[00:00:00] Dan Lenton: We've been building the core infrastructure and just getting the engineering building blocks in place. And we're now kind of shifting towards this mentality where we work backwards from user experience constantly. Every function we add, everything we do is directly answering the question, how is this improving the user experience?

‍

[00:00:23] Andrew Michael: This is Churn.fm, the podcast for subscription economy pros. Each week, we hear how the world's fastest growing companies are tackling churn and using retention to fuel their growth. 

‍

[00:00:35] Guest: How do you build a habit forming product?

‍

[00:00:37] Guest 2: We crossed over that magic threshold to negative churn.

‍

[00:00:41] Guest 3: You need to invest in customer success.

‍

[00:00:43] Guest 4: It always comes down to retention and engagement.

‍

[00:00:46] Guest 5: Completely bootstrap, profitable and growing.

‍

[00:00:49] Andrew Michael: Strategies, tactics and ideas brought together to help your business thrive in the subscription economy. I'm your host, Andrew Michael, and here's today's episode. Hey, Dan, welcome to the show. 

‍

[00:01:01] Dan Lenton: Thanks a lot for having me. Very exciting to be here. 

‍

[00:01:03] Andrew Michael: It's great to have you. For the listeners, Dan is the CEO and founder of Ivy, bringing all the best in class AI frameworks, infrastructure and hardware directly to you in one line of code. So to get started, maybe do your best if you could give us an overview of Ivy and perhaps what motivated you to start the company.

‍

[00:01:19] Dan Lenton: Sure, I mean that was the perfect overview to us. That was the perfect kind of pitch for Ivy's doing a few sentences already. Yeah, in terms of motivation. Basically so, during my PhD, I was increasingly frustrated with the fragmented AI stack. So in the lab I was in, where I was researching robotics and 3D vision, we had people who would often go and do internships at various different places. 

‍

[00:01:45] Dan Lenton: We had some that would go to OpenAI or DeepMind or Meta or Microsoft or Boston Dynamics or Nvidia. In fact, I feel as though everybody went somewhere different. I'm not sure why, but that's kind of what happened. What would happen is everybody would come back and people using different frameworks, whether it be back then, TensorFlow, PyTorch, Chainer, MXNet, Caffe, et cetera. 

‍

[00:02:08] Dan Lenton: But not only that, people were using different compile infrastructure and different hardware, with it being edge devices and so on. This was really frustrating. It was really hard to collaborate, really hard for new PhD students. And I got the sense that nobody was taking this problem really seriously in the field. So there was all this amazing research happening, but these fundamental frictions that made it really hard to collaborate across borders and across labs and across different companies and so on. 

‍

[00:02:36] Dan Lenton: So basically, I started down that rabbit hole trying to unify all of the stack. And I thought I'd spend a few weeks doing it and make a few tools to make it easier to research and get back to my research. But I never did. I've never emerged from that deep rabbit hole since and very happy for that to be the case. And to elaborate on what Ivy does, that's the motivation that kind of started me off on this journey. 

‍

[00:02:57] Dan Lenton: In terms of what Ivy does, I think there's two main, I mean, there's three things it can do. First of all, it can be an AI framework. So you can use it like PyTorch, the benefit being, or anything else, the benefit being it then serves lots of backends and works across hardware. But I think the main two immediate benefits of Ivy, that mean you don't need to write any new code, this is a long answer, but the main two benefits that mean you don't need to write any new code is that you can use any existing libraries or modules from another framework and bring them into your existing project. 

‍

[00:03:30] Dan Lenton: So you don't need to write lots of new stuff. You're using a TensorFlow project. You would like to use this library written for PyTorch and you can't use it today. Now with Ivy, with the one line of code, you can bring that library in from another framework and enter in your project. So one line of code, things become very easy to develop and build and research across the boundaries of these different tools. That's the first benefit, much more for, kind of builders and researchers. 

‍

[00:03:54] Dan Lenton: Now the second benefit that I think is perhaps more universal, particularly as we're moving into LLMs and a world where there's AI deployment on huge scales all over the place. The second benefit is that Ivy enables you to automatically optimize your pre-existing model and to get better runtime performance. So because we connect to everything so exhaustively, it means that you take your framework, which doesn't connect to all of the compiler technology. It doesn't connect to all of the new hardware vendors immediately. 

