AI transcript
0:00:01 (upbeat music)
0:00:04 Support for the show comes from Into the Mix,
0:00:07 a Ben and Jerry’s podcast about joy and justice
0:00:09 produced with Vox Creative.
0:00:12 Ainez Bordeaux is a self-described hellraiser,
0:00:15 and she became an activist after being caught up
0:00:16 in the criminal legal system
0:00:19 when she couldn’t afford her bond.
0:00:20 And without a trial,
0:00:23 Ainez was sent to a St. Louis detention facility
0:00:25 known as the Workhouse,
0:00:28 notorious for its poor living conditions.
0:00:31 Here how she and other advocates fought to shut it down
0:00:33 and won on the first episode
0:00:36 of this special three-part series out now.
0:00:39 Subscribe to Into the Mix, a Ben and Jerry’s podcast.
0:00:45 Support for the show comes from Into the Mix,
0:00:48 a Ben and Jerry’s podcast about joy and justice
0:00:50 produced with Vox Creative.
0:00:53 Ainez Bordeaux is a self-described hellraiser,
0:00:56 and she became an activist after being caught up
0:00:58 in the criminal legal system
0:01:00 when she couldn’t afford her bond.
0:01:01 And without a trial,
0:01:04 Ainez was sent to a St. Louis detention facility
0:01:06 known as the Workhouse,
0:01:09 notorious for its poor living conditions.
0:01:12 Here how she and other advocates fought to shut it down
0:01:14 and won on the first episode
0:01:17 of this special three-part series out now.
0:01:20 Subscribe to Into the Mix, a Ben and Jerry’s podcast.
0:01:24 – Scott, have you ever made a significant pivot
0:01:25 at one of your businesses?
0:01:28 And if so, how did you know is the right time
0:01:29 to make that change?
0:01:31 – Yeah, I’m not sure there is a significant business
0:01:35 that’s ever been built without something resembling a pivot
0:01:38 or iterating the business strategy.
0:01:42 So my first firm, Profit Brand Strategy,
0:01:44 was initially profit market research.
0:01:45 And we used to go out and do surveys
0:01:46 with the internet and computers
0:01:48 and try and find a different way to collect data.
0:01:53 And what we found is that what brands and clients appreciated
0:01:54 was the interpretation.
0:01:57 So we turned to Profit Brand Strategy
0:01:59 and we became a consulting firm at L2.
0:02:01 We were totally focused on the luxury space.
0:02:02 And then P&G called us and said,
0:02:04 “Would you ever do this for P&G?”
0:02:06 And the name of the company was Luxury Lab.
0:02:08 And I hung up the phone and said,
0:02:10 “Our new name is L2.”
0:02:15 And pivoted to or into consumer products at section,
0:02:17 my online education started up.
0:02:20 We initially thought we were gonna be the Netflix of business
0:02:23 that it would be short form videos for B2B.
0:02:25 And that was gonna be so expensive
0:02:26 to produce that kind of content.
0:02:29 So much of that content was available elsewhere
0:02:31 that we pivoted straight into online education
0:02:35 focusing on upskilling the enterprise or AI skills.
0:02:38 So I don’t think I’ve had a business where we didn’t pivot.
0:02:41 I think you just look at the data
0:02:45 and you are thoughtful about what are the opportunities?
0:02:49 And there’s nothing like facing the enemy.
0:02:51 There’s just no market research like launching a business
0:02:54 and seeing what people are actually willing to pay for
0:02:56 to inform your decision making.
0:02:59 And a lot of times clients will come to you and say,
0:03:01 “We’d love it if you could do this.”
0:03:05 And so you’ll get signals from the market.
0:03:09 And what I would suggest is have a solid board
0:03:10 and have your kitchen cabinet of people
0:03:14 that you can propose stuff to and bounce stuff off of them.
0:03:16 And also talk to your colleagues and your employees.
0:03:20 And I’m trying to think if we’ve done a pivot here,
0:03:24 I guess we’re sort of we had our adventures in television.
0:03:26 Now we’re kind of, I wouldn’t say all in on podcasts,
0:03:28 but I think we’re devoting the majority
0:03:32 of the human capital property media now to podcasts.
0:03:37 So the market is an unbelievable muse and advisor.
0:03:39 And you just wanna surround yourself with smart people
0:03:40 who can help you interpret the data
0:03:44 and make sure that you’re not speaking to yourself.
0:03:45 It’s very hard to read the label
0:03:47 from inside of the bottle sometimes.
0:03:50 (upbeat music)
0:03:54 – Welcome to First Time Founders.
0:03:59 One of the most promising use cases for AI is code generation.
0:04:02 That is doing the work of a software engineer.
0:04:06 Already in the US, an estimated nine in 10 software developers
0:04:08 are using AI coding tools.
0:04:11 My next guests created one of those tools.
0:04:13 And less than two years after launch,
0:04:16 it’s already one of the most popular AI coding assistants
0:04:18 in the world.
0:04:21 Last year, they had less than a thousand users.
0:04:24 Today, they have more than 600,000.
0:04:27 Now, after raising a $65 million funding round
0:04:31 at a $500 million valuation,
0:04:33 they’re looking to take over the industry.
0:04:37 Next up, compete with the likes of Microsoft and OpenAI.
0:04:40 This is my conversation with Varun Mohan,
0:04:42 CEO and co-founder of Codium,
0:04:44 and Jeff Wang, Codium’s head of business.
0:04:47 Varun, Jeff, welcome.
0:04:47 – Thanks for having us.
0:04:48 – How was the flight?
0:04:49 You just got in today, right?
0:04:50 – Yeah, we took a red eye
0:04:53 and we I think slept one hour each maybe.
0:04:55 And then when we got to the hotel earlier today,
0:04:56 maybe we slept another hour.
0:05:00 So we’re in perfect condition to do this podcast.
0:05:02 New York code spells are super fun.
0:05:04 You get like this matchbox room.
0:05:06 You know, you’re right next to the wall
0:05:08 every location you are, so perfect.
0:05:10 – Well, I appreciate you being here
0:05:11 and appreciate you joining,
0:05:15 despite getting one hour of sleep last night.
0:05:17 – It’s like two, I suppose.
0:05:19 – Yeah, yeah, good enough.
0:05:22 So you guys are the second AI company
0:05:23 that I’ve had on this podcast.
0:05:28 The previous one I had was an AI for finance company,
0:05:31 more or less replacing bankers,
0:05:35 though his argument is that he’s not replacing bankers.
0:05:37 I’m just gonna start this with the same question
0:05:39 that I asked him, which is,
0:05:42 are you guys at Codium replacing programmers?
0:05:46 – Our vision is actually to give developers
0:05:48 the ability to dream bigger.
0:05:50 And I know that that sounds very vague,
0:05:52 but I think there’s one way of looking at it
0:05:54 that we could just go out and replace
0:05:57 like low-skill labor or low-skill developers.
0:05:59 But I think that that’s not a very rich idea.
0:06:01 We wanna take the best developers
0:06:03 and make them 10 times as leveraged.
0:06:06 And the reason why we think that this is only going to lead
0:06:09 to more developers existing is unlike other professions,
0:06:11 there’s no limit to the amount of technology
0:06:12 the world can actually consume.
0:06:13 – Yeah. – Right?
0:06:15 I don’t think you would ever just be like,
0:06:18 hey, guys, stop making technology.
0:06:19 If we can actually make it so that
0:06:22 the next great invention happens 10 times faster,
0:06:23 we will just only get 10 times
0:06:25 the amount of invention in the world.
0:06:29 – So a program is generally fans of yours, would you say?
0:06:30 – I hope so.
0:06:33 Yeah, I mean, our message is not to replace developers,
0:06:34 I would say.
0:06:36 The Wave Room actually got me on board
0:06:40 ’cause he said like everyone that touches a product,
0:06:42 almost half the people continue using it.
0:06:45 So when you have that, you know there’s something there,
0:06:46 you know there’s some value, right?
0:06:47 And some excitement.
0:06:49 – Is that higher within the industry?
0:06:52 I mean, 50% retention basically of the customers.
0:06:56 That’s, do we know what it’s like for other AI tools
0:06:57 or is it too early?
0:06:59 – Yeah, I think for a lot of consumer products,
0:07:02 it’s usually in the, if you can get it above 10%
0:07:05 at the long tail that is considered very good.
0:07:07 That’s largely because consumers,
0:07:10 unlike companies, if they get bored of something,
0:07:12 they throw it away very quickly.
0:07:14 And these products are not collaborative also.
0:07:16 This is like a single player product.
0:07:18 So it’s actually very easy to churn off of the product
0:07:20 because you aren’t chatting with other people.
0:07:23 So it does mean it is providing a lot of leverage
0:07:25 and developers stay more in flow state
0:07:27 using products like Kodium.
0:07:29 But yeah, hopefully we can make it even better.
0:07:31 – I mean, you’re up against a lot of different tools.
0:07:34 And I mean, I feel like code generation
0:07:35 was one of the first things that people said,
0:07:37 “Oh, AI is going to take over this thing.”
0:07:41 But I feel like the biggest competitor in your space
0:07:43 is GitHub Copilot.
0:07:45 What does the competitive landscape look like
0:07:48 in the AI code generation space?
0:07:50 – There are reasons why we are able to compete directly
0:07:51 with Copilot.
0:07:54 One of them is that we have full repo context awareness.
0:07:56 So as the user is typing,
0:07:58 the results are highly personalized
0:08:01 and we’re seeing like a 30 to 40% boost in accuracy
0:08:03 just having that code base be available
0:08:06 and giving tools for the user to point to
0:08:09 what they’re working on and trying to figure out their intent.
0:08:12 So just the quality of our code suggestions
0:08:14 are very competitive.
0:08:16 And then the thing that we’re really kind of trying
0:08:19 to lean in on though is our ability to deploy
0:08:21 onto like a private server.
0:08:24 And people can host Kodium inside their company
0:08:26 or in their work environment.
0:08:29 And for example, if you’re like in the defense space
0:08:31 or like the finance or healthcare space,
0:08:33 they can’t use Copilot at all.
0:08:35 And that’s kind of where we are focusing right now,
0:08:38 our efforts, but then we have some things down the pipe
0:08:40 where we’ll just be competitive
0:08:41 even on the cloud front too.
0:08:43 – What does it look like for a programmer?
0:08:46 I mean, it sounds like you’re sort of typing in
0:08:49 and then it gives you a list of auto suggestions
0:08:51 for what comes next.
0:08:53 What does it actually look like if you’re a programmer?
0:08:56 And please explain it to me as if I’m five
0:08:59 ’cause I don’t program, I don’t, not a coder.
0:09:02 – Yeah, so a little bit about like the way software
0:09:03 sort of gets built.
0:09:06 So developers write code and what’s called an ID.
0:09:09 It’s this application that enables you to debug code.
0:09:11 So if there are bugs, you could run it.
0:09:14 You can actually see what the errors are, iterate on it, right?
0:09:18 And then right before the code gets pushed into production,
0:09:20 it goes through a review process.
0:09:22 And other people in the company take a look at the code
0:09:23 and actually review it.
0:09:26 And then after that, it goes and it actually gets deployed
0:09:28 into production and it’s on a website or whatever,
0:09:32 where end users can actually touch the product in the end.
0:09:36 And right now where codium is mostly focused is in the ID.
0:09:38 So that’s where developers actually write the code.
0:09:40 It provides value in multiple ways.
0:09:42 So as developers are writing code,
0:09:46 it fills in passively starts filling in more and more code.
0:09:48 And because of the fact that we actually do train
0:09:51 our own models for that passive AI,
0:09:54 we actually found that around 50% of all software
0:09:57 that is getting committed by a developer
0:10:00 is actually accepted and generated by codium.
0:10:02 So that’s the amount of leverage
0:10:05 that just autocomplete is providing to end users.
0:10:08 But to add to that, we also provide a couple of other pieces
0:10:10 of functionality that is super valuable,
0:10:13 despite what level you are as a programmer.
0:10:15 So you can even chat with your code base.
0:10:17 And this probably seems like a very basic piece
0:10:20 of functionality, but when you’re a new developer
0:10:21 and you’re coming into a company
0:10:23 and you have millions of lines of code,
0:10:25 it takes a while to actually onboard
0:10:26 onto that new code base.
0:10:29 And what we’re finding is even at the largest enterprises,
0:10:32 the time it takes to onboard onto a code base
0:10:35 goes down from four to six months to four to six weeks
0:10:37 with a product like codium.
0:10:40 – I wanna focus on how you started this company.
0:10:45 You were a software engineer at a self-driving vehicle company.
0:10:48 So you were kind of working on AI there, I feel like.
0:10:53 And you originally had an idea to start a company
0:10:55 not for writing code with AI,
0:11:00 but for something called GPU virtualization,
0:11:02 completely different company from codium.
0:11:03 It also had a completely different name.
0:11:05 The name was Exa Function.
0:11:09 Can you explain the first iteration of this company
0:11:11 before it became what is now known as codium?
0:11:12 – So to add a little color there.
0:11:14 So I graduated from MIT,
0:11:16 worked at this company called Nero.
0:11:18 It’s an autonomous goods delivery company.
0:11:19 And actually a lot of the learnings
0:11:20 that I had from autonomous vehicles
0:11:23 are actually making their way into the space
0:11:24 we’re in right now.
0:11:25 – Oh wow.
0:11:28 – And maybe to paint some clarity on that.
0:11:30 In 2015, TechRunch basically wrote,
0:11:32 this is the year of AVs.
0:11:34 And in 2024 now, the quote is,
0:11:36 is this the year of AVs?
0:11:38 And you can see how there’s probably going to be
0:11:39 a lot of parallels to generative AI
0:11:42 where we are going to severely overestimate
0:11:44 what is going to happen in a year.
0:11:46 And one of the cool parts about generative AI
0:11:48 is how easy it is to make a demo,
0:11:50 but it is tremendously hard
0:11:51 to make something production ready.
0:11:52 And if you make a claim
0:11:54 that you are going to get rid of a developer,
0:11:56 that is a massive, massive claim.
0:11:57 And in fact, I would actually argue
0:12:00 that is a harder problem than autonomous vehicles.
0:12:01 Because ultimately,
0:12:02 if you look at autonomous vehicles,
0:12:04 all you need to do is press the accelerator
0:12:06 or decelerator or turn a steering wheel.
0:12:07 – That’s a great point.
0:12:08 – Think about the number of different things
0:12:09 a developer actually needs to do.
0:12:11 So I actually led a team
0:12:13 to build large scale deep learning infrastructure.
0:12:16 So how do you run these models at scale
0:12:18 and sort of left the company
0:12:19 with this vision of deep learning
0:12:22 and the idea of running these large models
0:12:24 was going to affect many, many industries.
0:12:25 And we had a small team of people.
0:12:28 We had eight people managing upwards of 10,000 GPUs.
0:12:32 We managed close to 20 to 30% of an entire data center.
0:12:34 And we worked with a lot
0:12:36 of these large autonomous vehicle companies
0:12:37 when we started this company, Exafunction.
