He Grew Up in a Motel. Now He Builds the Infra Behind Your AI Coding Tools

Hi Tejas. Hi. For people who don't know what Morph is, can you fill them in? Yeah, so we train models and sub agents for for Cogen.

One of my hot takes was that like these general inference platforms that will accelerate any model you bring, but I don't think that like that they're going to serve it better than you can. You grew up in a Budget Inn with zero programming knowledge. You taught yourself to build a hotel website and it was generating 48,000 a year. People underestimate how good of a product feel it is when stuff is fast.

One of the problems with long-running agents is like And so we use Composio for Apollo. We use it for my Gmail. Very easy setup. Didn't have to do a lot.

I feel like the way I talk to my friends is such that they are closer to their one-year goal as opposed to like them feeling good about their day-to-day interaction. You have to be as a founder like able to change. What's a prediction you made about AI a few years ago that came true? The concept of What do founders get wrong about AI agents and infrastructure?

It's really hard to find code to good Hi Tejas. Hi. Thank you so much for coming on. m.

right now. Mhm. So super appreciate you coming in late. Yeah, I don't know.

I mean it's it's a late if you're not a founder, but it's like the middle of the day if you're a founder, right? What would you have been doing if you weren't here Oh, full machine psychosis in the cloud code. There's a parallel cloud codes. That's where it came from.

You just like Yeah. Wow, you have like a bunch of monitors. No, I have one monitor that's culture-wide and then I have like a foot pedal. Then I have a foot pedal for like WhisperFlow.

So my foot my foot pedal triggers WhisperFlow and then What is it? I have shortcuts to move between my terminals. Oh my gosh. So you code with WhisperFlow?

Yeah, yeah. I mean I I use it for voice when I'm using cloud code. Okay. I mean you can't actually say speak code into it.

You have to speak like speak words into cloud code and cloud code just go. A pedal for WhisperFlow? Yeah, I'm early. I'm early.

Like I like I feel like it's not worth the like the waste of a hand cuz the hand could do a lot of it could switch like switch windows. It could do a lot of like higher higher value things than like press the FN button. So Okay. I'm early on this.

I I think everyone's going to be doing the foot foot pedal pretty soon. What What When do you think that they'll figure it out? Why have they not figured it out? I don't know.

I think it's cuz the the product's actually really bad. Like it looks like something It's like made for sewing machines, I think. Oh, wow. Uh for like industrial factory workers doing sewing machines.

And so like it's like this thing with USB-A and it's like not It's janky. So someone's going to make like a good version. Like an Apple feeling version. Yeah.

Yeah. We're giving it out in my alpha for free to people. And then MorphFlow will acquire. Yeah.

Sweet. So speaking of founder, let's talk about your whole journey. Mhm. For people who don't know what Morph is, can you fill them in?

Yeah, so we train models and sub agents for for Cogen. So our three main products right now are FastApply, which is a model that's really fast at applying code edits. WarpGraph, which is a model for searching for code. And our latest one from two days ago was compaction.

So FlashCompact. It's a model that's very fast at like taking like 200k of context and compacting it down to like 50k, for example. Oh. And that that outputs like 33,000 tokens per second.

So we we train and we serve these ourselves. We like own our own inference engines. And that's why you get to be so fast. And you said two days ago.

So you just recently launched Launched the compaction model, yeah. Wow, that's super exciting. And then what's the biggest misconception people have about what you're building with Morph? I mean I guess often people think that like oh, either we don't train our own models or we don't serve them, but like we really believe that we have to own the full stack in order to be a compelling product.

I think like a lot of these open source products open source models just off the shelf are not super valuable just cuz like you don't have a speed speed boost compared to a frontier model. You don't have a price a price benefit that much. So like I think in order to be like compelling, you need to be better, faster, and cheaper. Otherwise it just ends up being like OpenAI's model or Anthropic's model is just like easier.

Uh and so that's where that's why we own the full stack from like the training stack to the serving stack. Okay. for all of our models. And did you realize that only after you like you um after you started kind of building, you realized that we have to like own the full stack or else I mean it was only at the beginning, but like one one of my hot takes was that like like these companies like uh Fireworks and Together, like these companies that sell you inference are like these general inference platforms that will accelerate any model you bring, but I don't think that like like that it's true that you can that they're going to serve it better than you can.

Um especially if you're like in a niche. Like if you're just running a chat model, yeah, they'll serve probably be able to serve it faster than you. But if you're doing some niche thing like uh like for us for uh compaction, like our inference engine is really wonky. Like if you're if you're doing something truly compelling, I think you should own it yourself.

And if if if the value is there, it's hard, but like I think it's worth it. Okay, why is it hard? Inference is just very finicky, especially for language models. Like there there's these open source things like vLLM uh and SGELang are sort of the the two most common ones.

And then Nvidia has one called TensorRT-LLM. Um and these abstract a lot of the the complexity away cuz they there's a thing called variable length computation. So like when you when you get something in ChatGPT, right? It's like you don't know how how long it the input is or how long the output is, right?

Like if you ask something easy, it'll be short output. If you ask something hard, it might be longer. So um these are all things you need to weigh when you're actually like building an inference engine. Like and so these these products just like abstract it away so you don't have to think about that stuff.

But like again, that's like uh those are all very general inference engines just like a Fireworks or Together uh is also very general. And so like I think you either take one of those or start your own from scratch. Um and like really like make it just for one task. Like for us for compaction, like we don't care about any other model other than our compaction model.