‍

[00:04:27] Dan Lenton: So it only connects to some of them, which means that you're limited when it comes to how efficient you can run your model. Because we connect everything to everything else, when you come to us with, or when you install Ivy, you have your model that's running at a certain speed. With one line of code, we can guarantee or almost guarantee in every case, we can make it faster for you. Maybe it's two, three, maybe ten times faster just by unlocking all of this technology that you're not immediately having access to. 

‍

[00:04:54] Dan Lenton: So basically, you develop more quickly and easily, run more efficiently and cheaply and do all of this with one line of code is really what Ivy is unlocking, as you very eloquently put at the beginning as well.

‍

[00:05:05] Andrew Michael: Very nice. And interesting how sort of that came out of a problem you were seeing as well with, between students going off to different companies, everybody working with their own different frameworks and then not effectively being able to collaborate across the board. The thing that caught my eye and the reason why I wanted to reach out today actually was I think in a relatively short space of time, Ivy has developed and become a really strong community. 

‍

[00:05:31] Andrew Michael: And just looking at the stats today, like there's over 12,500 get up stars, it's 1.2K and the open source contributors. There's over 5000 forks of Ivy and you have a, members in discord of a 15,000 now. I think one of the things when we think about churn and retention and in terms of business, one of the, like a really, really strong case you can make and where you can increase retention is by having a really strong community where people are with you from the ground up and they help you build and facilitate the growth of the organization. 

‍

[00:06:02] Andrew Michael: And it's very early on in your days as a company. But obviously I think I can see, like the steps and the process and the game plan here. And this is what I wanted to unpack today is the idea like, build a community, open source the frameworks and the technology. And then on top of that, build a business that's for those people that don't want to have to figure around and fiddle with open source tech and have a really made enterprise good to go service. Did it start out like this? Was it intentional? Like maybe start, from like early, early days.

‍

[00:06:34] Dan Lenton: I think if, I mean, let's say hypothetically that having that community helps with any potential, you know, churn of the company that emerges from this and that's a happy byproduct. But it certainly wasn't and isn't the main, you know, the main purpose of the community. And even the community itself is a byproduct of the open source project. I think they're all things that are, that kind of have emerged, not necessarily with total intent, but obviously very happy for it to emerge that way. 

‍

[00:07:03] Dan Lenton: So I guess starting from the beginning, basically, because of the nature of the Ivy project, it needs to be open source pretty much. And it was open source long before there was ever a company when it was myself doing this kind of, you know, committing away on my own for a year or so in the early history of the project. It was just myself and it was all open source. And it was the only way that it made sense because it's a framework like TensorFlow is open source. Pytorch is open source. All of this stuff is open source. If I want AI researchers to be able to use this, then it just needs to be open source. So that was never even a question. 

‍

[00:07:36] Dan Lenton: Now, I think what Ivy particularly benefits from, I mean, it suffers from and benefits from kind of in spite. I think it suffers from it actually, but it just happens to be a design feature of Ivy that it is incredibly modular and it's incredibly parallelizable. There are not many projects out there. I mean, I can't really think of any others that quite take this box in the same way. I mean, obviously everything's on a scale. But with Ivy, there are literally thousands of independent functions, all of which are basically the same kind of process repeated. 

‍

[00:08:08] Dan Lenton: So it's like what I did at the very beginning, winding right back to early 2022, which is when the company started, is, I made quite a lot of exhaustive YouTube series that explained how to contribute. And I went through and did like, you know, maybe six or seven functions in the front end and explained the process and just did a screen recording and showed how it works. I did these to the YouTube and et cetera. 

‍

[00:08:30] Dan Lenton: And then basically it's the same thing, but, you know, do it a thousand times more. And we've made huge progress and literally a thousand. That's not just me being exaggerating, literally a thousand times more or so. And then we start to make big progress on kind of getting the engineering workload done. So there's not many. Yeah. So basically we benefit because there is lots of, like low hanging fruit to pick up as a contributor. And then that's how we go from one contributor to, you know, twelve hundred contributors in the space of just over a year or so. It's because of that. 

