AI innovation, Y Combinator, and meeting Sam Altman! Yang Li of Cosine

Strategy & Tragedy: CEO Stories with Steph Melodia is the best business podcast for curious entrepreneurs featured in the UK's Top 20 charts for business shows.

Hosted by Stephanie Melodia, Strategy & Tragedy features candid interviews with entrepreneurs who have scaled - and failed - their businesses - sharing their lessons in entrepreneurship along the way. From Simon Squibb of 'What's Your Dream?' Internet fame to Lottie Whyte of Sunday Times Top 100 Fastest-Growing company, MyoMaster. From exited founders like Nick-Telson Sillett to subject matter experts like Alex Merry in the public speaking arena and Matt Lerner, the GOAT of Growth.

This is one of the best podcasts to listen to if you're looking for educational and inspirational content on Spotify, Apple, Google, Amazon, YouTube or watch the clips on Instagram, LinkedIn, TikTok, or YouTube Shorts

-

In this week's episode, Stephanie Melodia interviews Yang Li, the Co-Founder of Cosine multi-agent AI tool designed to do compound tasks in complex codebases with a 72% SOTA score on SWE-Lancer.

Watch on YouTube via the link below or keep reading for the transcript, where Steph and Yang cover:

• Is AI a bubble? And when will it burst? Yang's hot takes

• Meeting Tom Blomfield, Sam altman, Aaron Levie, and Michael Seibel

• Best advice for entrepreneurs in AI (hint: building for AI's "mediocre zone")

• HONEST thoughts on Y Combinator

• How we manufacture luck

SM: Yang, how can entrepreneurs win the AI game in 2025-26?

YL: Build something that is just outside of AI's grasp.

I think that we have been iterating our product for almost two years, and whenever we want to try to build something new, increase our feature set, increase it always felt just slightly beyond the possible.

But actually, by the time we get there, the AI has improved and a whole new technique and the whole landscape has changed. And suddenly you've unlocked a huge amount of value that your customers have just begun to think about. And you're already there.

SM: So by just beyond the grasp, is that kind of bleeding edge? Is that what you mean like making sure you're kind of like that at that forefront as much as possible?

YL: I think it's not bleeding edge in terms of the actual techniques of how to train AI. It's more the bleeding edge of like, what is AI good at today? And what is it kind of mediocre at? Build that feature set that supports the mediocre skills and tasks and functionality, because that is the next bit that is going to improve, right? The rate of improvement is phenomenal. Like I always say to all of my lots of AI founders that I know, have an eye on like what kind of sucks at the moment and build for that because it will come faster than you know.

SM: That's a great tip to even just exercise your futurist muscle, right?

It's like looking at like where are we already just kind of mediocre and improve upon that. It doesn't need to be like revolutionary groundbreaking.

YL: Yeah, like I think even within, you know, six months ago, video generation where AI models was considered like, oh, it's a bit of a party trick. Yeah. Right. It's not very good. It's not very realistic. It's very short clips, a couple of seconds. And now you're seeing on YouTube, people are doing, you know, short films. People are doing TV series that are completely AI generated.

Just the other day, I saw a complete Viking documentary series that was all AI generated, written by a real script writer. And so it was just very compelling, right? And that six months ago was not possible. And then actually, I guaranteed that these people were dreaming to say, if only I could build something. And they were preparing how to build an entire production team that is centred around this.

SM: So as a follow-up question to that, how do you connect the technological innovations with market validation? Because it's one thing to kind of be an inventor and a futurist and like really kind of push those technological boundaries. But connecting it with a market need is, of course, the difference between a hobby and a business.

YL: I don't think it has changed, right? With or without AI, how you figure out there is a market need for it hasn't changed, right? Like, I think there are definitely some places where AI is a game changer, where it wasn't possible before, or that industry wasn't invented before. For example, you know, there's plenty of people who are doing prompt optimization, right? Like that didn't exist before AI came along or LLMs became popular. But, you know, you it's an accelerant, right?

Think of it as it makes whatever that you were building before or whatever was a good idea going to go much faster, much better. The experience is going to be good and bad. Like there are challenges with AI, but I don't think validating a business idea has changed. And I don't think it will ever will change, right? You have to know the specifics of a market. You have to know the psychology of your buyer.

People don't necessarily always buy the best product, right? Technically, the best product doesn't always win. People are, whether it's brand loyalty or just impression or just how they feel, right? And overall, like not always, right? Like the specs of a product is what wins in the market. Those first principles are always constant.

SM: Hot take, is this a bubble? Is it going to burst soon? Is this just the beginning of a whole new era?