So we can make trade-offs that a general serving engine wouldn't. Okay. If that makes sense. Yeah, okay.

So when I first kind of reached out to you about interviewing, you told me something really crazy or something very something I've never heard before. You grew up in a Budget Inn motel in Anaheim. Literally lived there from birth until you were 18. And then at 16 with zero programming knowledge, you taught yourself to build a hotel website and it was generating 48,000 a year.

What made you decide to just figure it out? Um so my parents only did sixth grade. They There was really no no means to like build a website or even set up Wi-Fi. So I was just like if they if I didn't do it, then no one was.

And so like I saw around me that like people were like this is very late. When I was 16, it's like 12 years ago. So like hotel booking was like a thing for like 5 years before that. Uh many many years.

And uh with they were just really late. And so it was very obvious that they needed a website to do direct bookings with. And so yeah, it was like back when WordPress was a thing. Uh I didn't barely know how to code.

So I was like hacking something together with WordPress and then uh the WordPress template didn't have something I needed. So I had to go into WordPress internals and look at what MariaDB was. It was a nightmare, but like these days it's way easier, but Yeah. back then it was a it was a it was a struggle.

Are you almost glad that you you kind of went learned like you without AI cuz now it's easier. Are you glad it was harder for you then? Like it was like Yeah, I think so. I mean I think AI would be really nice for me cuz I was like genuinely pretty curious and like it's like helps you like learn a lot a lot faster.

But there's a lot of people where like they easily use it as a crutch instead of a learning aid. I think I would have used it as a learning aid and maybe made my own little software so much better. And then your parents, they just told you like hey, could you like build could you build this? They're they're They didn't believe that they could build it.

Like they don't really understand like how to how even like to turn on their computer almost sometimes. So it's like it's quite it's quite a gap that like oh like software is a thing that they don't understand. They don't know the word software really. So like Wow.

understand that software gets written and then it runs on a computer. So yeah. So you saw you saw the problem and you were like I'm just going to take it into my own hands Yeah, it was like why not? Why not?

And school is boring. So why not? How did you know about like coding? How did you find out about code like programming coding?

I mean for me I kind of knew that there was like some software stuff in the background. I mean I was a very hacky I mean maybe unethical hacking before that. Uh I think the first programming language I actually wrote anything in was like called Windows batch format. Um and there was a bug in Windows 7 where you can like copy files from someone's computer without them knowing if you plug in a USB with the malicious bat Windows BAT file on it.

Uh so that was like my first intro into programming. It was trying to get files off of a computer I wasn't supposed to with the with a USB. And then your LinkedIn title for that job is hotel tech child labor. Yeah, I don't take LinkedIn very seriously like everyone does.

There's just something in that framing. I mean you build real infrastructure under real pressure before you ever set foot into a CS classroom. How did that shape how you think about building now? Um I feel like I I never really even wasn't a CS classroom.

I did electrical engineering. So Okay. Um hardware. Did you start with hardware like electrical?

yeah. And so like I never really actually had traditional like software engineering principles ingrained. Sort of all self-taught. Wow.

Yeah, for software at least. So you went from hotel Wi-Fi access points to working at Tesla autopilot firmware to ML infrastructure to computer vision models. You touched almost every layer of the AI stack. Mhm.

What did working across all of that teach you that most AI founders don't know? Yeah, I mean even even like other ML engineers, I feel like I just have like the full breadth between like training optimization, serving, everything in between. Because how you train and how you serve are like actually very interconnected. Like a lot of teams often like um like consolidate training into one team and serving into one team.

And then like there's this mismatch cuz often times there's things you could leverage at inference time that you should. And so I feel like the ideal person to build the best product is someone who understands training, inference, and product. And like sometimes you get one and a half. Uh but like I think all three together can make you make something really compelling.

This wasn't Was this something like Was your strategy to really learn like every every part of it? No, I didn't really have a strategy. It was sort of just chasing curiosity, really. Chasing curiosity.

Yeah. Wow. I feel like it's almost I at the time I was wondering if it was toxic cuz like I was wondering like a lot of people like if you look at um Ilya Sutskever like a lot of these people like they go like really deep. Like they don't have a lot of breadth, but they're like they understand like everything down to like the finest detail.

And like these people have outside outside returns. So I was almost wondering if like the breadth was a mistake, but I almost just like couldn't look away with for the breath. Like everything is so interesting to me. When was the moment at Tesla where you knew it was time to go and build something?

I don't know. I don't want to I don't want to speak bad about Tesla. I I really liked my time at Tesla. It was really fun.

Especially autopilot was very fun. Um Tesla was like becoming a big cult a company before I left. And so like with a big company comes like restrictions on how fast you can build. Like you can't yolo stuff to prod anymore.

You have to like go through a stakeholder meeting. You have to do like this review and that review. It's very good. And all that stuff is fair.

It's just not something I wanted to do. And so Uh I feel like with smaller scrappier just like I can do move very fast is like the culture I liked. Which Tesla does have to be fair. Even at their scale is like Like if you compare Tesla and like an Amazon for example, Tesla at their scale still moves a lot faster.

But uh I I just didn't really have a lot of patience for like the whole Okay. business side of things. Was there something else that was like maybe an inspiration of yours like a mentor where it was like pushed you to like go and build something of your own or You were like you always knew you wanted to do this and you were like just waiting for the right Yeah. It's not more like I always knew I wanted to do this.