‍

[00:08:57] Dan Lenton: The second thing is that what we did, because I'm a believer that it is very good for people before they join to have a bit of work experience on the actual job. So everybody that applies, we've now had over thirty five thousand applicants to our roles since starting. So long answer. Let me just wind back. So another thing that enabled us to do this is that we, so I take the view that exceptional engineers are pretty universally spread across the world, across the globe and pretty much decoupled from, you know, the prestige of universities and so on. I think there's amazing engineers because you can learn online. So, you know, there's, so there's engineers everywhere. 

‍

[00:09:34] Dan Lenton: So we basically reached out to a ton of computer science departments, all over the world in every country. We put our job out on all kinds of job boards in all countries and we just got the huge influx of applicants all over. And then we have this process where if you want to progress to the next stage with interviews and video assessments and stuff, then you need to get a pull request accepted on the repository. And that's easy for us to do because, as I said, there's like a ton of open tasks. So that also helped to fuel the contributor, you know, the wave of initial contributors because people were doing so as part of the applications. 

‍

[00:10:07] Dan Lenton: And now we get contributors coming to the discord, stumbling across it and contributing, you know, completely decoupled from applications. But that helped to get the ball rolling, I think, early on when it was me and a discord server with one person, me and you know, everything was just one person. It's kind of hard to get the inertia. And I think that kind of helped the ball rolling. And then obviously, just to say, long answer and what we now do obviously is every engineer on the team has equal responsibility to review these pull requests. So I wouldn't be able to review 1200 pull requests in a year, but we have engineers every day, obviously at a rate of several a day getting new PRs merged basically. And that's helped, I think.

‍

[00:10:44] Andrew Michael: Super interesting. Just to recap as well a little bit. So you started out in the early days, open sourced the technology, like really providing something valuable to the community from that point on. You were sort of figuring out, okay, like, how do we grow this community now? How do we get our value? And one of the first things you did was really develop education and enablements, like really focusing on the developer experience and educating how they can actually contribute to the platform. A lot of times you see obviously like docs being a natural path, but you went a step further, I think, and actually started creating content around this for YouTube and guides and how to’s. I think I love that. 

‍

[00:11:26] Andrew Michael: And then in terms of growing the contributor side, you might even think of it as a little bit of a hack, but I think it's an awesome way to, like, A, as you said, find the best talents and be able to hire effectively. But then also show them and show all these engineers the power of what you're building through that platform, because at the early stages, like trying to get a little bit of momentum. I think what you also really need to find out is like, what is that value you're really delivering to those users and to those individuals and find that fascinating? Yeah.

‍

[00:11:57] Dan Lenton: Yeah, I think another thing it just quickly, just to quickly elaborate on that tiny bit is I think another benefit is that, yeah, like, I guess it's a bit of a filter because it might be the case that some people feel as though it's below them to make a PR or something. And there's certainly a lot of great candidates that apply that don't follow through with that. But I think it kind of puts everyone on the same level. Everyone's just going to make a PR, you know, even if you think it's too easy, you know, just make the PR. If it's a challenge, then maybe it's a good experience for you to catch up with others who are applying and kind of get, you know, also it's a bit of an educational thing because our team very much help new contributors and, you know, explain it along the way in the PR comment list and everything.

‍

[00:12:37] Dan Lenton: So it's also a bit of an onboarding process, I suppose, for people that are a bit further. But one thing is that it's not assessed because each PR is different. So we only treat it as, like an entry barrier, let's say. And once you've got past it, all applicants are treated equally again because, you know, you might get a really easy function. You might get a hard function. We acknowledge that. And, you know, therefore we don't want to kind of start trying to compare apples and oranges. Yeah, but anyway, but I think this has been a big help.

‍

[00:13:02] Andrew Michael: Absolutely. Super interesting. So then as a platform, like you've grown quite a large audience now and quite a big community. First stage, getting up developer content. Second stage, like filling up those contributors. Where are you at now, today? Like, how are you working with your community? How is it growing? And what are some of the key areas you're looking to help them with at the moment? 

‍

[00:13:30] Dan Lenton: Yeah, good question. So basically, well, speaking what should be only a few weeks before our first real official launch. So we're hoping to move away from everything being developer focused. We've basically been building out the core infrastructure. I mean, for the last 18 months, it's still not totally it. Well, it is in a usable state and we do have users. But there's a few, kind of just little things left, which then unlock a lot of value once we've got these like final things ironed out. So we're still effectively in some sense pre-release, which I also think is, which also gives me reassurance that we're building something that's useful or at least captured people's mind share. Because we have still a lot of engagement, a lot of stars, a lot of contributors and big community, even though there's not a product that everybody's using en masse. 