0:12:38 Because our mission was,
0:12:41 how do we make it easier to run deep learning models?
0:12:42 But what ended up happening?
0:12:44 This is where startups can always get disrupted.
0:12:46 And it sounds silly to get disrupted.
0:12:48 We were making seven figures in ARR,
0:12:51 but we realized actually most of the models
0:12:53 would probably become these transformer based models.
0:12:56 And these are the models that underpin the GPTs,
0:12:57 the models that open AI has.
0:12:59 – What is a transformer based model?
0:13:00 If you could explain for us, yeah.
0:13:02 – So the basic idea is,
0:13:04 so we now know of prompting, right?
0:13:06 You know, use chat GPT, you pass in a prompt
0:13:10 and notice how it streams tokens one at a time, right?
0:13:12 That’s actually a property of these models
0:13:13 that are called transformer models,
0:13:15 that they are what are called auto aggressive.
0:13:17 They actually generate one token at a time.
0:13:19 And this is very different
0:13:21 than a lot of other classification tasks in the past
0:13:22 where you pass in an input
0:13:25 and it just gives you the entire answer all in one shot.
0:13:28 But this actually like slowly generates the entire thing,
0:13:30 sort of one word, one token at a time.
0:13:31 And we started noticing actually
0:13:34 that that was how a lot of models were starting
0:13:37 to look like once open AI came out with GPT-3.
0:13:39 And the beautiful thing about the model
0:13:41 that is truly crazy is because of the way it is trained,
0:13:43 it can do it in an unsupervised way.
0:13:45 So one of the things that was very different
0:13:49 about models in the past is you needed a lot of label data.
0:13:51 But these models are trained on the entire public internet.
0:13:53 The label data is the internet.
0:13:55 So because of that, you suddenly got these models
0:14:00 that could take in basically trillions of tokens of code
0:14:03 or text and this was not possible in the past
0:14:05 and this created these new sort of generative models.
0:14:06 And in the middle of 2022,
0:14:09 we had this business that was a GPU virtualization business.
0:14:11 The idea was we made it simpler
0:14:14 to run applications on GPUs.
0:14:16 And we found out that most applications
0:14:19 would probably be these transformer models.
0:14:21 And if all of what we were doing
0:14:22 was running transformer models,
0:14:24 we would largely become a commodity
0:14:26 because they would become a race to the bottom, right?
0:14:28 It would be the equivalent of asking,
0:14:31 like they would ask Varun how cheaply can you run this model?
0:14:33 I’d say I can do it for a dollar.
0:14:34 Then they would ask Jeff how cheaply you can do it.
0:14:36 He’d say 50 cents and we’d go back and forth
0:14:37 until no one made any money.
0:14:40 And this is the commodification of this entire space.
0:14:43 But what we did see was we felt that this technology
0:14:45 would be like the early coming of the internet.
0:14:47 There would be a brand new set of applications
0:14:48 that would be created.
0:14:50 And we were early adopters of GitHub’s product,
0:14:52 GitHub Copilot, and we thought that that was just scratching
0:14:55 the tip of the iceberg of what the future would look like.
0:14:57 And that’s where Codium sort of came about.
0:15:00 But it was, as you can imagine, a very rough experience
0:15:03 because we basically said buy to all the revenue that we had
0:15:05 and we had to start all the way back down to zero.
0:15:10 So, I mean, that to me is like the mother of all pivots
0:15:14 where not only are you changing the entire business
0:15:16 as you know it, you’re also doing it at a time
0:15:17 when things are going really well.
0:15:20 I mean, you said seven figures ARR.
0:15:22 My understanding is that you’d also raise $22 million
0:15:24 for this company.
0:15:26 Like things are going right.
0:15:30 And then you turn around and you tell all your employees,
0:15:32 actually would scrap that.
0:15:34 We’re going to do a whole different thing.
0:15:38 If I were a software engineer who worked at Big Tech
0:15:40 and had quit to go work at ExoFunction
0:15:45 and the CEO told me that, I’d be a mixture of like pissed off,
0:15:46 freaked out, concerned.
0:15:49 How did you rally the team
0:15:52 and how were you able to make that pivot so successfully?
0:15:53 – So I’ll say a couple of things
0:15:55 about the composition of the team.
0:15:57 Largely they were people that we knew.
0:15:58 And that’s actually very important
0:16:00 because they would be people that would go
0:16:02 into the trenches with us.
0:16:04 They were people that knew the caliber of people
0:16:07 that both me and my co-founder were.
0:16:08 And also on top of that,
0:16:10 we picked a problem space
0:16:13 where we were all passionate about it.
0:16:15 And I 100% knew at the time
0:16:19 there were products like mid-journey that were taking off.
0:16:20 Small team of people, eight people
0:16:23 that were making tens of millions of dollars in revenue.
0:16:24 And frankly speaking,
0:16:26 when we decided to pivot the company,
0:16:27 we knew for a fact,
0:16:29 none of us were that passionate about image generation
0:16:31 despite the fact that it is a very cool area.
0:16:32 And if we had picked it,
0:16:35 our team wouldn’t have had I guess the mental fortitude
0:16:38 to dig deep enough to actually build the problem space.
0:16:41 And then I guess the sort of third part
0:16:43 is actually that we actually were able to take
0:16:45 a lot of the infrastructure expertise
0:16:46 that we had as a company
0:16:48 to actually go out and build the application
0:16:49 significantly faster.
0:16:51 We were very quickly able to train our own models
0:16:52 and run them at massive scale.
0:16:54 And right now, Codium is one of the top five
0:16:56 largest generative AI apps
0:16:57 in terms of text in the world.
0:17:00 And that largely is because the original composition
0:17:02 of the team was these people
0:17:05 who are effectively GPU infrastructure experts.
0:17:07 But all said and done, everything that I said,
0:17:11 it comes down to you need a very truth seeking company.
0:17:13 And at the time,
0:17:15 even if we were at seven figures in revenue,
0:17:18 I did not know how we would 10x the amount of revenue.
0:17:20 And we could continue to lie to ourselves
0:17:23 and have a slow but certain death as a company
0:17:25 as the technology commoditizes, right?
0:17:27 And we all run the same kind of models.
0:17:28 Or we could just say,
0:17:31 there is a high probability that we will die,
0:17:33 but there’s a space that we could be very passionate about
0:17:34 and it could be very big.
0:17:37 I think we just decided the latter was
0:17:39 the more rational choice it is very hard to make.
0:17:43 But in retrospect, it was the obvious choice, right?
0:17:45 – Did you have any data showing you
0:17:47 that you were going to crash and burn at that point?
0:17:48 Or was it just a hunch?
0:17:50 And part of the reason I asked that
0:17:53 is because if I were your investor,
0:17:55 I would wanna be like, oh yeah, yeah.
0:17:58 It’s very clear to me that you guys have to pivot.
0:18:01 – We were cashflow positive then.
0:18:02 We were cashflow positive then.
0:18:04 – So you just had a feeling?
0:18:06 – We had a feeling.
0:18:08 It’s just because, and this is a little bit of a curse
0:18:11 of being a venture capital, venture-based business.
0:18:13 If you’re making millions of dollars in revenue,
0:18:15 that is not a venture-backable business.
0:18:19 And if we can’t figure out a path to get that to 100,
0:18:21 then that is not a business that we could build.
0:18:25 We could continue to keep it as an eight to 10% team.
0:18:26 We have another mentality in the company
0:18:27 beyond being truth-seeking
0:18:29 that we are a very lean company.
0:18:30 By the time we raised our Series B,
0:18:32 we had barely spent our seed round.
0:18:35 And I think that’s just because we don’t think capital
0:18:38 is a limiting factor in building a good business.
0:18:41 It’s you have to build a great product that customers love.
0:18:43 And that is usually not just you had more money.
0:18:46 – We’ll be right back.
0:18:48 (upbeat music)
0:18:51 (upbeat music)
0:19:02 – Support for the show comes from Into the Mix,
0:19:05 a Ben and Jerry’s podcast about joy and justice
0:19:07 produced with Vox Creative.
0:19:10 Would you have $25,000 to post bill?
0:19:13 That’s how much Inez Bordeaux had to pay
0:19:15 when she was arrested in 2016.
0:19:17 And since she couldn’t afford it,
0:19:18 she was sent to the workhouse,
0:19:21 a pretrial detention center in St. Louis.
0:19:23 Inez and the other detainees weren’t locked up
0:19:26 because they’d been convicted,
0:19:29 but because they couldn’t afford their bail.
0:19:31 – Experiencing what I experienced
0:19:34 and watching other women go through it
0:19:38 and know that there were thousands before us
0:19:41 and there were thousands after us
0:19:46 who had experienced those same things.
0:19:48 That’s where I was radicalized.
0:19:50 – She spent a month at the workhouse
0:19:52 and witnessed abject conditions,
0:19:55 extreme heat and cold, mold and pest infestations
0:19:57 and poor medical care.
0:19:59 Eventually her charge was vacated,
0:20:03 but the experience changed her.
0:20:04 – They’re starting a campaign to close the workhouse.
0:20:06 Are you interested?
0:20:09 And I was like, hell yeah, hell yeah, I’m interested.
0:20:11 – Hear how she and other advocates
0:20:13 fought to shut it down and won.
0:20:16 – On the first episode of this special three-part series
0:20:18 out now, hosted by Ashley C. Ford,
0:20:21 subscribe to Into the Mix, a Ben and Jerry’s podcast.
0:20:34 – This ad was expressly recorded
0:20:37 to create a sense of simplicity.
0:20:42 Just a few simple sounds, no complexity.
0:20:46 – Hmm, like neutral.
0:20:51 Made with just vodka, soda and natural flavor.
0:20:54 Neutral, refreshingly simple.
0:20:59 – When summer brings the heat, McDonald’s brings the chill.
0:21:02 During summer drink days, enjoy a small iced coffee
0:21:05 or a refreshing Coca-Cola for $1 plus tax
0:21:07 and step up your summer today.
0:21:09 At participating McDonald’s restaurants in Canada,
0:21:10 prices exclude delivery.
0:21:13 (mouse clicks)
0:21:18 – We’re back with First Time Founders.
0:21:20 One of the things you mentioned
0:21:23 is that you are training your own models.
0:21:25 I mean, this is an AI application,
0:21:28 but most AI applications that I am aware of,
0:21:32 they’re purely building the application layer,
0:21:35 and that is they’re basically using someone else’s model,
0:21:38 usually open AI, and they’re tweaking
0:21:39 and they’re building off of that model
0:21:41 and creating their own application.
0:21:42 You guys are different.
0:21:44 It sounds like you guys are,
0:21:45 I don’t know what you’d call it,
0:21:48 full stack AI from the model to the application.
0:21:49 Is that right?
0:21:50 – That is correct.
0:21:52 Actually, even at the kernel layer,
0:21:55 we’ve done some, like we’ve written some code
0:21:57 even at the infrastructure layer
0:21:58 so that we could set the models on top
0:21:59 in an efficient manner.
0:22:01 And the reason we have to do that
0:22:03 is because of the latency issues we talked about.
0:22:05 For example, if you are a passive AI
0:22:07 and you take, let’s say one second
0:22:09 to show up the suggested code,
0:22:10 people are just going to stop using that.
0:22:12 They don’t want to get out of flow state
0:22:14 and pause and wait for results.
0:22:18 – How many other AI startups are doing,
0:22:19 what should we call it, the full stack,
0:22:23 the infrastructure to application?
0:22:24 Are there many others?
0:22:26 I just, I mean, off the top of my head,
0:22:29 I’m like anthropic like open AI,
0:22:32 but I guess they’re mainly kind of infrastructure layer, right?
0:22:34 I mean, isn’t this super rare?
0:22:36 – I think in my mind,
0:22:39 there’s maybe a couple of unique things about code
0:22:41 that make it so that you can actually do this here.
0:22:42 And you’re totally right.
0:22:44 Most of the large companies that are even successful
0:22:46 are largely built on top of an API,
0:22:49 but we genuinely felt to build a best in class app here.
0:22:51 We needed to become vertically integrated.
0:22:55 And for us, it was also not a complex thing for us to do
0:22:58 in that we have the technical talent inside the company
0:22:59 to actually go ahead and do that.
0:23:02 Maybe one of the unique aspects of code
0:23:05 on why we can do this is code can actually be run.
0:23:08 Like let’s say I am a illegal AI tool
0:23:10 and I’m redlining a bunch of documents.
0:23:12 The only way to know if that is good
0:23:14 is for a human to go in afterwards
0:23:16 and actually take a look at it.
0:23:17 For code, you can actually,
0:23:19 if you make an edit to a code base,
0:23:21 you can actually run the code and validate
0:23:22 it is doing the right thing
0:23:24 without a human in the loop at all.
0:23:26 And what that means is there are ways in which
0:23:29 you can close the loop in intelligent ways
0:23:30 that you can actually,
0:23:32 if you specialize on that application
0:23:34 and you are vertically integrated,
0:23:37 you can build an even better app for code.
0:23:40 And we’ve taken advantage of this in many, many ways.
0:23:42 And we realized if we didn’t do this,
0:23:43 we’d be shooting ourselves in the foot.
0:23:45 And I’ll give you even a simple example
0:23:47 of how this manifests itself.
0:23:49 Right now, I mentioned this.
0:23:51 Codium is, processes over a hundred billion tokens
0:23:53 of code every day, which is over 10 billion lines
0:23:54 of code every day.
0:23:56 If we passed that through OpenAI,
0:23:57 we would have gone bankrupt.
0:24:00 And there was a recent article from the Wall Street Journal.
0:24:01 – And why is that, sorry,
0:24:02 because you would have to pay for it.
0:24:03 – Because it would be too expensive.
0:24:04 – Exactly.
0:24:08 – And not the best for our particular application
0:24:09 on top of that.
0:24:11 And if you look, there was a recent article
0:24:13 on the Wall Street Journal about how GitHub Copilot
0:24:16 was spending tens of dollars per user per month.
0:24:18 And that’s actually because even GitHub,
0:24:20 Microsoft’s product is not vertically integrated.
0:24:22 They are relying on external models
0:24:24 to build their application.
0:24:27 And we view that as, hey, the model and the product
0:24:28 and the infrastructure are so critical
0:24:30 to delivering a great experience,
0:24:32 why would we not have control over every piece of that?
0:24:35 So my understanding is you guys don’t use
0:24:38 the actual data of your individual users.
0:24:42 So how is the model getting trained then?
0:24:45 – So we actually do sort of two different things.
0:24:48 And I’ll let Jeff add on how this affects our enterprises
0:24:50 and the customers that use the product.
0:24:52 So first of all, we use permissively licensed code
0:24:54 that is available on the public internet.
0:24:56 And we also attribute it on generation time.
0:24:59 So we actually take sort of copyright
0:25:01 and licensing very seriously as a company.
0:25:03 But on top of that, when we release product
0:25:04 from our user data, you’re right,
0:25:06 we don’t take the data from our users,
0:25:08 like let’s say when they’re auto-completing stuff
0:25:10 and copy that code and put it into our training set.
