Transcript: Building RelationalAI: AI Coprocessor for Snowflake | Molham Aref CEO & Founder

Host: My guest today is Moham, the CEO of Relational AI. He was the CEO of Logic Blocks and Predicttex before this. Relational AI recently raised series B of 75 million dollars from Tiger Global, Madrona, and Menlo Ventures.

They're widely known for solving use cases like rule-based reasoning, predictive analysis for large enterprise customers. So this episode will try to focus on, you know, what does modern data stack look like?

What are the AI applications that can be built and a lot more interesting things around those. With that, Moham, welcome to the show.

Guest: Thanks, Naraj. Pleasure to be here. Looking forward to the discussion.

Host: So, I think a good place to start would be, you know, you talk talk a little bit about your career and all the things that you've done before this, and how did you end up with Relational AI?

Guest: Uh, great. Uh, so I've been doing, um, uh, machine learning and AI type stuff in the enterprise, uh, under various labels, uh, since the early 90s. Uh, so it's over 30 years now.

Uh, it started out working on computer vision problems, uh, at AT&T, uh, as a young engineer coming out of Georgia Tech. And worked there for a couple of years, and then I joined a company that was commercializing neural networks.

Uh, company called HNC Software, Hack Neilson Neuro Computing, and worked on, uh, some of the early neural network systems. Um, in particular, in the area of credit card fraud detection. This is what HNC did. Uh, but other kinds of fraud as well.

Uh, as what what I focused on was more, uh, retail and supply chain and and demand prediction and so on.

Um, and, uh, was very fortunate to have, uh, a wonderful experience there, learning about, uh, the technology and all the things you have to do when you put together, what today we would call an intelligent application.

Uh, also, we had a very nice, uh, IPO in 1995. We, uh, bought a small company called Retail, that was working specifically in the retail industry and learned a lot from that experience.

We grew Retail quite substantially and spun it out in another IPO in 1999. And, um, we start thinking, oh, this is easy, anyone can do this. Uh, so in the early 2000s, I started getting involved in startups myself.

Um, you know, one was, um, um, uh, in there also, uh, actually trying to apply computer vision technology to a brick and mortar environment, a company called Brickstream. Uh, unfortunately, Brickstream was a good idea too early.

Learned, uh, a little bit about, you know, um, how too early is indistinguishable from wrong in the in the startup game. Uh, and then, uh, was involved in.

Host: Was it similar to what Amazon go stores was like? What were you trying to do and do.

Guest: Not, not nearly that sophisticated, but yes, sort of the idea is that you can put, uh, stereo cameras in the ceiling of a retail environment, retail bank, retail store, and start collecting information about what, uh, consumers are doing, you know, where they're dwelling, what products they look at, what products they don't look at.

Uh, again, within, uh, certain error tolerances, and then connecting the whole customer journey because you would get handed off from camera to camera as you walk through the store. Uh, you know, anonymously. We didn't know who people, uh, were.

We were just looking at the top of their heads. And, um, and so you build a, a picture of what the, uh, brick and mortar experience is like.

Um, and so at the time, this was pre-deep learning and everything was sort of handcrafted, uh, computer vision algorithms. It worked, uh, but it was, um, expensive. And, uh, a lot of the, um, the the challenge was, what do you do with that data?

And so we we weren't just solving a problem around the computer vision, we had to sort of justify the data. Um, so anyways, good experience, not a good, uh, commercial outcome.

Um, and then I helped, uh, um, start a company in the wireless network space.

Uh, where we did we built network simulation and optimization, uh, systems for, uh, AT&T, Singular, American Moville, um, a bunch of wireless operators and helped them migrate from the 2G systems they were on to 2.5G and 3G systems and helping them manage, uh, infrastructure and spectrum and a whole bunch of other stuff that you couldn't, uh, deploy without the benefit of very sophisticated, um, intelligence.