‍

[00:14:23] Dan Lenton: So the next stage is to get people using this. I mean, but simply, and the next day you get the launch out and to verify with concrete demos, what Ivy unlocks. We at the moment have this early pilot access where we've had, I think, getting towards a thousand signups and people using the pilot. But we haven't really even, I mean, I'm kind of doing it now. We haven't really announced it. Like we just kind of added it onto the website and just thought we'd see who clicks it. But really, we haven't properly got it ready like to make it fully polished. And now we're going to launch that properly. And I would imagine, get several thousand signups and actually then start to really make sure that the user experience is as good as it can be and work backwards from that. 

‍

[00:14:59] Dan Lenton: So I think we've been, like, building the core infrastructure and just getting the engineering building blocks in place. And we're now kind of shifting towards this mentality where we work backwards from the user experience constantly. Every function we add, everything we do is directly answering the question. How is this improving the user experience right now and so on? So I think that's basically it. Kind of leave, developer building phase and make everything just a bit more, kind of user oriented and get this out there in lots of hands and respond quickly to questions and so on.

‍

[00:15:28] Andrew Michael: That's interesting. Sort of the shift in terms of who the focus is and who the customer, if you want to call them inverted commas, becomes. 

‍

[00:15:39] Dan Lenton: Exactly. I also want to quickly say that we're not going to on that cut off the developer side. 

‍

[00:15:43] Andrew Michael: Yeah. 

‍

[00:15:44] Dan Lenton: We have all the developer-facing channels. We're going to open new channels.

‍

[00:15:46] Andrew Michael: I wouldn't expect it to be like, hey, thanks very much. See you later. Yeah. So, and then from that perspective, like in the early days, you must have received a ton of feedback, obviously from developers themselves, which helped iterate the platform and drive the direction. I'm interested in this concept then as well, because obviously on one end, you have this strong developer community that's actually building the platform alongside you. It's providing great feedback continuously. And you mentioned now that you're shifting to this phase. 

‍

[00:16:18] Andrew Michael: Now we actually launch the product at some point publicly and you anticipate users to start using the platform within their workflows, within their production environments. And this becomes a slightly different audience as well. Like, obviously, there's quite a lot of overlap, but there's also a contradiction in that. And I'm interested now, like, how is the team viewing this challenge in terms of gathering feedback and iterating on the product and understanding what that roadmap needs to look like? Because I could imagine there's some sort of tension as well between what the community wants and what's actually going to be best for users at the end of the day.

‍

[00:16:49] Dan Lenton: Great question. But the first thing is that actually we continue to have back and forth conversations with our community. If people are using it, we, you know, every now and then reach out and say, hey, we'd love to hear thoughts on Ivy. If you have time to hop on a call or want to explain your thoughts on this, we'd love to hear it. We've now had, I think, seven and a half thousand. Because again, not only does our amazing engineering team have distributed responsibilities on reviewing pull requests, everybody also has some recurring engagement with the community. Because I think it helps for us to all collectively have a refined intuition on what people are thinking, what people want, what their impressions are. 

‍

[00:17:23] Dan Lenton: And for that not to all just be kind of funneled through one person, but actually all of us kind of have an ongoing sense of that. So we keep our finger on the pulse of the kind of ongoing sentiment, let's say, to the extent that we can. What also obviously helps a lot is that I and our, I mean, kind of, I guess sales team, I don't think it's really sales because we're too early for that. But we're talking to lots of companies, let's say, and kind of hearing their thoughts and building some early proof concepts and things like this. And that obviously constantly feeds into our intuition about what people care about. 

‍

[00:17:59] Dan Lenton: And I think there's not that much misalignment between the internal team, at least people on our team are not interested in building an academic project. Because it's interesting that everyone's aligned that we want to make this useful. Like this is the most exciting, it was useful, obviously, because you want to be built as something that you know when you make your commit, then the next day there's a thousand people benefiting from it directly and everything. So we're all very much aligned to that. And I think, so I might be going off on a bit of a tangent here, but one thing just to actually say what this has led us to or what this has made me realize through having these ongoing conversations, remaining open. 

