The Next AI Investment Thesis

Julia Nimchinski:

Next up, please welcome Kathleen Estreich, early-stage investor and GTM operator, formerly Partner at Pear VC, our HSC regular. Super excited to have you back, Kathleen, and we have an exceptional group of investors today.

Kathleen Estreich:

Thanks for having me, I’m excited for the conversation!

Julia Nimchinski:

Awesome. Let’s get into it.

Kathleen Estreich:

All right, is everyone here? Let me see. All right, they’re all… gang’s all here. Awesome. Well, I’m excited to lead this discussion with all of you. Hopefully, the more… disperse beliefs we can share. Feel free to chime in if you have something that disagrees with one of the other panelists, so we can hear all the perspectives, but really excited to dive in. So we’re talking about the next AI investment thesis. I think anyone who’s investing right now knows it’s a pretty… crazy time, on a bunch of different fronts of, kind of, where are the opportunities, where are the moats, what, you know, what’s defensible, what’s not, so we’re gonna dive into all of those topics.

Before we dive in, or as we dive in, I would love for each of the panelists, rather than doing your intros, which are here, maybe each of you can talk about, you know, what is one thing you believe about AI investing right now that most of your peers don’t? So, we’ll start with your hot take, and then go from there as we dive into the additional topics. So, Jeremy, do you want to kick us off?

Jeremy Kaufmann:

Sure, I think so much of the… conversation around defensibility gets very academic, and there’s a lot of great, you know, ideas and theories, but I think people miss the basic premise that, like. at the seed round and the Series A, like, defensibility often doesn’t exist, but then it comes later, over time, due to technical reasons or go-to-market reasons, so sometimes I feel like… people engage in the conversation around defensibility almost, like, too early. Like, it’s great to talk about defensibility at the Series B, C, and D.

Like, I just use the example of lovable, or, like, so many people, you know, at the A said, you know, it’s not defensible, but, like, the sheer act of it existing, improving the product, getting better, distributing it. means that by the time it gets to the Series B or the Series C, suddenly we’re looking at seed or Series A opportunities in that same space, and my teammates at the IC are saying, wait, well, we can’t invest in X because of Lovable. And many of these people are the same people that two years ago said, Lovable doesn’t have a moat.

So I just find the timing dimension so interesting.

Kathleen Estreich:

Cool. Amir, what about you? What do you believe about AI investing right now that most of the peer group does not?

Amir Kabir:

Yeah, so, we have a differentiated thesis with the fund that I run, and, I believe that, you know, investors are still valuing AI primarily as a software opportunity, and we look at it from a risk infrastructure opportunity, and what I mean with that is that, you know, I think first of all, every technological wave makes something abundant and something valuable, and so I think we’re now in this AI wave, and intelligence is becoming abundant, and so as AI moves from you know, generating information to taking actions now. The scarce resource isn’t intelligence anymore, right? And we can see that on a daily basis, when we look at, like, open source models that are kind of getting on par with Entropic and OpenAI.

So what becomes valuable, what is, like, I think. you know. undervalue this trust? How can I verify what the system did? How do I understand what it did? How do I control it? How can I allow it to… what can I allow it to do it? And so we basically call it that We basically call that, software scaled by risk didn’t, and I think it’s a significant amount of value that will accrue to companies that close that gap.

Kathleen Estreich:

What about you, Andrew?

Andrew Brackin:

Yeah. I think maybe something similar to what Jeremy was saying, I think that there’s a lot of bearishness, right now, broadly just in the software landscape. Like, I think most investors are very afraid of the change that’s coming, and it does feel like a lot of change is coming. But I think one, one interesting thing that I’ve seen is just how slow labor replacement has been. VCs love talking about, you know, how fast models are improving, which they are, but it’s pretty amazing how little has changed in enterprises, legacy industries so far, and of course, much will change, but it’s just amazing how little has changed.

And so… Yeah, that’s, I think, maybe my contrarian view.

Kathleen Estreich:

Early innings, I guess. Yeah. Jessie.

Jessie Sheff:

Yeah, I’ll throw out more of, like, a macro one, just to take it in a different direction. I actually believe we’re gonna see an even bigger divergence in both the public markets and private markets for companies that are very AI-forward. I think even more likely, this could be an even starker split between public and private markets, where private market valuations continue to really climb while the public markets actually sell off, except for a few key players, and so I think from a macro perspective. we haven’t actually seen the end of, like, a divergence in how deeply people are gonna value AI-forward companies versus companies that haven’t been able to keep up.

Kathleen Estreich:

What about you, Joe?

Joe Dormani:

Maybe, kind of extending from some points that have already been made, Maybe the simple… take here, which might be somewhat contrarian, is that AI investing is, like, significantly harder than maybe investing in, like, SaaS or previous types of, technology waves. I actually think… you know, there are a lot of clear cases where, you know, AI companies are creating substantial value. And that is showing up in the typical types of metrics that we would see, you know, in a SaaS company. You know, usage is high, people are paying for the product, and it’s very clear that, you know, some of these products are delivering value.

I think… what’s a little bit different now is maybe there are more companies, and yeah, answering the question on sustainable differentiation is hard, but I actually think there is a lot of signal around companies that are working well, moving fast, and creating value, particularly in, you know, the enterprise space. The other thing that’s hard is maybe the being disciplined. I think it is rare to see these companies, but I do think When the companies are doing well, the signal around them is pretty loud.

Kathleen Estreich:

Well, we’ve got a broad spectrum here, so I’m excited to dig in, and we’ve got some early-stage investors, some growth-stage investors, so I think we’ll kind of cover the gamut here. Would love to, you know, kick off the first question. Beyond that is, you know, where are… where within the tech stack, and Amir, you mentioned this a little bit in your answer, like. what layer within the tech stack? There’s obviously, like, a lot of investment in infrastructure, there are some, you know, app layer is… is, to be determined in sort of the future there, but, you know, where do you think folks are undervaluing, and where do you think folks are overvaluing, within the stack of AI?