Host: How did you go from like complete machine learning vision in that space to like, you know, wireless networks? Like how did that idea came up?

Guest: So, you know, look, at the core, a lot of what we do in machine learning and AI is about modeling.

And, uh, you know, as we have deploy a handful of modeling techniques, uh, you know, simulation, um, machine learning, um, Gen AI more recently, uh, rule-based reasoning, uh, mathematical optimization, things like integer programming and and linear programming, uh, graph analytics.

So, the the the the AI toolbox has, uh, you know, half a dozen tools that you deploy in a variety of context. And it's it's perfectly normal for the domain to vary.

Uh, so whether you're modeling a wireless network or a retail supply chain, um, there are entities that are involved. Um, so in, you know, in a retail supply chain, you might have raw material in the form of fabric.

Uh, in the wireless industry, you have raw material in the form of spectrum. Uh, in the retail industry, you might change that fabric through manufacturing to make it into a t-shirt.

Uh, in the wireless industry, you take that raw spectrum and using, you know, Ericsson and Nokia equipment, you convert it into a wireless minute or a wireless kilobyte. And, and then you manage supply and demand accordingly.

So in both context, you have to predict how many customers are going to want this wireless minute or this t-shirt, and you try not to make too much of your product because if you make too much, um, you know, it's perishable.

Uh, it sort of loses, uh, uh, value. And in the wireless game, it instantaneously loses value if you make, uh, you know, something that no one consumes. Uh, in the retail, it depends, but it also, um, withers over time.

And there's just a lot of the same principles and what you need is a prediction typically of what the demand might be under various, uh, conditions and then how to satisfy that demand, uh, most optimally. So, I, you know, I'm not a wireless expert.

I'm still not a wireless, uh, cellular, uh, expert, but you learn enough about the core concepts in that domain. And of course, you work with, uh, you know, um, technical people who can simulate, uh, networks in the in the right way and so on.

And, uh, uh, but the value proposition is compelling regardless of industry. So, anyways, that company, uh, did reasonably well was acquired by Ericsson, and, uh, we didn't take any outside capital for that.

I I sort of helped bootstrap the company and so it wasn't a Silicon Valley outcome, but it was a nice outcome.

And, um, uh, after that, I went back into retail and, uh, helped build one of the first companies that, um, built retail solutions on the cloud. We went all in on the cloud, 2008, 2009, when that wasn't an obvious choice.

And leverage the cloud to be able to do a new class of machine learning, right?

Very super scalable, uh, compute infrastructure, uh, to be able to build better predictive models that predicted consumer demand under various scenarios, uh, much more accurately and and saved our customers hundreds of millions, if not billions of dollars, uh, because of the corresponding efficiency.

So, my whole career was spent, um, working at companies focused on building one or two intelligent applications. And in every situation, it was a mess.

Uh, it was very difficult because you had to combine different technology stacks, you know, the operational stack, with the BI stack, with the planning stack, with the prescriptive analytics, uh, uh, stack, with the predictive analytics stack.

And so you end up spending a lot of time and energy just gluing it all together. And and and that's fundamentally the reason, uh, uh, building intelligent applications is hard.

And I thought, hey, wouldn't it be cool to actually build a a generic technology on a platform that's very popular, very, uh, widely deployed, to make it so that you don't need, uh, so many components and so much duct tape and chewing gum and paper clips, uh, chew, you know, attaching it all together because each system has its own data management, it's its own programming model, its own limitations.

And, you know, most of the energy goes into the glue, not into the actual, you know, building a model of the enterprise and applying it properly.

Host: I think it would help for the listeners if you can talk a little bit more about, you know, what what does predictive analytics mean? Or prescriptive analytics mean? Or rule-based reasoning? Like what do these terminologies mean for a layman?

Guest: Yeah. So, uh, so broadly speaking, and this is sort of business speak a little bit. Uh, you have descriptive analytics. Uh, so what happened? It answers the question, what happened in my business? So business intelligence is a form of that.

