Jacob Effron of Redpoint joins Nick to discuss How Model Progress Shifts the Goalposts, Why The Death of Software Is Overstated, and How to Diligence Hypergrowth Without Getting Burned. In this episode we cover:
- Investing in AI and Vertical Applications
- Model Layer Advancements and Future Milestones
- Challenges and Opportunities in Agentic AI
- Investing in Tooling and Middleware
- Product Market Fit and Defensibility in AI Applications
- Verticals with Real Product Market Fit
- The Evolution of AI Investing Metrics
- Future Trends in AI and Robotics
Guest Links:
The host of The Full Ratchet is Nick Moran of New Stack Ventures, a venture capital firm committed to investing in founders outside of the Bay Area.
We’re proud to partner with Ramp, the modern finance automation platform. Book a demo and get $150—no strings attached.
Want to keep up to date with The Full Ratchet? Follow us on social. You can learn more about New Stack Ventures by visiting our LinkedIn and Twitter.
0:17
Jacob Effron joins us today from New York City. He’s an MD at Redpoint, an early- and mid-stage firm focused on AI, enterprise software, and healthcare. Before Redpoint, Jacob spent time at Flatiron Health, Off Grid Electric, and McKinsey. He has invested in companies including Abridge, Legora, Augment, Physical Intelligence, Ramp, and Garner. Jacob, welcome to the show!
0:41
Thanks so much for having me. Nick, really excited to be here. Yeah, excited to have
0:45
you. So tell us a bit about your backstory and your path to becoming a VC, yeah.
0:50
Well, I think you got the kind of basics there. You know, essentially started my career working at a few startups post consulting. Was over at Flatiron health on the product side, flatiron got acquired, and I was thinking about what I wanted to do with my life post acquisition. Had always been curious about venture and had had exposure to a bunch of different VCs and tech companies through that startup work, and had the good fortune of meeting the red point folks ended up joining five and a half years ago as a vice president initially, and it’s been a pretty amazing ride since then. Awesome. And
1:19
tell us more about the thesis at red point. We’ve featured some folks in the past, but would love to hear the update.
1:24
Yeah. So we have two funds at red point, an early stage fund, and then what we call our early growth fund. I sit on the early growth side. It sounds like an oxymoron, but what it basically means is Series B is kind of the majority of what we do, and the way we think of it is we want to be investing post product market fit, but at the earliest signs of inflection possible. So the goal is to be a half step before something becomes obvious in the numbers, we’re obviously looking for iconic, independent companies that will be the big names of tomorrow. You know, over the years, we’ve led rounds in companies like snowflake and Stripe and Twilio and HashiCorp, bridge, agora, some of the other ones you mentioned, but we’re pretty broad in the in the mandate we have, and trying to find those n of one special companies before they’re obvious to the market.
2:05
And how about, you know, approaching this new wave of AI? How are you thinking about that, and where are you putting most of your efforts?
2:14
Yeah, it’s certainly. The fun thing about AI is that it changes so quickly that I always joke we have, you know, a really strong opinion on that. And then, you know, if we do this podcast again in three months, maybe we’ll have changed 50% of that. And that’s the that’s the fun of investing in this era. I’d say, right now we focus kind of in two different areas. One is, we’ve been super excited about vertical AI applications and the ability to bring the capabilities of these models to a bunch of end industries like healthcare and legal and logistics. And I think there’s these companies that are basically partnering with the most important companies in those spaces and helping them figure out, how do we implement and adopt these models, and really being that translation and transformation layer between the capabilities of these models and what they can do in these end industries. And so we’ve invested a bunch in those spaces, and I think we’ll do a lot more. The second category, I’d say, is we’re really excited about the, you know, the kind of massive consumer and model bets, and just, you know, I think there’s obviously going to be tremendous transformation on, you know, consumer tools on the hardware side, and robotics, biology, and so we also spend a lot of time thinking about these huge categories where we think there will be, you know, Maybe $100 billion company birthed in those spaces. But it’s still really early and ambiguous just exactly how those spaces play out. And I think, you know, those have been kind of the two areas where we’ve been spending the most time as a fund.