Where do you think people are paying, or overpaying, and where do you think they’re kind of leaving some money on the table, and so maybe, Amir, you could kick this off. I think you guys cover a lot of, kind of, information.

Amir Kabir:

Yeah, so, as I mentioned, I think, going to what Andrew said, I think it’s another discussion point in terms of, we don’t see that really happening on the enterprise side, where, like, labor is better placed, or will be replaced. And I think it goes back to, you know, what I mentioned, is that most enterprises, are not able to replace labor yet because there’s not this risk infrastructure that is available to them to kind of manage and control the AI, right? So that’s one aspect of it, and I think, I feel like, you know. there is a lot of focus on the application layer, the application companies.

I think that’s… that’s one. Where differentiation, I think, is essentially model access plus some orchestration plus some nicer interface. And so, that’s one. I think the other one that is kind of emerging on the infrastructure side is obviously I think chips. I’ve seen a bunch of stuff there where people are really focused on to build the better and faster and, you know. cheaper chips. But again, I think from our perspective, a structurally undervalued layer, and I think it’s emerging faster, over the last, you know, month or so, two months. is kind of integrity or risk infrastructure around autonomous systems, right?

Identity, permission, verification, provenance, observability, evaluation, accountability. And going back what I said earlier, the way I looked at it is that you know, every technological wave meant something abundant and valuable. When we think about when the internet started, you know, data and information became abundant, and search became valuable, and a bunch of companies emerged around that. When cloud became, was with the next wave, you know, compute became abundant, and then security identity became, you know, valuable in a bunch of companies around it.

In the same way I look at it here on the AI front is that intelligence becoming abundant, and, you know, eventually we have better chips, that’s fine, but it’s not going to be, like, a thousand different chips, and we will have a bunch of models, but, you know, most of it is commoditized, but everything around that will become valuable, and that is important.

Kathleen Estreich:

Have you made any bets in that space yet? Or are you guys still looking for some bets?

Amir Kabir:

No, no, no, so we made a bunch of investments into, like, you know, obviously Agentic engineering, Agentic evaluation, verification, you know, AI security, evaluation, from, from AI agents, and so on and so forth. But again, I don’t think… it’s not going to be a winner-take-all. I mean, there’s going to be similar, I feel like, to… when we think about the cyberspace, right? I mean, you know, names that people know are Palo Alto Networks, you know, and a bunch of others, but there’s a bunch of other companies that became huge and were, like, sucked in by other enterprises, similar to what I mentioned before with the internet wave, I mean, when we look at, like, at Cloudflare, back in the days, nobody wanted to invest, because they were like, what is this company?

And they basically, you know, can basically shut off the internet if they want, right? And so that’s how I think about, you know.

Kathleen Estreich:

So we’ve come to learn.

Amir Kabir:

Yes, that’s how I think about companies around the AI wave as well.

Kathleen Estreich:

Cool. Jessie, you mentioned, kind of, the public-private bifurcation. Curious, kind of, you know, where you think there’s the biggest opportunity of investment now, and where you think, kind of, you know, the categories are less investable.

Jessie Sheff:

Yeah, I think that both the inference layer and then right above, sort of in post-training, is still very, very undervalued. If you look at companies like Base 10 and Fireworks, they cannot fulfill the demand that they have. And so I think that we’re, like, in inning one of seeing what those can do, and I think you’ll start to see people moving right from the inference stack, right above to post-training. So, like, most traditional enterprises haven’t moved beyond using, you know, Claude or ChatGPT for a couple questions, but I think we’re gonna move into, like, highly specific post-train models, where every enterprise is going to have their own stack.

So I think we’re really early days of that. That being said, I don’t think it’s the frontier models that are overvalued. I really believe we’re gonna see a Jevins Paradox here across the stack. I’m not sure I’m the right person to lead the next, like, pre-seed round at a billion for a Neolab, so maybe I’ll call that out, but I don’t think that we’re in 1999 or 2001. Like, I truly think this is… the late 1700s with the Industrial Revolution, and we’re kind of inning one here.

Kathleen Estreich:

Got it. Jeremy, you guys invest kind of Series A, Series B. Curious what you’re seeing, you know, what are you seeing as overvalued, undervalued, given the stage that you guys invest, which is a really interesting kind of inflection point for a business.

Jeremy Kaufmann:

Oh my god, I think it’s so interesting, like… I think our industry right now loves ideas, and loves, like, fast growth at $500 million in revenue, and there’s this, like, donut hole. Like, there’s a donut hole in the middle where the business is growing and working, but it’s not yet at $500 million in ARR, because turns out there are not a ton of businesses at $500 million in ARR, and people are acting as if it’s unfundable at $10 million, or $20 million, or $30 million, and I feel like, you know, me as an investor. like, I’m saying, I think the best round on the field right now is the Series C.

Like, or the Series B. Like, there’s… it seems like the middle has been abandoned by the venture investors, because ideas are cool, things that a billion in revenue are cool, nothing else is cool. So, yeah, I think that’s one of the most interesting places in venture right now.

Kathleen Estreich:

Yeah, it’s almost like if you are a team with an idea, you can get a higher valuation than a team with a product in market with.

Jeremy Kaufmann:

It’s Silicon Valley… it’s the Silicon Valley show in real life.

Kathleen Estreich:

I couldn’t watch that show, it was too real, so I think I need to… I’ve been told several times in the past few weeks I need to go back and watch it. But we’ll see. Joe, from your place, like, you guys are kind of a strategic investor into some of these businesses. Where do you see the opportunities? Or are there categories where you’re not even, you know, not even touching at this point?

Joe Dormani:

Yeah, look, I mean, we spent a lot of time thinking about data, orchestration, and, you know, the application layer. Like, I, I still think… You know, model intelligence is continuing to improve, and… you know, inference is super valuable, obviously, but I still think Where the stakes are very high, and it’s important to get answers correct the first time, and you don’t have the luxury of, like. you know, working over a long period of time, having the right data flow through the model, and having that data live in a structured way that is reliable is a really important form, or… Buffer on intelligence, so we put a lot of emphasis on companies that are able to either get access to or produce, you know, the right types of data that is important to their use cases.

