The VC's Lens: How AI Is Rewriting the Rules of Defensibility

0:00The chatbot era was very brief where everybody tried to build a chatbot, but I think today agents are essentially, as you said, first-class citizens. So as an example, a lot of enterprises were thinking of, hey, we have, let's say, insurance policies or healthcare data, and let's try to build a chatbot that can answer going through all of these. But the conversation has moved to, hey, how do we build actual claim processing or revenue cycle management using agents?

0:24Matt Paige:Welcome to the Talking AI podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI. Every company building AI right now is asking the same question. If the models keep getting better and anyone can access them, what actually makes us defensible? And it's the question that keeps founders up at night and the one that determines where billions in venture capital get deployed. And today's guest writes those checks. Avi Baradwaj is an investment director at Intel Capital, where he focuses on the software infrastructure layer of AI.

1:03Matt Paige:The company's building the tools, the platforms, and the systems that sit between the frontier models and the enterprise applications running on top of them. He's backed companies like Scale AI, Bree, True Foundry, and 12 Labs. He's got a front row seat to the questions that define this entire era that we're living in right now. when AI can do nearly everything, where does the moat actually live? Avi, welcome to Talking AI. Thanks, Matt. Great to be here. Yeah, and I'm excited for this chat because we have not had somebody on the VC side on the podcast yet. So I'm really interested in this perspective coming in because it's one I think a lot of people are very interested in with the dynamics in the market and everything going on.

1:44Matt Paige:But I want to start with the big question everyone in tech is wrestling with. So the frontier models, like we We say, keep getting better, more general purpose, more capable. And the conventional wisdom is that moats are being just demolished in real time. But as someone who's writing checks in this space, putting money where your mouth is, where do you actually see defensibility showing up right now? And is it where you expected it two years ago when we started in this crazy Gen. AI wave? Yeah, it's a great question. So I think it's a mistake to think that better models kill modes. In reality, I think what's happening is that they are commoditizing one layer of the puzzle.

2:25So as I'm sure most sophisticated software has multiple layers, the model being one of those layers, but there are many other important layers that make up an entire software platform. So what we are saying is that defensibility or the focus on defensibility is shifting to these other layers. So let me talk about it in a bit more detail. And maybe I'll talk about the application layer companies first, and then I'll talk about infrastructure layer companies. At the application layer, I'd say data modes are still relevant. It's just different from what it was before. So previously, just more data was important.

3:00I think given when you were building your own models, you just had to get a lot of different data. But today it has shifted that not just more data, but more unique data is important. Frontier models, they are largely trained on publicly available data. So they have data on language and coding and reasoning and so on, but they don't have data that's proprietary to an enterprise. So think of, let's say, triggers within your CRM system or domain-specific data in areas like healthcare and legal and so on, and real-world data in infrastructure deployments or industrial settings. These frontier model companies don't have that.

3:36So startups that focus on these types of unique datasets, I think, still have the edge to be able to beat out these frontier models. That's the first piece. Sorry. No, keep going.

3:46Matt Paige:Yeah, keep going on the different pieces. Yeah. Sure. Yeah. The second piece is what I would call workflow and system of action mode. So frontier model companies have pretty much standalone applications, right? Think of JackGPT, think of Claude sitting on your desktop and so on. but they don't have the ability to build out detailed workflows within enterprises. So a lot of enterprise workflows need to touch multiple systems, and there are like 50 to 100 steps in some of these instances. And you need some rigid system to be able to do all this. And you typically won't rely on a very flexible tool that some of these frontier models come up with.

4:24So startups, I think, are focusing more on this workflow and becoming the system of action as opposed to just providing one tool. That's the second piece. The third piece is around user experience and product. So in the SaaS world, there were just a bunch of screens for everything, right? So your work had to adapt to how that software was built out. And that created a lot of silos and data got segregated between these applications and so on. But with AI, startups have the ability to reimagine how work gets done. So there could be a lot of thinking about how to improve outputs, how to personalize these systems for certain workflows, how to automate things better and so on.

5:01So startups are focusing on that as well. The fourth bucket is around integration. So if you worked in a big enterprise, you know that there are 50 to 100 systems, both upstream and downstream from a specific task that you're doing. So how do you build integrations into all of these systems? And frontier model companies aren't interested or they don't have the resources to be building all these integrations.

