🚀 Add to Chrome – It’s Free - YouTube Summarizer
Category: Startup Insights
Tags: AIcultureinnovationpivotstartups
Entities: AtioGitHub CopilotGPT 3.5JP Morgan ChaseKodiumMITNuroVerunWindsurf
00:00
I think startups are basically like getting slapped in the face probably over and over again. That's basically it.
I also, you know, interestingly, I I actually like failures a lot. One of the things I hate the most is doing something and not knowing if it's working or not.
And it's actually very it's very freeing to know when something fails because that's very obvious. I
00:16
like when decisions are obvious. Like when something fails, it's obvious you need to do something new.
And the faster you fail, actually, the faster you can decide to do something new. And I think for a company like us, like we're at the frontier of the technology, we should be imagining we are failing at a lot of initiatives.
In fact, if everything is working, we're operating at less than
00:31
potential of the company. We should be failing.
Like that probably means we're not betting. We're not taking enough bets if we're not failing enough.
I I think it's like a required part of a company that you that you fail. I'd say like be more humble about your ideas.
And I wouldn't say that I was arrogant at the time, but probably a lot more idealistic about our ideas being
00:48
correct. And you need some amount of idealism as I said before, but at the same time, it's that there's no reason why an idea is great if you haven't like validated it.
be honest with yourselves about, you know, the viability of your idea and be willing to pivot very quickly. I feel like a lot of a lot of founders hate the word pivot, but it's
01:03
it's awesome. Like, it's awesome.
You know, what sucks more is like doing the wrong thing and just failing. Even right now, I would say one of my regrets is probably not doing the pivot 3 months earlier.
Hey, I'm Verun, CEO and co-founder of Windsurf. So, Windsurf is an AI powered ID uh that provides agentic capabilities
01:21
that enables developers and non-developers to build apps really quickly and also modify existing applications really quickly. A million developers have actually used the product.
We have hundreds of thousands of monthly active users on the on the product right now and it's growing, you know, exponentially faster with time.
01:37
So, I graduated from MIT in 2017 and after that I actually worked at an autonomous vehicle company in Mountain View called Nuro. there sort of got an early taste of what deep learning could do for other industries.
Started the company actually in 2021, so it's been 4
01:52
years. At the time, the company didn't do anything relating to code AI.
It actually built out GPU virtualization and compiler technology. We did that for over a year and a half and were able to get to a couple million in revenue.
We had eight employees at the time. One thing changed very very materially in
02:08
that time. In the middle of 2022, GPT 3.5 came out at that point and we thought that generative models would fundamentally transform many different industries and we needed to make kind of a pretty big decision at that point to pivot the company because we fundamentally felt everyone was going to run generative models for everything.
02:24
And at that point we believed a lot of the value would acrue to companies that were applications almost like in the early days of the internet the companies that proved to be very valuable were companies like Google and Amazon and we wanted to see what would it be like if we could build the next Google or Amazon. We were early adopters of a
02:41
product called GitHub Copilot and from there we actually built out a product called Kodium which was an extension that lived in all the major IDEs and we were able to get that to over a million users as well. Um, and we ended up serving some of the world's largest enterprises, companies like JP Morgan
02:56
Chase. Very quickly though, sort of middle of last year, we realized that we needed to control more of the experience, especially as these models became more and more agentic.
And that was why we decided to build our own ID, which was Windsurf. That has taken off very quickly as well.
03:11
[Music] It's actually kind of interesting. I think MIT is very different than Stanford.
I think most people don't go to MIT to kind of immediately start a company. I don't think my aspirations were to start a company.
I did progressively intern at smaller and
03:27
smaller companies. So I first interned at LinkedIn, then after that Quora uh and then data bricks.
Data bricks at that time was a very small engineering team. They were not even a unicorn at that time.
But I think I didn't actually end up working at any of these companies because I think the thing that motivated me was can I be a meaningful part of a
03:44
visionary company working with motivated people. That's always been what's driven me and that's sort of why immediately after MIT I decided to go work at an autonomous vehicle company.
