of service.

[00:12:17] What that means is you can 100% retire your spend commitment. You can run poolside entirely behind your firewall in your account. So your data doesn't leave somewhere else towards a model. The actual model and system comes to your data. And this really tight relationship with Amazon has also meant that you benefit from all of the other things that AWS offers. And so we feel very privileged to be in that position. I'm also very excited.

[00:12:44] And so we're really oriented right now on how do we create this incredible experience together for AI-assisted software development with poolside and AWS. And so it's our preferred cloud to be deploying on. And that's been an exciting journey so far. And you made a really important point there when you said that AI is not a cost-saving exercise. And, of course, there are a number crunchers and business leaders out there that when they see AI, their first reaction is,

[00:13:12] well, if we can do this quicker, we don't need X amount of people on our team. And, of course, the magic is having those developers alongside that AI colleague. Is that message being delivered still, do you think? Or how do you communicate that message? So I want to give you the non-politically correct answer here because I think there are organizations where it is a cost-saving measure. Yeah.

[00:13:35] And I think if we look at the world of large enterprises, the question is, are you an enterprise that competes with your peers to capture market by bringing more value to your customers? Right? Often in the form of software. And then it's absolutely not a cost-saving exercise. Then it's about, you know, scaling up your teams. And more so, AI is going to allow you in the future to scale up when you need to massively accelerate and scale down when you're in a place where it's not necessary. Now, there are, of course, businesses that don't operate in that dynamic.

[00:14:05] You know, more very, very old-school enterprises that are, you know, maybe bringing resources or commodities to a market where it's all about, like, you know, your small percentage of margin. But even there, I would argue very often software is to deliver value. Supermarkets are a great example of this.

[00:14:19] And so, yes, for the vast majority, you know, and I'd probably say 90% plus of enterprises in the world, look at the ability to bring AI to your organization and to your software engineers as a way to do more, to do it faster, to be more competitive, to actually, like, have this faster feedback loop between what your customers need and what you want to deliver. And seeing it as a cost-saving exercise is probably not the way you already operate today.

[00:14:48] You're not letting go of your engineers today. You're actually, and I think almost every enterprise is growing their engineering workforce. AI is a way for you to, if you want to today, you know, have 20%, 30% productivity increase right now. That's where AI for software development plays a role over the coming years that those percentages are going to start looking like multiples.

[00:15:08] And if we look at other concerns there, how do you address concerns around things like security, customization, especially when companies want to fine-tune large models on their own proprietary code bases? I suspect it's a question you get asked a lot. Quite a bit. And so one of the unique things about us is since from day zero, we decided to build four enterprises. We say, you know, we really, at the end of the day, our users are developers, and developers are developers no matter where they sit.

[00:15:33] But being able to bring such a thing as a model, context engine, a fine-tuning engine, the application experiences, is specifically in enterprise environments, actually requires quite a different set of decisions. And those are decisions we made from day zero. So around security, we bring the entire poolside stack inside your AWS account. Everything sits behind your security boundary in your firewall. No data leaves to anywhere else. So you can have your data locality guarantees.

[00:16:02] You can have your guarantees of your security boundaries. You can even build security boundaries in your organization. And then to your second part of your question, in terms of models learning from your data, learning from your code, the things that are specific to your organization. That can be libraries and APIs that you have spent years building. It can be 20 years worth of legacy systems that you're working on modernizing that you want the models to understand. And it can be your proprietary programming language, like you find in some enterprises. That's where the fine-tuning comes in.

[00:16:31] And fine-tuning is part of a bigger effort to make the models more relevant for you. When you're fine-tuning them on your knowledge and your data and your organization, and you're adding it to the context engine, it's not about making them more intelligent or more capable of software development. Think of this a little bit as if you hire an engineer, you can hire the world's most capable engineer. But as they're coming into your organization, they're going to have to spend some time onboarding. Yeah. Right? Reading your code bases, looking at your documentation, understanding how things are done here. How does these APIs work?

[00:17:01] That's fundamentally what we're doing. And we can do all of that behind the firewall. We can do that inside VPCs. We can also do that by scaling up elastically with AWS Bedrock, which is an amazing way to get going and starting with no infrastructure and then scaling up along as you go. Or it can be entirely in P5 machines on AWS. And look, we operate outside the realm of AWS as well. So we have customers that are running on AWS Cloud, but then they're running on-prem with Poolside as well.

[00:17:30] So we're really built to be where you need us to be and with an enterprise-first mindset. And you mentioned Bedrock there. And here at the AWS Summit, your team is showcasing tools that can supercharge developer productivity. So can you tell me a little bit more about what people can expect to see at the booth from, I think, one of your colleagues' talks coming up, right? Exactly. I think we had Sebastian's talk. I think it just happened, actually, maybe while we're recording this. Yeah. So it's definitely worth watching online.

[00:17:58] If you go to our booth at this AWS Summit or any other of the summits that we attend or conferences is you'll find our product. You'll find demos. You'll see people to explain to you both how do we work and operate inside enterprises, but also just get hands-on with it. Yeah. And I think that's important. Like, at the end of the day, we're all developers. And so, and even if you're an engineering leader that's, you know, been a while since you've been hands-on keyboard, all of these enterprise features, all of the things, that's the table stakes that matters.