And then, yeah, like, the structuring aspect of that, and the routing aspect of it, and making sure that the models and reasoning layers are able to perform the right analysis and, you know, have the right information in a way where it’s referenceable and, you know, creditable, you know, I think is a super important aspect of AI systems, particularly where, you know, you can have a kind of vertical-specific use case, and you’re able to accrue a ton of really hard-to-get context around a use case, which, you know, can create a feedback loop and flywheel to continue improving a product.

And then just the last on applications, like, I still think… I know it’s, like, very easy to create applications now, and it’s very popular to say applications have no value, but, like, at the end of the day, for when people are using software, like, they still have to have, like, an intuitive experience. They have to know what the buttons do, and where to click, and that has to feel good. for users, and maybe we’ll live in a world where everything is agents doing everything and people don’t do anything, but until then, I think product design and application usability are still, like, really important aspects of product market fit.

Kathleen Estreich:

Andrew, are you guys still investing in the app layer, gradient?

Andrew Brackin:

That’s a good question. I don’t know that the market is all that interested in the app layer right now, kind of to Jeremy’s point, like, the seed stage right now, you’re seeing, a lot of interesting stuff happening at, like, you know, in the infra world, you’re seeing a lot of Big ideas, raising crazy rounds, and a lot of funds piling in, and, you know, we’ll take a look at some of those if we’re… if we think it’s an exceptionally amazing team with a novel… you know, super novel technology or proposal, then we’re definitely interested to learn more.

But, it’s definitely a time with a lot of big ideas. people are afraid of defensibility, and so they’re kind of looking for, like, out-there, wacky bets. Just, like, amazing teams coming with crazy ideas that could be really big. And, to Jeremy’s point, a lot of the, like. multi-stage funds would rather plow capital into those types of opportunities, and so I do think you’re seeing you know, fewer of those Series A’s and Series B’s happening in some of the more down-the-fairway, teams, just growing a nice enterprise business in AI. So I think that’s kind of, like, the game that we’re… where we’re at right now.

But…

Kathleen Estreich:

How do you think about, like, the markets that, you know, the Frontier Labs may or, like, may go into or may not go into? Like, how are you, within your firm, kind of thinking about that? like, do you ask yourselves that question with the investments that you’re gonna do? Like, you know, would the big Frontier Labs do this or not? Like, what are… is it shaping, kind of, those categories? It sounds like, like, the more out there categories versus, you know, something that they, you know, I think… coding… stuff is really hard right now, because obviously, like, Codex and Claude are kind of dominating, but there are still the Devins of the world that are, you know, pretty high valuation, so it is possible in some ways, but… harder.

So curious how you, how, you’re thinking about that, of, like, is that a question you ask internally, or to, you know, to the founders?

Andrew Brackin:

Yeah, so I think, the market has swung back into a consensus, I think, that we’re fully on board with, which is, like. enterprises and engineers won’t just use frontier models, and they’re gonna want a post-trainer model, and etc, right? And so I think there’s tons of interesting infrastructure that needs to be built there, that we’re pretty excited about. And so that’s definitely a category we’re looking at, like, all of these different open source projects, infrastructure you could build there. And then… I think on the app side, I personally am just looking for things where wackier combinations of things.

So, like, in 2019, you get back a really good team building a SaaS product that… that technical portion was mattered less, it was more product, sales. I think now, if I’m looking at apps, I’m thinking about, like, is this a pretty… do they have someone technically incredibly capable, and do they have, like, an interesting go-to-market? Maybe there’s a services angle? Like, looking at fewer, just… down the fairway, what you would call SaaS AI things, like AI apps, and more interesting plays where maybe it’s a combination of, like, a regulated industry, a service, some other, you know. maybe a frontier tech team, right?

So I think if you have some combination of those things. There’s a lot of interesting… opportunities, but they’re not as… You know, they’re not, like, simple vertical software businesses where it’s just selling software to a customer.

Kathleen Estreich:

Yeah, the business model part is interesting. I guess, how much are you… maybe this is for you, Jessie, since you do slightly later stuff, like, how do you think about, kind of, pricing, business model, margin? Like, how are you evaluating some of these companies and deals that you’re doing today?

Jessie Sheff:

It’s very different than it used to be. 50% gross margins, and you’d run the other way. No, I mean, from a margin perspective, like, I was kind of raised in the growth equity world from Summit Partners, so, you know, I still have a bit of a queasy reaction when I see negative gross margins. But I think… I think a lot of the market has moved more towards, like, usage-based pricing is good, not bad.

And low gross margins, I think they make sense when they’re less about oh, I’m gonna drop my prices so I can beat out my competitors, and then I’ll raise them later, like, that’s not a good answer, but if it’s, hey, we’re investing a ton in our models and training and the infrastructure, and like, this is what, at scale, our business will look like, like, you can get a lot more comfort. But we’re really starting to see the move to, like, outcome-based pricing, and so I’m sure people will talk about this more, but in terms of, like, hey, we buy this SaaS package for $100,000.

It’s more like, okay, we’re gonna deliver you an outcome, we’re gonna deliver you a service, and it’s actually gonna be way more, because you don’t need to go buy from an integrator and, you know, have a consulting firm set it up for you. Like, we as the company are delivering both the software and the service. And so I think that the TAM for those are just much, much, much larger. And it’s funny to sort of see the wave of implementation used to mean stickiness, and then implementation was bad because it meant that you weren’t just selling software, and your, you know, gross margins were lower because of services, and now we’ve sort of swung back.

So I definitely think, you know, some of the more private equity and style firms have not gotten comfort yet, but Insight has done a good job at being in the middle there.

Kathleen Estreich:

Yeah, it is interesting. What’s old is new again, because I remember, yeah, when I was at Box, it was, like, services revenue, it was, like, very discounted, and now you’re an AI-native services company, and, you know, people are willing to underwrite that at the venture scale, which is, you know, what’s… what’s, to your point, what’s old is new, it just is a matter of, sort of. opinion. How do you make sure, though, that those companies that are kind of delivering those services can get the margin, like, that they don’t become what they’re trying to replace.