It wasn't when it became like super easy of a decision. It became an easy decision once I were like doing my story pointing meetings and business stakeholder meetings. I was like Yeah. Yeah.

Do you think you should be waiting for for when this decision feels easy or or when it's like you should do it even if it's if it's hard? Hard. Like you got to take that risk. Cuz maybe you easier you take less of a risk cuz you're like Yeah.

Yeah. That's a good question. Like I don't know. I feel like you should be compelled to do it.

And so like you shouldn't need to be convinced into doing into doing it. Otherwise you're probably going to like end up losing your motivation a few months in anyways. You wrote publicly last year the hottest AI tools were helping devs write code faster. This year the winners are agents editing code faster.

When did you see that shift happening and why did most people miss it? Well, I think people underestimate how good of a product feel it is when stuff is fast. Like when we work with our customers like roughly there's a one-to-one correlation with improving speed uh and conversion rates. Like for example, if you double speed, you will roughly double conversions um in the AI code gen space where like you're doing these AI app builders.

Uh but you can't do that at the cost of accuracy cuz then you'll also lose customers on the accuracy drop. So as long as you can keep accuracy constant. If you double your speed, you can roughly expect double conversion. Uh and that was becoming obvious to me as we were onboarding our first customers.

Yeah. Okay. So you saw that um through like customers like Yeah. Customers, just anecdotal experience with using AI tools AI writes code, but getting it cleanly into a real file is broken.

Tell me the story of the time you realized this is the actual problem worth solving. Yeah. So I mean Cursor was the first company to do a fast apply model. Uh and fast apply is basically a dedicated model that is trained at applying an edit into a file.

So like traditionally a model can do like a search and replace which is outputs a string to delete and a string to for what to replace it with. Um and this is sort of like wasteful from a token perspective because the frontier model is outputting what to delete which is like not really that hard. Uh and you mostly care about the new tokens, right? And so around around like middle of last year like models were really bad at doing this.

There was like 20% error rate on this. And so this this secondary model was like super critical and it was like very very easy to sell. Uh because there was really no other way. Uh Today search and replace is a bit better, but it's still very token wasteful.

Like you'll still lose there your 20 or 30%. Um but you won't have that 20% failure rate anymore. Your failure rate is maybe like 4 or 5% now. Um but you still have that wasted tokens.

And so my whole philosophy is that like there's these tasks where like you you want to use the frontier model's tokens for stuff that only it can do. So like novel reasoning, novel coding. Uh that's what you should use like the OpenAI GPT-4 or the Opus 416. Frontier?

Yeah. So a frontier model would be OpenAI's model or Anthropic's model. Why is it called frontier? Uh because it's on the front.

They're like the Yeah. Yeah. The the biggest models on the frontier recently. Oh.

Okay. Yeah. Sorry. That's the code gen term.

It's frontier model. Yeah. It's nice. I'm learning a lot.

So I learned that you have So you have four products. You have fast apply, warp, grab, and flash compact. For someone who doesn't live in this world, what does each one actually do in plain English? Yeah.

So fast apply is a model that applies edits to files very fast. Uh warp grab is something that searches for content between in files very fast. And so these are all very useful in the code gen space. Um and compaction is uh is a model that takes like a lot of text and shrinks it.

Um and this is a problem like when you're using a coding agent like Quad Code for example, like 10 to 15 messages in you'll hit like a spot where you need to compact because it's too long. Um the model's performance is tanking, so I need to shrink it. " And then it'll take like 2 minutes and write like a long summary of everything that happened. Uh for us our model basically chooses lines to delete.

Um And so that's what we train for and we train to do this very fast. And so like like in a in a product like if you've used something like lovable before, like if you're doing a compaction in that, waiting like 2 minutes is like a great spot for a user to churn. 5 seconds. So you're not going to have the user churn then.

And so we're basically solving these problems in uh the code gen space. Okay. Why why is there churn for Why haven't they like fixed it? Like why?

Uh well so so churn typically happens when the user has to wait. The user gets frustrated, right? Like if you're waiting for a page to load, I'm sure you've used that Google Chrome before and if something takes too long, you just like close it. There's a pretty high chance like I I think there's like a there's like a graph where they show like like every second there's like an exponential graph.

Like if the user has to wait 3 seconds, there's like a 50% chance he'd like just bounce and leave. Yeah. That's true. Um and so like that's what I mean by churn.

It's like when the user is waiting there, they're going to be like, "Ah, what's happening? I I have stuff to do. " Uh And so like all of these products we're doing is like solving problems like these in the code gen space. Fast apply runs at 4,500 tokens per second with eight times fewer errors than patch diffs.

For someone who hasn't used it, what does it actually feel like in practice? What changes? Yeah. So it's 10,000 tokens per second now 10,000?

This was like a few months ago. Nice. Okay. Yeah.

10,000. So like someone using a product like from our one of our customers like like anything or uh Framer or Webflow, they would never know our product exists cuz we sell to these companies building their agents. Um And so for them they should feel nothing. They should just feel like a good user experience.

So that's the ideal. Like ideally when we're selling B2B, no one even knows our product exists. They just feel like it's good. That's a tip for design stuff, right?

That there's a tip for um like the best design it feels like it doesn't ex- like doesn't exist. Yeah. Yeah. Cuz you know it's like very chaotic and all over the place.

