// transcript — 313 segments
0:00 Introduction and vision
0:02 Hi guys. How's everybody doing? This is awesome. This is really cool.
0:08 >> Great. Great to be here. >> exciting. Um gang
0:15 This is amazing and I and I just want to say that I'm really honored
0:19 to uh to be here. I love startups. I love the idea that you can take
0:28 something and you take an idea in your head and you put it in the world and you can make
0:34 the world a better place for other people. And um the ecosystem that supports that
0:43 happening in this world is not to be under-appreciated because
0:49 there are very few places in the world uh that have an ecosystem like this that
0:54 supports that. And it's the reason why America is the greatest place for innovation in the
0:58 entire [ __ ] world. So Let's go. >> Props to the people that put this
1:03 together. Cool. Um well Ashton, it's great to be here and doing this. Um
1:04 Investing in AI and Warp
1:09 you might not need an intro but do you want to introduce a little bit about
1:14 why you decided to invest in Warp, how you think about the space and how you're
1:18 broadly thinking about investing in AI companies today. Um maybe we can kick off there and then I
1:23 can talk a little bit about going from there about Warp. Yeah, awesome. And then maybe you can
1:27 tell people a little bit about yourself cuz your story is incredible. Um
1:32 What So uh I've I've been uh kicking around investing in startups for almost
1:37 two decades now. Um and I think um the one thing that every startup that
1:47 uh we find or invest in has in common is a founder that has shared moral values in what what is
1:57 their that that that the world should look like and and and what their vision
2:02 is for what the world should look like. Um a shared uh desire to make
2:08 life better for people uh and and generate enduring happiness
2:12 uh in the world. Um and then you know, when there are a lot of people in this world
2:19 when they come to an intersection, they think about going right or left. And
2:23 then there are a few people that when they get to that intersection, they see
2:27 up as an option. And I use this one of those people. Uh when you sit down with him and you talk
2:34 about the ecosystem, when you talk talk about um employee management systems in
2:39 general, when you talk about how to make companies better, more resilient, and
2:42 stronger, the way that he thinks about problems that he has within his company, but
2:48 problems that other companies have within their companies, he sees up as an option.
2:55 And uh there there have been these like moments in time that we've experienced and
3:01 invested through um what like when the iPhone arrived and all of a sudden there was this
3:07 smartphone in everyone's pockets, and everyone thought, "Well, you know, this
3:10 app ecosystem, you know, there are going to be some small companies that are
3:14 built in this app ecosystem, but the real companies aren't going to be these
3:18 app companies." And then and then turns out that there are a lot of these app
3:20 companies that we believed in early days that turned out to be the companies that uh
3:26 uh that are driving the markets today. Um and then when this AI revolution
3:35 happened and 3 years ago, we invested in OpenAI and Anthropic and these other
3:38 companies, you know, a lot of people said, "Well, you know, there's going to be like this
3:43 ecosystem that will exist on top of these AI agents or these LLMs
3:49 that on top of these multimodal models uh and leverage these things, but there
3:53 the real companies are going to be these companies that happen. And And And we
3:57 And we just have taken a contrarian belief to that and and believe that
4:03 there are verticals where there are needs for deep verticalized agentic uh
4:11 innovation that touch human beings and and the touch points for human beings
4:15 that that there is a true moat in that product that get that gets created. And
4:20 so, when we met a founder that matched the profile of a founder that we that we
4:23 were interested in and that was going after a problem that we were very familiar with. We've
4:31 invested in the space before. Um that we realized that uh that there's a moment in time that is
4:35 now that people who understand how to build on top of agentic infrastructure um and
4:43 are going to build from the ground up on top of agentic infrastructure are going
4:48 to be the people who have the leverage to build the best products for the
4:53 customers. And uh and that's what Warp is and um and that's who Ayush is. Well,
4:57 Ayush's background and Warp
4:58 thank you so much. That's very flattering description uh by the way. Uh
5:03 I really appreciate it. I think quick background myself. My name is Ayush. I'm
5:07 the founder and CEO of a company called Warp. Um as Ashton sort of set the stage
5:13 for me, um they invested in us about uh little less than a year ago as part of
5:17 our Series A. But, really, you know, I quick background myself. I grew up in a
5:22 small town in India uh in this very normal middle-class family. I didn't
5:25 know a single person who had gone to the US. So, obviously, I decided that I
5:30 wanted to study physics at MIT. Um behind the backs of my parents and
5:34 teachers and peers, uh there's a fun story there, but long story short, I
5:38 ended up becoming one of the first students out of this state in India of
5:42 about 250 million people to get into MIT for undergrad. There I studied computer
5:47 science, physics, math. I ended up specializing in doing research in
5:51 large-scale machine learning systems. Um and what was fun about this this was
5:55 like 2019, 2020, right before that pre-chat GPT moment had come out. Um so
6:00 I sort of quit my role and when that happened and decided that I think we
6:04 wanted to take a stab at one of the hardest problems in the space. And even
6:08 though back then the agents were quite brittle, we could see that the we could
6:12 use them to go after one of the most complex like tax compliance, deep like
6:19 employee management workflows where HR finance people are spending a ton of
6:25 manual time today. Uh so we took that bet. It was fairly contrarian at the
6:29 time. A lot of people told us to like not go down this one. It's like too
6:34 hard, too too like messy of a space. Uh but we sort of kept building. And so
6:37 fast forward today, um you know, we're about 50 people based out of New York.