And I think that’s, like, kind of the challenge, where you’re always going to kind of start there, because you have to build it out, but, like, most of them end up looking like services companies to start, to then automate a lot of the… back office staff, but do you ever… like, what margin can you expect from that, and how do you make sure that they don’t get stuck in the cycle of just being a services company without the AI native part of it?

Jessie Sheff:

Yeah, I mean, I think so much of it is, are they going in and customizing from day one, from. you know, from the start line for every single customer, and then trying to find repeatability. Like, I think the goal is get your product to 7… to get them 70-80% of the way there, and then have that customization be that final mile. Now, when you’re starting out, it might be closer to 50% of the way there, but, like, there has to be some base of the product that’s actually productized and reusable. And then it’s, okay, for every customer, we’ll customize more and more.

Some of that is going to be specific to the end customer, because that’s what they want, but most of it, hopefully, we can just, like, reintegrate back into our platform. I think companies actually, like, Basis is an example in accounting that’s done this really well, where, like, everything they’re learning from customers in terms of feature set that people are asking for are pretty, like, reusable, and it’s very quickly integrated back into the product, so how those four-deployed engineers are doing the final mile, and how that’s shrinking, I think, is really important. But I personally don’t think it’s, let’s go back this company that’s building from scratch for every single company, and hopefully they productize.

Like, you have to have some sort of a base, and then the final mile.

Kathleen Estreich:

Yeah. What percent, Jeremy, of the companies you guys are investing in are, like, for deployed engineers, kind of going after, you know, doing some bespoke work?

Jeremy Kaufmann:

Yeah, I think recently… probably… maybe… I mean, obviously, it’s higher in the app layer, right? Like, on the infrastructure layer, you don’t think… so, I mean, we’re roughly half apps, half infra. Let’s assume the infra isn’t within the apps. Maybe half, recently half. We did an investment in an AI wealth management platform called Range. You know, I know Andrew is an investor there as well. You know, that’s a classic… I’d use the word AI native services. There’s AI plus human consultants to help people with mid-level net worth who could never afford a fancy financial advisor. or do something there.

We’re in a recruiting marketplace called Powerform, so I’d say half and half on the app.

Kathleen Estreich:

Blood.

Jeremy Kaufmann:

I mean, it’s just so strange, I mean… It’s so… we venture capitalists have forced people now to almost say that they have engineers, even if they.

Kathleen Estreich:

The bingo card of 2020.

Jeremy Kaufmann:

2020.

Kathleen Estreich:

You know what?

Jeremy Kaufmann:

people are using the word, well, yes, of course I have a forward deployed, and then you have to ask yourself, well. wait, should you actually have a forward deployed? Like, if you’re talking to a real live person about their, you know, wealth, you probably want to have a person talking to that. It’s important. There are other cases where we hear, we’ve got a forward-deployed, you know, person doing something, and you’re like, no, that’s just because your product isn’t working. So I think you have to differentiate between the two.

Jessie Sheff:

not have a BS.

Kathleen Estreich:

Have you been…

Jessie Sheff:

ACVs when you have a forward-deployed engineer. I think that’s.

Kathleen Estreich:

Yeah, the business model.

Jessie Sheff:

everyone’s trying to do it, and I have portfolio companies where, like, they sell to the SMB, it’s $15. I’m like, this is not for you. Like, no.

Kathleen Estreich:

Yes. The business model alignment of the product, of all the things, has to make sense to actually, yeah, make money. Joe, you guys do a lot of stuff, obviously, in, like, legal space, like, how do you… how do you think about that, in terms of, like, the forward deployed, not, you know, just product for services? Like, how… where do you guys sit in… in this sort of continuum?

Joe Dormani:

It’s a great question, and it’s something we think about a lot. I think, like, maybe the first thing is… I think it is different for a company that is building a product for a professional services firm to sell to their clients, and an AI native firm that is just going directly to the clients and selling them an outcome, right? Like… If, if you’re building, for, like, a law firm, for example, and you’re building, like, an AI native system for them to help them embed their intellectual property, that could be a really, really sticky solution, right?

And so, you can really justify the cost of a forward-deployed engineer or team of engineers, because you can expect a really long lifetime value for that customer. So, you know, it’s definitely worth that upfront investment to go and do that. When you’re… the AI… and we’ve invested in AI-native… several types of AI-native services firms. I think the bet on customization is a little bit different. Like, you need to be able to… justify… a long-term relationship in another way. You need to be able to build things that are much more repeatable, that enable you to create the work product that maybe a traditional services firm would create.

You have to have equivalent quality, but you should have a structural advantage with how you could package and price that product. Because you’re able to take advantage of the full potential of AI, whereas maybe a traditional firm might be encumbered by the kind of historical structure of, like, a partnership model. Where all of the profits are distributed at the end of every year, and they don’t, necessarily reinvest, and so therefore, you know, like, they can’t really reduce prices because their practices have to grow, and that’s, you know, how you get to the next level in these firms, right?

So I think it’s… it’s different for those two types of, two types of models, but I think, you know, each one has merit, and, you know, we think there are probably some aspects of each portion of that model that need to be included in, what it kind of really means to serve a market from, like, a kind of end-to-end AI standpoint.

Kathleen Estreich:

Yeah, the legal tech space is interesting, because there are those, you know, products that are selling into the big law firms as the software layer, and then, as I’m sure you all have seen, there’s a bunch of AI-native law firm for, like, something very specific, and it’s interesting to see, because there’s a lot of historical context in terms of the business model, and Jeremy and I are both investors in a company called GCAI that is, you know, going after in-house teams, which I think is less of a business model disruption, because in-house counsel isn’t the partnership model, so just… but then if you look at the TAM for legal, you know, there’s the software, and then the services is just so much bigger than anything else.