Like you know that's actually good. like the eyes should be in design at least the eyes should be drawn places naturally. You should not think about it, right? So that's how you operate.

So you just like you don't want No one should know like I mean it'd be cool if I just had like a banner powered by Morph, but like the companies that use us aren't going to do that. You sh- You said agents are shifting from quick IDE edits to tasks running for hours or days. What does that change about the infrastructure underneath? I mean I I think I tweeted recently that all agents are becoming coding agents.

I mean this is sort of adjacent to what you're asking, but Yeah. um I think like when you have higher trust you can have agents do more and more. If you have higher trust in compaction, that happens. Cuz one of the problems with long-running agents is like let's say you have an agent run for 30 30 minutes.

It's it's going to need to compact at some point because it's too many tokens. Um and so having that process be good is one thing to to enable long-running agents. Um and why I think they exist is just because like Like for example, if I was to say like build me build me an a mobile app for Morph, um that's going to be a very long-running task. You have to have trust that that's going to work well.

Uh but we're getting a trust factor where the models are getting good enough where this can happen. So here it says you're you're actually a two-person team right now. Yeah. Yeah.

And were you you were on Times Square on like Oh, for the YC launch thing. Yeah. Yeah. And then trending on Open Router.

Mm. That was temporary. But yeah, the first week first week we launched. Yeah.

What's the thing you did early that you think most teams skip? m. every day. Every day.

I still I mean I still do. Like honestly it's bad. I feel like Quad Code has like this stranglehold on me where I just Like I I can't go to bed. Do Yeah, you can't go to bed until you finish.

Like a task from something like that. think like I'm I'm going to get like this one agent running that's going to run overnight. Um it just takes me so long to get to that point where I can like trigger like set off the long-running one. Why is it only two people right now?

Uh so part of my calculus when we're hiring is like So like a full-time hire would be like 200k, right? So roughly with like stock and insurance and everything. And so it's like should that 200k be spent on GPU compute or should it be spent on human compute? Uh and right now the highest alpha thing for us is to just like use more GPU compute.

Uh and so I mean this is this will probably change soon uh cuz especially now that we're like forward deployed bottlenecked. Like we have more integrations than we can possibly handle right now. So probably have to hire people for that now. Um but pre- like historically from today and before it was this was the the logic when we were talking to someone.

" And usually a GPU. It's just really hard because for us like the GPUs are so useful. I feel like this is a dumb question, but what's the what is GPUs? Uh GPU is like um Did you know the company Nvidia?

Yeah. So Nvidia makes GPUs. And the GPUs are used for training and inferencing the models. So inference is like when you Okay.

like when you use uh ChatGPT, that's the inference. Um and behind the scenes when they're training it, That's GPU. That's uh I mean they both use GPUs. Uh but they're different processes.

Like the training is like when you're It's like a whole different process with GPUs and when you're serving it into inference. That's also on GPUs. But that's what you're feeling in ChatGPT is inference. GPUs are very expensive.

Yes, very expensive. We spend like 8x more on GPUs than we do on salary. I guess it makes sense for founders. Like founders are they're making a lot less than their like spent.

Yeah, I mean, I have one hire and he he gets his phone done. But like other than that, it's just all GPUs mostly. You're using Composio for Slack Bot. Uh-huh.

Walk me through what the bot does and what problem does it solve for your team or your customers? So every time someone signs up for Morph we have like a So Apollo integrates into Composio. And so we use Composio for Apollo, we use it for my Gmail. And so we send a welcome email and we we do like an Apollo lookup for like that gets their LinkedIn link and stuff.

And then and then send a message in our Slack with all that. And that's all using Composio. I mean, it was cool. I mean, it was like very easy setup.

Didn't have to do a lot. Yeah. I feel like the onboarding to like get set up could be a little bit better cuz I feel like sometimes it's not clear in Composio in the dashboard for like where to go. But like overall, much better than having to do it myself.

Wait, why is it not How is it not clear? Well, cuz I'm using it via the API. And Composio has Ruby, right? It's like their Oh, we're disconnected.

Oh, I see. I see. Yeah. The dashboard is coming Yeah.

I I bet it'll be good now. Composio is like so big now. I bet like the UX will be a lot better now. Oh, yeah.

How did you like find out about Composio? I think it was like just like the first MCP marketplace, right? It was just like a very I was a very early user. I was like Okay, it's much better today than it was when I when I first started.

It was a really bad one I first started, but I still used it. Aw. It was a long time ago. I think I think I was one of the first users.

I think I saw like their launch probably and then I used Composio. What would building the Slack tool have looked like without Composio? So we would have to get like the Apollo API docs, built like a built Apollo tools. Apollo has like in Composio it exposes like 20 or so tools.

Um yeah, so it would have looked like building all of those. Which for us we're like too small of a team. So I think like that's that would be too much work. I do maintain a lot of other stuff.

I maintain like a Slack CLI, LinkedIn CLI, and Gmail CLI, and like all this stuff I do myself. But How long did it take you to ship the Slack Bot with Composio? I think like a day. And then without it would have taken a week I mean, I do a lot of stuff in parallel.

So I mean, maybe like maybe like 3 days two days two days two or 3 days. Okay, you like in parallel you like not multitask, but in parallel, yeah? Yeah, I mean, I have like the 6 to 8 cloud codes running generally at the same time. Okay.