YL: It is a bubble, but I always think that with any bubble, once it bursts, we are always at a better level than where we were, right? I think that there's a lot of enthusiasm, there's a lot of capital coming up, right? Even if you look at something, which I think a lot of people think of the bubble crypto a couple of years ago, I think everyone would say it was a huge bubble, way too much capital pouring in. But now, you know, stable coins has become such a hot topic, right? And that would not have come out if the crypto bubble probably didn't exist, right? People were inventing and trying and seeing what works, et cetera. And I think AI is the same.

I think that AI feels like an even bigger bubble, but the hope is that if and when it bursts, and I think it will burst, at least it will deflate at some point, we're going to quite significantly change how life looks like and make every part of industry and every part of life slightly better.

SM: Yeah, I'm with you. So let's connect this back to your business, The Day Job, and why we're talking about AI.

So you're, of course, the co-founder of of Cosine. For anyone who's unaware, give us the elevator pitch of Cosine.

YL: Fundamentally, we're AI dev tool. There's a few ways to look at the market. There are AI dev tools that basically help either the human go a little bit faster, right? So whether it's inline autocomplete or adding AI elements into the code editor. And that's great. Like all of our engineers internally use like AI enabled IDEs or basically code editors. And that's great, but what we see is that it fundamentally is still limited. The ceiling is much lower. So a human will only really open one task at a time. I think of it as Google Docs. You don't open three documents at the same time. You don't edit three at the same time. And it's the same with code, right? You're doing one task in one code base at any given time. And all these tools are making people's experience much better. It's people who learn to code are now becoming much more experienced much more quickly. The experienced coders and the output is becoming faster, but they're so limited that they're doing one at a time. The way that we see we're building our product fundamentally is that it's asynchronous and agentic. And what that really means is - well, go to your JIRA board, your linear board, your Asana board, wherever it may be, assign every single task, right? Whether that's two or whether that's 200, it doesn't matter to us. And our AI will go away and do those tasks. It will do things that every normal human developer will be doing. It will be researching your own code base. It will be accessing the internet to read documentation. It will use code execution to basically test applications every line of code it's written.

So as it's writing, it's being tested. And at the end, it will run something called CI, which is essentially you think of it as a checklist of things that should be validated inside a code before it's sent for a human review.

We're fundamentally trying to say, we're not trying to let people accelerate people doing one task at a time. Give us the task that you don't want to do or the task that you never get to, or just simply, you just want to have AI have a first attempt. And we are particularly good, right? Which is why we have a really special relationship with OpenAI and a few other sort of hyperscalers.

We spent the last two years training our model to be the best at navigating existing code. There's a lot of wonderful products in the market. If you want to start a brand new app, if you want to do a brand new project, right? Don't use us. There are better products at a moment for that. If you have existing code, if you have legacy code, you have messy code, you just have large projects. We are fundamentally way better than anything else in the market. And we can prove that by independent benchmarks, right? We are just outperforming any model, whether it's an open AI model or an anthropic model.

Fundamentally, what we have, our proprietary model, is just outperforming in existing code bases.

SM: Incredible. Well, congratulations on that. And linking this all back to the previous points around kind of the bubble, and you do predict that there is going to be at least a deflation of that bubble.

You started Cosine a few years ago, along with your co-founders, which we're definitely going to talk about as well.

Do you believe that you started at the right time, that you're positioned well in terms of the few years ago when you started the business and the selection of that niche to position you for the highest chance of success when that bubble deflates?

YL: Yeah, I think, you know, it wasn't by design. It wasn't that one day we woke up and thought, you know, based on the analysis and the reporting that we knew that that was the right time. I think that for us we were always playing with some of the older models, right? So, you know, this was GPT-2, right? This was way before ChatGPT. We were playing with older models like BERT, right? Non-LLM models as well. So we had always played as nerds, really, as understanding, well, we wanted to build something, but it was always... not quite good enough to do any of the fun things that we wanted to do, whether it was useful things or fun things. It wasn't quite there. It was when GPT-3 came out. And this was a few months before ChatGPT came out, right? So just the model in itself. We had an aha moment where we thought, right, like it's actually pretty smart. It was still limitations, but it was a point where we thought we've got to try, right? Like we didn't know how long we would go, but the timing was right for us. Like we had all...we were both all like stopped working for our previous startups or, you know, we had sold our previous startup. And so for us, it was the right time to say, let's just go play with something. And that it was just happenstance. Well, on time, it always say happenstance as well.

SM: Magic word there as well is, of course, the luck factor. And I'm obsessed with this kind of notion of hacking luck and actually you know, the more shots you take, the more chances you've got of succeeding and the fact that you guys were already kind of being playful and exploratory and then kind of as these shifts, these tides turned, you just so happened to be positioned in that right place to take advantage of that opportunity.

YL: Yeah, I know. Like, you know, I can't say that we were founders that had, you know, the stubbornness that we said, no, back when we started, like day one, that vision is completely different to what we're building today. And in fact, you know, our earliest investor essentially said, we'll give you the investment. But you have to change your idea because we hate it.