I thought that could be the future of where just robotics was going to go and I think for me and probably the other people at the company we want to work on the future of
04:01
technology and that's what truly motivates us. I think the motiv is just how do you build products in hard technology spaces where the technology is not there yet.
You know I I'll give like a quick just some interesting numbers. You know, when we started, obviously, deep learning was was pretty
04:16
popular. Uh, but the amount of compute that these models had access to grew exponentially year-over-year.
You know, just some numbers here. In 2017, the number of teraflops on a consumer grade GPU was 10.
By the end of 2022, it was actually 700. So, it's a massive increase of in compute that happened.
04:33
And I think what we learned was machine learning models were going to get more capable very quickly. And you should not bet on where the technology is today, but where it could be a couple years from now, right?
And if all you're doing is working on products that work today, you're going to be quickly irrelevant a year from now. And that was like a very
04:49
important lesson that we I think a lot of a lot of people that were in the autonomous vehicle space kind of learned, which is now if you were to look at autonomous vehicles, they're much more machine learning based than they were probably 5 or 6 years ago. Originally, we started out the name of the company was Exaf Function.
Part of the reason why we called it exunction
05:05
was our goal was to virtualize GPU computations to make it easier to run things on GPUs. And the the term exa function means we want to run an exa number of functions which is 10 to the 18 functions.
I think the biggest learning lesson for us was at the time was even if we were succeeding by some
05:21
metrics we were making some revenue just accepting that hey like the business might not be the best business and we need to go pivot was a very hard thing to do I would say right it's very hard when you have like employees you know you've already raised some amount of funding I think at the time we had raised over $28 million of funding to
05:37
basically start from scratch overnight uh but one of the learning lessons that I've sort of had and probably the other people at the company is every time you do a pivot you have an opport opportunity to maybe 10x the size of the company. And uh usually when you have an idea that you don't believe in and the ceiling is low, it's actually better to just scrap the entire thing than to like
05:53
try to incrementally claw and you know increase the value of what you're doing by like you know 20 30%. It's not going to matter in the grand scheme of things.
So we had gone to a couple million in revenue at that time. I think the hard part for us was it felt very ad hoc how we were adding revenue to the business.
06:09
It also felt like with the advent of these generative models, a lot of the complexity of running models would become commoditized. If everyone is going to run transformers, uh the transformer model architecture, what is the reason for us to have a platform that virtualizes arbitrary GPU computations?
It's not as important
06:25
anymore. So, it's a factor of two things.
We didn't understand how we could 10 or 100x sales, right? I think you build a company not to make a couple million in revenue, but billions of dollars in revenue.
So we didn't understand how to even get within an order of magnitude of that right that was one and two we also felt that the
06:41
problem we were solving with the advent of the generative models was going to get commoditized. So at that point it's kind of an easy thing once you believe that something is not going to be big you have no other option but to change your mind and do something new that is very painful because you did commit a lot of time and a lot of energy and passion to one thing.
You don't win an
06:58
award for doing the wrong thing for longer. I guess me and my co-founder did a walk over a weekend and I guess we decided over a weekend and we told the team on Monday.
Um and then everyone started working on the new thing starting Monday. Uh yeah, I guess we just you know one of the core beliefs that we have about startups is startups can if they're lucky do one thing really
07:15
well. We could not afford to have the company believe two things were important.
So we would need to rip the band-aid off and actually make it very clear to the rest of the company that the new thing was where we thought the future was. You know, it's interesting.
I would we were not worried at all when we did the pivot. We had basically written off the company is going to be a
07:31
zero and at that point anything is greater than zero. Ship whatever you can if it sucks you're in the same place that you are.
If it doesn't suck like great awesome only upside just honesty, transparency and int like intellectual honesty with the team is maybe a core
07:46
tenant of the company. We don't do things unless we like truly believe they are the right thing to do.
even if it like fits some narrative that the rest of the industry or our investors or convenient belief we believe things if from first principles they're correct and you know I guess one true fact was
08:03
after telling our employees there's a real chance that some of them would leave but I think like when you build a highquality culture where the people are are intellectually honest um good people don't want to leave if you if you chart a path for what the thing that does work is right because you know one
08:18
interesting thing about startups is people talk a lot about moes. What's the mode of a company?
