Why tech is racing to adopt AI coding

24:41And then everyone will more use those tools because I just think that like the interest and that stuff really differs amongst the population. But so, yeah, right now, commercially, I think that a lot of the more vibe-coding stuff falls more into like a mid-journey camp or like an entertainment camp, but something that some people get interested in for a bit and then kind of put it aside. And then, you know, some of it is in this professional camp of people that, you know, work on software for a living, but don't code right now. I think you're right, because when I worked at more traditional companies, whenever a new piece of software was introduced, everyone would get upset.

25:13So that's my case for most people not becoming like sort of pro-vibe coders. I like software, though. So I'm vibe code curious. Maybe two or three generations from now in Cursor, I'll be able to make myself something useful. You mentioned earlier that there are these kind of two main ways that people use Cursor. There is the I'm looking at code and you're helping me auto-complete things. And then there is the, I'm going to give you a task and walk away and come back and see what you've built. You told Ben Thompson recently that over the course of the next six months or a year, you think you can get to a place where maybe 20 or 25 % of a professional software engineer's job might be the latter use case of just handing off work to the computer and having the computer do the work end to end.

26:05Any updates to that number in the last month or so? And how high do you think that number can scale ultimately? I think these things are really hard to predict. I think some of the things that are blocking you from getting to 100%, one is having the models learn new things, like understand an entire code base, understand the context of an organization. But yeah, learn from their mistakes and really learn new things. I still think that the field doesn't have an amazing solution for that. The two candidate solutions are, one is you make the quote unquote context windows longer, which is these large language models, they have like a fixed window of text or images that they can see.

26:50And then there's a limit to that. And outside of that, it's just the model that came off the assembly line. And then that new kind of information that's put into the model's head, which is very different from humans because humans are going through the world and like, you know, your brain is changing all the time. You're getting new things. That's kind of persists with you. And like, obviously, you know, some memories fade away, but it persists with you somewhat. But so candidate solution number one to the continual learning problem is just make the context windows really big. Candidate solution number two is train the models.

27:18And so every time you want them to learn a new thing or a new capability, you go and collect some training data on that, and then you throw it into the models mix. And both of those have big issues, I think. But that's one thing that's stopping you. And I think that the rate of really consequential ideas in ML that are kind of like new paradigm shifts is pretty low industry-wide, even though that the rate of progress has been really fast over the past five years. And so ideas of the form of replacing long context or in context learning and like fine tuning with some other way of continual learning.

27:58I don't think that the field actually has an amazing track record of, of generating lots of ideas like that. I think it's sort of ideas on the rate of, you know, maybe one every three years. So I think that will take some time. I think the multimodal stuff will take time too. The reason that's important for programming is you want to play with the software and you want to be able to, you know, click buttons and actually, yeah, use the output. You want to be able to use tools also to help you make software tools that have GUIs. So for instance, observability solutions like Datadog are important for understanding how you can improve a professional piece of software.

28:30So that feels like it's needed. These models also, they can work coherently for minutes at a time, now even hours in some cases, but it's a different thing to work on a task for the equivalent of a human's weeks. And so just even architecturally, knowing if we're going to be coherent over sequences that long will be interesting to see. And that I think will be tricky, but there are all of these, you know, all, all of these technical blockers to getting to something that's a hundred percent. And there's many more that you could listen. There are also many unknown unknowns. And I think that in a year or so, even with just playing the game of going from a high level text instruction to changes throughout a code base, playing that really well.

29:09I think if in the bull case, you know, you could probably do over half of programming as it exists today. Yeah. Yeah. I see these studies that Meter puts out where they look at the average length of time that a software or that an AI model can do, and it does keep doubling at this really impressive rate. So I think the hurdles that you identify are super important, but when you pull back, it does seem like length of task is really improving. And ultimately, humans don't tend to work on discrete tasks that are all that long. So I do think it's getting easier for people to imagine an LLM putting in a full day's work.

29:48Yeah, I think that just forecasting these things is tricky. And one related field that can maybe be telling of how things will evolve here is just kind of the history of self-driving, which obviously has made leaps at the bounds of advancements. And in San Francisco, there are Waymos. There are commercial self-driving cars. my understanding is Tesla's also made big improvements. But I remember back in 2017 when people thought self-driving was going to be done and deployed within a year. And obviously, there are still big barriers to getting it out into the world. And that feels like as hard and as varied as driving is, it does feel like a much lower ceiling task than some of the stuff the field's talking about right now.