3:30
So, so let’s talk models for a minute. Jacob, yeah, lots of news lately at the model layer, some of which
3:36
was, you know, every week there’s some exciting news at the model layer, right there is,
3:41
and it’s funny, because, like you said before, if you rewind three months or six months, the the talking points are completely different. But so Google DeepMind is claiming historic breakthroughs with Gemini two, five, open AI’s launch of GPT five. You know, mixed fanfare around that, anthropic release, Claude Opus 4.1, with improved agentic capabilities. Lots happening in the space. We’ve seen some leaps, leaps forward. We’ve seen incremental improvements at the model layer. I’m curious, you know, in your estimation, what are the next major milestones for models?
4:15
Yeah, it’s a really interesting question, and I think I’d answer it in a few different ways. One is, you know, with the benchmarks that we’ve looked at before, clearly where the models have been amazing at hill climbing on in recent months is coding and math and kind of these easily verifiable domains that really lend themselves to reinforcement learning. And so I think there’s incredible milestones that we’ll see on the coding side. As these models continue to get better and better, I think the holy grail that folks talk about is, when do these models get to the point where they’re accelerating the work of AI researchers themselves a certain amount, and so that, I think, is probably the most interesting milestone, but certainly continued improvement in coding has a tremendous amount of applicability in real world use cases. The second area is kind of the broader swath of things that ever. One is using models for and I think what the general feeling in the ecosystem is is these benchmarks a have been saturated and B are only so helpful for like real world use cases, like hill climbing on some of these benchmarks doesn’t necessarily make my chat GPT query feel that much better, or make the ability of a doctor to answer a question or a lawyer to answer a question that much better. And so in many ways, the most interesting things for models to improve on are some of these evals that exist within the application companies themselves, right, like the legal AI companies. Know, the most important things for the models to get better on are the logistics AI companies. And so I think we’ll see models continue to make progress in all these domains. But it might not feel as clean as those graphs of the past, where it was like, you know, there was a standard set of benchmarks that were universal across everything a model could do, and it just felt like it was going up. I think it’ll look, you know, a combination of the fact that, one, these benchmarks look very different now across the domains where these models are applied, and two, it seems like in the past you would get all these general improvements as the models got better. But now it seems like you can have a model like GPT five that is definitely better on coding, but maybe not noticeably better on your average chat GPT query. And so I think the things I pay attention to are the overall coding benchmarks, and then within the specific domains where we’re seeing AI product market fit, how much better do the models get? Do
6:20
you think we’ll see big advances in agent AI over, over the coming year? Let’s say, you know, I use various tools for what we do here, and it feels like we’re, we’re still chicken, wiring and stitching things together to actually get workflows done. It’s, it’s not super efficient yet?
6:40
No, it’s, it’s the right question. I think, you know. And obviously I should have mentioned another axis that, you know, people are evaluating model progress on. It’s just the length of tasks that you can go, send a model to, go, to go off and do, and how reliable those tasks are. I think we’ll inevitably make progress on that. The question is, you know, to the tune of, to the tune of, what right? I think, I think that models are getting much better in the labs are super focused on these longer term tasks and around setting up specific environments to do reinforcement learning on such that these tasks can be executed, the extent to which, how easy that ends up being, how generalizable that ends up being, I think will determine, you know, the extent to which that diffuses into society. But I certainly expect in this next year, a lot of those longer term tasks and kind of end to end agents to get a lot better. And in many ways, they honestly could get a lot better. Even if the model stayed the same, we’ll probably keep building better scaffolding around them to, you know, allow more and more, right? I mean, what I always remind myself is, we’ve only had these latest cutting edge models for a short period of time, right? And there’s, there’s going to be endless experimentation with just these capability of models for quite a long period. Yeah, 100%
7:46
What’s your take on Chinese model companies? And you know, comparing that with us incumbents? How big of a threat do you view China’s efforts as? Yeah.