You're looking backwards and you're saying, okay, what were the sales of flat screen televisions in Boston last year? Okay, so, and you can aggregate the data by, you know, by region, by different time granularity, by different type of product.

Uh, BI is a big, uh, uh, you know, big driver of analytics in the enterprise, just trying to understand what happened. Okay.

Now, if you just have descriptive analytics, it's up to the human, uh, to look at that and and then project forward from that as to, you know, what what you'll sell in January in Boston or Philadelphia, uh, you know, how many flat screen TVs.

And so, there's a ton of data to look at and you can improve that with a variety of modeling techniques, you know, everything from time series forecasting to, uh, um, today graph neural networks, uh, to help you understand what drives the demand, and you know, various product and, uh, customer characteristics and location, store characteristics, uh, that might drive demand for a product.

Because if you can now predict what what is, um, demand going to be in January or February next year under various promotional scenarios and under various pri pri pricing scenarios, uh, now you can now leverage prescriptive analytic technology, right?

So descriptive, predictive, prescriptive is, I know the relationship between say price and demand, what should I set the price to, to maximize revenue or profit or market share, uh, or whatever it is that I want to make more of or make, uh, you know, whatever it is that I want to minimize.

And the technology you use for each of these, uh, tasks is different. Okay? And it's not always BI by the way for descriptive.

Sometimes you want to describe what happened in terms of understanding graph connectivity and understanding how things relate to each other in a supply chain network, which is a graph, or in a in a social network, uh, that, you know, you use to model your customers and so on because different customers will influence other customers and will share households and and so on.

So, graph analytics are typically, uh, descriptive analytics that you can use in in in informing, uh, predictive analytics, and the form of features and so on.

Uh, and then once you have that predictive model, then you can do, um, you know, the prescriptive.

And there are other other techniques around trying to understand the why and the how and so on, but generally I think broadly speaking, analytics, uh, falls into these three areas.

And, uh, anything that's forward-looking is typically considered in the overall bucket of, uh, AI, you know, um, Gen AI of course being a very powerful new technique that we have in our toolbox.

But at its core, it's predictive because it's trying to predict the next word in a in a sentence, right?

Host: Yeah.

I think that's that's probably the most obvious but not understood concept generally is that, you know, you're just predicting the next word and it's not the same as like how intelligence, you know, like human's intelligence is not the same as the predictive intelligence like LLMs have and people often tend to consume confused that, you know, and equate it to human brain also like, which is I mean, maybe the output might be the similar, but the way both things, you know, actually do it is quite different.

Um, so you you figured out, you know, it's very hard to actually deploy these, you know, solutions, and you started relational AI. What I mean, you started in 2017 before Gen AI.

Uh, what were the primary use cases you were targeting and types of customers you were wanted to target at that point and how did it evolve over time?

Guest: Yeah, so look, my my area of expertise and our team's area of expertise is in the enterprise, uh, deploying all these different techniques, uh, including by the way rule-based reasoning, which is symbolic reasoning, which was, um, you know, in the 80s considered sort of the future of AI, okay?

Uh, but the idea was, take all of these techniques and for any hard problem in the enterprise, you can, um, deploy all of them to solve it.

So we've we helped build applications today that have elements of Gen AI and elements of, uh, you know, graph neural networks and elements of integer programming and elements of graph analytics and elements of rule-based reasoning, okay?

And it's the it's the whatever is going to help, uh, with whatever part of the problem, um, because sometimes the symbolic techniques are actually the best ones, right?

I think there's growing realization that, you know, for really rich reasoning, um, you know, integer programming and and rule-based reasonings actually much more effective than, um, some statistical model, right?

And clearly, the statistical models and the probabilistic models and the generative models are also much, much better for some problems than the, um, than the other technology, right? And so I I strongly believe that the combination is what wins.