30:28So we will see. Yeah, interesting. I do want to sort of ask you about timeline stuff, but I'm going to wait until a little bit later. All right. Let me now ask you some of the famous decoder questions, Michael. How big is AnySphere today? How many employees do you have? We're roughly 150 people right now. Okay. And when you think about how big you want the company to be, are you somebody who envisions very big workforce or do you sort of like the smaller Nimbler team? We do like the Nimbler team. And I think the caveat there is we want to keep the team Nimbler for the scope of work that we're tackling, but that will still mean growing the team a lot.

31:05over the next couple of years. But yeah, I wonder if it will be possible to build, you know, a thriving technology company that does really important work with, you know, a maximum team size of maybe 2000 people or something like that, you know, something of the size of the New York Times. And, you know, we're excited to see if that is possible. But definitely we need to grow a lot from our current headcount. What is your org chart like? You have a few co-founders. How do you all divvy up your responsibilities? The two biggest areas of the org are engineering and the research side of things like R &D generally, and then the go-to-market side of things, so serving customers.

31:45This is a company that has really benefited from having a big set of co-founders and a big, very capable founding team. And so there's a lot of across that scope kind of dividing and conquering. In particular, I think that there's a really important set of people on the founding team who have done phenomenal work in building out that early part of the go-to-market side of things. And a lot of that is just entirely of people in the founding team is kind of entirely credited to like a subset of it. And so there's a lot of dividing and conquering across the business. At the same time, I think actually amongst, once you like zoom into the technical side of things, there's like an intense focus from the four co-founders on that and really putting kind of all eggs in, in one basket of that side of the business.

32:34And so, you know, I think that we're lucky enough to be in a time where there are really, really useful products to build in our space. And I think that the highest, you know, the highest order bid, the thing you cannot mess up is having, having the best product in the space. And so we've been able to be relatively lean in other parts of the business, especially relative to our scale, but also as a ratio to engineering and research and still be able to grow really far. What part of the business do you like keeping for yourself? Like where do you like getting your hands dirty and would you be mad if someone tried to take it away from you?

33:04I spend a lot of time trying to help how I can in growing the team. And we think hiring is incredibly important and especially the hiring of ICs. Individual contributors. Yeah, yeah, individual contributors. I think that one way technology companies die is that the best ICs start to feel disengaged, like they don't have control over the company, and the talent density lowers. And then I think that if you're working on technology, no matter how good the management layer is, if you have less than excellent people doing the real work, I think there's only so much you can do. I think that the dynamic range of what management can do is kind of limited.

33:45And so I like to help how I can by spending a bunch of time on hiring. And we actually got to maybe 75 people just with the co-founders hiring and without hiring functional recruiters. And so now I have fantastic people helping us hiring people on the recruiting side of things that work with us closely. Spend a bunch of time on that and then try to help how I can on the engineering and product side of things. And those are the two biggest areas of focus. And then there's a long list of long-tail things. And yeah, right. You're fairly young. I think you're 25 and have had to make a lot of really big decisions about raising money, making acquisitions, all those hiring decisions that you just made.

34:25How do you try to make decisions? Do you have a framework that you use or is everything ad hoc? Yeah, I'm not sure there's one framework. I think that some pretty common devices that help us is we try to do our best to farm from descent, kind of all up and down the group, the org. And this is not just for me, it's try to do it for kind of all decisions in the company of having increasingly like a very clear DRI. And then lots of people who are kind of inputs to the decision. Every decision is pretty unique. I think that, you know, other devices that are well-known, that are helpful, are kind of understanding how high stakes the decision is and how reversible it is.

35:08I think that especially when, you know, you're in a vertical like ours with the speed that it's moving, there's just a limit on the amount of time and the amount of information you can gather on each thing. And then, you know, other devices like, you know, clearly communicating the decision and using that as a way to kind of clarity for how you thought it through. We need to take another quick break. We'll be right back.

35:40Support for the show comes from Rippling. If you're a business owner, here's the truth. SaaS promised to make work easier, but now the average company is buried by hundreds of apps that silo your teams, slow you down, and simply don't work together. That's not SaaS. That's SAD. Software as a disservice. That's why you need Rippling. Rippling is the unified platform for global HR, payroll, IT, and finance. They've helped millions replace their mess of cobbled-together tools with one system designed to give leaders clarity, speed, and control. By uniting your employees, teams, and departments in one system, Rippling removes the bottlenecks, busywork, and silos your software created.

