Cursor 创始人复盘：不拼模型，专注人机交互界面

We don't need to compete with anthropic and open AI on models right now. The interface between the human and the model is the key thing.

If you looked at the competitive landscape, it was almost silly.

I asked Michael, I was like, what do you think about cloud code? And he said something to the extent, look, we are going after the biggest market in the world. You're always going to have formidable competitors. That does not scare us.

As founders, we all want to be ambitious and we want to push like to the maximum point the paro frontier, but like you have to kind of know what the curve is.

For us, the decision to invest in curs was actually pretty obvious. You remember Andre Karpathy was using it. It was clearly a phenomenon. It was clearly like a known brand.

They went and got all the users in record time. Now they have the asset, the data, the knowhow to build their own models. You really couldn't do the flip of that unless you started as a frontier lab. They were definitely a couple of oh moments like remember when

so to kick off I actually wanted to take us back to early 2024 and you know obviously you guys led the deal in I think it was May of 2024 that year. Um, but can you walk us through what was going on in the Q1, Q2 of that year and maybe how you guys started working with Curser well before uh we ever partnered officially with them?

So, you have to sort of transport yourself back to late 2023, early 2024. Um, I went back and looked it up. The leading models at the time were GPT40, Claude 3, and Llama 3. Like,

feels like forever ago.

Llama 3.

That was actually pretty good.

That was great. Yeah, it was a great model. um you know change in strategy obviously now um

the leading uh coding harness was copilot by like a mile right this this was like the consensus thing Microsoft was doing their Microsoft thing right they had copilot they had office you know they had vs code all of these things combined they were going to somehow win AI um and and so it was clear that AI coding was sort of working but it was a different universe compared to today right like like you had the sort of early signs that you could you could do meaningful work with AI, but agents didn't really exist.

You know, loops didn't really exist. Reasoning didn't really exist at the time. So, it's sort of sort of a crazy time.

That's there was a continuum of like companies going after coding, right? So, on one hand, you had those that like the only way to win coding is to build a base model and it has to be a coding specific one. And on the other hand, you had, you know, very preciently, by the way, companies just focus on agents,

right? And like cursor was somewhere in between. And and not only models, there were a million companies doing plugins, too. And both of these were very reasonable things to do done by very smart people, right? Like that that's what's so interesting about these new spaces. It's like smart entrepreneurs show up and do really interesting things. And in this case, model training was very a very reasonable thing to do because the models weren't that good at coding yet.

And there was sort of a theory that if we focus on this and do a coding specific model, it'll that's the breakthrough. people doing plugins on the other side were sort of saying hey like let's not toss out the whole IDE like that would be a crazy thing to do but like the interface between human and and model is really important so so plugin is a great place to go um you know it just kind of turned out that the thing that cursor was doing was the right thing to do right they like like us they were very bitter lesson pilled you know at the time they would have said oh there's no need for us to go train our own model of course that changed later for different reasons which we're going to get into later in this podcast but it's like we don't need to compete with anthropic and open AI on on models right now.

Um the interface between the human and the model is the key thing and this is something that that I remember Michael and Aman saying like a lot and saying just saying very clearly from the earliest days of of cursor saying code in the future will look like pseudo code like I remember Michael making this point a lot like when you look at old you know like computer science papers there's a section it's like here's the algorithm not in code but in like pseudo code which you know I'm sure many people um you know are used to and you know we don't exactly write in pseudo code now but probably you could like probably you could just take the pseudo code, you know, block from an old paper and and like the models now would implement that.

This is like really what the cursor guys thought from the very beginning. And they've proven to be exactly right about this, right? It's like this sort of natural languageish specification like

pairing the programmer's intent down to the minimum possible spec. That's what really sort of matters now. So, so anyway, so we're in we were in this world as Martine said, we, you know, we're talking to a ton of very smart um very interesting people working on coding in various ways. Um you know the cursor team kind of stood out because not only did they sort of align with the things that we thought about the world um more importantly smarter people than us you know you know engineers working at at the time the leading AI companies like OpenAI, MidJourney, Replicate a bunch of others.

Um you know all all kind of use this and subscribe to this model. Um and the team frankly just seemed kind of special right like they had an unusual degree of focus. Um, I remember we we we got Michael to come pitch our our GP group. Um, and 90% of the meeting was him saying no to things,

which which which is just like like for for like founders out there working on products that you love. Like

VCs will always ask you kind of dumb questions about other things you might do or things you're related to or whatever. And Michael literally just sat there and like very politely listened to all the questions and said, "Hm, interesting. No, like we're not going to do

Wait, wait, actually on that point, can I ask you guys a question?" Um because I remember that meeting so vividly when you brought uh Michael in for the A and one of the things that he said no to was actually and you know to your point on like there were a lot of plugins, right? And so there was this maybe more consensus thought that a VS Code fork versus the plugin into VS Code would one would be better for enterprise and it wasn't the former it would be the latter.

Um, and I recall I think it was you Martin or maybe it was both of you asking him, "Well, are you going to get enterprise customers with this medium?" Right. And he looked at all of us in the eye with high conviction and you know, well, I'm going to turn it back to you and like what did you think of his answer and um what gave you guys the conviction to take a pretty n I would say a pretty contrarian bet at the time.

They actually had a very reason articulation for why where they were in the design space makes the most sense. So for example um why wouldn't you build a coding specific foundation model. So what they said at the time uh was it turns out that you don't speak to these models in code. You speak to these as humans and natural languages. Therefore the the model has to like understand like the the breadth of human experience in language, right?

And so you know that means in order to build a coding model you're actually building you know a frontier language model and that's just a very hard thing to do. And then of course you know there are other companies doing it and there much more focused problems like t complete to do. So I think every one of their decisions was was very reasoned in our opinion. Um they were also very product focused um which you actually don't see a lot in AI.

So they they really believe