Uh, in the same way that I believe that me as a human being, you know, with my own intelligence, uh, I could be much more effective if I have access to a computer and a calculator and a library and and all of that, right?

There's no, it's not like a competition between me alone versus, you know, the calculator. It's us sort of, uh, working together to solve our problems, yeah?

Host: Yeah.

Guest: So, that's how we started. We had, um, you know, building what we're building, um, is not trivial. Uh, we've got some fundamental discoveries in the in the way that, um, relational reasoning works.

Uh, um, you know, my view is AI and machine learning in particular, drive us towards data centricity because the data sets involved when you do machine learning in AI are big and so the old architectures that are, you know, use the, uh, the databases sort of in a dumb way and then pull data out to put in a Java program that does some computation over it, stop working when you have to pull a terabyte of data out or 10 terabytes of data out to make a prediction about, uh, something, right?

And so we wanted to move all the the semantics, all the business logic, all of the model of your business into as close to the data as possible.

And so we built, um, a system that you can we take to market as an extension to data clouds like Snowflake.

Uh, we call it a co-processor, uh, you know, riffing on the GPU CPU thing, where we plug in inside Snowflake and, uh, we build a, um, a relational knowledge graph as a structure that facilitates inside of this environment, queries that do graph analytics, rule-based reasoning, predictive analytics, and prescriptive analytics, okay?

And so now it's all in one place. Your data, your SQL, uh, your Python for, you know, for orchestrating things, uh, and then all of these, uh, capabilities.

And so you can very easily, like we see 10 to 20 X reduction in complexity and code size, uh, combine, you know, build an application that starts off with a prediction and well, starts off with descriptive analytics and then eventually produces a prediction and eventually uses an integer program to optimize something, uh, at a fraction of the the complexity and the cost and the time and the effort.

So, that's what we do.

Host: What forced you to sort of build on top of or, you know, as you call it co-processor on Snowflake because, um, I mean, there are other Snowflake-like platforms, um, you know, I can, you could say databrix. Um, and how how did you generally I mean, what pushed the edge in Snowflake's direction when you were starting out?

Guest: Yeah, look, this is a very important decision to make.

I sort of I was around in the 90s when, you know, I happened to be at a company that picked Oracle and Unix as a platform, and we were competing against companies that picked Informix, uh, and Unix or Informix, uh, or or some other relational database on an AS 400, right?

If you don't get the platform decision right, uh, you can really jeopardize your go-to-market mission. Okay?

And so from my perspective, we talked to a lot of enterprises, and what we see in practice in the Fortune 500 and the Global 2000 is for SQL, for data management, Snowflake is, uh, by far the leader, okay?

We see them first, we see big query a distant second. Um, but we didn't see a lot of red shift for example, uh, in the big enterprises.

I think some of Redshift's, um, uh, history is around being on premise software that was cloud hosted, but wasn't really protected for, uh, the cloud and so it lacked, uh, really important features around, um, you know, scalability, you know, things that you get from separating storage from compute, that I think Redshift has now, but at the time didn't have and so it ceded the market to, uh, to Snowflake.

And so you get time travel versioning, zero copy cloning, consumption pricing, uh, you have all sorts of benefits when you're cloud native, um, uh, that you don't have if you're trying to be both on on premise and on the cloud.

Um, you know, databrix until sequently until recently didn't have a SQL offering. Um, and they're still, I mean, Databrix is everywhere with spark, but we still don't see them that much for SQL.

And then, uh, Microsoft, you know, had synapse and now has fabric. I think it's a great product, but it's also relatively recent.

And so for us it was a, uh, really obvious choice to build on Snowflake, uh, because, uh, they're they've got the traction. Okay.

Now, you know, uh, whenever someone does well in our industry, you get lots of people sort of going after their business, and so they're facing a ton of competition from a variety of sources.

Um, also the the industry is being continues unbundling of data infrastructure. So you have open table formats in Iceberg, you have open catalogs now and sort of a fight between the community and Polaris.