0:25:13 But we can see, hey, users are accepting
0:25:14 these types of suggestions more
0:25:16 and these types of suggestions less.
0:25:19 So we have preferences on what users and humans like.
0:25:21 And that actually informs us to actually build products
0:25:23 that are better and more willing to use.
0:25:25 And then that actually has a virtuous cycle
0:25:26 and that now people are willing to try
0:25:28 more complex things on our product
0:25:29 because the easier things they have,
0:25:31 high confidence that they work.
0:25:33 And suddenly the frontier of what we are actually able
0:25:36 to experience as we are able to give the user
0:25:37 increase more and more,
0:25:39 largely because we have a product that is so well beloved.
0:25:42 We now have over 600,000 users that use our product.
0:25:43 – Yeah, it’s unbelievable.
0:25:45 – I think one thing we might have glossed over
0:25:48 in the beginning was codeon made a very conscious decision
0:25:51 to make the product free for individuals.
0:25:53 And if you think about, you know,
0:25:57 Varun just said co-pilot loses $10 to $20 a month per user.
0:26:00 That’s after they’ve been paying a subscription to, right?
0:26:02 So having that infrastructure background,
0:26:05 being able to make it efficient to deploy these models
0:26:07 and then giving it out for free for individuals,
0:26:09 allowed us to build this very large user base,
0:26:12 probably the largest user base for a coding assistant.
0:26:14 – Yeah, I mean, just the data here,
0:26:17 you started out last year with less than 1,000 users,
0:26:19 by the end of the year, you had half a million.
0:26:21 – Yeah, and we have over a million,
0:26:23 millions of downloads across all the plugins.
0:26:24 And the reason why that’s very important
0:26:26 is because of what we just talked about.
0:26:27 If we roll out multiple models,
0:26:29 if we are changing the temperature
0:26:31 of the thresholds here and there,
0:26:33 we are getting so many signals
0:26:35 as to what is the appropriate settings to tweak.
0:26:38 And I think every hour we’re getting like a million signals.
0:26:41 So we can run all these experiments on these free users
0:26:43 of all these models we train
0:26:44 to really, really get the best results
0:26:45 that you could possibly get.
0:26:48 And then only when we’ve validated like, okay,
0:26:49 we’ve trained this model,
0:26:51 this is better than all the other ones we’ve trained.
0:26:53 These are the settings that makes the best results.
0:26:55 Then we could deploy that to our on-prem
0:26:57 or enterprise users, right?
0:26:59 Because we can’t, after we’ve deployed it,
0:27:00 we can’t really get that much more information from it.
0:27:01 It’s completely,
0:27:04 it can be hosted even in an air-gapped environment.
0:27:06 So I think that’s a big element
0:27:07 of why we are successful
0:27:09 of being able to deploy these models.
0:27:12 Because if somebody’s trying to start from scratch right now,
0:27:14 how are they gonna know that their model is good?
0:27:16 How are they gonna tweak the model, right?
0:27:18 Without a very large user base.
0:27:19 So that’s part of the secret sauce
0:27:21 is having that free user base.
0:27:24 – Yeah, and you, it’s still free,
0:27:26 which is what I find pretty fascinating.
0:27:28 And you said, I was just reading your state,
0:27:30 your like mission statement,
0:27:33 you are quote committed to having a free tier forever.
0:27:36 The natural next question is,
0:27:39 how is it gonna get properly monetized?
0:27:41 And how are you gonna maintain that,
0:27:43 that free tier into perpetuity?
0:27:45 – I think people underestimate kind of the demand
0:27:47 for both just the on-prem instance,
0:27:49 but also some of the teams features we add
0:27:50 on the SaaS product.
0:27:52 So we have a free individual tier,
0:27:56 but if you create a team and add on-board users to it,
0:27:57 that is a paid product.
0:27:59 And there are things like analytics
0:28:01 and actually bigger models and seat management
0:28:04 that are going to be better than the free product.
0:28:06 But we are committed to making the free product
0:28:08 the best coding assistant out there, no matter what.
0:28:10 So whatever other coding assistants come to market,
0:28:13 we will make sure our free product is still the best one.
0:28:14 We wanna make sure, you know,
0:28:16 disincentivize others from entering,
0:28:18 but we want people that always have the option.
0:28:20 We don’t want them to get forced to buy a co-pilot,
0:28:21 for example.
0:28:22 – And then maybe one thing to add to Jeff,
0:28:24 like we monetize enterprises, right?
0:28:27 So we have some of the largest Fortune 100s,
0:28:30 F-thousands, even over 10,000 developers on our product.
0:28:33 And those companies, obviously they want security guarantees,
0:28:34 they want personalization
0:28:36 for many, many repositories that exist.
0:28:37 And they also want support
0:28:40 across all source code management tools.
0:28:44 Less than 10% of Fortune 500 companies are on GitHub cloud.
0:28:45 And that is the competitor that we have,
0:28:46 that is GitHub co-pilot.
0:28:48 And they have committed to making their product
0:28:52 differentially better if you are on GitHub cloud, right?
0:28:55 So we want to take the approach of almost being Switzerland.
0:28:58 We don’t care what programming language you write.
0:29:00 We don’t care what ID is you use.
0:29:03 We don’t care where you store your source code.
0:29:06 And ultimately, we also just don’t really care
0:29:07 what seniority the developer is.
0:29:10 We will provide the maximum amount of leverage there.
0:29:13 Whereas a lot of the larger players in the space
0:29:15 are focused on being tied to another brand.
0:29:17 And the reason why we don’t think that that makes sense is
0:29:20 we think AI is such an up-leveler.
0:29:23 We think it deserves to be in a category of its own.
0:29:26 – You recently raised $65 million
0:29:29 in a round led by Kleiner Perkins.
0:29:32 It valued you at half a billion dollars.
0:29:35 Congratulations.
0:29:37 What is that money going to be used for?
0:29:39 – I think the way we would like to think about it is
0:29:41 we have ways of spending cash,
0:29:42 not only to train models,
0:29:44 but to also make it so that we can build
0:29:46 a better user experience for the end user.
0:29:48 But one of the cool things for us is our product
0:29:50 has such high ROI that we think that there will be
0:29:52 a real payback period on that.
0:29:55 Enterprises and companies will see enough value there
0:29:57 that they will be able to eat the cost
0:29:58 that we will need to spend upfront.
0:29:59 But also on top of that,
0:30:01 we want to spend a lot on making sure
0:30:03 that we can become better partners for our customers.
0:30:05 We’re onboarding some of the world’s largest companies
0:30:08 and we’re onboarding tens of thousands of developers.
0:30:09 That’s going to take a little bit of effort
0:30:11 to make sure that we do that properly.
0:30:13 One company that we’re working with right now
0:30:15 that we have a multi-year engagement with,
0:30:18 they account for 0.15% of all developers in the world,
0:30:20 it’s just that one company, right?
0:30:23 So, I’m a little bit of a different type of founder
0:30:24 in that I do not like the idea
0:30:26 of spending money unnecessarily,
0:30:28 but if it comes down to we are doing it
0:30:29 because we make our customers more successful
0:30:32 and our users more happy, we’ll do it any day.
0:30:35 – How have you guys restrained yourselves
0:30:36 in terms of spending?
0:30:40 Because I mean, yeah, the story, the narrative in AI
0:30:42 has been that it is an arms race.
0:30:45 And the thing that I hear about in the venture industry,
0:30:47 or at least what AI founders are being told,
0:30:50 is just go out and raise a shit ton of money
0:30:53 as much as is even possible.
0:30:56 One, because you just want to develop a war chest.
0:31:00 And two, you want to get the headlines.
0:31:02 You want to be the AI company that’s working
0:31:05 on co-generation, the AI company for finance,
0:31:06 whatever it is.
0:31:09 So I guess sort of two questions for me here.
0:31:12 One, is that accurate to your experience?
0:31:16 And two, how have you been so responsible
0:31:19 in terms of spending while you see all of these headlines
0:31:21 of other companies spending so much money
0:31:23 on talent and training their models?
0:31:26 – I think Varun mentioned earlier that we run very lean.
0:31:28 And the reason we’re able to do that
0:31:31 is we hire people that are generalists or ex-founders
0:31:33 and they’re capable of doing many job roles at once.
0:31:37 So our company size is probably like a little misleading.
0:31:39 It’s probably way more effective,
0:31:41 not just infrastructure, but the headcount also.
0:31:42 – What is the headcount?
0:31:44 – So right now we’ll be almost about 55
0:31:45 at the end of the month, I think.
0:31:47 And then the thing is like the people we hire
0:31:49 are able to just slot into different roles,
0:31:50 almost out of moments notice.
0:31:52 It’s like, oh, we don’t even have a marketing team,
0:31:53 for example.
0:31:55 The product’s growth has been organic,
0:31:57 but we do want to do some marketing experiments
0:31:59 to make sure that we are getting ourselves out there,
0:32:00 just as an example.
0:32:02 And then people, randomly someone will be like,
0:32:04 I have an idea, okay, go do it.
0:32:07 And then we all of a sudden have a Google ad strategy.
0:32:09 All of a sudden we have a bunch of blog posts.
0:32:12 We’re on podcasts like with you and Elson.
0:32:14 But I think that my point is,
0:32:16 part of the function of not spending the money
0:32:19 is just being very, I guess, practical
0:32:21 of where the money goes and running lean.
0:32:23 And I think when there is a moment that says like,
0:32:25 hey, we need to train a much bigger model
0:32:26 and we need to spend this much money,
0:32:28 we’re all for it, actually.
0:32:31 We actually are very conscious about what the ROI is
0:32:32 of everything we do.
0:32:35 – Yeah, how do you maintain that culture of leanness
0:32:38 and what would be your recommendation to other companies
0:32:41 that are maybe not as big as you, but trying to be?
0:32:44 – This is quite probably the hardest problem
0:32:46 that we’re trying to solve right now.
0:32:49 ‘Cause we wanna hire people that are very good,
0:32:52 very technical, maybe they’re generalists,
0:32:54 like I said earlier, but it’s very hard
0:32:55 to hire those people.
0:32:57 So actually, this is probably one of the things
0:32:59 we focus on in the next month
0:33:01 is what is our recruiting strategy?
0:33:03 How do we hire the best people?
0:33:05 Maybe part of that is getting Kodium’s brand name
0:33:06 much more aware.
0:33:09 Maybe it’s like a big push of user adoption.
0:33:11 Maybe we’re just gonna have to be scrappy
0:33:14 and be very creative of how we hire people.
0:33:15 For example, I don’t know if other companies
0:33:18 are just like pinging every ex-founder on LinkedIn,
0:33:19 but we are, right?
0:33:23 So we are trying to scale with creative means.
0:33:25 – One other thing is like as a company,
0:33:28 I think culture is, we have some cultural principles
0:33:30 and we run lean as one of them,
0:33:33 but who would wanna say you don’t run lean?
0:33:36 So I think, how do you actually live that out?
0:33:38 We are a five days a week in-person company,
0:33:40 so we don’t do remote work.
0:33:42 We don’t really do hybrid work either.
0:33:45 So people see what it’s like to work at the company
0:33:49 and until very recently, our CTO was ordering snacks.
0:33:51 And that’s not to say that’s a great use of his time.
0:33:53 It’s more just that no one is high enough
0:33:56 to not do some work to test a hypothesis.
0:33:58 And we don’t hire specialists at the company
0:34:01 until generalists outgrow that role.
0:34:03 And I’ll give you another example of this.
0:34:06 When we ran that GPU virtualization company,
0:34:07 even though we were making money,
0:34:09 we never hired a sales rep.
0:34:11 And that’s not because I don’t believe in enterprise selling.
0:34:13 No, we have a great VP of sales now at the company
0:34:15 and I think we might have in terms of talent density,
0:34:18 one of the strongest enterprise sales teams in the world,
0:34:20 actually, but why didn’t we hire someone?
0:34:23 I just didn’t believe that if we added one new person,
0:34:25 I would be setting them up for success
0:34:28 because the reality is if I could not get $1 of additional
0:34:33 sales, I cannot expect someone else to get $10.
0:34:35 So this is one of those things where we have a mentality
0:34:38 of we try to do it ourselves and then we try to eliminate
0:34:39 ourselves from the role.
0:34:41 We give away our Legos.
0:34:43 We let someone else take that over that understands
0:34:46 the role much more, but we don’t do things prematurely.
0:34:50 And I think there’s a tendency across people that the idea
0:34:52 of building a scalable organization is really valuable.
0:34:54 And I do see that that you do want to build
0:34:57 a scalable organization, but sometimes people get too excited
0:35:00 about this notion of org building or fundraising rather
0:35:03 than the idea of having customers, having users.
0:35:05 Because ultimately, people that join our company
0:35:07 don’t care about how much money we raised,
0:35:09 as long as we will survive.
0:35:11 And our customers genuinely don’t care.
0:35:12 Let’s look at it this way, right?
0:35:14 If you look at a company as big as JPMC
0:35:16 that makes hundreds of billions of dollars,
0:35:20 to them, does it make sense if we raised $100 or $200 million?
0:35:21 It all looks like peanuts to them.
0:35:23 It’s all like 10 basis points of the amount of revenue
0:35:25 that they make a year.
0:35:26 So to them, what they really care about
0:35:28 or companies of this size is,
0:35:30 are we the best partner for them?
0:35:32 Are we the best product to them?
0:35:34 And as long as we’re laser focused on that,
0:35:36 we should do whatever it takes to build that up.
0:35:39 We’ll be right back.
0:35:50 – Looking to buy or sell a used vehicle?
0:35:52 Forget the selection and protection
0:35:53 of BCAA auto marketplace.
0:35:55 Try some random guy off the internet.
0:35:57 – Some random guy.
0:35:59 – Yes, we offered zero security.
0:36:01 – I can pay in crypto.
0:36:02 – Plenty of questions.
0:36:04 – Is it still available?
0:36:06 – And zero convenience.
0:36:09 – 46 messages and I still make eye.
0:36:12 – BCAA auto marketplace helps you buy and sell vehicles.
0:36:17 Visit bcaa.com/marketplace and avoid some random guy.
0:36:21 – Pandora, be love.
0:36:24 What does be love mean to you?
0:36:26 – I definitely would say my be love role model
0:36:28 is for sure my sister.
0:36:30 Unconditional, infinite love.
0:36:32 Something that is never ending that you know
0:36:33 is always there.
0:36:34 – Never questioned.
0:36:36 – Never questioned.
0:36:38 No matter if you fall off a cliff,
0:36:41 she’s there to catch you, you know?
0:36:42 Be love.
0:36:44 – Shop now at pandora.net.
0:36:48 – This is an ad by BetterHelp.
0:36:51 What are your self-care non-negotiables?
0:36:53 It’s hard to make time for the things that keep you healthy
0:36:57 but being consistent with self-care is like working a muscle.
0:37:00 And when life gets crazy, that muscle keeps you strong.
0:37:02 Therapy is the ultimate self-care
0:37:04 and BetterHelp makes it easy to get started
0:37:07 with affordable online sessions you can do from anywhere.
0:37:10 Never skip therapy day with BetterHelp.