6:43 Uh growing quickly. We serve some of the fastest-growing AI companies with all of
6:48 their employee management needs. So whether that's payroll, deep tax
6:53 compliance, onboarding, benefits. We recently just launched our IT product
6:57 suite. So all things with like identity, access. We wrote our own device
7:01 management software. So we're taking this like very deep approach to
7:06 engineering-focused, product-focused on a space that's just like been untouched
7:11 by AI. And I think to kind of like Ashton's point, um we fundamentally believe that the last
7:17 generation or the last two decades of companies, you can't sort of reorient
7:23 them to be completely AI native, to just kind of bolt on on top. And that's where
7:26 we sort of come in. It's like fundamentally from day one, what does
7:32 the LLM agent native platform look like with deep like permissioning controls
7:36 when the stakes are this high, when you're dealing with taxes and money
7:39 The importance of architecture
7:39 movement, and so on. Okay, the thing the it's so the the thing that I would point
7:45 out relative to So, given the foundation of what the company's built on and the
7:48 way it's built, you heard all these products that just get like dinged off
7:51 that have been launched and there are more products that are pipelining that
7:54 are coming. You can't do that under the old architecture. You can't just crank out the next
8:02 product and the next product and the next product under the old architecture
8:06 of company building, but under the new architecture, you can. And and that that
8:13 is a core differentiator. And so, if if you really believe and there's a lot of
8:17 conversation around moats, which I think is is important to talk about. So, when
8:21 we look at these companies, you look at companies that are being being built,
8:25 one of the biggest thing investors are looking for all the time is where's the
8:28 moat? Like given that anybody and everybody can jump on cloud code and spit out the
8:36 software like HipChat, where's the moat? And and and it's a hard question and a
8:40 lot of people are like, "Well, it's brand." Or "No, no, no, it's taste." Or
8:44 what is No, it's the architecture with which you're building that allows you to
8:50 iterate on the front line the fastest. The true moat is the product.
8:57 It's actually your ability to service your customer better than anyone else. Like we've seen
9:03 lots of people like Let's just take the iPhone as an example.
9:08 The the the architecture that product and the precision of understanding the
9:13 customer and servicing the customer need is what made the iPhone continue to like
9:17 scale out. And and you can you can argue like Look when Airbnb came about, it was
9:22 like VRBO existed. It was called HomeAway at the time. But like the care
9:27 and the love and the attention to the consumer experience and the ability to
9:31 iterate product at the time in the way that they did in order to service the
9:36 customer need is what the differentiator was. You will not be able to keep pace
9:42 if you don't have and you're not built on top of this new architecture
9:46 that allows you to drop that many products in a year. Yeah. It's that
9:52 that's the differentiator. >> Yeah. I I think one that's a really
9:55 great point. I think we often joke internally that the we're
9:59 we're this very horizontal platform with lots of these products interlocking onto
10:02 each other. And we joke internally that like the number of products we have, they're
10:08 almost like independent several billion dollar companies each doing each one of
10:12 those products. Whereas for us, I think we've done that with a handful of engineers. We're a
10:18 small, much more elite, but a really small team. And I think that just wasn't possible
10:25 before. And so me, I think like I really sort of found the last sort of decade as
10:30 I was looking to start a company, I think the only reason I got really
10:33 excited is it's almost like if you're a technical and product-minded founder
10:37 now, like the time best time to build something is today. Whereas the last
10:41 decade almost felt like you were in this like late-game, end-game of SaaS
10:46 software where the only opportunities were these like very narrow point
10:51 solutions because all the budget was taken, and all the things were surface
10:54 area were taken. And there was nothing new that you could actually do. So I I I
10:59 genuinely feel like pre-AI, like I just wouldn't have been this excited to build
11:02 a startup and build a company. But I find it very exciting. So when you
11:06 think about building Warp and you think about your customers that you need to
11:12 service, what you need to give to them, how do you think about and it really is
11:17 like a new way of first principles thinking to some extent? As you're architecting your company,
11:25 where do you go where I can strip that away? The what like one of the greatest
11:32 this the whole sort of idea of this got is in a function like do more with less,
11:38 right? But there this is this is existed this simplification has existed forever.