7:58
I mean, the open source ecosystem in China is incredibly impressive. I think the top of a ton of leaderboards really, you know, some amazing models from deep sea quem and others. I think that what it’s caused, the most cutting edge stuff, is still happening in the US, but what it has caused is the better open source models are, the more pricing pressure there is on the, you know, leading foundation model companies, right? And there’s basically, you know, there’s always going to be some set of tasks that are the most cutting edge that you need the best model to do. And on the other extreme, there’ll be the most basic tasks that the models three years ago could have done. And those you can use any model for. And as open source models get better and better, the distribution of tasks that fall into the most complex first, hey, you can use an open source model for more and more fall into the open source side. And so I think, you know, at a time where we’ll see what happens with some of the Western open source models, there’s folks doing great work at meta and Mistral and other places. But for now, the you know, the best alternative to using anthropic, Google, open AI is these Chinese open source models. How that plays out in practice? You know, I don’t know whether, at the end of the day, Western companies will feel comfortable using a Chinese open source model, even if they can download the weights and, you know, run it in their own environment. And so it’ll be interesting to see. You know, right now, in terms of playing around with things, lots of folks are willing to play around with these models. I don’t know whether it’ll ultimately kind of permeate into large enterprise deployments.
9:21
I mean, does it become very sovereign and geographic in nature? Do you think over time? I think
9:28
this is definitely a trend we’re seeing, right? I think, you know, AI is so strategic and on some of these, you know, scenarios of continued model improvement, you know, pretty crucial to GDP growth, to competitive advantages of countries. But you’re seeing a bunch of folks say, Well, wait a second, I’m not sure I want to just outsource all of this incredible model development to just a few countries. And so, you know, I think if, if this were a completely non strategic space, you may see a few model providers really win and and be at the forefront. It’s very clear that there’s a ton of players. Both. Of, you know, governments, but also infrastructure players like Nvidia, that are highly incentivized for there to be many model providers that breaks down by, you know, by different geographic areas or other things. And so I think given those incentives, we’ll probably see a fair amount of model providers. Do
10:15
you think we hit certain constraints on the infrastructure, infrastructure side, you know, with the dispersion of compute demand and all the models that exists, are we hitting the upper end of sort of capacity,
10:30
at the edge in terms of, like, chips and energy data,
10:34
yeah, chips converted to data centers and compute, Yep, yeah. I mean,
10:38
obviously there’s, there’s, there seem to be bigger and bigger compute commitments made across the board. And so I think we’re still we’re still scaling, and I guess we’ll know a lot more in the next few years about just how long this scaling on reinforcement learning takes us. I think you know to the extent that there’s continued massive economic improvement and gain from it, capitalism has a great way of continuing to provide and so obviously, folks far smarter than I have opined on the energy limits of doing some of these things at scale. I think the question before the energy limits is just, how far does this current scaling wave get us right? And I think we’ll learn a lot about that in the next year or two, on the energy side. I think that is, that is the big question, but it’s really a second order question beyond you know, is it, is it going to be economically viable to make those really large investments? And I think we’ll see a lot more these next 1224, months, as people scale that up.
11:29
Interesting. You know, before we move on from agents, we touched on that a bit ago, do you have like a framework for how you think about agentic AI and what tooling agents will need?
11:40
Yeah, again, this is one that is constantly evolving. The challenge of it, right is there’s only so many companies that have have built agents and deployed them today, and so I think we’re still very early in understanding what the scaffolding is that’s required around these models. I guess I think about it in maybe three different categories. The first would be, the more kind of context and data you can feed into agents, the better they perform. And so that looks like connecting internal data sources, connecting external data, like the web, being able to search and add that data to agents. Basically, the more context you can feed into models, the better they do. And so I think there’s a huge amount of tooling on the data side. The second is, it seems pretty clear that the way a lot of these agents will run is in parallel, and you’ll be having, like, you know, hundreds of agents going off and doing tasks for you at a given time and coordinating among themselves. And so there’s clearly a massive orchestration and, like, coordination need that’s required for that. And then the third is, once you have these things in production, there’s clearly a need for observability monitoring. Where do these things break? Why do they break? How do you go about fix that and make that better? In the in the next ecosystem? So it’s a really like nascent ecosystem right now, but I think the key is just figuring out what kinds of those needs are going to be persistent. You know, as as an agent in a year is going to look very different than an agent today, than an agent in two years. But of all that stuff, I said, I think the key is investors, is which parts of that will will last, and obviously those categories intuitively make sense. You will need to be able to observe traces of your agents and figure out where they break. You’ll need to be able to connect them to data. You’ll need to be able to do some sort of orchestration. What the specifics of that end up being, it’s hard to say without more deployments of agents at scale.