0:37:12 Visit betterhelp.com to learn more.
0:37:14 That’s better H-E-L-P.com.
0:37:22 – We’re back with first time founders.
0:37:24 Sort of a more personal question.
0:37:29 I mean, you started this company in 2021
0:37:31 sort of just as the AI hype was bubbling up
0:37:35 and you now find yourself at the epicenter
0:37:37 of the hardest industry
0:37:39 and you are one of the hardest companies
0:37:41 in the hardest industry.
0:37:43 Just a personal question for both of you.
0:37:44 How does that feel?
0:37:47 Like what has it been like getting used to
0:37:50 being the guy in AI?
0:37:52 – I told this to the company.
0:37:56 I tell everyone just get ready to get destroyed.
0:37:59 Assume that something very bad is going to happen always.
0:38:02 And this is where us having gone through that pivot
0:38:03 is very critical.
0:38:05 Things are going very well for us as a company.
0:38:07 A lot of the reason why we haven’t spent a lot of money
0:38:09 is now we make money, which is a unique property
0:38:12 about a lot of companies apparently in this space
0:38:13 where most companies talk about vision
0:38:15 rather than actually building a product that people use.
0:38:18 But I just, I tell everyone get ready
0:38:20 for something really bad to happen.
0:38:22 And this is why it’s like very important
0:38:26 that we hire people that are truly in it for the long run.
0:38:28 And I tell people this when they joined the company,
0:38:30 I think we could be a company
0:38:32 that’s worth over a hundred billion dollars.
0:38:33 I think we can.
0:38:35 And that’s largely because the total adjustable market
0:38:37 of what we’re building
0:38:38 and the amount of impact that this can have
0:38:41 given how important technology is could be massive.
0:38:44 But also a series of bad decisions that we make
0:38:46 could completely kill the company.
0:38:48 And that will happen very fast.
0:38:49 And I think this is where us building
0:38:51 a very truth seeking company.
0:38:53 And that actually is very hard
0:38:55 because people want to believe
0:38:56 that what they’re doing is correct.
0:38:59 And they want to embrace psychological safety.
0:39:01 And what I tell people is, hey,
0:39:04 if you feel something is wrong,
0:39:07 lean into what you think is wrong and tell everyone.
0:39:10 Tell everyone, because we should not have pockets of people
0:39:12 that want to report up the chain
0:39:14 and tell their manager or tell me,
0:39:16 we have a very flat company or tell me,
0:39:17 things are going fine.
0:39:19 I would much rather hear everything is on fire
0:39:21 and have paranoid people at the company
0:39:23 than people who are just happily going to work.
0:39:25 And this is why I think startups
0:39:27 are so much harder than big companies.
0:39:29 It’s actually not that, you know,
0:39:31 you’re taking a massive risk on the monetary side.
0:39:34 You still can make a six figure salary, right?
0:39:36 These are not people that are living hand to mouth.
0:39:40 But the really hard part is the lack of psychological safety.
0:39:42 We make a bunch of people at the company
0:39:43 make a series of bad decisions
0:39:46 and the entire thing can go to zero.
0:39:48 Whereas if you’re at Google,
0:39:49 you have very little accountability.
0:39:51 If your team doesn’t perform well,
0:39:55 Google makes so much money that you are a rounding error.
0:39:57 You will be shuffled to some other part of the company.
0:39:59 You never really need to deal with the impact
0:40:02 of your decisions and consequences of your decisions.
0:40:04 And that’s why it’s always a little bit of a funny statement
0:40:06 when someone at a big company is like,
0:40:08 I work at a startup at a big company.
0:40:10 No, no you don’t.
0:40:14 Imagine the idea of you potentially losing your job
0:40:16 every quarter or every month.
0:40:17 Exactly.
0:40:19 One thing that the way I think about it
0:40:21 is radical transparency.
0:40:23 And we have a lot of conversations within our company
0:40:26 of let’s be super transparent about everything.
0:40:28 And even in my personal life,
0:40:30 I’m like, I’m just gonna be like totally upfront.
0:40:35 That’s the cleanest, most kind of hygienic way to operate.
0:40:41 Do you ever feel that there could be too much truth?
0:40:44 Do you feel that there’s a possibility
0:40:46 that if you’re encouraging everyone
0:40:48 to tell the truth, tell the truth, be transparent,
0:40:52 tell me everything, that they’ll kind of go overboard?
0:40:54 I think this is the hardest part about a startup.
0:40:57 The two things that I say is a startup is really hard
0:41:00 because you need to be both irrationally optimistic.
0:41:02 Because if you’re not optimistic,
0:41:04 the answer is always Microsoft is gonna beat you, right?
0:41:04 Right.
0:41:06 It’s the biggest company of all time.
0:41:07 They have the most capital.
0:41:08 They have the most people.
0:41:09 They have the most distribution.
0:41:10 Why does any company win?
0:41:11 But clearly that’s wrong, right?
0:41:12 There are companies that have beat Microsoft
0:41:13 at different areas.
0:41:14 We always say this,
0:41:16 oh, they have the most capital they’re gonna win.
0:41:18 Yes, yes, it’s very easy from my perspective to say that.
0:41:19 Exactly, right?
0:41:22 And then HBS, I try to think about what Harvard Business
0:41:25 School would say 10 years from now for us as a company.
0:41:28 And I tell people this, it’s like, we fail.
0:41:31 And what they write is in a world
0:41:34 in which technology was changing,
0:41:36 Microsoft had all the distribution in the world.
0:41:38 They had this property called GitHub.
0:41:40 It was inevitable that they would win.
0:41:41 And because of that, they won.
0:41:42 They won the entire space.
0:41:45 And all of these startups, it was a fool’s errand.
0:41:46 Why did they even try?
0:41:48 And then the world in which we win,
0:41:50 they were going to write in a world
0:41:54 in which the technology was getting disrupted so materially.
0:41:56 There were a set of companies,
0:41:57 and one of which was Codium,
0:42:00 that had such a technological advantage
0:42:02 that there was no way a slow-moving gorilla
0:42:03 like Microsoft could even compete.
0:42:05 So what I think is very hard–
0:42:06 – Very hard, yeah.
0:42:07 – Yeah, exactly.
0:42:08 They will write whatever the future looks like
0:42:10 and the history will be written by the victor.
0:42:12 And no one will know exactly
0:42:14 what the pages of the book look like.
0:42:18 And I think the hardest part about a startup is,
0:42:19 you do need to be irrationally optimistic
0:42:21 to believe that you can win.
0:42:24 Because by default, if you don’t, you will definitely lose.
0:42:26 But then also uncompromisingly realistic,
0:42:29 which is that sometimes, actually,
0:42:33 there’s no point continuing in a particular direction.
0:42:36 This is where I think it comes down to what you just said,
0:42:38 which is that if you were too truth-seeking
0:42:41 and everyone is paranoid all the time,
0:42:43 it could lead to paralysis.
0:42:45 And I think there’s a fine line there.
0:42:48 There’s a fine line where one time we lost a deal,
0:42:49 and we actually, every day for dinner,
0:42:52 talked about it for two weeks in a row for multiple hours.
0:42:54 We talked about the implications as a company.
0:42:55 And that was very useful.
0:42:58 But if we extended that out to an entire year,
0:43:00 we would not go anywhere.
0:43:01 – Yeah.
0:43:02 – So you’re totally right.
0:43:03 There’s a level here.
0:43:05 But what I do feel most companies do
0:43:08 is probably err on the side of not being truth-seeking enough.
0:43:11 They become too complacent with what they’re doing.
0:43:14 And I think the thing to really think about,
0:43:15 and this is not true at a big company,
0:43:17 which is why you really need to think about it
0:43:19 at a startup is you do not win an award
0:43:21 for doing the wrong thing for longer.
0:43:24 So the sooner you can rip the band-aid off,
0:43:25 the better your company will be.
0:43:27 You will be way happier that you did it.
0:43:30 It is going to be incredibly painful for a week,
0:43:31 but just do it.
0:43:32 – I love that.
0:43:34 You guys have been generous with your time,
0:43:36 so we’ll begin to wrap up here.
0:43:39 Jeff, I’ll ask you this question.
0:43:41 What do you think has been the greatest challenge
0:43:44 that this company has faced in the past couple of years?
0:43:48 – I think the pace at which things switch
0:43:50 is really challenging.
0:43:53 We relied on some technology features
0:43:54 to be the selling point.
0:43:56 And then our competitors would come out
0:43:58 and it’s like the same thing all of a sudden.
0:44:01 And every time that there’s a news release
0:44:02 for one of our competitors,
0:44:04 we immediately go into like a code red
0:44:06 and go into a competence room.
0:44:08 Reverse engineer what they’re doing.
0:44:09 And this is constant.
0:44:12 This is like every month this is happening.
0:44:13 The question is like, will this last forever?
0:44:16 Like will we just always be panicking?
0:44:18 And the question that answers probably yes, right?
0:44:21 – I think for people listening to this podcast,
0:44:26 you too are the closest thing to AI experts as we’ve had.
0:44:30 You’re kind of on the front lines of this.
0:44:33 What would be your advice to anyone
0:44:35 who is working in the AI industry
0:44:37 or who wants to work in the AI industry?
0:44:39 And I think that doesn’t just have to be founders
0:44:43 but engineers, product managers,
0:44:44 business operators, et cetera.
0:44:46 What’s the most important thing
0:44:49 that they should understand about AI right now?
0:44:51 – There’s an interesting property about AI right now
0:44:54 in that it is actually imperfect.
0:44:56 You know, when you use the products,
0:44:58 sometimes says the wrong thing.
0:45:00 And despite that, it is actually very useful
0:45:01 in some domains.
0:45:03 That’s not like anything in the past.
0:45:05 When you used the internet
0:45:08 and you ordered something off Amazon in 2002,
0:45:10 it’s not like they would ship you something incorrectly
0:45:12 or maybe they would do that, but that’s not,
0:45:13 there was no expectation going in.
0:45:16 You would buy one book and you would get a different one.
0:45:18 And somehow that is still fine right now,
0:45:20 which is a cool part of the technology,
0:45:22 which is that it actually gets perfect.
0:45:24 It’s going to usher in a brand new set
0:45:26 of applications as well.
0:45:29 But I think the key thing to really think about
0:45:31 is not to think about, go back from a demo
0:45:33 and try to build that today.
0:45:36 So think about what products can you build today
0:45:38 that are actually imperfect,
0:45:41 but still can generate a lot of value.
0:45:43 And that is a lot harder than you would think
0:45:46 because in a lot of domains, if you are imperfect,
0:45:49 like let’s say you’re reviewing a legal document
0:45:51 and it actually completely reviews the document,
0:45:53 but 10% of the time it’s wrong.
0:45:55 You can’t give that to an end customer.
0:45:58 So actually thinking about the trade-offs of
0:46:02 how good does the quality need to be to ship your product?
0:46:04 How fast does the experience need to be?
0:46:06 In a world in which the quality is not perfect,
0:46:07 it better be fast.
0:46:09 There’s no world in which I’m going to wait 24 hours
0:46:12 for something, and the quality is imperfect, right?
0:46:16 And then if the quality is imperfect and it is fast,
0:46:18 how quickly can I correct it?
0:46:20 And these are all important factors of a product.
0:46:22 If you don’t hit the sweet spot here,
0:46:24 you will have a product that’s a cool demo,
0:46:26 but no one will ever use it.
0:46:28 I think this is the most important thing
0:46:29 to really think about.
0:46:32 Adding AI to just any field that exists
0:46:34 doesn’t just suddenly make the product usable.
0:46:37 It needs to be a product that is useful in its own right.
0:46:40 – I think if you give away a product for free
0:46:42 and everybody keeps using it
0:46:45 and a lot of people keep trying to get it,
0:46:47 then you know there’s something of value
0:46:48 and there’s product market fit.
0:46:50 And I think if you look at a lot of these AI tools
0:46:52 and AI demos, maybe they’re not realizing that.
0:46:53 Maybe they’re just building something
0:46:56 that is a cool demo or it’s like a really,
0:46:58 it pushes the limits of the technology,
0:46:59 but it’s not actually building something
0:47:02 that people want or will find value from.
0:47:03 I think that’s maybe our biggest message
0:47:05 to other founders or other people
0:47:06 that are building products in the area.
0:47:08 It’s just think like if I gave this away for free,
0:47:11 is everybody gonna wanna use it and keep using it?
0:47:13 I think that’s maybe something people miss.
0:47:15 – Do you think maybe that’s the model that,
0:47:17 I mean, I was gonna say software founders,
0:47:19 but maybe all founders that you should just start
0:47:22 with giving the product out for free
0:47:24 and see where things take you from there.
0:47:26 – It worked really well for chat GBT, right?
0:47:27 – Yeah, yeah, exactly.
0:47:28 – And it worked well for you, yeah.
0:47:30 – That’s actually an interesting principle
0:47:31 that Jeff just said.
0:47:33 Obviously, in some scenarios,
0:47:35 the reason why chat GBT could do that and open it,
0:47:37 it could do that is they own the infrastructure.
0:47:38 And if another company did that,
0:47:39 they would have gone bankrupt.
0:47:43 So you do need advantages in particular places
0:47:44 to be able to do that, but at the very least,
0:47:47 if you did burn money and you gave it away for free,
0:47:49 if no one runs to use the product,
0:47:51 you’re probably in a world of trouble.
0:47:55 – That’s a great place to end.
0:47:57 Varun Mohan is the founder and CEO of Codium.
0:48:00 Jeff Wang is the company’s head of business.
0:48:01 Guys, thank you for joining us.
0:48:02 That was awesome.
0:48:03 – Yeah, thanks for having us.
0:48:04 – Thanks for having us.
0:48:09 – Our producer is Claire Miller,
0:48:11 our associate producer is Allison Weiss,
0:48:13 and our engineer is Benjamin Spencer.
0:48:15 Jason Stavis and Catherine Dillon
0:48:17 are our executive producers.
0:48:18 Thank you for listening to First Time Founders
0:48:20 from the Vox Media Podcast Network.
0:48:22 Tune in tomorrow for ProfG Markets.
0:48:25 (upbeat music)
0:48:50 – Support for the show comes from Into the Mix,
0:48:53 a Ben and Jerry’s podcast about joy and justice
0:48:55 produced with Vox Creative.
0:48:57 Into the Mix is back for a new season
0:48:59 and welcomes you in with four new stories
0:49:02 that take listeners beyond the headlines
0:49:04 and into the lives of ordinary people
0:49:07 fighting for justice in their communities.
0:49:09 Starting with Ainez Bordeaux,
0:49:10 an activist and St. Louis native
0:49:13 who fought to shut down the workhouse,
0:49:15 a notorious pretrial detention center
0:49:18 that she says functioned like a debtor’s prison.
0:49:21 Subscribe to Into the Mix, a Ben and Jerry’s podcast
0:49:23 to listen to the first episode
0:49:25 of this special three-part series out now.
0:49:35 [BLANK_AUDIO]
0:00:04 Support for the show comes from Into the Mix,
0:00:07 a Ben and Jerry’s podcast about joy and justice
0:00:09 produced with Vox Creative.