Um, and there there's there are even sort of even the SQL engine itself is becoming unbundled and you have, um, some, uh, you know, credible open source, uh, SQL engines that can run on Iceberg and and and using these open catalogs.

So, you know, I I think, um, uh, you know, there's a lot, uh, a lot of competition here and we'll see how it all evolves and, uh, you know, my my bet's still on Snowflake.

I think they're going to do very well in the space, but, uh, you know, we'll we'll see how it goes.

Host: Um, you know, one of the challenges you often hear working in B2B is how do you get to those like first five customers? that's usually like a real big challenge because B2B is a sort of a different beast in terms of how do you approach customers, you know, how do you sell them.

Um, you know, it if if you don't do it right, it might look like you're operating a custom software creating company where you're just, right? Initially a lot of successful B2B companies look like in that format.

Uh, so can you tell us a little bit about, you know, your journey of how to find, how did you find your first five customers and, you know, what did they, what did it take to get to those first five customers?

Guest: Yeah, look, it gets progressively easier as you work in in in in a B2B space, uh, more, right?

When I when I was, um, you know, earlier in my career working at HNC, I didn't source the customer at all. we had a whole organization and, um, at the time, you know, with credit card fraud detection, um, so, you know, the when HNC started, they were just selling neural networks, okay?

And you go to a bank and say, buy my neural network and the bank goes, what's the neural network and why would I buy it? And at some point they realized, that's not really effective.

Uh, let's go to a bank and tell them we're solving a a problem they have in their language.

And so if you go to a bank that issues credit cards, and you say, hey, you're losing $500 million a year in in credit card fraud, and if you use our system, you'll only be losing 200 million a year, any banker is going to understand that, anybody's going to understand what you're claiming, okay?

And then it just becomes a a a matter of proving that you can actually, uh, create those savings.

And so you do a POC or a POV and you show how accurate your models are and you translate that accuracy into, uh, you know, um, profit or risk reduction or whatever it is, uh, that matters to your customer.

And so anyways, I just learned the importance of, um, learning the language of the industry, uh, that I'm selling into and learning the language of, uh, the customer. Um, so there's a ton of research on this as well.

Like the the you the folks that are most effective in sales are not like the slick talkers and, uh, you know, that's sort of a stereotype is a misconception. It's the people who can actually bring content to a conversation with the prospect.

So the prospect doesn't feel they're wasting their time, you know, and you can teach them something and you tailor it to their business, right?

And that gives you the opportunity to kind of earn time and with more time, you earn, you know, um, business, money. Uh, you get revenue. And so that is a thing that I learned in various, you know, various stages of my career.

Um, you know, fortunately after, uh, you know, being at it for a while, you you get a multiplicative effect from that.

Uh, so we had some, you know, early customers at Relational AI that used to be customers of mine 15 years ago, 20 years ago to different startup. But, uh, because of the good work we did for them then, uh, there was a level of trust.

And so when we walk in and we say, hey, we're early, but we're working on this class of problem and this is the team. Um, do you want to be a design partner and development partner for this? Um, um, you know, we made it easier for them to say yes.

Host: So, what what was the because you mentioned you know, you wanted to show some value, you know, for their problem. Um, because these applications or these techniques are sort of universally applicable in different fields, right?

Guest: Right.

Host: So, which field did you pick initially and what was the value that you were offering them?

Guest: So, again, starting from like graduating college, okay?

I I like computer vision, uh, and I just stumbled into an internship at AT&T at the time to work on it and at the end of the summer they offered me a job and it was a whopping $40,000 a year or something, which seemed like an infinity of money.

And so that that sort of stumbled in computer vision that way or stumbled into my first industrial use cases through my interest in computer vision. And AT&T was applying computer vision to, um, you know, reading documents, okay?

And then applying them to, uh, even back then we were looking at, uh, very expensive cameras to start processing video.