5:22Matt Paige:Quick break in the pod. If you're listening to this podcast, chances are you've been thinking about how to actually use AI inside your business. And that's exactly why we built the AI Opportunity Finder. It's a free tool that helps you uncover high impact, tailored AI use cases based on your business, your goals, your pain points, and your industry. No fluff, no generic use cases, just real ideas that fit your business and the rank by ROI potential. It takes about three minutes to run and it's like having your own personal AI strategist for free. If you want to try it for free, check out the link in the show notes or go to hatchworks.com backslash AI dash opportunity dash finder.

6:01So startups have the ability or the gap to essentially go into this and build these integrations. And maybe lastly on the application side is around trust and compliance. Enterprises care a lot about how to deploy applications safely and at scale. And they would be hesitant to essentially give their entire reins to a frontier model tool that employees can essentially use as they want. So there is some level of trust and compliance that startups can do to essentially win over enterprises. So these are on the application side, and I'll talk a little bit about infra side, but it's somewhat similar.

6:36So I'll go through that quickly. So on the infra layer, the ultimate thing is that models aren't the end goal. They're just one component of a bigger platform. So tools that help builders build these complex systems within enterprises. So those are the types of companies that I'm seeing and other VCs are investing in. So how do you help enterprises understand these workflows, understand internal data, leverage these feedback loops and create data flywheels and so on. So those are the types of enterprise companies that are winning. And in addition to building all this, you still need to evaluate, you still need to benchmark how things are performing, need to make sure that security and identity are taken care of and so on.

7:14So those are still not in the purview of frontier models, if I may say so.

7:18Matt Paige:So for everybody listening, just take the transcript from this first part of the podcast, Just feed it into Collado Chat GPT and you have your investment thesis. Go see where everything's going. That was like, it's so great. But like data, workflows, taste, reimagining how work gets done. Like super context rich right there. I'm curious that like, I don't know a bit of an aside, but Jack Dorsey and Block, they cut 40 % of their workforce recently, right? And they were one of the few people I think that actually came out with a thesis. a perspective on where the organization of the future is going.

7:57Matt Paige:And they called it, he called it from hierarchy to intelligence. He put out this blog. I don't know if you saw it recently is in the past couple of weeks, but I saw it was literally this morning, I was watching this video from YC and they were going deep into this. But the reason I ask, I'm curious your thoughts on not only the things companies are building their products, their services, but like the company themselves, because in the YC article, I'll just kind of quickly hit the high points. And I'm curious your take for both startups and existing enterprises, but they hit on AI. It's not just a tool, but it should be the operating layer, like an intelligence layer that companies have.

8:35Matt Paige:Having closed loops within your company, because many times today you have a lot of open loops and input goes in and then something happens. And then you don't actually have the feedback loop that comes back in. But with AI, you literally can build that system in and it's a system that's self-improving in a sense. And then making the company queryable was the other thing they talked about where you could basically ask a question of anything and AI has context of your entire organization. And they talk about the human middleware and token maxing and all this stuff. But I'm curious your thoughts on this from the org perspective.

9:12Matt Paige:And you're typically looking at more startups, newer companies, so they're not having to deal with all the bloat, but any thoughts on this hypothesis that's starting to emerge? Yeah, absolutely. I think some of the big takeaways around being able to do more with fewer people, being able to build products that have a faster feedback loop, as opposed to in the previous version, I think a lot of time got spent collecting data, labeling data, building proprietary machine learning models. I was a data scientist before I became a VC. Oh, really? Yeah. And a big chunk of the work was essentially clearing data, getting data, labeling data, and so on.

9:48And even after you did all of that stuff, you got a fairly reasonable model, not the best model. But the difference today is that frontier models are essentially, because they're commoditizing intelligence, any developer anywhere in the world can immediately get access to the best model out there, right? So they can essentially embed intelligence or top of the state-of-the-art intelligence into their products. And that allows them to start at a different footing than what they would have started a few years ago. So all the additional gains that they get from that point is through, as I said, unique data, workflow, and so on.

10:21But they're just starting from a different point than what they would have done a few years ago.

10:26Matt Paige:Yeah, it's funny. So I used to work on a team of data scientists. I did not have the PhD in data science, but I was the conduit of pulling out the insights and analytics and things like that. But, and this was before generative AI existed. And I used to always notice there was like this invisible wall between what everyday humans could do and what PhD data scientists could do. And they were just doing these amazing things. But I also noticed the speed at which they did them was very slow, right? It's just completely different now. With data scientists as a role, how do you think that's evolving out of curiosity?