‍

[00:18:33] Dan Lenton: This is also just a quick, advice to anybody building a company, which is obvious advice. But of course you need to stay open minded. You might have an initial idea for the way things will go, but the field, particularly a field like AI, changes incredibly rapidly and you need to go with the flow and adapt. And it will definitely be different every year a bit than what you thought before. And that's kind of almost essential for you to have any chance of surviving, I think, because of how quickly things move. 

‍

[00:18:56] Dan Lenton: And at least in our case, we were always focused on the frameworks and the infrastructure and the hardware. But we were perhaps earlier more focused on developer tools and enabling researchers, especially. I mean, I come from a research background, you obviously often build for yourself, researchers, especially to be able to develop and create an experiment across these boundaries. But actually, it's become increasingly apparent that deployment is a huge pain because people are deploying very inefficiently on scale. 

‍

[00:19:23] Dan Lenton: And therefore we recently opened up this whole new division, let's say, in what we're building this whole new area on deployment where we have. I mean, I won't go into the details, but we have all these deployment backends, which are the C++ Compiler [Instructure] backends where we want to make Ivy really, really fast on all hardware. And it was something that wasn't necessarily the focus even three months ago. And that's come from community conversations and sales calls and everything. 

‍

[00:19:47] Dan Lenton: But I would say you mentioned this kind of maybe slight difference between what the community wants, what the users want and everything. I think there's more alignment than I would have necessarily thought, because we're building something for lots of people. It's probably true that researchers and students and the kind of people that are most interested in Ivy in the community now might be a bit more on the building side, because they've kind of come on board with the old GitHub and the old messaging that was like, you know, unify the frameworks. So they might have this problem when they're doing research and there's maybe, you know, maybe on the enterprise side is a bit more deployment focused. So it might be a bit of a disconnect. 

‍

[00:20:24] Dan Lenton: But what I would say is that actually, thankfully for us, it all builds on the same stack. So I think even though there's a bit of a different [inaudible] on addressing these two problems and kind of molding Ivy to fix them, there's not that much. Actually, it's a pretty fundamental piece of technology, I think. And it kind of becomes a case of messaging and what demos you highlight rather than, like hugely changing what we're building in both cases.

‍

[00:20:47] Andrew Michael: For sure. But I think it's also then the other sort of what features matter to the different types of audiences. And when you're working more enterprise environment like security and privacy and user management and all these things start to become more of requests, which take away from, like the core concepts and the core offering of the product. So it's always a sort of, like, balance between when and why to do these things.

‍

[00:21:12] Dan Lenton: No, exactly. And we end up being spread a bit thin trying to cover them all. But I guess that's just part of doing stuff, I guess. 

‍

[00:21:19] Andrew Michael: Part of a startup. Yep. So I'm also interested in, like this model, obviously starting out with an open source project, building and growing the community and then building a business on top. I mean, it's been done quite a few different times and extremely successfully. So like, the likes of automatic come to mind, Red Hat, GitLab, MongoDB. All of these companies started out as open source projects and then went on and built on businesses off the back of that. What does that look like for you now, today? 

‍

[00:21:48] Andrew Michael: So you sort of got step one and you've got some good momentum building there. You're obviously going to continue to build and develop that community and offer services for them. You're going to launch the product to users. Like how are you thinking about the business side of things now? So how are you going to turn this community that you've developed and built now into business?

‍

[00:22:06] Dan Lenton: So I think the first thing is that we're not, particularly with the community, we're not thinking that we want to turn them all into customers or something, for example. I think, really the benefit, at least in my mind, the benefit of having a community, particularly when it's research and academic. So Ivy will remain totally free to use for all noncommercial reasons, actually. Parts of it are gated and you need to create an account and get a dashboard for a few of the more kind of, enterprise focused features. 

‍

[00:22:34] Dan Lenton: But basically, I think the value there comes from in terms of how this actually benefits your chances of making enterprise sales and actually getting the customers that are also essential alongside the open source stuff that you're building. I think Developer Mindshare is huge. There's several companies where the revenue has lagged, let's say. But what has been at the forefront is just buy in. Like this is the standard. I mean, it's like, Hugging Face, as an example, not necessarily kind of jaw dropping revenue, I don't think. But it doesn't matter. That comes. 