SM: So what was the original idea?

The original idea is actually, well, building new projects, right? So we were very early. I don't know if we're the first, right? I don't think we'll ever know if they're first, but we were very early in terms of giving a set of a prompt, right? Or a set of instructions at the beginning and building a mobile app. So you could build Uber for dogs. You could build Facebook for cats. You could do these things. At the time, the context window of GPT-3 was so small, we were daisy-chaining the instructions. So we were doing what we could with a limited model, then passing the output into a brand new session, and then so on and so forth. It was very janky. A lot of the time it was broken, but it was the idea was essentially so, well, anyone who wants to build, it would build the front end, the back end. It would also work out the business logic, right? So the understanding you go from catalog to details to checkout, it understood the logic as well. That was the original idea. But at the time, we were challenged to say, how often do people start brand new projects? And I still think that to me, that is a question mark. Maybe I'm drinking my own Kool-Aid these days, but I think that I still think that for most developers, at least for work, they very rarely start a code base from scratch. They very rarely might prototype someone, but ultimately when they are trying to put it into production, into the main product that they're building, they're existing code exists, right? And they can't work in a vacuum. And that is often where I think a lot of AI tools start headbutting against the wall.

SM: So this speaks to, obviously, we all know about like the depth of the pain point, but you've got the frequency of the pain point to look at here as well, right? From that, how do we actually convert this idea into a viable, scalable commercial business? And in order to do that, it's not just that market size but it's that frequency of use it is and I think it's like where you see a lot of AI tools are at least promoting marketing themselves to for doing a certain skill people that are adjacent to that profession right so co-pilot thing.

YL: I think it's like you know you look at again just as my industry you look at these app generators or AI dev tools that help you build brand new projects

It's often targeted to designers or product managers or project managers or marketers, right? So people who are like, I have this idea, I want to test it, but engineering resources are always limited. Engineering resources never has enough capacity to basically build my experiments and want to do AP testing and all these other ideas. Communication. And so now these tools are like, well, you've got a Figma design or you've got this idea that you can articulate in text. Well, why don't we take this from zero to one or maybe even 0.5 so that then you have a stronger case to what if it's functional, great, then you can do experiments. Even if it's not fully functional, you can now bring that, visualize that idea and show it to your developers and say, hey, look, this is what I meant. Do you agree? And you can have that internal stakeholder management and get more support for what you're doing. And I think that's where you see a lot of the marketing material targeted as.

you can now code yeah even if you don't know the whole vibe coding thing right we've seen like Lovable is in like all the headlines at the moment

SM: So you mentioned about your early investor kind of encouraging you to pivot - was this pre or post YC because when you talk about Y Combinator as well?

YL: Yeah this was at YC. It was Tom Blomfeld and Michael Seibel oh just okay uh heads up guys there's gonna be quite a bit of name dropping in this interviews

Just casually mention Tom Blomfield. Well, I don't think we did anything. I think we were just, it was just half and half. There were so many partners at YC, right? The funny story is just that when I met Tom finally in person in San Francisco I told him that all the way back when I was graduating from uni I interviewed with him at GoCardless for a job and he rejected me and it was for it was a graduate sales job and he's just like he obviously doesn't remember me I hold no grudge that he doesn't remember me was he in the interview he was in the interview so you had to it was actually the second time you met him in person then yeah I met him a few times at other events but like nothing nothing that he would remember me but like I just I did tell him the story over a beer and he was just like yeah I don't I don't I don't remember it and I was like fair enough

SM: What an amazing full circle moment there. That's hilarious. All right.

So talk to us about getting accepted onto YC. What was that experience like? And what are your views on it now?

Because it's not what it once was, but we'd love to hear your take on it.

YL: I'm a fan boy. I am like, I'm just going to caveat to all your listeners. I am through and through. I think that you can, you can criticise anything and anyone and any organization there's always flaws and faults i still think they're on top of their game like there's just simply there's no other alternative what makes them so good them i i think it's three things right if it was practical maybe maybe the last one no i see probably don't really want me to talk about but i think it's a it's a basically open secret but the first two i think number one is the it's...

-

Tune in to Strategy & Tragedy wherever you get your podcasts to listen to the rest of the interview.

In the meantime, please hit that subscribe button if you haven't already. Thank you. Really hope you've enjoyed this. Episodes are weekly, so make sure to tune in next time.

Strategy & Tragedy: CEO Stories with Steph Melodia is the best podcast for curious entrepreneurs and ambitious founders. Learn from those a few steps ahead of you in these candid interviews of the highs and lows of scaling and failing business.

Watch the full episode here, subscribe on Spotify, Apple Podcasts, Google Podcasts, Amazon Music.

Thank you for listening and for supporting the show!