So then I learned a little bit about banking, I learned a little bit about, uh, the retail industry and retail in general, uh, by being exposed to the use cases, uh, for vision, okay?

And then when I went again to HNC, it was because I was interested in neural networks. I wasn't really particularly interested in industries. Uh, and the group I was attached to was selling into retail and into, uh, manufacturing in CPG.

And so you start think learning about forecasting problems. You learn about supply chain, and you learn about merchandising. And, uh, and then you learn again the the the the language of the industry that way.

And then you start reflecting on that and going, okay, even in a new industry, like when I started doing, um, cellular and Delco, um, you know again what's important to learn the language and you just, you know, it's not that hard.

It's not, uh, it's not something that you couldn't do if you commit yourself to it. Um, and, uh, you know, you improve over time as you engage with more and more customers and learn more and more about the problems, uh, uh, that they have.

So, it it takes time that, you know, I often see folks who just want to do the technical work, you know, in the abstract, and, uh, they can build, you know, amazing technology, amazing products, but not necessarily commercially successful ventures because you've got to do the hard work of seeing it from the eyes of your customer.

Host: So, um, right now, what type of customers are you mainly catering to? It's like is there a sweet spot of, you know, this is the type of customer that really, you know, is the fit for Relational AI? Like is there that customer profile evolve over time?

Guest: Yeah, so Relational AI, remember, is more of a horizontal infrastructure play, right? We are, uh, a co-processor to Snowflake. And there we've had to sort of learn new, um, new things about going to market.

And so in in the context of Relational AI, instead of going to the line of business executive and going and speaking to a supply chain person about supply chain or to a banker about banking or to an HR person about HR or marketing person about marketing, they all have, they all could benefit from AI.

They could all benefit from, um, the power of of AI. At our scale, we can't speak that language, uh, be experts in all of the possible, uh, uh, you know, business, uh, areas.

So, uh, part of what is working for us in terms of our strategy is we're targeting the Snowflake customer and there's usually a CDO or CIO or CTO or someone senior who understands infrastructure and data management.

And to to that person, we seem like magic because, you know, you think about it from their perspective, they've spent the last two or three years moving all their data from hundreds, thousands of databases into Snowflake.

They've turned off terdata and Hadoop and exdata and Natiza and, uh, and maybe Redshift, and they're generally very happy with that move because they get all this capability that they didn't have, uh, you know, in the old, uh, in the old infrastructure.

And, uh, the last thing they want to do is take that data and pull it back out from Snowflake to put it in a point solution for graphs or a point solution for rules or a point solution for predictive or a point solution for prescriptive.

It's all in there, it's all secure, it's all governed. Uh, uh, you know, it it's just, uh, you know, they don't want to go back in the other direction. So we come along and we say, keep it all there.

We plug in into that environment, we run inside the security perimeter, same governance. You don't have to worry about data synchronization because our knowledge graph is just a set of views effectively on the underlying Snowflake tables.

And we're we're cloud native, so we're architected like Snowflake. We separate storage from compute, so you get the versioning and the zero copy cloning and the consumption pricing and all that goodness.

And we're relational, which is the same paradigm as, uh, Snowflake, which is a, a conversation we can have, uh, maybe some other time because it's also really interesting, really deep of how to build these, uh, support for these workloads in a relational, uh, setting.

And so you get, uh, you get something that feels, uh, like magic, right? And so there we're understanding the buyer as someone who is on Snowflake and likes it. Okay?

And then we're coming to them, we're meeting them where they are, uh, and we show up with sort of integration and with SQL and Python, which are the languages of of, uh, that kind of platform, and, uh, we meet the the developer where, uh, they are.

And that makes everything easier.

If we take a point of view of, you know, we're awesome and you come to us and, uh, you should, you know, see our awesomeness and that's why you should sort of walk away from everything you've done, uh, uh, you know, and all the technology choices you've made, you can understand why that wouldn't be successful.