‍

[00:23:06] Dan Lenton: And what really matters is everyone knows if you have a state of the art model, it's a Hugging Face. And everybody adds their models there. And this kind of has huge, huge value that kind of unlocks itself over time. So I would say that we don't kind of think right now, how can we quickly turn all of these people that love Ivy into how do we make them more customers? We think, well, maybe some of them are doing research, maybe some of them are doing PhDs. 

‍

[00:23:26] Dan Lenton: One day they're going to be working in a company and if they want to keep using Ivy and if they have the Mindshare that this is the standard and this is the tool for the job, and we manage to cement ourselves into that place in people's minds, then when the time comes, whether it be in one year or three years or whatever long, then the thinking is that it then pays off at that point. And of course, in parallel to that, we are talking with enterprises who already might be interested in using this in an enterprise setting. 

‍

[00:23:54] Dan Lenton: So yeah, I think, yeah, I mean, that's kind of a half answer. But basically, it's kind of like what we first of all want to solve before we start rushing into any of this is like, is this useful? If people have this for free, are they going to keep using it? And like, let's first of all get that solved in the hobbyist stuff. And then obviously, kind of, we can build from there, is the thinking. But also what I would just quickly say is I say for free. Sorry, there's a bit of a delay as well. We keep talking to Richard there a bit. 

‍

[00:24:19] Dan Lenton: The thing I was just going to say is that it is for free, but not everything in Ivy is totally open source actually. So the ability to convert between frameworks and the ability to make your code deploy really efficiently is free for all noncommercial stuff. But it does just require an account basically.

‍

[00:24:33] Andrew Michael: I think that's also like going circling back to the beginning where you said you're not sure how the community leads to churn and increases retention. But I think you've alluded to a couple of different things now. One is first of all, just adoption and activation. And it becomes a really, really strong growth channel. So like if you build up a successful community, you become the standard. You first of all, like, create this incredible growth loop whereby new students are being educated. They're coming in. They get familiar with your tool of service. They move to a new company. They recommend Ivy as the tool of choice. And it just becomes a sort of self-fulfilling prophecy in a way. 

‍

[00:25:15] Andrew Michael: And ultimately then you become really, really sticky because I think a product like this as well, it is, like it's not one of the first things you're going to want to take out. Because if it becomes a core component of your stack and the way you work and operate, it becomes very, very difficult as well to come out. So I think you already have that, like, strong point of retention once it's in. And building the community is that way to get it in now. And it's that way to sort of foster that community, keep it as a standard. So I think it's all very, very exciting what you're building there and how it's growing. 

‍

[00:25:48] Dan Lenton: Fingers crossed it comes off as you say. 

‍

[00:25:50] Andrew Michael: As I say, yeah. And you're not at the moment. So obviously, like you don't want to convert all of the community. You still want to have this really good open source freemium component to the business because it's what spreads the word of mouth and encourage new users. There's a lot of other great companies as well. I think they've taken this approach specifically with students as well. So in the research space, like a couple of good companies come to mind.

‍

[00:26:12] Dan Lenton: Yeah. And it was actually a good decision whether we do it for all students or hobbyists as well. And in my mind, it made me kind of think about even some of the people that have applied to us, not all people who are learning about software engineering are necessarily officially students. There's a lot of people self-educating. There's a lot of people doing some other job, but just kind of in their spare time trying to change, career path. And it felt like blocking them off from the kind of freemium felt a bit unfair. So we've taken an approach that means that people in big enterprises can just log in with their Gmail account and do whatever they want to some extent. 

‍

[00:26:47] Dan Lenton: But then if you really want the fully on-prem, you know, dedicated engineer, all the proper security stuff that an enterprise needs, of course, this doesn't come with the freemium. So I'm thinking is even if anybody can sign up with a Gmail and get a lot of value, we still have a hook that means enterprises and also a license. I mean, the license will then mean that it can't be done for commercial reasons and stuff. So that's the thinking at least. We want to not put up too many guardrails.

‍

[00:27:08] Andrew Michael: Sounds good. Everything, moving, looking like it's going in your direction. What's keeping you up at night?

‍

[00:27:13] Dan Lenton: I don't know. That's a good question. I guess like–

‍

[00:27:21] Andrew Michael: What's really keeping you up at night? 