So, to Jessie’s point of, like. the TAM expansion in a lot of these verticals is so much bigger than I think was possible a few years ago. So, you know, we’re in the early innings, I think. I think legal is further along than some others, but I’m curious, like, what other markets are you looking at that have kind of similar dynamics, where you’re seeing, you know, some interesting stuff happening? Are there any areas that you’re, like, particularly interested in in some of these kind of larger markets?

Jeremy Kaufmann:

You’re asking me, Kathleen?

Kathleen Estreich:

Sure. Go for it.

Jeremy Kaufmann:

Oh, I wasn’t trying…

Kathleen Estreich:

in the group! I’m sorry, I didn’t direct it at anyone. Go ahead. Let’s hear from you, and then it looks like Amir has something to say after you.

Jeremy Kaufmann:

Yeah, I mean, I guess, like, at the app layer, I think usually the pattern is There’s some new characteristic that the next generation of the model enables, and then because that next generation of the model has enabled something. then suddenly some set of work becomes doable, and then some set of companies, you know, at the app layer benefit from that new ability. So, law moved first because co-pilots were good enough. now there’s newer categories emerging where you… a co-pilot wasn’t really that helpful, but you needed a longer-running agent.

So, look, I… I think there’s more complex flows, clearly, in financial services, in biology, like, I just… I can’t imagine… I know the app layer is a little out of favor right now, but it’s just hard for me to imagine as the next improvement comes out, you know, the next generation of models we see, you know, world labs create new 3D capabilities. You have to imagine as new capabilities maybe in the creative space emerge. There’s gonna be a whole lot more you can do. Like, I don’t have an investment yet in the AI creative space.

I just can’t imagine that over the next year or two, the models don’t get so much better that there’s new, great stuff you can do.

Amir Kabir:

Yeah, I think from my perspective, anything in the regulated markets, is fair game, right? But at the same time, it’s, I don’t think, as easy as people think to break into those regulated markets. I know legal, obviously, is also a regulated market. But I believe there, as Jeremy mentioned, it was more so like a co-pilot situation that is now evolving. But when you, when you think about insurance, banking, healthcare, there is no maybe. There can only be yes or no, because if it’s a maybe, someone might die, and if it’s a maybe, the loan will approve for $10 million instead of $1 million.

And if it’s a maybe, the insurance product maybe never pays out, although it has to pay out, right? So, I think those regulated markets are much more… sensitive, when it comes down to these AI, both application infrastructure layer. And I believe there, actually, if you built the mode, it’s harder to replicate or replace it.

Kathleen Estreich:

Yeah. When you think about, I guess, the subject of this, you know, summit is… is around, kind of, the harness, and I’m curious, you know, every few months, there’s a new hot thing in AI. First, you know, it was the… the kind of AI wrappers, then the agents, and now it’s the harness, that everyone’s talking about. Where do you think companies can differentiate there? Like, do you think that… it’s, like, favoring startups? Is it favoring the enterprises? You know, is this something that enterprises are just going to end up having to do, internally? Can they leverage, you know, external companies to help with some of that, you know, internal… context, permissions, orchestration, logic.

How are you thinking about, kind of, this new era with, with, How that fits into the investment ecosystem. Like…

Andrew Brackin:

I think, you know, Cognition today released some evals on their new model that, as a post-trained version of Kimi, you know, post-trained on that data, and it’s interesting that we’ve been seeing that models seem to do better when paired with a harness and trained, you know, designed for a harness, right? And so that kind of goes counter to some of the belief that we’ll just have, you know, a model, and you’ll just throw everything in, and then the model will magically do everything you need, which could… could happen, we’ll see.

But, I think that we’re maybe now, rather than trending toward, you know, the… the AI wrapper world, we’re seeing more of these, Application layer companies really… act a little more like research labs and leverage their data to offer a better product experience, you know, create models that are cheaper than frontier models. And so, that’s definitely, you know, an exciting opportunity. We’ve seen it work really well in coding. There’s clearly a lot happening in terms of Moving to open models, and so I think that’s… that’s an opportunity for startups, is… it’ll be interesting to see other types of harnesses you can build.

Obviously, Instinct, kind of a consumer harness that’s. taking it off, there are some other general productivity harnesses. I think we’re early in terms of other, you know, use cases and products that people will build.

Kathleen Estreich:

Jessie, any thoughts on this, and from where you sit?

Jessie Sheff:

No, I mean, I think, like, you know, we’re… we’re investors in factory, which I think is a really good example of, kind of, insights thesis in the space. I think, you know, we went from a world where, okay, the frontier models are gonna do everything. We swung back, to… there was a lot of room around them, but now we’re seeing companies like Factory and Blitzy and Cognition try and own more and more of the stack. And so, interestingly enough, 12 months ago, when we were asking the question. What in the app layer, whether it’s law, insurance. finance do the Frontier models want to own?

Now we’re asking the question, what will be, like, subsumed by the harness? Like, what part.

Kathleen Estreich:

to put.

Jessie Sheff:

training, what parts of memory, what parts of context, and so that’s kind of our new question, the new version of what in the app layer the frontier models want to subsume, and so I think time will tell, kind of, what Is important for it to be owned by a third party versus owned by the harness itself, like a factory? And to Amir’s point, I think a lot of the risk and security And, you know, code checking and verification and agent drift and governance and security, I personally think that will all be third party, and you’ll need… and you’ll want that external validation.

But for things in the harness, I think that there’s a lot more room for the harness to grow and own.

Kathleen Estreich:

Do you think that will be owned by one third part… like, one third party or standard, or do you think it will be kind of enterprise-specific? Like, how universal do you think that will be? How big do you think that opportunity is?

Jessie Sheff:

I don’t think it will be enterprise-specific, but I think we will see, like, one company, or, you know, one to three companies emerge that are very strong at, like, governance, and. one company that’s very strong at orchestration, and one to three companies that are very strong at security. I think the companies and platforms that are trying to do, like, governance, orchestration, security, like, it’s just too much. So I think going really deep in one of those, that’s where I start to think we’ll see some real winners emerge when you need that, like, third-party verification on top.