You might get a few clients who are, you know, um, early adopter types, but you wouldn't get to, um, scale, you know, Snowflake has over 10,000 customers now.

So there's a lot, uh, a lot of people could, you know, are in principle, um, uniquely served by our solution.

Host: I think, um, you know, when I see Relational AI, I I see this phenomenon of, um, companies building on top of other companies.

Like, um, you know, for example, there are billion dollar, I mean, there could be a billion dollar e-commerce company built on top of Shopify. I mean, there are unicorns, uh, you know, that are coming out of Shopify.

Um, you know, the three big clouds have enabled companies like, you know, Snowflake and, you know, Netflix, you could say is built on AWS, right?

Um, and probably the entire world's tech is is a certain percentage is built on top of, you know, all the three clouds.

And once the three clouds are established, then there are these set of platform companies, platform as a service companies that have evolved on top of it which are itself billion dollar businesses like Snowflake.

And then I see you are sort of building even on top of Snowflake. Um, it like, you know, and you're becoming a substantial first business.

Um, I often wonder like because if you're building an e-commerce company on Shopify, Shopify will never sell your e-commerce product.

But if you're building, uh, let's say an email service in Shopify, um, Shopify could cannibalize your business, which happened with, uh, I forgot the name of the company.

I think it was MailChimp or MailChimp acquired a company which was doing an email thing on Shopify and, you know, they eventually launched their own email thing and Shopify has also an invested in another email marketing company Klavio.

Um, so when you think of because you're building on top of Snowflake, and this is a problem probably Snowflake itself had when they are building on top of AWS or Azure is and they all have competing products.

So my question is how do you think about, you know, competing or getting cannibalized by Snowflake itself? Um, so love to listen your thoughts on that.

Guest: Yeah, totally not a new phenomenon in our industry, right? So, um, we've always, so we've had a history of things becoming platforms.

You know, used to be the hardware was the main thing and then people got tired of, uh, being held hostage by the hardware vendors.

And so operating systems came along and you build your applications and operating system and that abstracts over the hardware. And then people got tired of sort of being held hostage by, uh, you know, the different operating system providers.

And so Oracle came along in the 80s and 90s and said, the idea is like you build your application on top of Oracle and Oracle runs on all these operating systems. Uh, and then, you know, uh, we have Windows and, uh, you know, versus DOS and so on.

And so there's always this tension between the the platform and the thing running on the platform.

If if the thing running on the platform starts to generate a lot of, uh, value, then the platform provider can try to make it into a feature or capability into the platform. Okay? Uh, and you see this all the time.

Like every reinvent, I'm sure a bunch of startups go to Las Vegas for the Amazon, uh, uh, user conference and, uh, and a few of them come back in tears because Amazon just announced, you know, what they do as a feature.

Uh, and by the way, Amazon doesn't always win when that happens. You know, sometimes they announce something as a feature and it's terrible and the market still goes with, uh, you know, with the someone that's not, uh, that's not Amazon.

So yeah, you have to be really good and you have to have, you know, a substantial enough solution where it's either, you know, very difficult or very expensive to copy or it just takes time and the copy is always going to seem, uh, inferior relative, uh, to you.

But if you, you know, if you don't have anything that deep, uh, you know, look at vector databases, for example, right? Um, very hot, uh, six months.

Uh, uh, you know, uh, right after, uh, Gen AI kind of became came to the fore and some of those companies got massive funding, but now it's a feature in everything, right? Every every database has a vector capability built into it.

Um, you know, all certainly all the big providers, you know, still Snowflake, Amazon, Google, Microsoft, Databrix and even a niche database technologies now included. So, it's a you got to sort of walk a line, uh, there.

Um, and sometimes, uh, platforms get so powerful that, uh, that they win just because they attached their capability to the platform. I mean, that's in a way contributed to the problems that Microsoft had in the

---