13:16
You mentioned this the scaffolding, and you know, it applies to each layer of AI. You know, I found it difficult to evaluate tooling and middleware horizontal AI, as everything is evolving so quickly. So, so how do you approach kind of the tooling related investments in companies being built? Yeah,
13:39
I think, you know, I think you’re totally right that basically things change so fast. And so, you know, one of the the main things you can bet on is just the velocity of the team, right, and the ability to respond to whatever the like latest way of doing things are. And so that, I think ends up being a paramount importance when we look at some of these companies, you know, that being said, I think there’s, there’s a few other things that we look for. The first is, like I was saying earlier, a need that feels persistent, that you’re like, okay, even if the models get 10 times better, they will still need to access internet data, or they’ll still need to be some way to monitor why these models broke in some way. And so what needs are kind of persistent across time, and what needs are maybe just temporary today, based on what the models are, you know, have as limitations. And then the second thing that we look for is you ultimately want to land with these customers today and then kind of grow with them over time. And so what are the needs that customers have today? I think it can be more challenging to predict what needs customers will have in the future that they don’t have today, and then sit in an ivory tower and not kind of deploy it with customers. And learn from those deployments. And so I think we always look for, you know, ways to land with these customers and kind of grow with them and evolve with them
14:51
over time. Kind of takes the idea maze concept to a whole new level. Yeah, it’s,
14:55
you’re constantly The fun thing about AI, and, you know, operating investing is every three six months, is. Sort of existential event for the company, right? You have to reinvent yourself in some way. The teams that do it well are absolutely incredible, but it’s certainly challenging,
15:08
amazing. It’s like the book hard thing about hard things Ben Horowitz kind of condensed into three month increments,
15:15
exactly. I would say it’s like dog years in AI world, like what would be normally in a seven year cycle in a normal company happens in like a year in an AI company. So,
15:24
Jacob, you mentioned bridge agora. You’ve done some investments at the the application layer. Tell us more about how you’re approaching and thinking about investing there.
15:35
Yeah. I mean, lots of exciting companies right now. You know, a really interesting time for applications. I think a few things we think about on the application side, first and foremost, will come as no surprise, but it’s like a really effective wedge into these companies. And they say, you know, product market fit when you see it, but the way some of these businesses are scaling, you can really tell what’s an effective wedge. That’s really a 10 times better experience with AI and something that will really be adopted at scale. Versus, you know, maybe something people are experimenting with, but not absolutely in love. And I think in some time, in some ways, the best ability to track that down is talking to the end users themselves. You know, what do doctors or lawyers really love using? Versus, you know, I don’t know, some innovation person thought it would be a good idea to to adopt. And so we’re, we’re always focused on that, like, really effective wedge, and actually getting usage in these companies. The second thing we then think about is, there’s, there’s a lot of great wedges in AI, which ones lend themselves to a breath, you know, broad, breadth of product, and the ability to kind of continue building product for, I don’t know, decades and decades, right? There’s some wedges that are that are great wedges, but then you built the wedge and it’s like, okay, well, there’s not that much else to do. And in many ways, those should probably be kind of features of other platforms, right? Whereas there are some wedges where you’re like, God, if you could have a compelling AI tool like this in the hands of every doctor and lawyer. I mean, the possibilities are endless of things you could do on top. And so I think that that gets us really excited. And then the third bucket, which is everyone’s favorite, everyone’s favorite topic, is defensibility. We don’t, in the early days, think, okay, there has to be some insane moat in a series a company, but I think you want to see a path over time to what that defensibility might look like. And a lot of times that’s connecting different stakeholders in an ecosystem, or, you know, interesting partnerships you’re able to leverage, or just the fact that, you know, funding and velocity kind of compounds, and you’re able to build out the broadest product. And so I think those, those three things really dictate where we spend most of our time at the application layer. Is
17:33
there kind of a data set of characteristics or a shape to the data strategy that kind of informs that path to defensibility over time. Yeah, you
17:45
know, I’ve gone back and forth on this a lot. I think in the early days of AI, everybody wanted there to be some really clean story of, like, well, the only person that has this data will then be the only person that can train this model. And, you know, that’s, that’s a huge advantage. I don’t think things have played out that way. I mean, I do think people are doing reinforcement learning on a bunch of data they have. I think everything I’ve ever seen on fine tuning in RL is you don’t need that much data to make it work. And so I think certainly yes, you want to have some data to make your to fine tune smaller models that are cheaper and faster to do some more cutting edge stuff, but it’s not like one person who has, you know, most of the data automatically wins. I do think what these companies do incredibly well is they’re really smart about their data strategy. So what does that look like? It’s being really smart about what are the evals that matter in a given domain, what really makes a product delightful to use versus not, and how do you then orchestrate and architect and whatever the set of tools are, whether it’s building your own model or bringing some together, or adding a model call at the end of something, finding a way to really Hill Climb on these evals and provide a much better product experience. I think that probably will determine more which of these companies were successful than sitting on top of some data that no one else has love
19:05
it so still at the application layer, what? What areas in your estimation have real product market fit versus just experimental revenue?