So you're always doing multiple tasks at once. Yeah, because you can because cloud code takes so long to to run. I don't know if you've ever used cloud code. You should.

It's really cool. Even if you're not technical, it's really good. Uh So like you can like tell one one cloud code to start working on one thing. You can tell another one to start doing another thing.

And while that one like takes like 10 minutes to work, you can still do a bunch of stuff in parallel. Okay. Cuz I use cloud like for like for other stuff just like and it does take a little bit to cuz it like I it helped me create a storyboard. So you used it on the web app?

Yeah. On mine? Yeah. You Oh, you can like So cloud code is in the terminal.

Terminal on the terminal. Yeah. Are you using Composio for anything beyond beyond Slack? Or maybe what tool kit are you most excited about the most?

From Composio? Yeah. Typically I just use it to like bring in like cuz they have like their hundreds of integrations. Just plug it in and then I have the Composio via their Composio API.

And that's what I've been the That's what I use it for. So I don't use it outside of the Slack Bot. But maybe I should. Okay.

As a founder who has been on both sides, building infra yourself at Tesla and now buying it from Composio, what's your framework for the build versus buy decision? It's like that Rick Rubin meme. You just like vibe it. Do you know what I'm talking about?

Wait. No, what what do you mean? Oh, there's some memes are always hard to put into words. It's like Do you know who Rick Rubin is?

No. It's it's not it's not worth it. Is he in tech? No, he's like this this guy that is known for being like a taste I mean, you're trying to be a tastemaker, right?

Yeah. Tastemaker storytelling. Rick Rubin is well known for like being the the tastemaker type guy. Okay.

Um and he has like this meme of him like closing his eyes on the computer just like showing him Like he just Yeah. Uh that's what I'm saying. Like I just sort of vibe it. But like essentially like we're a small team.

So I just whether I build or buy it's sort of almost like an impulse. Like if I feel like it's going to be more than like an hour of work Okay. perhaps I'll go buy. But increasingly my decisions are becoming more build just because like to have cloud have the full power to configure stuff.

But despite that the Composio Slack app is still useful for us. What would you tell any other technical founder who's on the fence about using Composio versus rolling their own tool kit? Um I mean, I think there's the pros and cons of like both options. But I think uh especially if you're going to be doing like multiple integrations that are already like Apollo for example.

Um I don't know if the LinkedIn connection does like connection requests and messages and stuff like this. Haven't tried it out. Well, like a lot a lot of this stuff like LinkedIn scraping has antibotting and whatnot. So like a lot of that work would just be too much work to do yourself.

Especially if it's not internal tools and whatnot. It's like just not If you're going to build, I see that the use case for building if you're like it's center to your business. Like if you're building like a like a Gmail app. Like you can't use Composio to like to Composio Composio's Gmail integration for your own Gmail agent that's externally you're selling.

But if you're just building something quick for like a Slack app or some internal tool where you need it, it's like it's it's pretty good. dev. They're real companies using Morph in production. What are they building with that surprises you?

I mean, I work with so many of them like so little surprises me now. But like I I I mean, in the beginning there like Create was one of our earliest customers. I think one thing that surprised me is like how they store their repos, stuff like this. I mean, I probably shouldn't share what their exact stack is.

But like um like a lot of the stuff that they're building like in charge of which is just so early. Like you would imagine that there's a good way to store like thousands of repos, but there's not. Like GitHub doesn't let you. Uh so like they all go looking for some like hacky like some like serialize it and you store it on S3.

Some will just like dump it into their database. Um others will just like fix it in their sandbox and like freeze the sandbox. Um and there's trade-offs to all of them. Yeah, it's like when you're working on the frontier, there's nothing built for you.

So they just have to do a bunch of hack job solutions a lot of the time. Which is just the nature of it. Um how did they become your customers? Uh for Create it was Twitter.

Continue was over email. Data Button was in person. I flew there. Uh Where what Where did you find them?

They're in Norway in Oslo. Oslo? Yeah. Yeah.

Yeah, I was coming back from China from another customer. And so I was like I could just do that. Um In person meetups with the customer, yeah. That's how it's done.

I mean, when you're a small startup, it's hard to like Sure, you're you're like a dependency and they're trusting their production app with you. So like the trust from just like an email or like a LinkedIn message is like almost not going to be there. Uh whereas like when you're in person you get like this dense understanding of who this person is and like would they be there for you if like that kind of thing. So I think that the trust goes a long way especially as a startup.

Your your job is to make some some someone else that's building like a bigger product uh trust you as a cuz they're going to trust their what their app with you essentially, right? They're building They were building a bigger product, yeah. I mean, they they were building a much bigger app at the time. Well, okay.

You said your models power AI 20 in the background so your customers can be in the headlines. Is that a deliberate choice to be invisible? Oh, yeah. You You were talking about this already.

Yeah, pretty much. I mean, my job is like their agent needs to be really good and really fast. Okay. Their customers don't need to know about Morph.

But like I need I want Morph to be so compelling that everyone needs to use us from the from the agent perspective. You went through YC as a solo founder. Most people say don't do that and that is just very hard to get into YC as a solo founder. What What was that actually like?

Like yeah. Honestly, I feel like I let it get to my head. Like I feel like when I got into YC as a solo founder, I was like oh yeah, I'm the I got into YC as a solo founder. But it was such a stupid stupid way to think of things because like it's not anything.