But really, what is the mode of a company that has 10 people? You know, if you do have a moat, it's very shallow at best, right?
The number of engineering years that went into your product is already very small. So, the real moat is you work on the
08:34
right thing with enough people for long enough. That's like the only moat a startup ultimately has.
And that's like why you believe in a team. Obviously, when a company becomes thousands of people, hopefully the moat is not the current business.
It would be sad if the moat was the company can just pivot and do the next thing. In that regard, I
08:49
think if you genuinely believe the motive of a of a company is betting on the right things and and hiring a great group of people, then if you are intellectually honest with your team and tell them we don't believe in the strategy we just had, it's should be totally fine with everyone.
09:06
This episode is sponsored by Atio, the AI native CRM for the next era of companies. Connect your email and ATIO instantly builds your CRM right before your eye with every company, contact, and interaction you've ever had enriched and organized.
That's ATIO. And here's
09:22
what makes it even more game-changing. You can build AI powered automations and use its research agents to tackle some of your most complex business processes, freeing you to focus on what matters most, building your company.
Join thousands of companies who are already using to power their businesses. Visit
09:40
the link in the description to begin your twoe free trial with Atio. I think we were early adopters of a product like GitHub copilot.
So we and we were we actually built a lot of
09:56
infrastructure. So at the time we thought that was the tip of the iceberg of what the technology could actually build.
We thought a lot more could be built in this category. Right?
If all people were doing was autocompleting, we could potentially see a world in which entire PRs would get generated. And at that time, we were like, okay, the time
10:12
it takes to build applications probably has gone down by a double digit percent. Why don't we set the goal of the company to reduce the time it takes to build technology by 99%.
It's a very ambitious goal, not something we'd be able to do in the next year or two. But if we set a very ambitious goal, there's probably a lot of work to be done in this space.
We
10:27
believe that technology is always going to get better. We believe the models are going to get better.
Because of that, we just believed there was a lot of opportunity here. Because we were also an infrastructure company at the time, we were able to train our own models and run models ourselves, which enabled us to build an extension that was entirely free in all the major IDs.
And uh that
10:44
was able to very quickly get to hundreds of thousands of users. It was a combination of what our skill set was, also a belief of where we thought the future was going.
I think it took us less than 2 months to build the MVP and ship it out. I think as a company we're probably more of the opinion of if you are going to do something hard um it is
11:00
actually important to have intermediate goals that you can actually give to other people to validate your hypothesis. That's maybe something that we learned a lot of us at the company were previously from autonomous vehicles and I think one of the biggest mistakes that autonomous vehicle companies might have made is maybe to no fault of their
11:16
own was they built a hard technology where it's very hard to validate if the technology is good in the middle. Right?
So for us it's always been it is actually a feature if you can have a very very like visionary goal but have very tractable intermediate steps and so we actually picked a very tractable
11:33
intermediate step to deploy which was can we build a basic VS code extension that had autocomplete capabilities uh with our own model and that was that was the intermediate goal that we had and we believed you know maybe it was like a little bit of necessity we just pivoted the company people were not working on
11:48
anything else there was an existential dread what the value of the product was so we very quickly built it. I think a basic form of product market fit was when our inbound for the product from like companies was was more than like any of us could handle and we did need to hire a real go to market team.
I
12:04
think at that point we realized hey we're actually solving a real pain for the market and we are providing a lot of value. So that was that happened very quickly like within months of the product coming out we had a lot of companies reaching out to how could they run the product uh for themselves and and use it for their large code bases.
12:20
Yeah, I think the word product market fit is probably something I don't like a lot because it it creates a lot of like overconfidence. I think what what you have as product market fit in one day, if you don't continue to innovate and you have competitors, is very quickly going to become a commodity where you don't have product market fit.