‍

[00:27:22] Dan Lenton: I guess, everything. No, I know. I'm thinking of it. I guess if everything coalesces around one piece of tech, then what we're doing becomes a little bit less useful, perhaps. I mean, I get, I'm interpreting that about very company specific. There's all kinds of things that keep me up at night about the state of the world. But Ivy, let's say, then I would say if everything was to completely center around video, for example, if everything was to center around, I don't know, like one particular compiler tech, let's say Python, for example, or XLA or something, then maybe Ivy becomes less useful. 

‍

[00:28:02] Dan Lenton: Now, I think we've been able to shift and provide value as the field evolves. So I'm not too worried about that. And also, I don't see any signs that the compiler infrastructure world or the hardware landscape is becoming super monolithic. And I actually think the GPU shortage, as one example, is an indicator that we need to be able to use long tail technology to get this working and not least CPUs. Actually, there's amazing research, not even particularly new research, but there's research that shows that combining things like quantization and pruning and even tensorization and so on, lots of methods to create sparse networks, enable them to be run in this like depth first approach where you don't do it layer by layer, but you can actually kind of explore deep, kind of neuron paths basically asynchronously and which is really well suited for CPU architecture. 

‍

[00:28:53] Dan Lenton: So I think there's going to always be a need to get all of this stuff working on like a shortage of hardware. And that in itself also means that you want your stuff to work across hardware so you're not locked in to one vendor that might run out or one cloud that might suddenly stop having access and so on. So I mean, yeah, but anyway, I'm kind of turning that into a positive from the question. I think we benefit when there's fragmentation, all layers of the stack. And obviously if that diminishes, then we need to kind of make sure we're providing value in such a scenario.

‍

[00:29:23] Andrew Michael: Yeah. And to summarize some of your gardening metaphors there as well, like energy, like what you're sort of seeing at the moment is, your concern is if everybody gravitates to a single hardware provider or a single compiler, this sort of makes the offering redundant. But that's obviously not going to happen because like we see in the market today, there's a big shortage is, improvements being done every day on ways you can optimize these models to work better on hardware. And there's different ways of doing that as well. And really like that's not going to stop anytime soon. I think it's evidence as well. Like if you've seen the open source community, if you see in different way like, programming languages have evolved. 

‍

[00:30:01] Andrew Michael: There was probably always a time, it's not concerned and it was like, Oh, what if everything becomes Python or what everything becomes JavaScript or C plus and there's probably always these tensions and arguments. But ultimately it's the communities like yours, like others that are growing and developing that are going to fight to keep them alive. 

‍

[00:30:15] Dan Lenton: That makes sense. And also, I think one thing I would say as well, even in that world where there is more coalescing than we might imagine, even then, Ivy provides, I mean, a lot of value because there's companies we're talking with that are using. I mean, this company's talking, whether they’re using Pandas and [psychic learn] or even Julia and Matlab and so on. And there's always going to be a long tail where it's really valuable for them to keep up with and integrate with all the cutting edge stuff. 

‍

[00:30:43] Dan Lenton: And even like five or 10 percent of the AI market, if we're the only tool that bring… keeps them on board, that's a big market to go after still. And so even in that world, it's not kind of existential, I don't think. But this is something that might, could diminish things a bit. But again, as you say, I don't see any signals of that really, really being the case.

‍

[00:30:58] Andrew Michael: I think as well, you mentioned, like a lot of different names, technologies and frameworks and things now. And I think one of the things that always struck me was, like how bad engineers are at naming things like you can definitely tell that a market wasn't involved in, like naming half of the frameworks and stacks out there. I do have, like Ivy and I like, it's all even more Unified AI. Like Unified itself, it makes, like, so much sense. It's easy to remember and it's a great domain. So good on you for getting that right. 

‍

[00:31:26] Andrew Michael: Last question then, as we wrap up today, like maybe, what's the thing that's exciting you most about the future and the direction that you're taking today?

‍

[00:31:33] Dan Lenton: I don't know. Yeah, good question. I think what excites me most is a world where we really achieve what we've set out to because we're not there yet quite, where we're getting closer and it won't be too much longer, but it's not all of the demos that we have in our mind. You know, some of them are still out there in our imaginations. And I think the day where you can really take any code, it can be NumPy code written 10 years ago. It can be TensorFlow one code from some state of your paper or some like, proper canonical piece of work that is the original implementation there from this super famous paper. 