11:43 Like Coco Chanel, one of the great designers of all time, would line up the
11:47 models to go out on the runway, look at them, and take one thing away.
11:51 And then go, "Okay, now it's perfect. Now you can go." The idea like in
11:55 studying Steve Jobs to play Steve Jobs, I played him in the movie in case
11:59 anybody ever saw it. But one of one of the things that he did and
12:01 when he talked to these engineers and folks that one of the things he
12:05 constantly did is like how what what can we take away from the screen? How do we
12:08 simplify this? How do we simplify the user experience? When you're building
12:12 your company, what do you think about taking away? Yeah. So,
12:14 The subtraction economy
12:16 as we were thinking about the title for this talk, I think we sort of settled on
12:20 this unusual phrase the subtraction economy. But really what we were
12:25 thinking about is if the last generation, the last couple decades of
12:29 software were about adding more dashboard, software, tooling, vendor
12:35 that human still had to manually hand crank, I think the next generation is
12:39 going to be about how much of the tedious manual operational work can you
12:44 just completely remove from the picture. So, we started with one of the hardest
12:49 areas in this space, which is in the US thousands of tax jurisdictions and
12:53 state, local, federal tax compliance, very high stakes, very low fault
12:57 tolerant, and we just pointed all of our energy towards solving that. And as what
13:01 what we're finding is this entire category of like the finance and HR
13:06 suite that is running manual reconciliations, manual work, we're able
13:11 to just remove that drop down from like 10-20 hours a month for one person to
13:15 couple minutes. Um and that's leading to this like new kind of generation of companies, the
13:22 fast-growing companies we see, they're able to do with one HR superpowered ops
13:28 generalist, let's just call them, what typically would have required 15-20
13:33 people teams in the past at the rate at which we're growing. So, I think we're
13:35 really excited about that. >> So, this is this is a trend that we've
13:41 seen as investors as well, where it it used to be um or the humans are going to input
13:46 information in. Um the company is going to aggregate data, and then we're going to create
13:53 dashboards that tell the humans what to do or what needs fixed.
13:56 >> Next. And then what we're seeing now is the humans are putting information in or
14:02 some AI note-taker is putting information in, you're doing your data
14:05 aggregation. Then it's coming and saying, "Hey, we need to fix this. Are you good if I fix
14:11 it for you?" And then it's actually running and and so the human in the loop
14:16 is actually just permissively driving the agent to make the change.
14:21 Like we've seen it in ITSM. Uh we have a company called Serval that we invested
14:25 in that's like an ITSM ticketing solution that just delivers it whoa whoa
14:28 whoa whoa. There's the whistle whistle. Um uh and we um it basically says, "All right, here's
14:35 here's a ticket that needs resolved. This is the way you can resolve it here.
14:38 Now it's resolved, and now you just move on and you move through that flow much
14:42 faster." We've seen it with CRM. There's a company called Day AI that we invested
14:46 in where the last thing sales professionals want to do at the end of
14:50 the day is update the CRM and pound a bunch of information into it so that it
14:53 can it it actually just does that for you. It does all the data entry based on your
14:59 conversations and flows. And so it's it's taking that next step and it's
15:04 going, "What did the dashboard used to prompt you to do? And now we can do that
15:09 for you as well," which extends out the the scale of the product.