But… Work is changing so fast, so who knows?

Kathleen Estreich:

Amir, any thoughts on that?

Amir Kabir:

Yeah, I think one thing we haven’t maybe touched base on is that you know, for these models to be good and great, data is, like, huge, right? And I think we’re getting to a bottleneck here, too, when it comes down to data. Specifically. We haven’t talked about it, about, around robotics and autonomy, right? But also, in regulated markets, and I was just speaking to, you know, some C-suites on the regulated side, on the insurance side and banking side. you know, whatever data they have, it’s kind of moat, right? And all of these model providers, they work I know that Entropic works with a bunch of insurance companies, you know, OpenAI works with a bunch of banking companies.

It’s basically, you know. they have the keys to their hands. I mean, the model is one thing, but the data portion is the most important to really train the model to the specific use case, and again, specific to these regulated markets. And so business logic and policies and proprietary contexts are really, really important. And so it’s really hard to you know, going back to, like, defensibility, right? I mean, when I talk to a company that wants to sell into these regulated markets, and like, oh, we have, like, you know, better X, Y, and Z, but I’m like, hey, but how do you get to, like, you know, to the next level here?

And they need the business logic and data. I think one, one workaround here to an app, and I made an investment, is into a company that creates, like, synthetic data, synthetic business logic data, right? And that’s fine, and I think that can be helpful, but I still believe, like, you know, the real, the real data is the most valuable.

Kathleen Estreich:

God. Do you believe there… like… who’s gonna win by getting that data? Like, where do you see that kind of data market going? Obviously, there’s, like. you know, a handful of players in that today. Like, are they gonna just… are they gonna become more verticalized data brokers that people are… Are, you know, like, what does the future of, you know, that data capture look like?

Amir Kabir:

I mean, as I mentioned, I think we’re lacking fundamentally data around robotics and haptics and, like, you know, humanoid robots and whatnot. I think that’s a huge market, that is evolving as we speak. And I think an analogy there would be probably Waymo. Because they started, I think, 12 years ago, and collected all of that data to really be able to do what they do today, and I think 12 years ago, people were not looking at that that way, but they needed to start to kind of collect their data to be able to operate these Waymos the way that they do today.

And so… I know that there’s, as you mentioned, there’s a bunch of companies working on data and providing data. But I still believe, like, enterprises who own their own data are the most valuable, right? I know maybe that might be a controversial thing to say, but I think we’re also at the point right now where I think enterprises are more awake when it comes down to AI of this kind of technological wave, and there might be a lot of M&A happening to kind of ingest it into the in-house kind of, you know, framework.

Kathleen Estreich:

Andrew, have you guys made any bets at the, like, the seed stage, or some of these kind of data capture… I know within robotics, like, we looked at a bunch when I was at Pear around, you know, house cleaner, like, the gloves, where they just are capturing data and then, like, selling it to robotics companies. Like, have you guys made any bets there, or how are you kind of thinking about this data capture?

Andrew Brackin:

Yeah, we’ve looked at many. We were actually an early investor in LabelBox, which was kind of founded in the last generation, and is now a fairly scaled business, but, data vendor to a lot of the labs, so, you know. That’s an exciting business. We looked at a lot of the egocentric data companies, the world model companies trying to generate data for physical AI. you know, it’s an exciting category. It will be interesting to see how durable these businesses are. I think It will be interesting to see how durable the day, like, you know. data labeling and RL environment businesses.

We actually do have a couple companies in that space, but it’s kind of an open question, I think, in terms of, like, what data will be valuable to models, model labs. I think Google buying the Spirit Airlines business.

Kathleen Estreich:

Yeah, I thought that…

Andrew Brackin:

Interesting. That’s super.

Kathleen Estreich:

We’re interested in… yeah.

Andrew Brackin:

Yeah, going around.

Amir Kabir:

Go back to data. Goes back to data.

Kathleen Estreich:

They’re going around the data back.

Andrew Brackin:

tremendous.

Kathleen Estreich:

I thought that that was super interesting, because it’s like. I wonder how many… to Jessie’s point of, like, some of these, public companies that are not doing well, I wonder if they just end up, you know, selling their data and getting acquired in so that they become part of an AI-native company. There’s something interesting there, because it is, you know, many decades of data that is hard to replicate.

Andrew Brackin:

Yeah, and then I’m personally spending a ton of time, a little bit, off-topic, but on bio data. I think that’s where Anthropic is openly looking for opportunities, and I think that there’s very little data today that exists, and so that’s kind of an interesting space where I’ve been looking.

Kathleen Estreich:

Joe, as a… like, at Thomson Reuters, you guys obviously have a ton of data, over many decades. Like, how do you guys think about data in terms of what you’re interested in investing in, maybe even acquiring? How do you kind of think about that from… from where you sit?

Joe Dormani:

Yeah, I mean… you know, some of the things I was saying, just to start off the conversation, just about the importance of data and where value accrues, I think You know, there are a lot of different types of data, and different types of data have different relevance, in different types of workflows, so… Like, there’s the sort of authoritative data that is, you know, about a domain, you know, policies or regulations or, like, case law or, you know, rules, standards that exist in regulated markets. Like, this is important data, but it’s, like, insufficient by itself. You know, there’s other data about internally, like, what subject matter experts know. about what they do and how they do it, what they’ve seen before that looks similar to, you know, the work at hand.

This is sort of, like, the tacit knowledge or institutional knowledge that a firm has. Like, you know, there’s a lot of, I think, intellectual property or proprietary aspects of that type of data, and I think you kind of need that to be commingled with the sort of external authoritative data sources to really create you know, kind of a valuable AI system with reliable outputs that can truly, kind of. you know, turn a process, you know, from manual to autonomous, and then, you know, there’s data that can be accumulated from the, you know, the use of a product, or, like, we’re seeing emerging behaviors in AI systems where, you know, agent-to-agent interaction, negotiation, and debate is sort of like this new sort of data that just didn’t exist on Earth before, you know, a couple of years ago, and it’s starting to be used to kind of reinforce the overall quality of the system when good outcomes are generated from agents.