‍

[00:32:10] Dan Lenton: You can just pluck these things off and run them on the latest hardware in your latest experiments pipelines and everything just becomes commodity. Like because even today, like researchers today, who are comparing to, you know, these canonical pieces of work have all these real implementation efforts. And also there's all these huge inefficiencies. The day that you can take this model from this framework and run it on your iPhone straight away or run it on your, whatever kind of edge device or alarm clock or something, kind of crazy, where all of these things are interconnected. 

‍

[00:32:41] Dan Lenton: I think the reason it excites me is because obviously it means the success of Ivy, but it just means a huge loss. It just means a huge efficiency boost, both in terms of energy consumption, which obviously has impacts for the environment, but also just in terms of developer hours and in terms of companies becoming more efficient and so on. Something about what we're doing, if we get it right, then the whole AI developer experience can become just so much more seamless, so much less painful and people which all AI infra companies are striving to one way or another. But people can again have another tool that enables them to focus on the maths, focus on the core innovation and technology and just kind of stop worrying about everything else. 

‍

[00:33:25] Dan Lenton: And like that percentage of time worrying about the details of the implementation and so on is particularly if you're doing, like theoretical contributions. It's just a waste of time. It's just like, painful and error prone and less time being creative and innovative and more time, you know, painstakingly copying and pasting this network over to this one line by line or something like this. So I think the innovation unlock and what that means for the AI field, the energy consumption savings and cost savings that enables are all, and I think just to wind that up, I think just in general, I've always been excited about just building one thing that's quite fundamental. I was so excited when I got my, like, first PR merged into TensorFlow.

‍

[00:34:01] Dan Lenton: For example, I've always been really excited when I find a big open source project and even just make a tiny change that makes things more efficient. I was like, wow, that's now like thousands of people that are going to benefit from my little bit of C++ or something. And it's kind of that same feeling, but taken further, because if this does get widely used and we're really helping the field, then, you know, even if we're not at the applied level, like, it's just like, really exciting. Yeah. And also just to quickly round it up with one that's not just self focused, not just inward focused in terms of the field of AI. The thing I'm very excited about as well is actually medicine and how [RL] can unlock really amazing things such as our [fold] and everything. 

‍

[00:34:38] Dan Lenton: I think that's actually more exciting than the chatbot wave potentially. And also I know that maybe deep mind, in recent months have kind of lost the limelight a little bit with the whole LLM boom. But I think that [RL] work they're doing has a huge amount of implications for unlocking new algorithms and mathematics and computer science. And I think that's going to take us down a path where AI compilers can be end to end driven. And if we can learn to compile code in a way that is, like, hard to understand and can combine assembly, logic and memory writing in a non human understandable way, that's like super, super efficient taking these shortcuts. I was thinking that's really exciting. So just to kind of put a few points on what excites me about the field. This is what I would say as well. 

‍

[00:35:23] Andrew Michael: Very nice. It sounds like an exciting future, Dan. I think for the listeners, I might need to put together a glossary for today's episode just for some terms and terminology you may not be familiar with. 

‍

[00:35:36] Dan Lenton: I can help you with that if you want. 

‍

[00:35:38] Andrew Michael: Thanks. Thanks. Yeah, it was great having you on the show today and sounds like you're off to an exciting future. I wish you best of luck now going forward. And yeah, thanks again for joining. 

‍

[00:35:46] Dan Lenton: Fingers crossed. Likewise. Thanks for having me, Andrew. Real pleasure to be here. Thanks. 

‍

[00:35:51] Andrew Michael: And that's a wrap for the show today with me, Andrew Michael. I really hope you enjoyed it and you're able to pull out something valuable for your business. To keep up to date with Churn.fm and be notified about new episodes, blog posts and more, subscribe to our mailing list by visiting Churn.fm. Also, don't forget to subscribe to our show on iTunes, Google Play or wherever you listen to your podcasts. If you have any feedback, good or bad, I would love to hear from you. And you can provide your blunt, direct feedback by sending it to Andrew@churn.fm. Lastly, but most importantly, if you enjoyed this episode, please share it and leave a review as it really helps get the word out and grow the community. Thanks again for listening. See you again next week.

‍