00:29:41 · Matthew Fitzpatrick
[00:29:41] S0: We use all we use all of the kind of exist... Main models. So the way I would think about we're we're doing there is data does tend to be the biggest impediment, particularly in in large enterprises, to building anything with AI. So, you know, let's say you're a business and you've got four ERP systems, a CRM system, an ecommerce platform. You've got five, six, seven, eight different systems that don't all reconcile each other. So you wanna look at, like, the definition of a customer. It's... You have overlap or kind of components of that that don't tie across the various systems. A lot of what our tech does, Neuron brings together data from those different systems until we call ontologies or data objects. And then from that, you can use some combination of machine learning, one of the LMs, etcetera, to actually build out the models. So I I would say we use all the available tech. We we do... You know, we leverage every model out there. We will leverage a lot of tools out there, like Databricks, Snowflake, etcetera. So a lot of what we're delivering is a working AI system that starts with data, builds out a workflow, uses various AI models, etcetera, to actually kind of do automation and and predict outcomes. And then we also have kind of quality assurance and testing on the back end of that, which is equally important. So the observability and telemetry to say the model's working. Here's where errors arise. And that's Synapse is the mod... Is the is the platform we use to do that, which is another commodity of our text
00:31:11 · Guillaume Moubeche
[00:31:11] S1: there. And the the reason I'm asking this is when you see how models are evolving, there is a... You know, like, if you look at the story of of LLMs and also, like, if you look at some companies, if you take... Not sure if you've heard of them, Jasper, they grew extremely fast at some point, you know, they were like an OpenAI rapper. And what they were doing is like a lot of prompt engineering on top, you know, of of this LLM that at the time, you know, was was not that great, but with all the prompting could give, like, results that people had never seen before and that this distraction. However, as the models, you know, got better, basically, what they were building started to kind of shrink because the model could actually produce it. And where I'm saying, you know, the most value is whenever you can produce, you know, like a software or whatever model that whenever there's gonna be an update in the original model, you actually get a better, like, experience, you know, for for the customers. So from your perspective, how exactly do you create this kind of of experience for the customers so that whenever there is an update, you make sure that everything you've built is actually getting better and not kind of worse. You know what I mean?
00:32:33 · Matthew Fitzpatrick
[00:32:33] S0: Yeah. So Axon, which is our agentic platform, platform, is configured such that when you do context encoding, you can swap in and out every new model release that happens. So I think that is... The the... There is a fallacy of kind of like, let me wait till the models have advanced more before I do anything. And and the answer is if you if you architect that correctly, you can set that up so that you can easily swap to the newer version of the model. The other thing I'd say, though, is, like, the concern you raised in that question, though, is right, which is if you if you take something that's publicly available data and you're just adding that to an LM, it's gonna get commoditized pretty quickly. And so I tend to think of the world as, like, you have all the LMs, which are, you know, this all massive, Then you... And even some of the smaller LM builders. And then you have kind of the application layers that are being built on top of them. And in some cases, those are very, very, very value add. In some case, they are, as you're saying, wrappers. But those are usually businesses that are taking predominantly third party available data and configuring it around an LM. We're different than most of those. We're what I would call a last mile business, which is most of the context we're adding is very unique to that company. So it's not data that exists anywhere else. So it might be, you know, specific patient data in a health care context or specific inventory data in a manufacturing context. But that's data that is never gonna be... That an LLM will never have in it. And so actually, a lot of what we're doing is we're configuring that and and kind of encoding that knowledge into an LLM, which is actually... Obviously, I'm biased because it's what we do, but I think that last mile is actually where a lot of the value will come in enterprise over the next decade.
00:34:12 · Guillaume Moubeche, Matthew Fitzpatrick
[00:34:12] S1: Yeah. I agree. And I think like to your point, you know, of some enterprise, I mean, even if you take like some of the software business right now, potentially, you know, like their valuation went down a lot because as you mentioned, like, it's not that the business is gonna shrink, it's just the business will grow probably like at a at a slower pace. However, for some of them, they are sitting, you know, on a gold mine of data that they haven't exploited yet, and probably like this is the chance for them, you know, to to create something unique. [00:34:47] S0: Yeah. Look. I think of the existing software players, there will definitely be several of them that evolve themselves around new AI tooling and come out much stronger than they started. So I I don't. I think in any... You know, I I use the example in Mary Meeker's predictions of some of the incumbents did well, some did not. I think it will be the same thing with software. If there will be businesses that will legitimately transform and have a bunch of really interesting data and infrastructure and relationships already in place, and they'll probably do really well. [00:35:16] S1: Do you have, like, a example off top of your mind of different kind of moats that software businesses could could build?
00:35:16 · Guillaume Moubeche, Matthew Fitzpatrick
[00:35:16] S1: Do you have, like, a example off top of your mind of different kind of moats that software businesses could could build? [00:35:25] S0: So I think the biggest... You know, I think there... One is is businesses that have kind of a regulatory moats in some ways. You could call that in any kind of regulated entity. You know, I think actually systems of record that have to go through audits. Like, those are things that are actually gonna be pretty painful to change. And so, you know, I I I, for example, don't think systems of record disappear anytime soon because, actually, they're something a lot of people are gonna wanna have for a long time. Think the other big mode I would call out are actually two others. I think data moats, as you said, are a really interesting one. So if you if you have a lot of proprietary data on interactions, pricing, any of that, you know, AI is locally powered by data. So if you have that, there's a lot of value to that. And then the third one I would call out that I think will be probably the most important in the next couple of years is kind of the marketplace network moat. So if you've built something that has interactions between numerous parties on your platform, like an ecommerce site, I don't think you're going anywhere. Because I actually think in many cases, you know, multi... Multisided marketplaces are some of those powerful businesses in the next paradigm.
00:36:29 · Guillaume Moubeche
[00:36:29] S1: Yeah. I agree. I agree. And another thing, because we we haven't touched based on this so far, but, you know, like, everyone's talking about AI taking jobs. What's like... Or like even AI as a existential risk. So from my understanding, you think this is wrong, but can you maybe, like, elaborate?