19:14
Yeah, I think the the killer ones have been, I mean, coding has been amazing, obviously incredible tools, customer support, also really exciting. I think healthcare, legal, we’ve seen incredible growth. Voice has been really, really impressive. I think these models have gotten a lot better. In zones, in categories like logistics, you’re starting to see just incredible value from these models. You know, in other areas, they’re still they’re still early, right? I think go to market has probably been one of those categories where I’d say we’re still early. And, you know, maybe need another step change of model improvements to really drive a ton of value. There. Got it.
19:48
And what are the verticals? I mean, you mentioned legal, you mentioned you did Shiv and abridge in healthcare. And are there other verticals that you think are going to become high. Over the next three years or so, and why?
20:03
Yeah, it’s a good, a good question, maybe to go back on what I liked about healthcare and legal. I think those, you know, those are really great markets, not first and foremost, they have a great wedge and great ROI today for these customers. But also they’re, you know, categories in general, where you can see quality really matters, right? You don’t want to go to the hospital that’s like, Hey, we’re using the like, 50% is good solution. But don’t worry, it’s like 10% of the cost. Like, you’re gonna be like, what? Or, you know, you and I at venture funds, we don’t want to be using the law firms that are like, we’re using the mediocre AI law tool. But don’t worry, it’s cheap for us. That’s not true of every category, right? And so I think there’s, it’s a space where quality matters. And I think, you know, there’s obviously access issues that exist in both healthcare and law, and you can imagine, over time, AI really playing a role in in those categories. I’d say
20:53
to that end, Jacob, is it possible that lower stakes segments are gonna potentially adopt AI quickly in the future, because, you know, if agents are really good at making low stakes decisions and the recourse for making the wrong decision isn’t that high, then
21:11
they’ll adopt faster. Yeah, I think about this all the time, and I think there’s totally an argument that’s the case. And in many ways, maybe customer support, you know, is the space that some people see as lower stakes. I think, you know, with every good thing comes maybe not so good thing. The challenge of some of these lower stakes areas is people may end up being more price sensitive, right? And there may be, you know, someone that comes in as the lower cost vendor, and it’s okay to take the one that’s, you know, 60% is good, or 70% is good, and it might not, you know, it might not be. And I think what you’ve seen those companies do is maybe you land with the lower stakes, you know, just like FAQ docs, but you expand to things that are much higher value. And I think those, those companies definitely are doing a good job of that. But I think, you know, in many ways, one of the challenges is some of the easiest stuff to, yeah, to insert AI into, is stuff that was already outsourced or wasn’t cared about. You know, too much. And again, can be a great insertion point to go build other things. But if you just stay doing that, I think there’s a real risk of, you know, massive price competition.
22:10
So Jacob, I don’t know if you heard but t2, d3 is dead.
22:14
I did. I heard him on say that on 20 BC, if anyone has some t2, d3 businesses that they want to shoot my way because it’s dead everywhere else, I’m certainly happy to spend a lot of time
22:25
me as well, right? So, triple twice. Double three times has been kind of the heuristic, right? You want to see an early stage business triple the first two years, double the next three years. And if you see that profile, you know there’s a lot of venture dollars chasing it. But now, like Hamath said, you know, 100 million ARR in less than three years is now the new goal. I mean, is this just hyperbole, or do you think this is a realistic new, new benchmark?