36:23Automated, perfectly in sync, and seriously simple to use, Rippling gives your company one source of truth for your people, their data, and everything they touch. With Rippling, you can run your entire HR, IT, and finance operations as one. Or pick and choose the products that best fill the gaps in your software stack. And right now, you can get six months free when you go to rippling.com slash decoder. Learn more at rippling.com slash decoder. That's rippling.com slash decoder for six months free. Terms and conditions apply.

37:06We're back with AnySphere CEO Michael Truel. We just covered the core decoder questions, but now I wanted to talk a bit more about hiring, In particular, how the AI talent wars are affecting a company like AnySphere. Well, let's talk a little bit more about hiring since you brought it up. There has been talk that OpenAI had considered acquiring you. And I have to ask, given his recent spending spree, has Mark Zuckerberg invited you to his house in Tahoe? No, no. No? He's not coming around with his$200 million signing bonuses saying, Michael, why don't you kind of come over here? We're building super intelligence.

37:45This for us is kind of life's work territory. So yeah, I feel really lucky to have the technology lineup, kind of the initial founding team lineup, the people that have decided to join us, and the way things have gone on the product to have the pieces in place to execute on this ambitious goal of automating programming. And time will tell if we're going to be the ones to do that. But as people who have been programming for a long time and working on AI for almost as much, being able to reinvent programming and help people build whatever they want to build on computers with AI kind of feels perfect for us.

38:16And it feels like one of the best commercial applications of this technology too. So I think that if you can succeed in that, you can also push the field forward in big ways for other verticals and other industries too. And so now. Yeah, it sounds like you really want to stay independent. Yeah. Has Meta's recent hiring spree made it noticeably harder for you to recruit lately? Um, no, I, not really. The research team, we try to keep fairly small. I mean, the whole, whole company is kind of small relative to what it's doing, but the research team, especially people think through hiring decisions in different ways.

38:54And, you know, what we have to offer is most appealing to people who want to be a part of an especially small team, working on something, focus kind of solving problems with AI out in the real world. I want to be in a place too, that marries, like we're kind of this weird company. You talked about some products that are being made by some of the great folks that work on like kind of the API models. But I think that we're like this weird experiment in a company that's smack dab in between the foundation model labs and normal software companies where we try to be really excellent at both the product side of things and the model side of things under one roof and have those feed into each other.

39:29And so we appeal to, I think, a certain type of ML researcher or ML engineer and for them, I think it's kind of about being part of this and a little bit less, you know, like some of the other things. One last hiring question. It was reported this week that two folks who used to run Cloud Code that you'd recruited to come over to Cursor left back after a couple of weeks. Can you speak at all to what happened there? Kat and Boris are awesome and I think that they have a lot left to do on Cloud Code And they're really, as I understand it, just the people behind that. And that is their creation. And as someone who's been working on something for three and a half years from inception, kind of understand the ownership that comes with that.

40:17And they have a lot left to do. And they were excited about that. And so they've decided to stay. Cool. It seems like you were mentioning this interesting position that you sit in in between the big labs and other startup companies who are using your software. How do you describe Cursor's culture when you're recruiting people? Perhaps unsurprisingly, we are process skeptical and kind of hierarchy skeptical. And so, you know, we need to, as we do more and more ambitious things, like more and more coordination is required. But for like a certain level of thing, you know, the scope of the company, we try to be like pretty light on each of those.

40:48I think it's a very intellectually honest group. It feels very low stakes to criticize things and just be open very publicly about feedback on work. It's a very intellectually curious group. You know, I think that people are interested in doing this work, you know, for the end goal of automating programming and separate from any work-life balance things, because we want this to be a place that's all levels of work-life balance can do great work. It's a place where I think no one really treats it like, quote, you know, so far, like just a job. Like they're really, really excited about this. And I think it's kind of a special time to be building technology.

41:23And so one should try to seek out a role where you can, don't treat it like just a job. I think it's very focused and understated. I think from the outside, partially because of how little communication we do with the outside world and we need to get much better at that. I think mostly people know Cursor as, oh, that thing that grew really fast and kind of know about top-level metrics and things like that for just how fast the adoption has been. And internally, we've thought that it's really important to hire people who are, while they might be very ambitious, are very humble and pretty understated.