It means nothing. Like the the job is to build a product everyone needs or that people need and be like genuinely useful. That's the hard part. That's worthy of like respect.

Not like getting into like this made up thing and be like a made up like being a solo founder is like and getting into YC is like not actually a thing. It's not it's not people say it's like they won't accept you if you're not if you're not you don't have a co-founder. That's That's just It's rare. It's like like there was only like eight of us in our batch.

But like it it's not actually anything concrete, right? Like getting customers and like building a good product that's useful. That Those are the hard things that are like worthy of a celebration and respect. Not this made up thing trite thing that I was proud of.

That And so like I feel bad that I was like I let that get to my head cuz I feel like it's not actually a thing. I mean, it's definitely a lot harder. I would say like Yeah, but I think maybe I was Maybe it's worth like a good job. But like not like a not to be so proud for many weeks.

I think that that was a mistake of mine. What did you have to figure out that a co-founder would have to handle? Um So like as a as a solo founder like when like when I come here for example, like there's no one leading the company. Or when there's a when I do fundraising for example, like no one's leading the company, right?

So that stuff is a lot different. So like my calculus on stuff I can spend time on is a bit different because someone has a co-founder, right? When they're fundraising other the CTO could just like handle things. Uh but I don't have that.

So like when I when I'm doing things I have to like think of different structures and systems. I make trade-offs like another two-person or three-person co-founder setup wouldn't really understand. Like for example, if like some customer comes in now that like isn't worth the time. Sometimes I have to like Like I the way I do outbound for example is like within my cloud code and like I have like a bunch of LinkedIns and I have a bunch of emails.

Uh and like maybe it would be better to do this all manually, but like the trade-off is like I have to take from like both on the coding side, go to market side. I think many two-founder or three-founder setups would say it's like not optimal or insane. And what was the choice behind behind being a solo? I mean, I had like a a bar that I wanted to have met and like I only really wanted I think like being a co-founder with someone random rarely works out.

Uh and so that was sort of why. I mean, I worked with one of my close friends and like I think it didn't work out from a technical perspective cuz we had like different technical bars, but Um yeah. You're very careful like who you're like choosing to be like on a team and Yeah, both with hiring as well as At what point did you decide you needed people around you? I mean, I think once we got to like 100 Slack.

Once I got to 100 Slack channels, I was like I don't know what I can do now. Like you know, like it's getting out of hand. We're at like 210 or 215 or something now and it's like Um even for us two, it's a lot. But like I I I expect a lot out of everyone.

Like I I share all of my CLI my CLIs or setups and like I think one person is capable of a lot more than your people think. I think we're we're all we're still picking up like assumptions from last year and the year before like what one person is capable of, but I think one person is capable of around like I don't know, around like 10 to 15x more than like a year and a half ago. Uh and I think the expectation today is maybe 2x that person, uh but I think it should be a lot higher. More prep I did before this interview was reach out to you It's one thing Molly O'Shea does.

Mhm. He reaches out to the the person she's interviewing, their teammates and friends, and asks them to or do diligence. Yeah. So what I asked that some people about you is what's something about Tejas that most people don't see from the outside, and how would you describe Tejas, who is he through your POV?

And I got some answers. We're going We're going to read them out. It's going to be brutal. I think it's going to be bad.

It's not going to be good. Okay, first quote. Tejas is very intensely him. If you're closer to him, you're still the same.

No secrets. Just all the traits are more amplified amplified. What traits define you? Like if someone got really close to you, would they see more?

What would they see more of? I think the filter comes off. Like I'm Like I'm not afraid to tell someone like that your code is bad. Like I'll I'll tell them like this this PR is like ass.

Like I'm not merging this and like you should be embarrassed of this. And so like some people think it's kind of mean, but like I I only really say that if I think that a person is capable of more. If I didn't think the person capable more, I just kind of like give up. It's like it's not worth the breath if they're not capable of it.

Like you're just being mean for no reason. But I feel like that the nicest thing you could do for someone is be mean where it's like net ROI positive. Like Like there's people that are like nice in the short term, which is like you're nice in your day-to-day. Um but is that really nice in the long term?

Cuz like if you're that person and that person ends up becoming like not a good engineer a year later um or like they don't get If it's if their goal is not being a better engineer, whatever their goal is. Like My I I feel like the way I talk to my friends is such that they are closer to their one-year goal as opposed to like them feeling good about their day-to-day interaction with me. That's really good. Oh, okay.

So some people think like they they take away from that is like he's a meanie. But like But like at the end of the year, I think it's a different story. Okay, very critical feedback is very important. Like and if someone's giving you that feedback, it doesn't mean that they're trying to put you down.

They just generally care. Like they want you to succeed. Yeah. Well, I don't know if I want everyone to succeed.

I I I would only do it if I think that person could get better. Or like if they're capable of more. Okay. Like if you see me and I'm not giving you like good feedback or like like if they're I generally just don't think you're capable of more than you think.

So that But I I bet that person in the short term is going to say Tejas is nice, but like Oh. The capable people I feel like can like actually take the feedback and grow from it and actually get net better. That's true. Yeah.

Not that I know everything, but yeah. He's very hardcore. Codes all day, every day until he beca- became a CEO and had to do other things. He seems to get a kick out of improving things, both code, products, but also people.

Mhm. You code That's a good take. Yeah. Yeah.