22:54
You know, it’s probably somewhere in the middle. The reality is, those are still great businesses. And, you know, are Emily fundable the you know, the reality, though, too, is we’ve seen some of these AI businesses grow so fast that I think it’s reasonable to say there’s just such so much gravity in the market, like pulling these AI solutions in. And if you have something with product market fit, there’s so many people that want to use it that it might be reasonable to say, given that we’ve seen what real AI, product market fit is something, you know, that isn’t growing at those insane rates, has some product market fit, but maybe not the same scale that we’ve seen elsewhere. And I think, you know, maybe what Haman is getting at, which I think is right, is in the past, you know, the gold standard of product market fit was the, you know, t2, d3, right? And that was an indication that you have as good product market fit as you’re going to find. I think if you’re building an AI app, and you have that, now you have a good business, but it’s very, you know, it’s very positive. It’s not like you have found maybe the most killer wedge or the best market to adopt some of these things. And so I think in the past, you might have done a business that was growing, you know, the way some of those businesses were just on the metrics and, you know, you’re like, those metrics are great. So some people may have invested in it, because it’s like, well, this is, you know, top, top quality product market fit. I think today you have to be a bit more thoughtful about, you know, okay, maybe it’s not completely flying off the shelf the way some of these other businesses are. Do we think in the future that will change? Like, is it, you know, is this industry poised to adopt AI, maybe just a bit later, maybe as the models get better, or diffusion takes longer, is this team building a product that today kind of works but will work even better in the future? There’s certainly plenty of reasons to get very excited about investing in businesses that don’t have the crazy growth profile, but I think Haman is right there. They’re not the deepest product market fit in AI, if they’re growing that
24:43
way today. Well, to some degree, I feel like there’s, there’s a big risk investing large amounts of money in super fast growers without having, you know, some sort of track record of retention, customer usage. I mean, don’t get me wrong. Long this growth is is outstanding, right? But there are companies like, like Wiz that grew really fast, but the retention is pretty high. It’s really sticky. You know, great, great product. And then there’s like consumer examples, like cameo, where the growth was outstanding, and then, you know, fell off a bit. So these are things that may or may not be be able to be diligenced, but when you have that at fast of a ramp, I’ve just seen a lot of thin layer AI get adopted quickly, and then the companies collapse. Yeah, that’s
25:33
why I think those, those questions we were talking about earlier, about the kind of overall category are so important. But, you know, I actually think trying to answer those questions is the fun of it, right, in terms of where we play, you know, let’s say we play the businesses out two, three years, and they’re growing super fast, and the retention looks amazing. Those businesses are going to price perfectly, right, like they’re there. And so I think part of the fun of what we try to do is, yeah, like, what are all the leading indicators and all the people you can talk to, and all the stones you can turn over to get a sense of, you know, what of this stuff is flash in the pan, and what of this stuff really seems recurring, but I’ve never had more fun doing the job. I think that’s like such a fun challenge to go do and try and figure out
26:12
that is the job,
26:14
right? Exactly. That’s
26:16
where you get your alpha, okay? So, you know, I noticed this tweet from David haber of Andreessen. He said, In the past, many vertical software companies were started by domain experts who added technical capacity later. Now it’s the inverse. Ai native startups are led by technical founders who hire for context early your take. Agree or disagree with David, I
26:37
don’t think I would have worded it as strongly. Obviously, there’s, there’s a ton of, I think this archetype is very in vogue right now of the really high velocity, kind of cracked young team. And there’s, and there are some amazing teams doing this out there. I back some of them. They’re, they’re phenomenal, but I think that there’s some great domain experts building these spaces as well. And you know, the way I think about it is, there’s a big question with all these businesses. Of you found this wedge. You get to some level of scale, and then what do you do next? Right? You know, ultimately, with all the excitement around investing in AI, you really have to believe that these businesses get to, you know, hundreds of millions of dollars and then have a great growth story from there, right? It can’t just be like, cool, we got there. Mission accomplished. Like, those businesses aren’t worth nearly as much as the as the prices that are being put on them. And so I think one advantage that folks with domain expertise have as CEOs is an intuitive notion of what to go build from, okay, you’ve gotten to scale. What are all the ways you can bring the ecosystem together and the different players and align incentives and start to build both defensibility and like next acts around these companies. So I think both archetypes are going to be tremendously successful. But as I think, as more of these next innings play out, it may look a little bit more, you know, 5050, than it looks, you know, 95, five
27:55
today. It’s funny, because we did this analysis two years ago or something, all the unicorns that were minted in the breakdown of domain versus, you know, tech first. Oh, so you have