0:00:12 Ainez Bordeaux is a self-described hellraiser,
0:00:15 and she became an activist after being caught up
0:00:16 in the criminal legal system
0:00:19 when she couldn’t afford her bond.
0:00:20 And without a trial,
0:00:23 Ainez was sent to a St. Louis detention facility
0:00:25 known as the Workhouse,
0:00:28 notorious for its poor living conditions.
0:00:31 Here how she and other advocates fought to shut it down
0:00:33 and won on the first episode
0:00:36 of this special three-part series out now.
0:00:39 Subscribe to Into the Mix, a Ben and Jerry’s podcast.
0:00:45 Support for the show comes from Into the Mix,
0:00:48 a Ben and Jerry’s podcast about joy and justice
0:00:50 produced with Vox Creative.
0:00:53 Ainez Bordeaux is a self-described hellraiser,
0:00:56 and she became an activist after being caught up
0:00:58 in the criminal legal system
0:01:00 when she couldn’t afford her bond.
0:01:01 And without a trial,
0:01:04 Ainez was sent to a St. Louis detention facility
0:01:06 known as the Workhouse,
0:01:09 notorious for its poor living conditions.
0:01:12 Here how she and other advocates fought to shut it down
0:01:14 and won on the first episode
0:01:17 of this special three-part series out now.
0:01:20 Subscribe to Into the Mix, a Ben and Jerry’s podcast.
0:01:24 – Scott, have you ever made a significant pivot
0:01:25 at one of your businesses?
0:01:28 And if so, how did you know is the right time
0:01:29 to make that change?
0:01:31 – Yeah, I’m not sure there is a significant business
0:01:35 that’s ever been built without something resembling a pivot
0:01:38 or iterating the business strategy.
0:01:42 So my first firm, Profit Brand Strategy,
0:01:44 was initially profit market research.
0:01:45 And we used to go out and do surveys
0:01:46 with the internet and computers
0:01:48 and try and find a different way to collect data.
0:01:53 And what we found is that what brands and clients appreciated
0:01:54 was the interpretation.
0:01:57 So we turned to Profit Brand Strategy
0:01:59 and we became a consulting firm at L2.
0:02:01 We were totally focused on the luxury space.
0:02:02 And then P&G called us and said,
0:02:04 “Would you ever do this for P&G?”
0:02:06 And the name of the company was Luxury Lab.
0:02:08 And I hung up the phone and said,
0:02:10 “Our new name is L2.”
0:02:15 And pivoted to or into consumer products at section,
0:02:17 my online education started up.
0:02:20 We initially thought we were gonna be the Netflix of business
0:02:23 that it would be short form videos for B2B.
0:02:25 And that was gonna be so expensive
0:02:26 to produce that kind of content.
0:02:29 So much of that content was available elsewhere
0:02:31 that we pivoted straight into online education
0:02:35 focusing on upskilling the enterprise or AI skills.
0:02:38 So I don’t think I’ve had a business where we didn’t pivot.
0:02:41 I think you just look at the data
0:02:45 and you are thoughtful about what are the opportunities?
0:02:49 And there’s nothing like facing the enemy.
0:02:51 There’s just no market research like launching a business
0:02:54 and seeing what people are actually willing to pay for
0:02:56 to inform your decision making.
0:02:59 And a lot of times clients will come to you and say,
0:03:01 “We’d love it if you could do this.”
0:03:05 And so you’ll get signals from the market.
0:03:09 And what I would suggest is have a solid board
0:03:10 and have your kitchen cabinet of people
0:03:14 that you can propose stuff to and bounce stuff off of them.
0:03:16 And also talk to your colleagues and your employees.
0:03:20 And I’m trying to think if we’ve done a pivot here,
0:03:24 I guess we’re sort of we had our adventures in television.
0:03:26 Now we’re kind of, I wouldn’t say all in on podcasts,
0:03:28 but I think we’re devoting the majority
0:03:32 of the human capital property media now to podcasts.
0:03:37 So the market is an unbelievable muse and advisor.
0:03:39 And you just wanna surround yourself with smart people
0:03:40 who can help you interpret the data
0:03:44 and make sure that you’re not speaking to yourself.
0:03:45 It’s very hard to read the label
0:03:47 from inside of the bottle sometimes.
0:03:50 (upbeat music)
0:03:54 – Welcome to First Time Founders.
0:03:59 One of the most promising use cases for AI is code generation.
0:04:02 That is doing the work of a software engineer.
0:04:06 Already in the US, an estimated nine in 10 software developers
0:04:08 are using AI coding tools.
0:04:11 My next guests created one of those tools.
0:04:13 And less than two years after launch,
0:04:16 it’s already one of the most popular AI coding assistants
0:04:18 in the world.
0:04:21 Last year, they had less than a thousand users.
0:04:24 Today, they have more than 600,000.
0:04:27 Now, after raising a $65 million funding round
0:04:31 at a $500 million valuation,
0:04:33 they’re looking to take over the industry.
0:04:37 Next up, compete with the likes of Microsoft and OpenAI.
0:04:40 This is my conversation with Varun Mohan,
0:04:42 CEO and co-founder of Codium,
0:04:44 and Jeff Wang, Codium’s head of business.
0:04:47 Varun, Jeff, welcome.
0:04:47 – Thanks for having us.
0:04:48 – How was the flight?
0:04:49 You just got in today, right?
0:04:50 – Yeah, we took a red eye
0:04:53 and we I think slept one hour each maybe.
0:04:55 And then when we got to the hotel earlier today,
0:04:56 maybe we slept another hour.
0:05:00 So we’re in perfect condition to do this podcast.
0:05:02 New York code spells are super fun.
0:05:04 You get like this matchbox room.
0:05:06 You know, you’re right next to the wall
0:05:08 every location you are, so perfect.
0:05:10 – Well, I appreciate you being here
0:05:11 and appreciate you joining,
0:05:15 despite getting one hour of sleep last night.
0:05:17 – It’s like two, I suppose.
0:05:19 – Yeah, yeah, good enough.
0:05:22 So you guys are the second AI company
0:05:23 that I’ve had on this podcast.
0:05:28 The previous one I had was an AI for finance company,
0:05:31 more or less replacing bankers,
0:05:35 though his argument is that he’s not replacing bankers.
0:05:37 I’m just gonna start this with the same question
0:05:39 that I asked him, which is,
0:05:42 are you guys at Codium replacing programmers?
0:05:46 – Our vision is actually to give developers
0:05:48 the ability to dream bigger.
0:05:50 And I know that that sounds very vague,
0:05:52 but I think there’s one way of looking at it
0:05:54 that we could just go out and replace
0:05:57 like low-skill labor or low-skill developers.
0:05:59 But I think that that’s not a very rich idea.
0:06:01 We wanna take the best developers
0:06:03 and make them 10 times as leveraged.
0:06:06 And the reason why we think that this is only going to lead
0:06:09 to more developers existing is unlike other professions,
0:06:11 there’s no limit to the amount of technology
0:06:12 the world can actually consume.
0:06:13 – Yeah. – Right?
0:06:15 I don’t think you would ever just be like,
0:06:18 hey, guys, stop making technology.
0:06:19 If we can actually make it so that
0:06:22 the next great invention happens 10 times faster,
0:06:23 we will just only get 10 times
0:06:25 the amount of invention in the world.
0:06:29 – So a program is generally fans of yours, would you say?
0:06:30 – I hope so.
0:06:33 Yeah, I mean, our message is not to replace developers,
0:06:34 I would say.
0:06:36 The Wave Room actually got me on board
0:06:40 ’cause he said like everyone that touches a product,
0:06:42 almost half the people continue using it.
0:06:45 So when you have that, you know there’s something there,
0:06:46 you know there’s some value, right?
0:06:47 And some excitement.
0:06:49 – Is that higher within the industry?
0:06:52 I mean, 50% retention basically of the customers.
0:06:56 That’s, do we know what it’s like for other AI tools
0:06:57 or is it too early?
0:06:59 – Yeah, I think for a lot of consumer products,
0:07:02 it’s usually in the, if you can get it above 10%
0:07:05 at the long tail that is considered very good.
0:07:07 That’s largely because consumers,
0:07:10 unlike companies, if they get bored of something,
0:07:12 they throw it away very quickly.
0:07:14 And these products are not collaborative also.
0:07:16 This is like a single player product.
0:07:18 So it’s actually very easy to churn off of the product
0:07:20 because you aren’t chatting with other people.
0:07:23 So it does mean it is providing a lot of leverage
0:07:25 and developers stay more in flow state
0:07:27 using products like Kodium.
0:07:29 But yeah, hopefully we can make it even better.
0:07:31 – I mean, you’re up against a lot of different tools.
0:07:34 And I mean, I feel like code generation
0:07:35 was one of the first things that people said,
0:07:37 “Oh, AI is going to take over this thing.”
0:07:41 But I feel like the biggest competitor in your space
0:07:43 is GitHub Copilot.
0:07:45 What does the competitive landscape look like
0:07:48 in the AI code generation space?
0:07:50 – There are reasons why we are able to compete directly
0:07:51 with Copilot.
0:07:54 One of them is that we have full repo context awareness.
0:07:56 So as the user is typing,
0:07:58 the results are highly personalized
0:08:01 and we’re seeing like a 30 to 40% boost in accuracy
0:08:03 just having that code base be available
0:08:06 and giving tools for the user to point to
0:08:09 what they’re working on and trying to figure out their intent.
0:08:12 So just the quality of our code suggestions
0:08:14 are very competitive.
0:08:16 And then the thing that we’re really kind of trying
0:08:19 to lean in on though is our ability to deploy
0:08:21 onto like a private server.
0:08:24 And people can host Kodium inside their company
0:08:26 or in their work environment.
0:08:29 And for example, if you’re like in the defense space
0:08:31 or like the finance or healthcare space,
0:08:33 they can’t use Copilot at all.
0:08:35 And that’s kind of where we are focusing right now,
0:08:38 our efforts, but then we have some things down the pipe
0:08:40 where we’ll just be competitive
0:08:41 even on the cloud front too.
0:08:43 – What does it look like for a programmer?
0:08:46 I mean, it sounds like you’re sort of typing in
0:08:49 and then it gives you a list of auto suggestions
0:08:51 for what comes next.
0:08:53 What does it actually look like if you’re a programmer?
0:08:56 And please explain it to me as if I’m five
0:08:59 ’cause I don’t program, I don’t, not a coder.
0:09:02 – Yeah, so a little bit about like the way software
0:09:03 sort of gets built.
0:09:06 So developers write code and what’s called an ID.
0:09:09 It’s this application that enables you to debug code.
0:09:11 So if there are bugs, you could run it.
0:09:14 You can actually see what the errors are, iterate on it, right?
0:09:18 And then right before the code gets pushed into production,
0:09:20 it goes through a review process.
0:09:22 And other people in the company take a look at the code
0:09:23 and actually review it.
0:09:26 And then after that, it goes and it actually gets deployed
0:09:28 into production and it’s on a website or whatever,
0:09:32 where end users can actually touch the product in the end.
0:09:36 And right now where codium is mostly focused is in the ID.
0:09:38 So that’s where developers actually write the code.
0:09:40 It provides value in multiple ways.
0:09:42 So as developers are writing code,
0:09:46 it fills in passively starts filling in more and more code.
0:09:48 And because of the fact that we actually do train
0:09:51 our own models for that passive AI,
0:09:54 we actually found that around 50% of all software
0:09:57 that is getting committed by a developer
0:10:00 is actually accepted and generated by codium.
0:10:02 So that’s the amount of leverage
0:10:05 that just autocomplete is providing to end users.
0:10:08 But to add to that, we also provide a couple of other pieces
0:10:10 of functionality that is super valuable,
0:10:13 despite what level you are as a programmer.
0:10:15 So you can even chat with your code base.
0:10:17 And this probably seems like a very basic piece
0:10:20 of functionality, but when you’re a new developer
0:10:21 and you’re coming into a company
0:10:23 and you have millions of lines of code,
0:10:25 it takes a while to actually onboard
0:10:26 onto that new code base.
0:10:29 And what we’re finding is even at the largest enterprises,
0:10:32 the time it takes to onboard onto a code base
0:10:35 goes down from four to six months to four to six weeks
0:10:37 with a product like codium.
0:10:40 – I wanna focus on how you started this company.
0:10:45 You were a software engineer at a self-driving vehicle company.
0:10:48 So you were kind of working on AI there, I feel like.
0:10:53 And you originally had an idea to start a company
0:10:55 not for writing code with AI,
0:11:00 but for something called GPU virtualization,
0:11:02 completely different company from codium.
0:11:03 It also had a completely different name.
0:11:05 The name was Exa Function.
0:11:09 Can you explain the first iteration of this company
0:11:11 before it became what is now known as codium?
0:11:12 – So to add a little color there.
0:11:14 So I graduated from MIT,
0:11:16 worked at this company called Nero.
0:11:18 It’s an autonomous goods delivery company.
0:11:19 And actually a lot of the learnings
0:11:20 that I had from autonomous vehicles
0:11:23 are actually making their way into the space
0:11:24 we’re in right now.
0:11:25 – Oh wow.
0:11:28 – And maybe to paint some clarity on that.
0:11:30 In 2015, TechRunch basically wrote,
0:11:32 this is the year of AVs.
0:11:34 And in 2024 now, the quote is,
0:11:36 is this the year of AVs?
0:11:38 And you can see how there’s probably going to be
0:11:39 a lot of parallels to generative AI
0:11:42 where we are going to severely overestimate
0:11:44 what is going to happen in a year.
0:11:46 And one of the cool parts about generative AI
0:11:48 is how easy it is to make a demo,
0:11:50 but it is tremendously hard
0:11:51 to make something production ready.
0:11:52 And if you make a claim
0:11:54 that you are going to get rid of a developer,
0:11:56 that is a massive, massive claim.
0:11:57 And in fact, I would actually argue
0:12:00 that is a harder problem than autonomous vehicles.
0:12:01 Because ultimately,
0:12:02 if you look at autonomous vehicles,
0:12:04 all you need to do is press the accelerator
0:12:06 or decelerator or turn a steering wheel.
0:12:07 – That’s a great point.
0:12:08 – Think about the number of different things
0:12:09 a developer actually needs to do.
0:12:11 So I actually led a team
0:12:13 to build large scale deep learning infrastructure.
0:12:16 So how do you run these models at scale
0:12:18 and sort of left the company
0:12:19 with this vision of deep learning
0:12:22 and the idea of running these large models
0:12:24 was going to affect many, many industries.
0:12:25 And we had a small team of people.
0:12:28 We had eight people managing upwards of 10,000 GPUs.
0:12:32 We managed close to 20 to 30% of an entire data center.
0:12:34 And we worked with a lot
0:12:36 of these large autonomous vehicle companies
0:12:37 when we started this company, Exafunction.
0:12:38 Because our mission was,
0:12:41 how do we make it easier to run deep learning models?
0:12:42 But what ended up happening?
0:12:44 This is where startups can always get disrupted.
0:12:46 And it sounds silly to get disrupted.