You agree with that, yeah? Yeah. Yeah, I do. I mean, you code every every day until being a CEO forced you to stop.

What did that feel like when the job changed and you couldn't just be in the code anymore? I'm very much still in the code. You're still in Uh but I mean, I like I have to do the other stuff now. Like I have to like when there's a contract, I can't just like Okay.

code my way through the contract. I've tried though. Like I've tried to like give like a docx file to something I built. Like edit the docx file.

I'm like some things you just have to end up doing like a DocuSign or something. Oh, yeah. So code code can get you quite far, but it can't get you everywhere. Like it doesn't build the relationships with people.

It doesn't do the DocuSigns. It doesn't do the security review either. Do you want to code? Like you want to have coding as part of your job?

Like for a lot like No, I mean, I think like the ideal founder from like zero to 100 million or zero to 500 million and then 500 to 2 billion. Like they look they're almost like completely different people. Like you have to be as a founder like able to change. Uh cuz they're like completely different people.

Like Cuz like from 500 million to 2 billion, you're scaling up your go-to-market. Your job becomes hiring and fundraising as opposed to like right now it's like making sure the product is very compelling for us. Uh I think that in order to have good product sense in or in the domain I'm in, I think I do need to have my hands in like training and in cuz like right now like last week kernels came out for A6000 on uh A6000 Pro. And so like that enables a certain product that we couldn't have been exist exist before because blank.

And things like that I need to know about a high level, but I don't need to be like coding as much as I am today. I think maybe like today it's like 80% of my time coding. Uh or maybe 70. And then I think maybe it goes down to like 25 or 20 or something.

He's a hate hater of half-assing things. He likes surrounding himself people that are authentic and very intensely care about something. He believes actions speak louder than words or intentions. Only looks at the result of your actions.

Very true. Very true. Only looks at the result of your actions. So not words, not intentions.

Has that ever cost you a relationship? Yeah, I think so. When you notice that maybe someone is kind of just kind of words, you feel like you try to like stay away from those people? No, I mean, I just think like what when someone does an action or chooses to do something, it speaks much more loudly than like what they say they're going to do.

Okay, last quote. He thinks very deeply about his work, in this case the future of AI, and so has surprisingly accurate predictions. And known him for a few years now, so I I can verify. This person can verify.

What's a prediction you made about AI a few years ago that came true? Well, I mean, basically all my predictions about code gen has sort of has been coming true that like you'll like the the concept of vibe coding where you'll be able to build like web apps with code It's from from English. 5. I mean, it's If you're like in the ML space, it's not that hot of a take, but it seems like a hot take if you're outside of it.

Um I mean, I was generally bullish on AI-driven approaches even back when everything was confident that it's an L S T M. So I was working like um Like I was I was in AI around 2019 as well, so. By by by 2019 definition, today's AI is definitely AGI. Like uh but yeah, I was generally kind of early on that as well.

And what's one you're making right now that people aren't taking seriously? I'm going to start to take it more seriously. It's the concept the concept that all agents are going to be coding agents. I think we're in this like weird period in between where not all agents are coding agents.

Um but uh the stacks are flushing out. Every agent you're going to use is going to I I don't think everyone's going to know. Like I think you're going to use like an agent and you're just not going to know there's any code. Like like like you're going to say like do this thing with my Google Sheet.

You're going to say um do do this research. And um I think it's going to go start writing code and doing some cool stuff and you're not you're not going to see the word Python anywhere. You're just going to see magic. Here's the thing you wanted.

Yeah, that's where things are going. I think cuz cuz code isn't nearly unbounded. Like when you when you build a structure inside of tools, everything is bounded and code gives you this sort of like things with no ceiling. And all that you're limited by is model capabilities to code.

And models are getting so good at coding that like that ceiling is almost like you can't see. What's on the road map that you're most excited about? The thing that keeps you up at night in a good way. I'd say right now we're sort of tapped out on launches for the model side at least.

Compactions, uh we're just working on like scaling and integrations. Uh we have a release for product in like a month or so, I think. Fine. Uh I probably shouldn't say too much about it, but it's it's very cool.

I think it's uh I think it'll make us a lot of enemies. That I don't know. Like big company enemies. Like it's competitive with some Yeah, I'm not the vague post.

Like it's uh that's what I'm I'm excited about is launching that. I want to see how it goes. But nervous though cuz we haven't really launched a product before. Cuz it's uh Like we've we've just launched a whole bunch of models, right?

As APIs. Okay. But so this would be like an actual product. For actual first product, that's that's so huge.

Oh my god. What do founders get wrong about AI agent infrastructure? Oh, a lot. Like I work with a lot of the YC startups.

They got a lot wrong with it. Like in the beginning like they're just trying to get stuff working. I think it's really hard to vibe code to good harness right now. Like if you say like Claude make me like this agent that's good and does this these tools, it's like quite bad.

It's really bad. Infrastructure like some people will start storing like for coding agents they'll start get GitHub repos in a Postgres database and Uh it it there's there's just a lot of stuff to get wrong right now because the it's so early in the space for agents. And like all the stacks are sort of built around traditional software and people like hotchpodging things together to get it to work. I think one of the things they get wrong is like they focus too much on what people want today uh versus what's capable for tomorrow.