So, you need to be very paranoid. I always like
12:36
to tell the company we're probably going to fail. And I think the good thing about telling the company we're probably going to fail is it creates enough sort of energy in the company to find the next thing that keeps you as a differentiated product and and survive in a in a space.
But I think in general
12:52
like product like I feel like if you have pull from the market, it's very hard to fake it. It's it's very obvious when you have it and very obvious when you don't.
And I'll tell you this like even when we were making a couple million in revenue with the GPU virtualization business is very obvious we didn't have product market fit. It's not a scalable company.
you don't really
13:08
understand how you can grow revenue and find find a way to like provide more value at scale. So very quickly I think a lot of large companies were starting to reach out to us for a variety of reasons for security reasons because we were able to run models ourselves.
We were able to run it in a secure way. The other thing is they wanted the systems
13:25
to work with their complicated sort of code bases. Right?
Their code bases are not simple. Some of our customers have tens of millions of lines of code in a single codebase.
and we started building more and more technology to make it possible that we would give very personalized suggestions to them regardless of where they stored their code. I guess very quickly we started to
13:42
get to 100 customers. It went from going from zero to 100 customers probably within months um at the company.
I think the feedback that was kind of interesting to us is how important just small details were right to the entire user experience. We started off as an infrastructure company and we needed to kind of become a product company which
13:58
is a big kind of change. You know basic things like the latency of the product how quickly we provide suggestions was quite important.
You know just basic things if we do too much computation on the user's machine and it kind of like makes the machine too slow that's a unacceptable user experience that is not something we would think about if we
14:13
were purely a server-based application. So there were kind of new things as a company we needed to learn because the space we were operating in was fundamentally different.
I think basically what happens is once a company gets bigger it's possible that you start losing the reason why you had a good product to start with right you start
14:29
listening actually in some ways too much customer obsession might be bad in that if you do things that all of your customers ask you to do sometimes you might like incrementally iterate to actually build a really bad product and it's actually better off to completely change the paradigm in a way that your customers don't really ask for because
14:46
you you suspect it could be helpful. So there's like a little bit of a healthy amount of don't just listen to random people about what you should build.
Listen to your customers, but don't listen to your customers in how you should build it. There could be a way you could build it in a way that is very transformational, but not something that they would even expect.
One of the
15:02
things about our product is we have a pretty large individual product and we have ways in which our users can communicate with us, you know, either online on social media. Uh, and I guess I I look through it fairly frequently to understand what the pain points our users have about the product.
Also at
15:17
the same time I guess like everyone at the company uses the product almost day-to-day which is maybe a unique aspect of the product right there are probably a lot of AI tools out there but the people building the tools are not using them all the time whereas the people that build our products literally use our product to build the product whether it be windsurf everyone builds
15:34
software on windsurf for windsurf which is a unique way for us to actually get feedback a lot of our feedback is internal and if no one at the company likes a part of the product it's unlikely that a lot of people outside of the company are going to like the
15:50
No, I I always try to run the company as like the smallest company it can be. But I think the goal of a company is not to be a small company.
I think the goal of a company is to actually provide the most value in a space. If the right way to do that is be a bigger company.
I think that's like the right thing you should sort of do. One of the operating
16:06
principles we sort of have at the company is we try to run the company fairly lean. And that doesn't mean we have very few people.
It just means that for a given amount of ambition, we are the smallest company we could possibly be. And that means that people are largely speaking underwater most of the time.
Hopefully like we don't have large
16:22
amounts of people that have too little to do, right? The goal is everyone should have too much to do.
And because of that, that forces rapid prioritization inside the company with a lot of urgency that triggers whether or not we should hire people. And honestly, that itself is a good forcing function of how you should scale a company.
You
16:38
scale a company when everyone is underwater and the moment people are no longer underwater, we don't scale the company anymore. I think in in terms of just some of the challenges obviously like as the company gets bigger, communication gets harder.
So there are some processes in place in terms of you want to ship a feature, it's no longer
16:55
just one person did everything end to end. There are like many different parties involved.