11:04Matt Paige:because there's still tons of value, but is it different in your opinion from where it was pre-generative AI to now? Yeah, absolutely. I think the way I would think about it is the work of a data scientist had, there was a lot of grunt work. And this, again, data scientist, data engineer, I think there are multiple roles that sort of blend into each other. So A, they had to do a lot of this grunt work on getting data, cleaning data, building models, testing, and a lot of those models would be useless and they had to throw it out. And then the other aspect was they had to build models for things that are pretty trivial today, like a sentiment analysis model or classification model and so on.

11:43They had to do all that by hand, whereas today they can focus on higher value work. Because when all of these sort of basic things that are common across companies are getting commoditized in the model layer, you can focus on building higher value models. You can focus on things that sort of get you from the 80 threshold to closer to 100, right? So I think it is definitely interesting. I've talked to a lot of my data scientist friends and they are really energized by the fact that they don't need to do a lot of the grunt work. Most of the data pipeline work can be automated today and they're spending more of their time on higher value work, which I think is great for everyone.

12:19Matt Paige:Yeah. And I think the interesting thing there is that's true for any role out there, right? Because I think engineers are at the forefront of everything. It's the canary in the coal mine. But literally this can be applied to any function in business. I just think it's going to take time for people that are less technical to realize that. But I love seeing that, that aha moment in people's minds where they're like, oh, wait, I can do that. And then it starts to trigger a million other things they can, they can do in a sense. But what, what, right. It's actually a good segue there. A good question.

12:53Matt Paige:within your own company, like Intel is obviously massive. Intel Capital is, and correct me for a while, but it's a smaller subsidiary or piece of Intel. But like, how are you all leveraging AI in your business? Any unique instances or examples, use cases of AI? I think the audience loves hearing, okay, how are the guests actually using AI in their business, in their personal life, in their personal day-to-day within their work life? Yeah, absolutely. So we've been using AI for more than a couple of years now, and it evolved pretty rapidly based on newer tools coming out. So at a high level, the job of a VC is I can classify it in four buckets.

13:34So it's seeing, it's picking, it's winning and supporting. So seeing is essentially making sure you're seeing all the startups in your focus area. Picking is like doing the due diligence and making sure you invest in the right companies. Winning is how do you compete with others who also want to invest in these companies? How do you make sure that the startup takes your money versus someone else's money? And then supporting is once you've made the investment, how do you do portfolio governance? How do you help your founders achieve whatever is great for everyone? So given this workflow, I've seen a lot of AI adoption both on my side, our team side as well, and broadly in the industry in the seeing and picking.

14:14So that's the first two buckets. I think the winning and supporting is more ad hoc and more hands-on as opposed to seeing and picking. So in the seeing and picking buckets, we use a lot of tools like Floodco. I use AlphaSense. We use Affinity and a bunch of other early stage products. So I'll give you a couple of examples on how I use these. So being an infra investor, I spend a lot of my time reading research papers, technical blogs, hacker news, Reddit threads, Discord channels, and so on. So that used to be a pretty big time sink. Even if you're a fast reader, it takes a lot of time to digest all that's coming out today.

14:51Now I have an agent on Cloud Cowork that goes through a lot of these threads, summarizes findings, identifies what companies are being talked about, looks them up on PitchBook, and then enters it directly into our CRM. And every morning I can go into it and I can see what new companies have emerged and what the conversation among the technical community is about these companies. So it's significantly reduced the amount of work that I have to put in to learn about some of these newer companies. In the picking side, again, I use Cloud Co-Work within Excel to help with modeling and analysis. So it used to take me hours to do some of the financial analysis, but it can be done in minutes today.

15:29So it's really quick and pretty accurate. So I'm really excited about all these other tools that will help accelerate some of the work that I've been doing.

15:37Matt Paige:Yeah, it's funny. I was, it's like things like that. You're just like, I think we're taking for granted how easy stuff's getting. We were doing our quarterly planning for Q2 and I'm sure you're used to this. many listeners as well. You do your quarterly planning. You got to look back at how the last quarter went, what happened. But this time around, I just fed all the data to Claude. I can't remember if I was using coworker, Claude code, whatever it was, but I fed all of our data. I connected it with our CRM and I just said, Hey, do a deep dive on the past quarter, any insights and help me plan for next quarter.