15:14 >> Yeah. It's It's almost like if you assume that the job of software,
15:20 like truly the the the true promise of software was to free humans up from
15:25 tedious manual work. I think we're finally able to kind of do that with AI
15:30 and agents because this exact point of around like how can we just do that in the
15:35 background for you? And use and reserve human input and human judgment for that
15:42 very like tie-breaking, for that the the exact context that the exacts or the
15:47 CHRO or the CIO holds in their head, only bring them in whenever needed and
15:50 the system kind of runs in the background. I think there's like a funny
15:54 tweet the other day it's like finally we can just let the agents run and just go
15:59 take a walk to the park. So I think it's it's really just I think we're
16:02 approaching that point too. We're not there yet. We're but we're getting
16:05 close. We're we're getting close. In the meantime though, I think we're
16:09 just getting a lot more done and I think that's really fantastic. So when you
16:11 Building efficient teams
16:13 think about architecting the organization inside Warp, and you think about scaling out the
16:22 structure of of what you're building, is are you hiring less people? Are you like
16:29 and and is there a certain profile of person that you're looking for and
16:33 capabilities within like how are you determining, hey this is a good team
16:37 member for Warp because they're going to understand that next level of
16:40 efficiency? That's a great question. I I really So we we're we were just about 10 or so
16:49 people up until a couple months ago and everyone always used to think it's like
16:53 this must be a much larger company because of all the products and like the
16:56 customers, how many they have. And we're now come up to like very
17:00 quickly 40, 50. And I think we are scaling the team, but I think the kinds of people we are now
17:06 really sort of wanting for are are very different from pre sort of AI pre-agent
17:11 sort of era where I think we're just much more focused on high slope people
17:14 now. And so Uh, it's it's can you just really quickly adapt and use these tools to
17:22 sort of like orchestrate the amount of work that previously would have taken
17:27 several days to now you can actually do that in maybe an hour. Um, that's what's
17:31 really exciting maybe on the engineering and product side. I think we're seeing
17:35 that on operational side as well. For a company like us, what we're doing, it
17:41 the public comps that we have, we want don't need to name them, but often if
17:44 you look at their head counts, they're like 30, 40% of thousands and thousands
17:48 of people that are staffed in like support, tax, operation function. Um,
17:53 and because we sort of built from day one to try and automate that
17:57 fundamentally with software and agents, we have one and a half person on our
18:01 internal ops and tax teams. So, one one person is part-time. That's why I say
18:04 one and a half. See, it's funny because you you you you you think about them as a
18:10 competitive set and I as an investor I'm not sure that I think about
18:15 the other companies that are doing similar functions as a competitive set.
18:20 In so much as I don't think that um, Steve Jobs looked at Sony Walkman as
18:26 a competitor for the iPhone. I don't think that Steve Jobs looked at Nokia's
18:31 handset cell phone as a competitor for the iPhone. I don't think that Steve
18:34 Jobs looked at, you know, all the other functions that an iPhone I don't think
18:35 Adoption and industry shifts
18:38 you looked at any of those incumbents as absolute competitors.
18:43 I think it's like category defining in a way. I think look, there's still going
18:49 to be companies that are existing enterprises that may not want to
18:55 transition or maybe just are uncomfortable transitioning into an
19:02 agentic world or an AI world that those companies are going to have
19:06 tons and tons of customers to serve. But I I think there's just going to be a
19:09 different kind of company builder and I think you're built to service a
19:13 different kind of company builder that actually wants those and needs those
19:17 tools within their organization and wants the people that are in
19:23 human resources to actually be interacting with humans instead of
19:28 interacting with a bunch of labor that is is in some way shape or form
19:33 servicing a human. But I think that I think that that human to human contact
19:38 is going to become more accentuated inside of those orgs as a result.
19:43 The other impediment that I'm that we're seeing a lot is is truthfully um just employees having a
19:52 willingness to automate their workflows. Um where they go, "Wait, if I auto make my
19:56 workflow, I you're not going to I'm not going to be useful to anyone." And you
19:59 go, "Actually, no. If you automate your workflow, you're going to be more useful
20:03 because you become a person that knows how to automate workflows, which creates
20:08 a more efficient organization overall and and on a go forward. I think that
20:13 that that right now the tools exist to do this, but what we're faced with is a diffusion
20:19 problem where the adoption of those tools is not happening at the same rate
20:23 that it at the rate that that it could and should to create the efficiencies in
20:27 the outcomes. Yeah, this this definitely will take 5 years, 10 years, 15 years, 20 years to
20:33 really kind of play out and diffuse. That's just like the lesson from the
20:38 last big platforms change as always. And to your point, I think we internally
20:42 sort of phrase this as Warp is built for purpose built for ambitious companies.