And other is I think they they ask their users too much for solutions. Like the user what the the users don't know what the AI can do. Like they have their problems which I think you should listen to, but I don't think you should listen to their ideas because they don't really know like what AI can do cuz cuz like what user would have asked for ChatGPT, right? 5 was there or three was there, no one was asking for that.

Like you ask any user uh they would not have said like oh I want this chat and I want like a chat box with it. Yeah, no one said no one said no one said that. No one no one would have said that like No one would have known that was capable. Yeah.

That's why I think like the best the best products and best founders right now are the people that are have their hands in ML as well as product. Yeah. And then you can sort of merge the two and build something like really good. Yeah.

I almost I read something in a book and it mentioned that like you build something that almost sounds absurd. Like something that like people today would would be like what the heck? Like no, this is not possible almost. But something, yeah.

But like maybe So maybe it's a it's okay that if people question you, like it's not the end of the world. If people are questioning your idea, like just keep going on that. What do you think? Or like they don't believe in the the solution like Yeah.

I mean I think if you launch it and people don't use it. Yeah. I mean I think if you launch it and people don't use it Okay. it's got to be a bad idea.

It's probably a bad idea. Okay. So it's a launch factor you got it. But you don't need inspiration from your users to build a product.

Like you I think one of my hot takes is the YC advice is like talk to your users, get your ideas from your users and just do what do what they want, do what they say. Uh I sort of think like the only part you should take from that is like listen to their problems. Uh but definitely don't listen to their solutions and just like cuz the the founder today like you use LLMs all the time you get intuition of like Claude can do this, but it can't do this. And it it's bad at this, but it's good at this.

And this part is like And so you should just sort of trust that intuition as opposed to talk to like listen to your what your users ask for too much. That is a hot take, but it's such good advice. That is so good. I mean at least it's been good for me.

I mean it took me many years cuz like I after YC I pivoted a lot. Uh and I think part of it was cuz this I was hard for me to step away from the advice, but I think like how some of the YC advice is really good especially around like co-founder disputes and fundraising and stuff like this, but some of it I don't know about especially AI products. It's like a lot of the YouTube videos especially are a bit out of date now. You grew up building the tech for a motel and now you're building infrastructure that agents use to write code autonomously.

What's the thread between those two things? Uh I mean I like just seeing opportunity really. I mean like at the at the motel I saw opportunity because we didn't have a website. Here I see opportunity because I think no one's looking at like both speed and inference optimization as a as a way to optimize models.

I think today a lot of people they they they know that they can train a custom model uh but they typically will just go to a Fireworks or a Together to go serve it. Uh and we're like okay, we're going to own the full stack. We're going to own the um the RL, we're going to train it, we're going to serve it ourselves on our own custom inference engines and we're going to extract every bit of value we can this way. Cuz I think again, to be compelling we have to be better, faster, and cheaper.

Um and maybe we could be like a little bit better if we trained it and just served it to someone general, but then we wouldn't be faster cheaper um and yeah. I I think that like just training one model one time is not really good recipe because you'll just OpenAI and drop it. Just release the next one the next one you'll just be displaced. Like you need some sort of system where you're going to keep retraining um for us at least we retrain every week.

Or you need online RL. So online RL is like a sort of new type of way of approaching these problems where like your model self improves based on data from people using the product. Like this better, faster, cheaper. Like what do you what do you like which one do you shoot like first?

Like which one you go like you're Um for training models? Can you do everything like better, cheaper, faster? Is that possible? mean yeah, it is.

That that's sort of what our whole research is is like does that main model with our model as a tool does that outperform that model alone? Um and so we eval on that. And of course obviously it needs to be better with our model otherwise it's not worth using. Um and better, faster, cheaper means you have to use less tokens from that main model which is like a Claude for example.

Um and so it ends up being that faster if you're not using if you're using less tokens from that as well. And then better is more accuracy, cheaper is based on how we price. Um yeah. And faster is from our inference speedups.

What's one thing you know now that you wish you had known when you were starting more? To prioritize sales more. I think like I I know that and I know this now. Um but in the beginning I think I just I I had that I had this tendency to like just go into the the basement code and just uh lock out all users and hope that they find my website.

Uh but this is uh I I'm much much I'm much reformed now. Like I I I see the light for for sales go to market, its importance, how to do it, how to do it well. You're learning you're diving deeper into like marketing, go to market, yeah. of my job now is that.

I mean I don't know if it's most of my job. It's definitely a large part of my job is that. The whole go to market motion. Yeah, cuz you're like I am able to I I think I've reached the limits of how much I can push into code.

Like go to market into code. Like I put my Slack, my Gmail, my LinkedIn, and everything into my uh um into my Claude code. So I can I can do my market, but to some extent like the in-person stuff I can't Claude code my way into in-person stuff. That's true.

Something yeah. Right now. Yeah. Maybe I can Claude code my humanoid to go to sales calls.

I was going to say I was going to say yeah, like you double you like But would they yeah, would they want to speak to a humanoid humanoid version of you? No, probably not. Now they're going to be like who is this like person like That would definitely definitely lose the deal that way. Yeah.

Yeah. And it might be cool to do that but I don't know if it's optimal. Yeah, I don't know if it would be cool. Yeah.

If you jumped on to it early Yeah, like you know how all the startups do the growth hacking like bring donuts or something? Yeah. I'll send you my humanoid instead. Yeah.

Be a good good growth hack. That's yeah, that's crazy. Wow. Okay, this is amazing conversation.

Thank you so much for hopping on.