16:12Matt Paige:And it generated this amazing presentation and it had a lot of nuanced takes that I wouldn't have gotten to myself. Right. And I also had a presentation that was ready to go in the meeting, which beforehand I would have spent a week formatting crap and thinking back to the days of having to format a PowerPoint presentation. I will never go back there. but there's just so many examples there. I think part of the, I think the insight here for people listening is just start, just start doing something, just start playing with it. Cause then you'll start to get more familiar with the tools. I think connectors are a massive thing with whatever tool you're using, connect it to where your data lives back to the point earlier, you're mentioning around the data side of things.

16:58Matt Paige:Yeah. I got like several ideas floating in my head in your industry specifically where you could apply AI because at the end of the day, it's pattern matching. That's a key element to what you do in essence. And AI is so good at pattern matching. Absolutely. And I think there are a lot of newer tools coming out as well, like OpenClaw, the whole autonomous agents and so on. I'm a bit hesitant to use it in my work setting. I've tried it on my personal machine, but I think as security and sort of trust improves, we'll start using a lot more of these tools within our workflow. But yeah, it's pretty fascinating how much of this can be automated.

17:34And to your point earlier, this will allow us to do more higher value things as opposed to, as you said, like fixing the PowerPoint presentation or doing things that are essentially groundwork. Totally.

17:45Matt Paige:And you've backed a lot of companies building specialized foundation models, like video understanding with 12 labs, image generation, Bria, time series, robotics, all these different things. But I'm curious to your take, there's some pundits out there that'll say, oh, this is just a head start. The Frontier players, Quad and OpenAI, they'll drop some feature and it'll completely obliterate some company or the market will have a massive reaction. We're seeing some of that in the markets, but what's like your take of like the hype versus reality when QuadCode does their latest announcement? I think the latest one is like QuadDesign and then I think it was Figma just was hurting the next day, but where do you see the hype versus reality and how should people think about this on the market side of things?

18:33Yeah, absolutely. So I think it's definitely possible for some types of specialized use cases to get subsumed by the bigger frontier model companies. Like one example is a couple of years ago, I think one of the prevailing wisdoms was that for coding, you needed a dedicated model that can do coding. And as it turns out today, you don't need that for most green field coding. You may still need that for some lower level coding, or you may need it for the code migration and a few use cases. But for most green field coding, I think it turns out that big generalized models are pretty good there. But I feel like there are still areas where specialized foundation models can be helpful.

19:12I think your question had two aspects. One was around specialized foundation models, and the other was around what happens if a tool like flawed design comes out affecting something like Figma, right? So on the specialized model side, I feel like there are certain fundamental things that make certain areas pretty attractive for specialized models. So I'll give you a couple of examples. So one is with unique datasets, again, going back to data, like our portfolio company, Bria, for example, only trains on licensed images with the permission of the license holder. So they cannot violate people's copyrights.

19:43And there is a universe of customers that cares a lot about this. So that's the universe that would use something like a Bria model. And another portfolio company of ours is Field AI, which trains on live deployments in industrial sites, which none of the frontier model companies are doing today. So it's a very unique type of data set that they're getting. The second angle is around unique model approaches. So as I'm sure most of the big frontier model companies are trained on standard transformer-based LLMs. But there are other approaches like state-space machines or hybrids like Jamba and so on that could be better for use cases like edge devices because they are low memory or they are energy efficient or they can do real-time inference and so on.

20:25So there is that niche where big model providers won't change their entire architecture to fix this use case. So it's a fundamentally different approach. And then the third bucket could be like unit economics. Like there are lots of use cases where frontier models can do a good job, but the cost per token is so high that it may not make sense to use these big generalized models. I was talking to a startup recently and they use a model called the segment anything model to pick out check boxes from PDFs. So apparently they have better accuracy and they have a lower cost as opposed to using a big frontier model, right?

21:01So I think we see the world moving towards having dozens or tens of different foundations, like specialized foundation models for different use cases. And like I said before, the trick is in building big platforms in which models are only one component. Like you would use one or a handful of models, but there would be other layers in the stack that would make it pretty sophisticated. So I think that's our view on generally specialized foundation models. And your question on Figma, I think is really interesting. Thankfully, the private market companies don't gyrate quite closely with public market stock.

21:35Otherwise, it would have been crazy to manage our investment thesis. So we need to have our investment thesis needs to be looking out five or 10 years out and thinking about what will happen in the medium to long term. But there are trends, I think, that vary based on what's happening in the public markets. So you have to take that into account. So again, this goes back to a lot of what I already said, which is, and this is not an original thought that I've had, but a lot of smart people have said this, is that try to build things that improve as model improves. Don't try to build things that sort of fit the gap of, hey, the model companies are doing these things, but they can't do, let's say, A, B, or C.