Uh but I think we added more process in place to make sure that even that could happen much faster. If you run the company in a way where everyone has their own set of priorities, I think it's it's cool.
Everyone's kind of having a good time.
17:11
But I think, you know, this boils down to why do companies work? I think they work if you do one or two things really well.
Maybe even one. Forget about two, maybe even one.
The problem is then actually if you were to imagine in a lot of cases, it's very hard to just have infinitely many people doing one thing
17:27
really well. So that naturally reduces the number of people that you have at the company.
Yeah. I guess I guess the the issue is naturally when you have more people, they're going to be more priorities, right?
Like people are not malicious, but they want to have something to do. If there are enough people to do what the original intent of
17:42
the company was, they're going to find other things to do and that's going to cause problems internally as well. One of the things that I've sort of done is I still interview everyone that joins the company, make sure that they go through a culture fit.
I think that's actually like quite important to me to make sure that we are adding people that are going to like maintain and preserve
18:00
and maybe even even just embrace the culture we have inside the company. So I think that's like honestly honestly like a very key thing that I've not given up even as the company has scaled you know close to 200 people.
I I think actually it's the same principles that made the company work when it's small. I I think big companies are not that different than small companies.
Maybe it's hard to
18:16
operate like a small company when you're a big company but I think if you were to ask every big company they want to operate like a startup. I think the very hard thing is once again intellectual honesty.
It's very hard as the company gets bigger for for a large group of people to analyze what they're doing and to say, "Hey, we should stop working on this or should work on something else."
18:32
It gets very hard because people feel some sort of safety as the company gets bigger. But the reality is like no company is really safe regardless of the size.
And that's how every company should be operating. They should be operating as if almost existential dread.
Not in that it's causing paralysis, but in that it causes the
18:47
agency to kind of figure out what the next thing to do is. Yeah.
I think people get they feel safety, psychological safety by feeling that what they're working on is the most important thing, right? I think, you know, people it's very hard to keep in your head, hey, like let me go work hard on something, but all the while it may
19:03
not be that useful for the company. And I think as a company gets bigger, it's very hard to have people that just still don't have the psychological safety.
Maybe one interesting thing about the company is I think companies have like two sides attached to them, which is that you need some form of irrational optimism. The reason why you need irrational optimism is without some
19:20
amount of optimism, there's no reason a startup ever wins at anything. A big company has more capital, more resources, more distribution.
So you need to believe you can do something better, right? With great people, uh you can somehow build something that's generational.
You have to believe that. Uh which is in most cases not true,
19:36
right? So you somehow need to believe that.
But also at the same time, sometimes that's just not going to be the case. And you need to be very realistic.
And I think this comes with, you know, uncompromising realism. Like you need a healthy amount of both.
You need you need the optimism to go out and do something when everyone else believes it's not that valuable because if
19:51
everyone else believed it was valuable, the big company would do it and you would still lose. But all the while you also need to agree most ideas are bad ideas and you need to kill your ideas fairly quickly as well.
This is this is like this tension that a company is always going through. It progressively gets harder and harder.
The company gets bigger. Uh but I think it's a strength
20:06
if a company can do that regardless of what size it is. Yeah.
I think the only thing is that I would sort of like to say is build for where you think the technology is going, not for today, right? And don't build probably technologies in places that like are as good as they will be given
20:22
where the technology is today. It feels like it's going to be very hard to differentiate in the long term.
And that's a hard thing to do, right? Uh that means you you you actually need to be willing to work on something that doesn't feel like it's working for maybe, you know, an indeterminate period of time.
I think what this means is these kind of short-term heristics that
20:37
you invest deeply in that make a model today work are probably actually have massive diminishing returns because the next model is going to be much better and the heristics are going to be unnecessary. I think the things that are important are how do you actually take advantage of the fact that you do have a lot of users and build better
20:52
experiences for them and actually because they are using your product you're able to build better and better experiences for them and these experiences are better and learned in a way they're actually learned from the way people use the product and I think you should be focusing on that not on kind of cosmetic things that make a
21:08
model kind of magically do something that probably is going to happen anyways in 6 months.