0:12:48 We were making seven figures in ARR,
0:12:51 but we realized actually most of the models
0:12:53 would probably become these transformer based models.
0:12:56 And these are the models that underpin the GPTs,
0:12:57 the models that open AI has.
0:12:59 – What is a transformer based model?
0:13:00 If you could explain for us, yeah.
0:13:02 – So the basic idea is,
0:13:04 so we now know of prompting, right?
0:13:06 You know, use chat GPT, you pass in a prompt
0:13:10 and notice how it streams tokens one at a time, right?
0:13:12 That’s actually a property of these models
0:13:13 that are called transformer models,
0:13:15 that they are what are called auto aggressive.
0:13:17 They actually generate one token at a time.
0:13:19 And this is very different
0:13:21 than a lot of other classification tasks in the past
0:13:22 where you pass in an input
0:13:25 and it just gives you the entire answer all in one shot.
0:13:28 But this actually like slowly generates the entire thing,
0:13:30 sort of one word, one token at a time.
0:13:31 And we started noticing actually
0:13:34 that that was how a lot of models were starting
0:13:37 to look like once open AI came out with GPT-3.
0:13:39 And the beautiful thing about the model
0:13:41 that is truly crazy is because of the way it is trained,
0:13:43 it can do it in an unsupervised way.
0:13:45 So one of the things that was very different
0:13:49 about models in the past is you needed a lot of label data.
0:13:51 But these models are trained on the entire public internet.
0:13:53 The label data is the internet.
0:13:55 So because of that, you suddenly got these models
0:14:00 that could take in basically trillions of tokens of code
0:14:03 or text and this was not possible in the past
0:14:05 and this created these new sort of generative models.
0:14:06 And in the middle of 2022,
0:14:09 we had this business that was a GPU virtualization business.
0:14:11 The idea was we made it simpler
0:14:14 to run applications on GPUs.
0:14:16 And we found out that most applications
0:14:19 would probably be these transformer models.
0:14:21 And if all of what we were doing
0:14:22 was running transformer models,
0:14:24 we would largely become a commodity
0:14:26 because they would become a race to the bottom, right?
0:14:28 It would be the equivalent of asking,
0:14:31 like they would ask Varun how cheaply can you run this model?
0:14:33 I’d say I can do it for a dollar.
0:14:34 Then they would ask Jeff how cheaply you can do it.
0:14:36 He’d say 50 cents and we’d go back and forth
0:14:37 until no one made any money.
0:14:40 And this is the commodification of this entire space.
0:14:43 But what we did see was we felt that this technology
0:14:45 would be like the early coming of the internet.
0:14:47 There would be a brand new set of applications
0:14:48 that would be created.
0:14:50 And we were early adopters of GitHub’s product,
0:14:52 GitHub Copilot, and we thought that that was just scratching
0:14:55 the tip of the iceberg of what the future would look like.
0:14:57 And that’s where Codium sort of came about.
0:15:00 But it was, as you can imagine, a very rough experience
0:15:03 because we basically said buy to all the revenue that we had
0:15:05 and we had to start all the way back down to zero.
0:15:10 So, I mean, that to me is like the mother of all pivots
0:15:14 where not only are you changing the entire business
0:15:16 as you know it, you’re also doing it at a time
0:15:17 when things are going really well.
0:15:20 I mean, you said seven figures ARR.
0:15:22 My understanding is that you’d also raise $22 million
0:15:24 for this company.
0:15:26 Like things are going right.
0:15:30 And then you turn around and you tell all your employees,
0:15:32 actually would scrap that.
0:15:34 We’re going to do a whole different thing.
0:15:38 If I were a software engineer who worked at Big Tech
0:15:40 and had quit to go work at ExoFunction
0:15:45 and the CEO told me that, I’d be a mixture of like pissed off,
0:15:46 freaked out, concerned.
0:15:49 How did you rally the team
0:15:52 and how were you able to make that pivot so successfully?
0:15:53 – So I’ll say a couple of things
0:15:55 about the composition of the team.
0:15:57 Largely they were people that we knew.
0:15:58 And that’s actually very important
0:16:00 because they would be people that would go
0:16:02 into the trenches with us.
0:16:04 They were people that knew the caliber of people
0:16:07 that both me and my co-founder were.
0:16:08 And also on top of that,
0:16:10 we picked a problem space
0:16:13 where we were all passionate about it.
0:16:15 And I 100% knew at the time
0:16:19 there were products like mid-journey that were taking off.
0:16:20 Small team of people, eight people
0:16:23 that were making tens of millions of dollars in revenue.
0:16:24 And frankly speaking,
0:16:26 when we decided to pivot the company,
0:16:27 we knew for a fact,
0:16:29 none of us were that passionate about image generation
0:16:31 despite the fact that it is a very cool area.
0:16:32 And if we had picked it,
0:16:35 our team wouldn’t have had I guess the mental fortitude
0:16:38 to dig deep enough to actually build the problem space.
0:16:41 And then I guess the sort of third part
0:16:43 is actually that we actually were able to take
0:16:45 a lot of the infrastructure expertise
0:16:46 that we had as a company
0:16:48 to actually go out and build the application
0:16:49 significantly faster.
0:16:51 We were very quickly able to train our own models
0:16:52 and run them at massive scale.
0:16:54 And right now, Codium is one of the top five
0:16:56 largest generative AI apps
0:16:57 in terms of text in the world.
0:17:00 And that largely is because the original composition
0:17:02 of the team was these people
0:17:05 who are effectively GPU infrastructure experts.
0:17:07 But all said and done, everything that I said,
0:17:11 it comes down to you need a very truth seeking company.
0:17:13 And at the time,
0:17:15 even if we were at seven figures in revenue,
0:17:18 I did not know how we would 10x the amount of revenue.
0:17:20 And we could continue to lie to ourselves
0:17:23 and have a slow but certain death as a company
0:17:25 as the technology commoditizes, right?
0:17:27 And we all run the same kind of models.
0:17:28 Or we could just say,
0:17:31 there is a high probability that we will die,
0:17:33 but there’s a space that we could be very passionate about
0:17:34 and it could be very big.
0:17:37 I think we just decided the latter was
0:17:39 the more rational choice it is very hard to make.
0:17:43 But in retrospect, it was the obvious choice, right?
0:17:45 – Did you have any data showing you
0:17:47 that you were going to crash and burn at that point?
0:17:48 Or was it just a hunch?
0:17:50 And part of the reason I asked that
0:17:53 is because if I were your investor,
0:17:55 I would wanna be like, oh yeah, yeah.
0:17:58 It’s very clear to me that you guys have to pivot.
0:18:01 – We were cashflow positive then.
0:18:02 We were cashflow positive then.
0:18:04 – So you just had a feeling?
0:18:06 – We had a feeling.
0:18:08 It’s just because, and this is a little bit of a curse
0:18:11 of being a venture capital, venture-based business.
0:18:13 If you’re making millions of dollars in revenue,
0:18:15 that is not a venture-backable business.
0:18:19 And if we can’t figure out a path to get that to 100,
0:18:21 then that is not a business that we could build.
0:18:25 We could continue to keep it as an eight to 10% team.
0:18:26 We have another mentality in the company
0:18:27 beyond being truth-seeking
0:18:29 that we are a very lean company.
0:18:30 By the time we raised our Series B,
0:18:32 we had barely spent our seed round.
0:18:35 And I think that’s just because we don’t think capital
0:18:38 is a limiting factor in building a good business.
0:18:41 It’s you have to build a great product that customers love.
0:18:43 And that is usually not just you had more money.
0:18:46 – We’ll be right back.
0:18:48 (upbeat music)
0:18:51 (upbeat music)
0:19:02 – Support for the show comes from Into the Mix,
0:19:05 a Ben and Jerry’s podcast about joy and justice
0:19:07 produced with Vox Creative.
0:19:10 Would you have $25,000 to post bill?
0:19:13 That’s how much Inez Bordeaux had to pay
0:19:15 when she was arrested in 2016.
0:19:17 And since she couldn’t afford it,
0:19:18 she was sent to the workhouse,
0:19:21 a pretrial detention center in St. Louis.
0:19:23 Inez and the other detainees weren’t locked up
0:19:26 because they’d been convicted,
0:19:29 but because they couldn’t afford their bail.
0:19:31 – Experiencing what I experienced
0:19:34 and watching other women go through it
0:19:38 and know that there were thousands before us
0:19:41 and there were thousands after us
0:19:46 who had experienced those same things.
0:19:48 That’s where I was radicalized.
0:19:50 – She spent a month at the workhouse
0:19:52 and witnessed abject conditions,
0:19:55 extreme heat and cold, mold and pest infestations
0:19:57 and poor medical care.
0:19:59 Eventually her charge was vacated,
0:20:03 but the experience changed her.
0:20:04 – They’re starting a campaign to close the workhouse.
0:20:06 Are you interested?
0:20:09 And I was like, hell yeah, hell yeah, I’m interested.
0:20:11 – Hear how she and other advocates
0:20:13 fought to shut it down and won.
0:20:16 – On the first episode of this special three-part series
0:20:18 out now, hosted by Ashley C. Ford,
0:20:21 subscribe to Into the Mix, a Ben and Jerry’s podcast.
0:20:34 – This ad was expressly recorded
0:20:37 to create a sense of simplicity.
0:20:42 Just a few simple sounds, no complexity.
0:20:46 – Hmm, like neutral.
0:20:51 Made with just vodka, soda and natural flavor.
0:20:54 Neutral, refreshingly simple.
0:20:59 – When summer brings the heat, McDonald’s brings the chill.
0:21:02 During summer drink days, enjoy a small iced coffee
0:21:05 or a refreshing Coca-Cola for $1 plus tax
0:21:07 and step up your summer today.
0:21:09 At participating McDonald’s restaurants in Canada,
0:21:10 prices exclude delivery.
0:21:13 (mouse clicks)
0:21:18 – We’re back with First Time Founders.
0:21:20 One of the things you mentioned
0:21:23 is that you are training your own models.
0:21:25 I mean, this is an AI application,
0:21:28 but most AI applications that I am aware of,
0:21:32 they’re purely building the application layer,
0:21:35 and that is they’re basically using someone else’s model,
0:21:38 usually open AI, and they’re tweaking
0:21:39 and they’re building off of that model
0:21:41 and creating their own application.
0:21:42 You guys are different.
0:21:44 It sounds like you guys are,
0:21:45 I don’t know what you’d call it,
0:21:48 full stack AI from the model to the application.
0:21:49 Is that right?
0:21:50 – That is correct.
0:21:52 Actually, even at the kernel layer,
0:21:55 we’ve done some, like we’ve written some code
0:21:57 even at the infrastructure layer
0:21:58 so that we could set the models on top
0:21:59 in an efficient manner.
0:22:01 And the reason we have to do that
0:22:03 is because of the latency issues we talked about.
0:22:05 For example, if you are a passive AI
0:22:07 and you take, let’s say one second
0:22:09 to show up the suggested code,
0:22:10 people are just going to stop using that.
0:22:12 They don’t want to get out of flow state
0:22:14 and pause and wait for results.
0:22:18 – How many other AI startups are doing,
0:22:19 what should we call it, the full stack,
0:22:23 the infrastructure to application?
0:22:24 Are there many others?
0:22:26 I just, I mean, off the top of my head,
0:22:29 I’m like anthropic like open AI,
0:22:32 but I guess they’re mainly kind of infrastructure layer, right?
0:22:34 I mean, isn’t this super rare?
0:22:36 – I think in my mind,
0:22:39 there’s maybe a couple of unique things about code
0:22:41 that make it so that you can actually do this here.
0:22:42 And you’re totally right.
0:22:44 Most of the large companies that are even successful
0:22:46 are largely built on top of an API,
0:22:49 but we genuinely felt to build a best in class app here.
0:22:51 We needed to become vertically integrated.
0:22:55 And for us, it was also not a complex thing for us to do
0:22:58 in that we have the technical talent inside the company
0:22:59 to actually go ahead and do that.
0:23:02 Maybe one of the unique aspects of code
0:23:05 on why we can do this is code can actually be run.
0:23:08 Like let’s say I am a illegal AI tool
0:23:10 and I’m redlining a bunch of documents.
0:23:12 The only way to know if that is good
0:23:14 is for a human to go in afterwards
0:23:16 and actually take a look at it.
0:23:17 For code, you can actually,
0:23:19 if you make an edit to a code base,
0:23:21 you can actually run the code and validate
0:23:22 it is doing the right thing
0:23:24 without a human in the loop at all.
0:23:26 And what that means is there are ways in which
0:23:29 you can close the loop in intelligent ways
0:23:30 that you can actually,
0:23:32 if you specialize on that application
0:23:34 and you are vertically integrated,
0:23:37 you can build an even better app for code.
0:23:40 And we’ve taken advantage of this in many, many ways.
0:23:42 And we realized if we didn’t do this,
0:23:43 we’d be shooting ourselves in the foot.
0:23:45 And I’ll give you even a simple example
0:23:47 of how this manifests itself.
0:23:49 Right now, I mentioned this.
0:23:51 Codium is, processes over a hundred billion tokens
0:23:53 of code every day, which is over 10 billion lines
0:23:54 of code every day.
0:23:56 If we passed that through OpenAI,
0:23:57 we would have gone bankrupt.
0:24:00 And there was a recent article from the Wall Street Journal.
0:24:01 – And why is that, sorry,
0:24:02 because you would have to pay for it.
0:24:03 – Because it would be too expensive.
0:24:04 – Exactly.
0:24:08 – And not the best for our particular application
0:24:09 on top of that.
0:24:11 And if you look, there was a recent article
0:24:13 on the Wall Street Journal about how GitHub Copilot
0:24:16 was spending tens of dollars per user per month.
0:24:18 And that’s actually because even GitHub,
0:24:20 Microsoft’s product is not vertically integrated.
0:24:22 They are relying on external models
0:24:24 to build their application.
0:24:27 And we view that as, hey, the model and the product
0:24:28 and the infrastructure are so critical
0:24:30 to delivering a great experience,
0:24:32 why would we not have control over every piece of that?
0:24:35 So my understanding is you guys don’t use
0:24:38 the actual data of your individual users.
0:24:42 So how is the model getting trained then?
0:24:45 – So we actually do sort of two different things.
0:24:48 And I’ll let Jeff add on how this affects our enterprises
0:24:50 and the customers that use the product.
0:24:52 So first of all, we use permissively licensed code
0:24:54 that is available on the public internet.
0:24:56 And we also attribute it on generation time.
0:24:59 So we actually take sort of copyright
0:25:01 and licensing very seriously as a company.
0:25:03 But on top of that, when we release product
0:25:04 from our user data, you’re right,
0:25:06 we don’t take the data from our users,
0:25:08 like let’s say when they’re auto-completing stuff
0:25:10 and copy that code and put it into our training set.
0:25:13 But we can see, hey, users are accepting
0:25:14 these types of suggestions more
0:25:16 and these types of suggestions less.
0:25:19 So we have preferences on what users and humans like.
0:25:21 And that actually informs us to actually build products
0:25:23 that are better and more willing to use.