28:05
the actual data. I’m just pontificating. So what’s the actual data set? Well, it was
28:09
5050, funny, right? Well, there we go. But it changed it that was in terms of successes, but on origination, so like company formation and investment, it was much more skewed, even back in the SAS days to technical founders, I think we looked at YC and like nine out of 10 YC companies were tech founders entering a domain they had no experience in. So almost the premise of his point, I don’t know that it’s accurate, because he’s contending that in vertical software, the common knowledge was to back a domain expert. And I don’t have that
28:41
much, but I do think he’s on to something. One thing I’ve been shocked by in AI is just, you know, in the past, you’d have these companies that were started in healthcare or law and, you know, it would be really, really hard to get in a room with like, the head of one of these hospitals or law firms. And now, I think, just given how everyone’s trying to react to this. And given some of the great products people are building, it is definitely way easier for founders that are newers to a space, if they’re really sharp and have a great product to get in some of these rooms that, in the past were just reserved for the subject matter experts, and
29:13
so that’s been in C suites are getting that pressure to totally you know, how are you incorporating AI and and I think they’re often looking for answers. I mean, some, you know, have their own frameworks and their own strategy around it, but there, there are many that are, are looking for the tech folks to kind of show them what that roadmap could look like.
29:37
Yeah, I think looking for answers today, and also a partner on the like 10 year journey that is diffusing this stuff in society and reacting to every you know, improvements and changes in capabilities.
29:48
So Jacob, if we zoom out a bit in and look back in five years, five years from now, we look back, what’s one big AI trend you think the industry is wild, wildly over. Estimating, and what’s one that you think is is underestimated?
30:04
It’s funny to take this one. I asked this on I run an AI podcast, and I always ask this in my questions. And the nice thing about running one, you know, myself, is I never have to actually answer it. So I’ll try and do it okay this time. But, you know, I think on the overestimating side, I mean, there’s so much talk about, just like the death of software and the idea that everybody will be able to generate their own UIs on the fly and ways of doing things, I feel like that’s pretty overestimated. You know, obviously, making software is one part of the battle, but maintaining it and coordinating it across a bunch of different people that use it, and having continuity as people move from, you know, one job to another and common ways of doing things. You know, I think the existing software providers, they’ll obviously be changes, and AI native incumbents, or AI native startups, will disrupt some incumbents. But I think there will, you know, there will be dominant software players in a lot of these industries. It would be my take under hyped or underestimated. I think it’s the progress in robotics right now, obviously, like, you know, the llms and image and video are much more visceral. And I think robotics is going on this progress line. That’s not, you know, it’s not like all of us are interacting with robots every day, but the research progress in the last year or two, as well as the feeling from folks on the inside, like, Whoa, it feels like this is starting to be different, you know, over a five year time horizon. And obviously, I hope this is true, because we’re investors in physical intelligence, but I hope we look back and say, wow, it was kind of in retrospect, like what people were feeling between GPT, you know, around GPT two time like, wow, this. The people closest to it really were feeling like this is starting to happen. And I hope that five years from now, there’s a ton of really interesting use cases that are built on top of some of these model
31:41
breakthroughs, anything we didn’t cover today, you host your own AI podcast, like any hot takes, or any, like, really good perspectives that you’ve come across or come up with yourself that we didn’t touch on, that you think would be beneficial for the audience. I
31:57
mean, I think, like the, you know, there’s always the question of what the model companies will do, and you know, how that affects things on the investing side. And I think, you know, the topic we probably were, I was just alluding to, but we haven’t talked about a ton, is, I think there’s kind of this adjacent set of model capabilities that rely on, you know, different data, like So robotics, biology, material sciences, that I think, is is just poised for fascinating development in the next 510, years. And it’s not in the it benefits from, like the overall advantages advances in llms, but it’s not kind of directly in the in the path of those companies. And so I think there’s a ton of really interesting work and research going on there, and just super excited about about some of the developments there. Obviously, I feel like everyone in AI is excited about biology because it’s one of the most like mission driven impacts these models can have. But across these spaces, it really feels both that there’s some real research advances happening, and then candidly, also there’s just free capital flow into these spaces now in a way that lets a lot of these organizations really scale compute and data collection. And so I’m closely following this, and really hope we see some some strong model progress over the next
33:05
years. Jacob, if we could feature anyone here on the show, who do you think we should interview and what topic would you like to hear them speak about?