[00:18:27] Then you need the developers to love the experience. And I think this is one of the things that the booth really shows where you can go and play and see the experience. And then I think there's a third part, which is using it as a leader. It can be a CEO or it can be a CIO or CTO. So you have to start making decisions today about how you think about AI as part of your organization. Is it something where we send our data off over to the cloud to someone else and increasingly all of our organization's data?

[00:18:53] Or is AI something that we want to truly make part of our organization, a model that becomes ours, you know? And I think this is a part that we feel very strongly about. AI will increasingly get more capable. It's why you see us make these extremely large investments in more compute scale and more data to make the models more capable. But while we do that, that base layer is what you in an enterprise should be having behind your security domain. It should become yours.

[00:19:20] And it should be learning from your organization and from your data and from your developers' behavior over time. So I don't believe in a world where we have just one single AI that everyone sends all of their data to. I think every enterprise will have lots of versions of their models. And before you join me on the podcast today, I was doing a little research on you. One of the exciting things that I read that you said is that you predicted we could see human-level AI in knowledge work within the next 18 to 36 months.

[00:19:47] So what milestones are you tracking to know that we're getting closer? So I think it's the reason you hear such a specific definition from me. And often if you hear me say it, it's, you know, human-level intelligence and capabilities and knowledge work done behind a laptop. So the things that we are doing behind a screen. And that's, yes, writing code and working on marketing material and doing research and all of the things that, frankly, we do as knowledge work.

[00:20:14] And so at some level, you want to look at the core capabilities of reasoning of the models. Another area, you want to look at their ability to process and understand knowledge and the world. But really, I think the best measure overall is the length of time horizon of a task that AI can do. Right? In the beginning, it was they could answer a chat message. Then it could start answering, you know, tasks that took two or three steps.

[00:20:39] Now we're seeing areas where it can start doing tasks that, you know, would have taken a person 20 or 30 minutes to do. Now, I think the moment we say, you know, mission accomplished on that definition is when you see a fully capable agent that can operate, you know, in the same environment you do in your laptop. And is able to really go end to end over the course of multiple days, weeks, months in terms of doing the same job and tasks that we do as humans.

[00:21:07] And so I think we're actually really every individual user of AI is actually pretty capable of judging this progress. We, of course, in our organization have lots of evaluations and benchmarks and things that we look at to see how this trajectory goes step by step. But I think even as end users, we are highly capable of understanding where AI is. And if you look at where models are today and the length of tasks that they can help you solve versus, you know, six months ago or a year, it's, of course, already quite different. Yeah. Exciting times ahead.

[00:21:35] And finally, what's next for Pulsar, whether in terms of product development, global expansion or a broader mission you're driving towards? Any teasers you can offer there? Well, we try to look everything in our business kind of through two lenses. In one hand, the building of capabilities. So what you're going to see from us this year and next year and all the following years, just increasingly more capable models that get, you know, from a world that has gone towards early agentic to fully agentic to one day fully autonomous.

[00:22:03] And then on the go to market side, it's what you're going to see from us is just an increasing more amount of capabilities and features that are built specifically for the enterprise. We want to be, you know, and we want to earn the right, frankly, to become the trusted partner to all of the global, you know, enterprises around the world as they are going through this journey of today, assisting their software developers with their, you know, AI models.

[00:22:29] And over time, really become the partner that can also help them power part of their digital workforce. Fantastic. And for anybody listening or watching that want to find out more information about anything we talked about, or keep up to speed with some of those developments, where would you like to point everyone? So probably a good starting point is to go to our website. It's poolside.ai. And I think that's a fantastic place to be going to. And of course, you know, we're at Poolside.ai on Twitter and we're ex these days, but that's probably the best place to get going.

[00:22:57] Well, as I said, we're here talking together today at the AWS Summit. So many great new tech tools, emerging technologies. Great seeing how they're all evolving and coming to fruition, how they're all converging. But at the heart of it all, more than anything, I'm going to have to channel my inner Steve Borman out and say it's all about developers, developers, developers, developers. And that is a great message to send. So thanks for joining me today. Thank you so much, Neil. So as we wrap up our conversation today from the AWS Summit, what stood out most is how fast this field is evolving.

[00:23:27] Not just in terms of model size or performance, but in how AI is being embedded direct into enterprise workflows, developer environments and even organizational structures. And my guest vision of a future where AI doesn't replace developers, let's dispel that myth, but amplify them. And where enterprises bring the model to their data, not the other way around. Well, that isn't just a roadmap for poolside.

[00:23:54] It's a signal of how the next phase of enterprise AI adoption could unfold. So are we truly on a path towards human level AI and knowledge work? And if we are, the time to think strategically is right now. So how are you and your business going to harness these capabilities? How will your teams adapt? I'd love to hear your thoughts on this. Could your development teams benefit from a co-pilot team, whatever label you want to give it, with reinforcement learned experiences?

[00:24:24] And would you welcome an AI teammate trained on your propriety systems? How would that work? Let me know. Join the conversation on social or email me directly. Techblogwriteroutlook.com LinkedIn X Instagram And remember, there are eight different podcasts on my network now. So we have techtalksnetwork.com You'll find eight different shows all focusing on niche areas on how technology can impact your business. So thank you for listening.

[00:24:53] And until next time, keep asking those big questions. And I'll speak with you all again very soon. Bye for now.