0:25:25 And then that actually has a virtuous cycle
0:25:26 and that now people are willing to try
0:25:28 more complex things on our product
0:25:29 because the easier things they have,
0:25:31 high confidence that they work.
0:25:33 And suddenly the frontier of what we are actually able
0:25:36 to experience as we are able to give the user
0:25:37 increase more and more,
0:25:39 largely because we have a product that is so well beloved.
0:25:42 We now have over 600,000 users that use our product.
0:25:43 – Yeah, it’s unbelievable.
0:25:45 – I think one thing we might have glossed over
0:25:48 in the beginning was codeon made a very conscious decision
0:25:51 to make the product free for individuals.
0:25:53 And if you think about, you know,
0:25:57 Varun just said co-pilot loses $10 to $20 a month per user.
0:26:00 That’s after they’ve been paying a subscription to, right?
0:26:02 So having that infrastructure background,
0:26:05 being able to make it efficient to deploy these models
0:26:07 and then giving it out for free for individuals,
0:26:09 allowed us to build this very large user base,
0:26:12 probably the largest user base for a coding assistant.
0:26:14 – Yeah, I mean, just the data here,
0:26:17 you started out last year with less than 1,000 users,
0:26:19 by the end of the year, you had half a million.
0:26:21 – Yeah, and we have over a million,
0:26:23 millions of downloads across all the plugins.
0:26:24 And the reason why that’s very important
0:26:26 is because of what we just talked about.
0:26:27 If we roll out multiple models,
0:26:29 if we are changing the temperature
0:26:31 of the thresholds here and there,
0:26:33 we are getting so many signals
0:26:35 as to what is the appropriate settings to tweak.
0:26:38 And I think every hour we’re getting like a million signals.
0:26:41 So we can run all these experiments on these free users
0:26:43 of all these models we train
0:26:44 to really, really get the best results
0:26:45 that you could possibly get.
0:26:48 And then only when we’ve validated like, okay,
0:26:49 we’ve trained this model,
0:26:51 this is better than all the other ones we’ve trained.
0:26:53 These are the settings that makes the best results.
0:26:55 Then we could deploy that to our on-prem
0:26:57 or enterprise users, right?
0:26:59 Because we can’t, after we’ve deployed it,
0:27:00 we can’t really get that much more information from it.
0:27:01 It’s completely,
0:27:04 it can be hosted even in an air-gapped environment.
0:27:06 So I think that’s a big element
0:27:07 of why we are successful
0:27:09 of being able to deploy these models.
0:27:12 Because if somebody’s trying to start from scratch right now,
0:27:14 how are they gonna know that their model is good?
0:27:16 How are they gonna tweak the model, right?
0:27:18 Without a very large user base.
0:27:19 So that’s part of the secret sauce
0:27:21 is having that free user base.
0:27:24 – Yeah, and you, it’s still free,
0:27:26 which is what I find pretty fascinating.
0:27:28 And you said, I was just reading your state,
0:27:30 your like mission statement,
0:27:33 you are quote committed to having a free tier forever.
0:27:36 The natural next question is,
0:27:39 how is it gonna get properly monetized?
0:27:41 And how are you gonna maintain that,
0:27:43 that free tier into perpetuity?
0:27:45 – I think people underestimate kind of the demand
0:27:47 for both just the on-prem instance,
0:27:49 but also some of the teams features we add
0:27:50 on the SaaS product.
0:27:52 So we have a free individual tier,
0:27:56 but if you create a team and add on-board users to it,
0:27:57 that is a paid product.
0:27:59 And there are things like analytics
0:28:01 and actually bigger models and seat management
0:28:04 that are going to be better than the free product.
0:28:06 But we are committed to making the free product
0:28:08 the best coding assistant out there, no matter what.
0:28:10 So whatever other coding assistants come to market,
0:28:13 we will make sure our free product is still the best one.
0:28:14 We wanna make sure, you know,
0:28:16 disincentivize others from entering,
0:28:18 but we want people that always have the option.
0:28:20 We don’t want them to get forced to buy a co-pilot,
0:28:21 for example.
0:28:22 – And then maybe one thing to add to Jeff,
0:28:24 like we monetize enterprises, right?
0:28:27 So we have some of the largest Fortune 100s,
0:28:30 F-thousands, even over 10,000 developers on our product.
0:28:33 And those companies, obviously they want security guarantees,
0:28:34 they want personalization
0:28:36 for many, many repositories that exist.
0:28:37 And they also want support
0:28:40 across all source code management tools.
0:28:44 Less than 10% of Fortune 500 companies are on GitHub cloud.
0:28:45 And that is the competitor that we have,
0:28:46 that is GitHub co-pilot.
0:28:48 And they have committed to making their product
0:28:52 differentially better if you are on GitHub cloud, right?
0:28:55 So we want to take the approach of almost being Switzerland.
0:28:58 We don’t care what programming language you write.
0:29:00 We don’t care what ID is you use.
0:29:03 We don’t care where you store your source code.
0:29:06 And ultimately, we also just don’t really care
0:29:07 what seniority the developer is.
0:29:10 We will provide the maximum amount of leverage there.
0:29:13 Whereas a lot of the larger players in the space
0:29:15 are focused on being tied to another brand.
0:29:17 And the reason why we don’t think that that makes sense is
0:29:20 we think AI is such an up-leveler.
0:29:23 We think it deserves to be in a category of its own.
0:29:26 – You recently raised $65 million
0:29:29 in a round led by Kleiner Perkins.
0:29:32 It valued you at half a billion dollars.
0:29:35 Congratulations.
0:29:37 What is that money going to be used for?
0:29:39 – I think the way we would like to think about it is
0:29:41 we have ways of spending cash,
0:29:42 not only to train models,
0:29:44 but to also make it so that we can build
0:29:46 a better user experience for the end user.
0:29:48 But one of the cool things for us is our product
0:29:50 has such high ROI that we think that there will be
0:29:52 a real payback period on that.
0:29:55 Enterprises and companies will see enough value there
0:29:57 that they will be able to eat the cost
0:29:58 that we will need to spend upfront.
0:29:59 But also on top of that,
0:30:01 we want to spend a lot on making sure
0:30:03 that we can become better partners for our customers.
0:30:05 We’re onboarding some of the world’s largest companies
0:30:08 and we’re onboarding tens of thousands of developers.
0:30:09 That’s going to take a little bit of effort
0:30:11 to make sure that we do that properly.
0:30:13 One company that we’re working with right now
0:30:15 that we have a multi-year engagement with,
0:30:18 they account for 0.15% of all developers in the world,
0:30:20 it’s just that one company, right?
0:30:23 So, I’m a little bit of a different type of founder
0:30:24 in that I do not like the idea
0:30:26 of spending money unnecessarily,
0:30:28 but if it comes down to we are doing it
0:30:29 because we make our customers more successful
0:30:32 and our users more happy, we’ll do it any day.
0:30:35 – How have you guys restrained yourselves
0:30:36 in terms of spending?
0:30:40 Because I mean, yeah, the story, the narrative in AI
0:30:42 has been that it is an arms race.
0:30:45 And the thing that I hear about in the venture industry,
0:30:47 or at least what AI founders are being told,
0:30:50 is just go out and raise a shit ton of money
0:30:53 as much as is even possible.
0:30:56 One, because you just want to develop a war chest.
0:31:00 And two, you want to get the headlines.
0:31:02 You want to be the AI company that’s working
0:31:05 on co-generation, the AI company for finance,
0:31:06 whatever it is.
0:31:09 So I guess sort of two questions for me here.
0:31:12 One, is that accurate to your experience?
0:31:16 And two, how have you been so responsible
0:31:19 in terms of spending while you see all of these headlines
0:31:21 of other companies spending so much money
0:31:23 on talent and training their models?
0:31:26 – I think Varun mentioned earlier that we run very lean.
0:31:28 And the reason we’re able to do that
0:31:31 is we hire people that are generalists or ex-founders
0:31:33 and they’re capable of doing many job roles at once.
0:31:37 So our company size is probably like a little misleading.
0:31:39 It’s probably way more effective,
0:31:41 not just infrastructure, but the headcount also.
0:31:42 – What is the headcount?
0:31:44 – So right now we’ll be almost about 55
0:31:45 at the end of the month, I think.
0:31:47 And then the thing is like the people we hire
0:31:49 are able to just slot into different roles,
0:31:50 almost out of moments notice.
0:31:52 It’s like, oh, we don’t even have a marketing team,
0:31:53 for example.
0:31:55 The product’s growth has been organic,
0:31:57 but we do want to do some marketing experiments
0:31:59 to make sure that we are getting ourselves out there,
0:32:00 just as an example.
0:32:02 And then people, randomly someone will be like,
0:32:04 I have an idea, okay, go do it.
0:32:07 And then we all of a sudden have a Google ad strategy.
0:32:09 All of a sudden we have a bunch of blog posts.
0:32:12 We’re on podcasts like with you and Elson.
0:32:14 But I think that my point is,
0:32:16 part of the function of not spending the money
0:32:19 is just being very, I guess, practical
0:32:21 of where the money goes and running lean.
0:32:23 And I think when there is a moment that says like,
0:32:25 hey, we need to train a much bigger model
0:32:26 and we need to spend this much money,
0:32:28 we’re all for it, actually.
0:32:31 We actually are very conscious about what the ROI is
0:32:32 of everything we do.
0:32:35 – Yeah, how do you maintain that culture of leanness
0:32:38 and what would be your recommendation to other companies
0:32:41 that are maybe not as big as you, but trying to be?
0:32:44 – This is quite probably the hardest problem
0:32:46 that we’re trying to solve right now.
0:32:49 ‘Cause we wanna hire people that are very good,
0:32:52 very technical, maybe they’re generalists,
0:32:54 like I said earlier, but it’s very hard
0:32:55 to hire those people.
0:32:57 So actually, this is probably one of the things
0:32:59 we focus on in the next month
0:33:01 is what is our recruiting strategy?
0:33:03 How do we hire the best people?
0:33:05 Maybe part of that is getting Kodium’s brand name
0:33:06 much more aware.
0:33:09 Maybe it’s like a big push of user adoption.
0:33:11 Maybe we’re just gonna have to be scrappy
0:33:14 and be very creative of how we hire people.
0:33:15 For example, I don’t know if other companies
0:33:18 are just like pinging every ex-founder on LinkedIn,
0:33:19 but we are, right?
0:33:23 So we are trying to scale with creative means.
0:33:25 – One other thing is like as a company,
0:33:28 I think culture is, we have some cultural principles
0:33:30 and we run lean as one of them,
0:33:33 but who would wanna say you don’t run lean?
0:33:36 So I think, how do you actually live that out?
0:33:38 We are a five days a week in-person company,
0:33:40 so we don’t do remote work.
0:33:42 We don’t really do hybrid work either.
0:33:45 So people see what it’s like to work at the company
0:33:49 and until very recently, our CTO was ordering snacks.
0:33:51 And that’s not to say that’s a great use of his time.
0:33:53 It’s more just that no one is high enough
0:33:56 to not do some work to test a hypothesis.
0:33:58 And we don’t hire specialists at the company
0:34:01 until generalists outgrow that role.
0:34:03 And I’ll give you another example of this.
0:34:06 When we ran that GPU virtualization company,
0:34:07 even though we were making money,
0:34:09 we never hired a sales rep.
0:34:11 And that’s not because I don’t believe in enterprise selling.
0:34:13 No, we have a great VP of sales now at the company
0:34:15 and I think we might have in terms of talent density,
0:34:18 one of the strongest enterprise sales teams in the world,
0:34:20 actually, but why didn’t we hire someone?
0:34:23 I just didn’t believe that if we added one new person,
0:34:25 I would be setting them up for success
0:34:28 because the reality is if I could not get $1 of additional
0:34:33 sales, I cannot expect someone else to get $10.
0:34:35 So this is one of those things where we have a mentality
0:34:38 of we try to do it ourselves and then we try to eliminate
0:34:39 ourselves from the role.
0:34:41 We give away our Legos.
0:34:43 We let someone else take that over that understands
0:34:46 the role much more, but we don’t do things prematurely.
0:34:50 And I think there’s a tendency across people that the idea
0:34:52 of building a scalable organization is really valuable.
0:34:54 And I do see that that you do want to build
0:34:57 a scalable organization, but sometimes people get too excited
0:35:00 about this notion of org building or fundraising rather
0:35:03 than the idea of having customers, having users.
0:35:05 Because ultimately, people that join our company
0:35:07 don’t care about how much money we raised,
0:35:09 as long as we will survive.
0:35:11 And our customers genuinely don’t care.
0:35:12 Let’s look at it this way, right?
0:35:14 If you look at a company as big as JPMC
0:35:16 that makes hundreds of billions of dollars,
0:35:20 to them, does it make sense if we raised $100 or $200 million?
0:35:21 It all looks like peanuts to them.
0:35:23 It’s all like 10 basis points of the amount of revenue
0:35:25 that they make a year.
0:35:26 So to them, what they really care about
0:35:28 or companies of this size is,
0:35:30 are we the best partner for them?
0:35:32 Are we the best product to them?
0:35:34 And as long as we’re laser focused on that,
0:35:36 we should do whatever it takes to build that up.
0:35:39 We’ll be right back.
0:35:50 – Looking to buy or sell a used vehicle?
0:35:52 Forget the selection and protection
0:35:53 of BCAA auto marketplace.
0:35:55 Try some random guy off the internet.
0:35:57 – Some random guy.
0:35:59 – Yes, we offered zero security.
0:36:01 – I can pay in crypto.
0:36:02 – Plenty of questions.
0:36:04 – Is it still available?
0:36:06 – And zero convenience.
0:36:09 – 46 messages and I still make eye.
0:36:12 – BCAA auto marketplace helps you buy and sell vehicles.
0:36:17 Visit bcaa.com/marketplace and avoid some random guy.
0:36:21 – Pandora, be love.
0:36:24 What does be love mean to you?
0:36:26 – I definitely would say my be love role model
0:36:28 is for sure my sister.
0:36:30 Unconditional, infinite love.
0:36:32 Something that is never ending that you know
0:36:33 is always there.
0:36:34 – Never questioned.
0:36:36 – Never questioned.
0:36:38 No matter if you fall off a cliff,
0:36:41 she’s there to catch you, you know?
0:36:42 Be love.
0:36:44 – Shop now at pandora.net.
0:36:48 – This is an ad by BetterHelp.
0:36:51 What are your self-care non-negotiables?
0:36:53 It’s hard to make time for the things that keep you healthy
0:36:57 but being consistent with self-care is like working a muscle.
0:37:00 And when life gets crazy, that muscle keeps you strong.
0:37:02 Therapy is the ultimate self-care
0:37:04 and BetterHelp makes it easy to get started
0:37:07 with affordable online sessions you can do from anywhere.
0:37:10 Never skip therapy day with BetterHelp.
0:37:12 Visit betterhelp.com to learn more.
0:37:14 That’s better H-E-L-P.com.
0:37:22 – We’re back with first time founders.
0:37:24 Sort of a more personal question.