And so, I kind of think, like, you really need the right data from all the different sources for specific use cases, but you also need the right systems and the right technology, the right assets to get the most value out of all those different data sources, put them all together, and, you know, have them flowing through the right set of processes to to really have kind of, like, a scalable, defensible kind of AI system. So, it’s an important thing that we pay a ton of attention to it.

Kathleen Estreich:

Jeremy, you started the conversation talking about, you know, like, moats and these AI companies, and are there any motes? You know, it used to be network effects was a big one. You know, data, we just, you know, spent the last few minutes talking about. Do you think there are still modes and beyond? Or that… and if so, like, what are they, and how have they changed, and what… how do you think they’re going to change in the next year or two? It’s funny, I used to think that, you know, product… product used to be, you know, a… you know, the best product wins, it’s like, no, you actually have to, like, think about distribution and go to market.

Now, you know, it’s just a totally different world where it’s easy to build things, but it’s hard to, you know, get distribution, and then we talked a little bit here about, like, the stickiness within that, like, will… Will your customers stay once you acquire them? And, yeah, just curious kind of how you think about, are there still votes? If so, which ones get you most excited, and which ones are you kind of skeptical of at this moment?

Jeremy Kaufmann:

Yeah, so… so my… my first comment was. I think people focus on that a little bit too… like, I do believe they exist, I just think it’s very hard at the seed or the A to talk about them. Yeah, so, I mean, we have this slide internally, basically saying, look, I think 10 years ago. scale’s largely a software-focused fund, there’s gonna be more businesses that we just call Software Plus. Like, 10 years ago, you would say Software Plus would consist of maybe hardware, maybe it would consist of payments, software plus payments, and vertical SaaS. and then kind of plain software.

One of my partners uses the word plain vanilla soft force, plain vanilla SaaS. So we think just, you know, in this decade. fewer plain vanilla SaaS, more SaaS plus. So, you know, Jessie and I are investors in a company that has a hardware sensor in, like, a care home in the healthcare space, so we’ve done a lot in sensors, sensors in trucks, sensors in, you know. the physical environment, I think the sensor businesses are in an interesting place. Then you’ve got businesses that are kind of learning from the AI interactions, like Kathleen, I know you used to work at, you know, like, Intercom, for example.

That’s a business and customer support where there’s learnings back and forth, you know, Harvey’s learning back and forth, GCAI is learning back and forth. And then, yeah, there’s still gonna be the classic businesses like marketplaces, classic businesses like data network effects, and then classic just out-execution on go-to-market. So, I think that still holds. It’s just, from an investor lens, it’s just super hard to guess that at the seed or.

Kathleen Estreich:

Are you evaluating founders any differently on some of these mentions? Like, I think that, I’ve long believed that, like, founders need to be good at product and think about go-to- or at least have good instincts around go-to-market, and I think, you know, you need someone who can build a great product and is technical and can do that, but also, you can’t be blind on the business side. I’m curious if you have different criteria as you’re evaluating founders that maybe is different today than it was You know, 5 years ago.

Jeremy Kaufmann:

Yeah, I mean, I just think the YC data on average age of a founder by year is super interesting. Like, in the last 3 years, just how much younger the founders are, and I know that’s been, like, a… a vibes-based sentiment, but it was just interesting to really see that in the YC data. And then I think the other trend is just… because people have to scale these businesses faster, you… just, like, in terms of attracting attention, attracting hires, like, you just have this very narrow window of time to execute, and I think that maybe those are just two characteristics that are a bit different.

Kathleen Estreich:

Yeah, that’s interesting. What about you, Andrew? Are you guys using any other additional criteria or different criteria today than you were a few years ago?

Andrew Brackin:

Yeah, I continue to agree with a lot of what Jeremy’s saying, but, I think just going back to my point earlier. fewer vanilla SaaS, vanilla software businesses, and the technical bar for building a software, even though it’s easier to build software, you actually want a technically more capable team than ever. Like, are they gonna be able to out-execute the market, and move in incredibly quickly? And so I think, Yeah, that’s… I mean, I think back in the day, you could back a kind of, like, great sales leader, and, like, a pretty decent engineering team, and maybe outsourced engineering team, that would be fine, but I think There are a few of those kind of, like, simple software businesses to build, and so you probably want Something more novel, more interesting, unique team dynamic.

And probably experience in the market matters less, so, like… You don’t need the, you know. seasoned person, seasoned, like, me in insurance, like, I used to work at New Front, like, I’m less valuable now. I think you can do more with less of a experienced team in the industry.

Kathleen Estreich:

Interesting. Jessie, what are you guys looking at at the growth stage?

Jessie Sheff:

Yeah, I mean, I think we… like, I kind of still have this view around moats, where, like, I’m looking for one type of answer when I’m evaluating a business at the growth stage. Whether that’s some sort of a proprietary data, some sort of a feedback loop, some sort of distribution that gives them, like, an extra edge over other… others in the market. So, for example, like. Filevine is an investment in the legal tech space that we led, and this is a pre-Chat GBT company where they had massive distribution to a ton of lawyers. They were the core case management system for all lawyers.

And we were seeing all of these competitors pop up, you know, Harvey, LaGora, Even Up in particular, you know, doing demand letters. And Filevine was, like, in beta, starting to build these AI tools, and this was 2 years ago when we thought, okay, well, if you had the distribution. you’ll be able to… and you have all of this customer data that you can build your AI tools around, like, you will be able to be successful. I’ve been really surprised that we haven’t seen more of that. And so what Filevine did is they bolted on AI tools on top of their core platform.

Now they have, you know. between 5 and 20, let’s call it, AI solutions that they’re offering to lawyers and upselling at, you know, 2, 3 times the ACV, and it’s been really, really interesting to see, like, the distribution flywheel work. And then we have another business called ExaCare in Healthcare, where they’re doing admissions for senior care, and you’re able to give feedback to the model and say, like, yes and no, and the model gets smarter for every customer, but it also gets smarter for specific customers based on what they will allow in their own environment.