20:46 Um and what we mean by that is like our customers, we're we've seen customers
20:52 grow from uh 10-person companies that just signed up when they were really
20:55 young and getting started to now 100 plus in a matter of like 9 months. So,
20:59 that's a 10x head count growth all on Warp within a very, very short period of
21:04 time. And all that time what we were able to see is like they have maybe one,
21:08 one and a half people on their finance, HR teams and because they're so
21:12 leveraged on these systems that are automating a large chunks of that manual
21:16 work that they would have to do. And I think there's just this like K-shaped
21:21 variance happening in the startups as well and the companies that are just
21:25 growing a lot faster today. So I think like the to your point, I
21:28 think that these larger sort of mid-market or enterprisey companies that
21:34 will that are not as amenable to this new kind of workflow, this change, it
21:37 will probably take a long time and in the meantime there'll be great companies
21:40 serving them as well. >> Okay, I'm going to I'm going to throw
21:42 you a curveball and you don't know this The So get ready. And and then I
21:44 Managing risk and reliability
21:55 The big the big knock right now that a lot of people have, a lot of
21:58 enterprises have um a lot of market companies have uh on using AI-based infrastructure
22:12 What does the error rate look like for the outputs inside of uh your
22:16 products? Um and how have you dealt with it and how are you reducing it over time? Yeah.
22:24 The This is the number one thing that we thought about actually when we were
22:27 starting Warp is we have to be because we're in such a high-trust, low-fault
22:37 are not filed correctly on time, the IRS is going to come knocking and that's not
22:40 going to be fun for anybody, right? So I think what we sort of
22:44 from day one we were very, very thoughtful on how can we how can we have
22:48 these like evals and validations and human in the loop systems where we can
22:54 pair world-class domain expertise in tax, finance, accounting functions with
22:59 our AI system. So the error rate question, what we've seen is as the models keep getting better,
23:06 we've gotten a lot of free benefits, but also the more compliance workflows that
23:10 we complete, the more edge cases we handle, that all feeds back into our
23:13 systems and we just improve continuously, right? And so, this was
23:18 why I think like more people should be thinking about these things very
23:20 long-term. Like we've done this for now 2 and 1/2 years to get to a point where
23:25 we're ready to kind of start like going this like full throttle, but it took us
23:29 a decent amount of time like just in the grind mode trying to figure out all
23:34 these like nuances with uh let's say Ohio tax department and and Arizona
23:39 labor registration system. Uh but that's all finally coming together now. It's
23:43 really funny. I was I was contemplating this the other day um because
23:49 I was like, "Well, it's only, you know, 70% reliable on blah blah blah." And I
23:54 was like, "Yeah, but like I remember reading a Malcolm Gladwell book and being like,
23:59 'Well, that must be a fact.' And then years later being like, 'Well,
24:03 actually.' Uh and then Malcolm Gladwell coming forward and being like, 'Nah, I
24:07 was wrong.' Uh or reading like, you know, Good to Great and being like,
24:10 'Well, that must be how you grow a company.' And then going like, 'How are
24:14 those companies doing today?' Um and realizing that it wasn't true. And so,
24:18 and it's funny because you can go you can search something uh on Google, you
24:22 can search something on and you can find a lot of misinformation on your Google
24:27 results or people just trying to sell you things. Uh or you can read a book
24:31 and find out years later that in fact the marshmallow test isn't accurate when
24:36 you build an entire framework for hiring around a marshmallow test. Um and and
24:41 thousands of books quote it as a factual piece of information.
24:46 And and so, so much of what we know and and believe in the world uh has uh doubt
24:54 built in to it. Um and and the pursuit to build something on top of a system
24:59 that is already extraordinarily reliable, but will only become more reliable.
25:05 Um it it would seem to me that that that is a that that is a persistent challenge
25:10 uh that we all have to face uh as we continue to build on top of
25:13 this ecosystem. Yeah, I think we're at time here, but uh it's really great to be here do this
25:18 conversation. Thank you so much Ashton for sitting down. Um maybe any
25:23 That's it. >> [cheering and applause]