0:37:29 I mean, you started this company in 2021
0:37:31 sort of just as the AI hype was bubbling up
0:37:35 and you now find yourself at the epicenter
0:37:37 of the hardest industry
0:37:39 and you are one of the hardest companies
0:37:41 in the hardest industry.
0:37:43 Just a personal question for both of you.
0:37:44 How does that feel?
0:37:47 Like what has it been like getting used to
0:37:50 being the guy in AI?
0:37:52 – I told this to the company.
0:37:56 I tell everyone just get ready to get destroyed.
0:37:59 Assume that something very bad is going to happen always.
0:38:02 And this is where us having gone through that pivot
0:38:03 is very critical.
0:38:05 Things are going very well for us as a company.
0:38:07 A lot of the reason why we haven’t spent a lot of money
0:38:09 is now we make money, which is a unique property
0:38:12 about a lot of companies apparently in this space
0:38:13 where most companies talk about vision
0:38:15 rather than actually building a product that people use.
0:38:18 But I just, I tell everyone get ready
0:38:20 for something really bad to happen.
0:38:22 And this is why it’s like very important
0:38:26 that we hire people that are truly in it for the long run.
0:38:28 And I tell people this when they joined the company,
0:38:30 I think we could be a company
0:38:32 that’s worth over a hundred billion dollars.
0:38:33 I think we can.
0:38:35 And that’s largely because the total adjustable market
0:38:37 of what we’re building
0:38:38 and the amount of impact that this can have
0:38:41 given how important technology is could be massive.
0:38:44 But also a series of bad decisions that we make
0:38:46 could completely kill the company.
0:38:48 And that will happen very fast.
0:38:49 And I think this is where us building
0:38:51 a very truth seeking company.
0:38:53 And that actually is very hard
0:38:55 because people want to believe
0:38:56 that what they’re doing is correct.
0:38:59 And they want to embrace psychological safety.
0:39:01 And what I tell people is, hey,
0:39:04 if you feel something is wrong,
0:39:07 lean into what you think is wrong and tell everyone.
0:39:10 Tell everyone, because we should not have pockets of people
0:39:12 that want to report up the chain
0:39:14 and tell their manager or tell me,
0:39:16 we have a very flat company or tell me,
0:39:17 things are going fine.
0:39:19 I would much rather hear everything is on fire
0:39:21 and have paranoid people at the company
0:39:23 than people who are just happily going to work.
0:39:25 And this is why I think startups
0:39:27 are so much harder than big companies.
0:39:29 It’s actually not that, you know,
0:39:31 you’re taking a massive risk on the monetary side.
0:39:34 You still can make a six figure salary, right?
0:39:36 These are not people that are living hand to mouth.
0:39:40 But the really hard part is the lack of psychological safety.
0:39:42 We make a bunch of people at the company
0:39:43 make a series of bad decisions
0:39:46 and the entire thing can go to zero.
0:39:48 Whereas if you’re at Google,
0:39:49 you have very little accountability.
0:39:51 If your team doesn’t perform well,
0:39:55 Google makes so much money that you are a rounding error.
0:39:57 You will be shuffled to some other part of the company.
0:39:59 You never really need to deal with the impact
0:40:02 of your decisions and consequences of your decisions.
0:40:04 And that’s why it’s always a little bit of a funny statement
0:40:06 when someone at a big company is like,
0:40:08 I work at a startup at a big company.
0:40:10 No, no you don’t.
0:40:14 Imagine the idea of you potentially losing your job
0:40:16 every quarter or every month.
0:40:17 Exactly.
0:40:19 One thing that the way I think about it
0:40:21 is radical transparency.
0:40:23 And we have a lot of conversations within our company
0:40:26 of let’s be super transparent about everything.
0:40:28 And even in my personal life,
0:40:30 I’m like, I’m just gonna be like totally upfront.
0:40:35 That’s the cleanest, most kind of hygienic way to operate.
0:40:41 Do you ever feel that there could be too much truth?
0:40:44 Do you feel that there’s a possibility
0:40:46 that if you’re encouraging everyone
0:40:48 to tell the truth, tell the truth, be transparent,
0:40:52 tell me everything, that they’ll kind of go overboard?
0:40:54 I think this is the hardest part about a startup.
0:40:57 The two things that I say is a startup is really hard
0:41:00 because you need to be both irrationally optimistic.
0:41:02 Because if you’re not optimistic,
0:41:04 the answer is always Microsoft is gonna beat you, right?
0:41:04 Right.
0:41:06 It’s the biggest company of all time.
0:41:07 They have the most capital.
0:41:08 They have the most people.
0:41:09 They have the most distribution.
0:41:10 Why does any company win?
0:41:11 But clearly that’s wrong, right?
0:41:12 There are companies that have beat Microsoft
0:41:13 at different areas.
0:41:14 We always say this,
0:41:16 oh, they have the most capital they’re gonna win.
0:41:18 Yes, yes, it’s very easy from my perspective to say that.
0:41:19 Exactly, right?
0:41:22 And then HBS, I try to think about what Harvard Business
0:41:25 School would say 10 years from now for us as a company.
0:41:28 And I tell people this, it’s like, we fail.
0:41:31 And what they write is in a world
0:41:34 in which technology was changing,
0:41:36 Microsoft had all the distribution in the world.
0:41:38 They had this property called GitHub.
0:41:40 It was inevitable that they would win.
0:41:41 And because of that, they won.
0:41:42 They won the entire space.
0:41:45 And all of these startups, it was a fool’s errand.
0:41:46 Why did they even try?
0:41:48 And then the world in which we win,
0:41:50 they were going to write in a world
0:41:54 in which the technology was getting disrupted so materially.
0:41:56 There were a set of companies,
0:41:57 and one of which was Codium,
0:42:00 that had such a technological advantage
0:42:02 that there was no way a slow-moving gorilla
0:42:03 like Microsoft could even compete.
0:42:05 So what I think is very hard–
0:42:06 – Very hard, yeah.
0:42:07 – Yeah, exactly.
0:42:08 They will write whatever the future looks like
0:42:10 and the history will be written by the victor.
0:42:12 And no one will know exactly
0:42:14 what the pages of the book look like.
0:42:18 And I think the hardest part about a startup is,
0:42:19 you do need to be irrationally optimistic
0:42:21 to believe that you can win.
0:42:24 Because by default, if you don’t, you will definitely lose.
0:42:26 But then also uncompromisingly realistic,
0:42:29 which is that sometimes, actually,
0:42:33 there’s no point continuing in a particular direction.
0:42:36 This is where I think it comes down to what you just said,
0:42:38 which is that if you were too truth-seeking
0:42:41 and everyone is paranoid all the time,
0:42:43 it could lead to paralysis.
0:42:45 And I think there’s a fine line there.
0:42:48 There’s a fine line where one time we lost a deal,
0:42:49 and we actually, every day for dinner,
0:42:52 talked about it for two weeks in a row for multiple hours.
0:42:54 We talked about the implications as a company.
0:42:55 And that was very useful.
0:42:58 But if we extended that out to an entire year,
0:43:00 we would not go anywhere.
0:43:01 – Yeah.
0:43:02 – So you’re totally right.
0:43:03 There’s a level here.
0:43:05 But what I do feel most companies do
0:43:08 is probably err on the side of not being truth-seeking enough.
0:43:11 They become too complacent with what they’re doing.
0:43:14 And I think the thing to really think about,
0:43:15 and this is not true at a big company,
0:43:17 which is why you really need to think about it
0:43:19 at a startup is you do not win an award
0:43:21 for doing the wrong thing for longer.
0:43:24 So the sooner you can rip the band-aid off,
0:43:25 the better your company will be.
0:43:27 You will be way happier that you did it.
0:43:30 It is going to be incredibly painful for a week,
0:43:31 but just do it.
0:43:32 – I love that.
0:43:34 You guys have been generous with your time,
0:43:36 so we’ll begin to wrap up here.
0:43:39 Jeff, I’ll ask you this question.
0:43:41 What do you think has been the greatest challenge
0:43:44 that this company has faced in the past couple of years?
0:43:48 – I think the pace at which things switch
0:43:50 is really challenging.
0:43:53 We relied on some technology features
0:43:54 to be the selling point.
0:43:56 And then our competitors would come out
0:43:58 and it’s like the same thing all of a sudden.
0:44:01 And every time that there’s a news release
0:44:02 for one of our competitors,
0:44:04 we immediately go into like a code red
0:44:06 and go into a competence room.
0:44:08 Reverse engineer what they’re doing.
0:44:09 And this is constant.
0:44:12 This is like every month this is happening.
0:44:13 The question is like, will this last forever?
0:44:16 Like will we just always be panicking?
0:44:18 And the question that answers probably yes, right?
0:44:21 – I think for people listening to this podcast,
0:44:26 you too are the closest thing to AI experts as we’ve had.
0:44:30 You’re kind of on the front lines of this.
0:44:33 What would be your advice to anyone
0:44:35 who is working in the AI industry
0:44:37 or who wants to work in the AI industry?
0:44:39 And I think that doesn’t just have to be founders
0:44:43 but engineers, product managers,
0:44:44 business operators, et cetera.
0:44:46 What’s the most important thing
0:44:49 that they should understand about AI right now?
0:44:51 – There’s an interesting property about AI right now
0:44:54 in that it is actually imperfect.
0:44:56 You know, when you use the products,
0:44:58 sometimes says the wrong thing.
0:45:00 And despite that, it is actually very useful
0:45:01 in some domains.
0:45:03 That’s not like anything in the past.
0:45:05 When you used the internet
0:45:08 and you ordered something off Amazon in 2002,
0:45:10 it’s not like they would ship you something incorrectly
0:45:12 or maybe they would do that, but that’s not,
0:45:13 there was no expectation going in.
0:45:16 You would buy one book and you would get a different one.
0:45:18 And somehow that is still fine right now,
0:45:20 which is a cool part of the technology,
0:45:22 which is that it actually gets perfect.
0:45:24 It’s going to usher in a brand new set
0:45:26 of applications as well.
0:45:29 But I think the key thing to really think about
0:45:31 is not to think about, go back from a demo
0:45:33 and try to build that today.
0:45:36 So think about what products can you build today
0:45:38 that are actually imperfect,
0:45:41 but still can generate a lot of value.
0:45:43 And that is a lot harder than you would think
0:45:46 because in a lot of domains, if you are imperfect,
0:45:49 like let’s say you’re reviewing a legal document
0:45:51 and it actually completely reviews the document,
0:45:53 but 10% of the time it’s wrong.
0:45:55 You can’t give that to an end customer.
0:45:58 So actually thinking about the trade-offs of
0:46:02 how good does the quality need to be to ship your product?
0:46:04 How fast does the experience need to be?
0:46:06 In a world in which the quality is not perfect,
0:46:07 it better be fast.
0:46:09 There’s no world in which I’m going to wait 24 hours
0:46:12 for something, and the quality is imperfect, right?
0:46:16 And then if the quality is imperfect and it is fast,
0:46:18 how quickly can I correct it?
0:46:20 And these are all important factors of a product.
0:46:22 If you don’t hit the sweet spot here,
0:46:24 you will have a product that’s a cool demo,
0:46:26 but no one will ever use it.
0:46:28 I think this is the most important thing
0:46:29 to really think about.
0:46:32 Adding AI to just any field that exists
0:46:34 doesn’t just suddenly make the product usable.
0:46:37 It needs to be a product that is useful in its own right.
0:46:40 – I think if you give away a product for free
0:46:42 and everybody keeps using it
0:46:45 and a lot of people keep trying to get it,
0:46:47 then you know there’s something of value
0:46:48 and there’s product market fit.
0:46:50 And I think if you look at a lot of these AI tools
0:46:52 and AI demos, maybe they’re not realizing that.
0:46:53 Maybe they’re just building something
0:46:56 that is a cool demo or it’s like a really,
0:46:58 it pushes the limits of the technology,
0:46:59 but it’s not actually building something
0:47:02 that people want or will find value from.
0:47:03 I think that’s maybe our biggest message
0:47:05 to other founders or other people
0:47:06 that are building products in the area.
0:47:08 It’s just think like if I gave this away for free,
0:47:11 is everybody gonna wanna use it and keep using it?
0:47:13 I think that’s maybe something people miss.
0:47:15 – Do you think maybe that’s the model that,
0:47:17 I mean, I was gonna say software founders,
0:47:19 but maybe all founders that you should just start
0:47:22 with giving the product out for free
0:47:24 and see where things take you from there.
0:47:26 – It worked really well for chat GBT, right?
0:47:27 – Yeah, yeah, exactly.
0:47:28 – And it worked well for you, yeah.
0:47:30 – That’s actually an interesting principle
0:47:31 that Jeff just said.
0:47:33 Obviously, in some scenarios,
0:47:35 the reason why chat GBT could do that and open it,
0:47:37 it could do that is they own the infrastructure.
0:47:38 And if another company did that,
0:47:39 they would have gone bankrupt.
0:47:43 So you do need advantages in particular places
0:47:44 to be able to do that, but at the very least,
0:47:47 if you did burn money and you gave it away for free,
0:47:49 if no one runs to use the product,
0:47:51 you’re probably in a world of trouble.
0:47:55 – That’s a great place to end.
0:47:57 Varun Mohan is the founder and CEO of Codium.
0:48:00 Jeff Wang is the company’s head of business.
0:48:01 Guys, thank you for joining us.
0:48:02 That was awesome.
0:48:03 – Yeah, thanks for having us.
0:48:04 – Thanks for having us.
0:48:09 – Our producer is Claire Miller,
0:48:11 our associate producer is Allison Weiss,
0:48:13 and our engineer is Benjamin Spencer.
0:48:15 Jason Stavis and Catherine Dillon
0:48:17 are our executive producers.
0:48:18 Thank you for listening to First Time Founders
0:48:20 from the Vox Media Podcast Network.
0:48:22 Tune in tomorrow for ProfG Markets.
0:48:25 (upbeat music)
0:48:50 – Support for the show comes from Into the Mix,
0:48:53 a Ben and Jerry’s podcast about joy and justice
0:48:55 produced with Vox Creative.
0:48:57 Into the Mix is back for a new season
0:48:59 and welcomes you in with four new stories
0:49:02 that take listeners beyond the headlines
0:49:04 and into the lives of ordinary people
0:49:07 fighting for justice in their communities.
0:49:09 Starting with Ainez Bordeaux,
0:49:10 an activist and St. Louis native
0:49:13 who fought to shut down the workhouse,
0:49:15 a notorious pretrial detention center
0:49:18 that she says functioned like a debtor’s prison.
0:49:21 Subscribe to Into the Mix, a Ben and Jerry’s podcast
0:49:23 to listen to the first episode
0:49:25 of this special three-part series out now.
0:49:35 [BLANK_AUDIO]
Ed speaks with Varun Mohan and Jeff Wang from Codeium, an AI code generator. They discuss the importance of being a lean company, how their product stacks up against competitors and why having a level of paranoia has been imperative to their success.
Follow the podcast across socials @profgpod:
Learn more about your ad choices. Visit podcastchoices.com/adchoices