And so, something where there’s, like. The distribution, or the feedback loop because we have the most customers, or in insurance, if you have a bunch of the top carriers willing to share their data so you can train the model on their data, like. That’s sort of an edge on the application side, I’d say is really, really important. Otherwise, it’s been really hard to invest in the application side. So, like, we look for things like that.

Kathleen Estreich:

Yeah, it’s interesting on the… the kind of incumbents who were pre-AI, so many of them are not making that transition, where it’s like, you… it’s yours to lose, because you own the customer, and so everyone… you know, you have, in theory, that relationship, but if they don’t make that hard pivot, I think it’s really hard.

Jessie Sheff:

You know, we’ve debated this so much internally, and I think it’s such an interesting point. I think that, I mean, I can name on one hand how many companies have done this well, Filevine being one of them, and I think it’s because they were, like, you know, not a 20-year-old startup pre-jet… They were… they were more recent, and you’re small enough, and nimble enough, and you see the opportunity, you hire the right engineers, and you just go for it. But, like, if you look at something like a Salesforce, they just released their first actual, like, AI product. with Anthropic, and who knows if that will actually work, like, time will only tell.

We use Salesforce at Insight, and, like, we have logs with founders going back 7 years, and, like, we can’t even summarize them. Like, it’s almost wild how some of the incumbents just haven’t pivoted at all. And so I think that we’re gonna start to see strategics come in really aggressively to start to acquire some of these AI-native startups in order to bolt on to their solution, or be their shiny new toy. But it has been… it has been, like, the biggest shock for me in the past few years that we haven’t seen more Firebinds.

Kathleen Estreich:

Yeah, the Salesforce example is interesting, because they also own Slack, and that has all your internal data, and all these agents are running… it just feels like a…

Jessie Sheff:

Yeah.

Kathleen Estreich:

A missed opportunity, or they’re not moving as quickly as one would hope.

Jessie Sheff:

Yeah.

Andrew Brackin:

I don’t know if you just saw, Meta moved from Google Chat… they moved from their own chat to Google Chat, and then now they’re moving to Slack, because they’re making the bet that Slack will be the best platform for agents. So that’s, again, kind of an interesting, data point in pro app corner.

Kathleen Estreich:

I’m just gonna wrap up whatever.

Andrew Brackin:

Whatever you want to call it.

Kathleen Estreich:

Alright, we’re coming up on, the last 5 minutes, so I want to kind of get to the prediction, portion of this panel, so I have… I have two questions. So this first question, maybe everyone can go fairly fast, and then the second question. What do you believe about AI… what did you believe about AI 12 months ago that you have changed your mind on in the last 12 months?

Andrew Brackin:

I think we all… many of us believe that frontier models would be the answer to, to everything in AI, and now data is looking more interesting, in terms of, Not relying on one model for everything.

Jessie Sheff:

Yeah, mine’s very similar, but to give it a slightly different pivot, because you stole mine, I think… I think it’s just, like, the prevalence of open source, like, actually how quickly that market has grown, and how… you know, a year ago, you would not have thought it would be obvious at all that people would really trust open source, and now I think that that’s the direction the market’s moving in.

Kathleen Estreich:

Amir?

Amir Kabir:

I think for me it was similar, like, I think many believe that these model providers have an edge that is much bigger, and probably much stronger. But the way, as Jessie mentioned, too, open source has caught up, I think, people didn’t believe in, and I didn’t believe in, and I think, another… thing that happened, too, is that, you know, countries that people never thought about came out with models that aren’t par, like Japan is one that came from the left field with the open source model that is really strong. So, I think that’s, Something that people thought that would hold on longer, and maybe that’s the reason why, you know… said companies want to go IPO now because, you know, open source is catching up pretty fast.

Kathleen Estreich:

Jeremy?

Jeremy Kaufmann:

Just continue to be amazed by how quickly Consens… consensus and just general thoughts on things change by the week. It’s just… it’s truly… I mean, we’ve seen it for the last two years, we saw it this year, and I’ll keep on seeing it.

Kathleen Estreich:

Joe.

Joe Dormani:

Maybe bits and pieces of what others have said, which is… and I go back and forth on this, so maybe it’ll change, but vertical AI is an interesting venture investment, and should be something we spend time on. Maybe a year ago, I was a little bit more negative, and today I’m back on the train.

Kathleen Estreich:

What… what changed in the last year?

Joe Dormani:

I, I, I think… you know, I think there’s been a lot of proof points around where… company… where AI companies can build differentiation, more confidence in, you know, the different types of data that matters and how it matters in use cases. you know, the ability for software… for these AI companies to build more comprehensive end-to-end workflows, which are, you know, very specific to the way Different industries work, and, like. just the… the amount of things on OpenAI and Anthropic’s plate that they’re gonna have to do to remain competitive, like, it’s just gonna be very hard for them to compete in, sort of, hand-to-hand combat versus, like, someone who’s just very, very focused in some of these… some of these verticals.

Kathleen Estreich:

Cool. Well, we took too much time on that one, so we’re not… we don’t have time for the second one, so we’ll have to stay tuned. But, it was awesome having all of you on the panel, and I… I learned a lot, and it was very interesting to hear the various perspectives and various stages that you all invest, so thanks for your candid insights and for sharing them with, everyone who attended.

Amir Kabir:

Thanks for moderating, that was great, thanks for having us.

Julia Nimchinski:

Phenomenal, panel, thank you so much. Thank you, Kathleen. What’s the best way to support you?

Kathleen Estreich:

Me? I don’t know, build cool, cool companies. I don’t know. If you’re working on something cool, we’d love to hear about it.

Julia Nimchinski:

How about yourself, Amir and Joe?

Amir Kabir:

Same, always open for, like, you know, new, interesting founders that are… Have novel thoughts and are thinking about, you know, what the future might look like.

Joe Dormani:

Yeah, always open to a good conversation. Feel free to reach out anytime.