33:11
Well, one thing I’ve always wanted to do is there’s this company in healthcare epic, which is like an incredible success, you know, adopted in almost every health system. And the founder, Judy. She doesn’t give many interviews, but the story of how that company is built is, is just incredible. I think a choir did a whole episode on on Epic, but they didn’t get to sit down with Judy. And, like, I really hope at some point somebody, maybe you get to sit down with Judy and just get the like, true story of how that business was built, because it’s just, it’s an amazing story.
33:39
Awesome, full tear down, Jacob, what book, article or video would you recommend to listeners?
33:44
You know, one that I found really relevant, I think it’s actually behind me right now, is this book, engines that move markets, which is about basically past big changes in technology, you know, railroads, electricity, computing, the internet, and basically goes through each of those and the lessons that were learned from investing in those cycles. And I don’t think anything could be more relevant to the AI wave right now. I mean, obviously AI is different in some way, but history always rhymes, and I found it just incredibly helpful to have some of that historical context.
34:15
You know, as you read a book like that, there’s a lot of fear around AI, particularly around jobs or just change. Is there anything that stands out about previous waves and maybe the perception at the time versus what ultimately happened that you would maybe share with regards to what we’re seeing now in this this transformation? Yeah,
34:36
I mean, I think, you know, it’s been said before, but obviously in some of these previous waves, there was, you know, dislocation in job markets, and things that people used to do were no longer relevant. You know, I think over time, there were amazing new jobs that were created, and, you know, ultimately people moved into those things. That’s not to discredit, the transition period is highly uncertain, right? And so I think as we think about this period, I’m confident there will be a whole new. Set of roles that are enabled by this technology. And, you know, as a species, we’ll be happy that we’re not doing a lot of the rote work that that these models do for us. But, you know, I think we have to be thoughtful about how we manage that transition. It’s not just like a seamless, overnight thing to move everything, everyone you know, from one area to
35:14
another. Jacob, do you have any habits or behaviors that are a secret weapon? I
35:19
don’t know about secret weapons, but maybe to the to the conversation we’re having earlier. I do think I just, like, love learning about the stuff. And so I think, you know, I think that’s super helpful, right? Like when, when a new thoughtful podcast comes out, or an article, or I just get really excited about, about consuming it. And, you know, I always joke that I would, you know, maybe don’t tell red point this, but I would like, probably spend my time in a pretty similar way if I could, if, I could, if, even if I wasn’t working here for free, like, I just love learning about this stuff. And, you know, it’s such a privilege to get to meet, you know, people at the cutting edge of their industries, and get to learn from them and work with them. And so maybe, if anything, I just say, the amount of energy I get from, like, doing the job, I just find it’s just such a privilege to get to do I
35:58
love that. I love that since, the early days of my career, I’ve always said, like, I’m getting paid training. I’m getting paid to learn. It’s amazing. You know, you can suffer the tough parts when they’re paying you to teach, to learn on the job. And then finally, your Jacob, what’s the best way for listeners to connect with you and follow along with red point?
36:17
Yeah. So my email is just jacob@redpoint.com and then we are. We’re super active on all social media platforms, and so feel free to follow us anywhere. We have a great Tiktok presence, and, you know, a more formal LinkedIn one too. So whatever, whatever your flavor is, definitely check us
36:34
out. Love it. He is Jacob Efron, the firm is red point, and he has a podcast as well on AI. So check that out, Jacob, thanks so much for joining us today this. Yeah,
36:43
thanks so much, Nick. This is a ton of
36:48
fun. All right, that’ll wrap up today’s interview. If you enjoyed the episode or a previous one, let the guests know about it. Share your thoughts on social or shoot them an email. Let them know what particularly resonated with you? I can’t tell you how much I appreciate that some of the smartest folks in venture are willing to take the time and share their insights with us. If you feel the same, a compliment goes a long way. Okay, that’s a wrap for today. Until next time, remember to over prepare, choose carefully and invest confidently. Thanks so much for listening.