41:58and pretty focused and level-headed because there's noise left and right. And I think that, yeah, just having kind of clear focus and putting your head down is actually really, really important for people being happy in this space and also just for the execution of the team. Yeah, those are some things to describe the current group. You mentioned communicating with the outside world. I think, you know, your cursor's history is mostly just a history of delighting its customers. But you did have this moment recently where you change the way you price things and folks got pretty mad. And basically you just move from a set fee to more usage-based pricing.

42:34And some people ran over their limits without realizing it. What did you learn from that experience? Yeah, I think that there was a lot to learn from that and a lot on our end that we need to improve on. To set the stage to the way cursor pricing has worked even back when cursor first started is by and large, you sign up, for a subscription, and then you get an allotment of a certain number of times you can use the AI over the course of your subscription term. And the pricing evolves, features were added, features were changed, kind of like up and down that limit has, or like, you know, there are different ways like you have been able to pay down that limit or not pay down that limit over time.

43:15And what's happened in parallel is kind of using the AI once, what that means, the value that gives people and the underlying costs in some cases has changed a lot. One big switch there for us is that increasingly when you quote unquote use the AI, the AI is working for longer and longer and longer. And so you called out that chart that you've seen where it's showing the kind of max time an AI can work. And it's gone from seconds to minutes to hours at this point, and it's gone up very fast. We're kind of front lines of that, where now when you ask the AI to go do something or answer a question, it can work for a very, very, very long time.

43:50And that changes the value it can give to you. You can go from just asking a simple programming question to having it write 300 lines of code for you. And that also changes the underlying costs. And in particular, less the median and more the variance of those costs. So yeah, we bundled together a series of pricing changes. And the one that garnered the most attention was switching from a world where kind of the monthly allotment is in requests to it's in the underlying compute that you're spending. And one thing to knit on what you said is that actually usage-based had been a big component of Cursor before, because over the life of Cursor, people have just used AI more and more and more and more.

44:29Then they started running out of limits, and we wanted to give people a way to kind of burst past that. What this did is it changed kind of like the structure of also how that usage pricing worked, where it's not on a request basis, it's on the underlying compute basis. And definitely that could have been communicated legions better. There's a lot we learn from that experience and a lot we need to show up on in the future. Yeah. I think it's hard for consumers in particular to understand usage-based pricing because they're used to Spotify and Netflix where they pay their 10 or 20 bucks a month and it's all you can eat.

45:02But the economics of AI just don't really work that way. It will be interesting to see how things play out in our space in particular. because I think that for the consumer chat app market, so far at least there's been... Yeah, it would be interesting to see how the curves of just how compute per user over time has gone up. But I wouldn't be that surprised if it's been pretty flat over the past 18 months or so, where the original GPT-4, I'm not privy to any inside information, but it seems like there have been big gains from a model size perspective where you can actually miniaturize models and get the same level of intelligence.

45:38And so I think that the model that most professional users are using in something like a ChatGPT has actually maybe gone smaller over time. The compute usage has gone down. But in our space, yeah, I think that there's just for one user, I think the compute is probably going to go up. And there's a world in which the token costs don't go down fast enough. And it starts to become a little bit more like AWS costs and a little bit less like Percy productivity software and still remains to be seen. But one thing to note is that we do think it's really, really, really important to offer users choice.

46:09And so we want to be the best way to code with AI if you just want to turn on all the dials and just get the best, most expensive experience. We also want to be the best way to code with AI if you want to just pay for a predictable subscription and get the best thing that that price can offer you. Even for the main individual plan, the$20 pro plan, the vast majority of those users don't hit their monthly limits. and so aren't hit with a message saying you need to turn on usage pricing or not. That's the kind of AI user I am. I never hit my list. It makes me feel like I need to be using it more.

46:39I'm trying to - There is a really, really big difference between the top 5 % and a median user. So some people are very, very, very AI forward. Well, coming to my last couple of questions here, I want to try to get at how AGI-pilled you are Because when we were talking earlier, you're sort of identifying all these very real technical problems in building more advanced systems that are just truly unsolved problems in AI. You know, the size of the context we do, giving these systems longer memory, helping them learn the way that a human might be able to learn. We don't know how to do that yet. And yet there are lots of folks in the industry who believe that by 2027, 2028, the world looks very, very different.

47:27it. So where do you sort of plot yourself on the spectrum of people who think that everything is absolutely about to change and we're sort of at the start of a process that's going to take decades? I think we're kind of this bet on the messy middle where we do think it's going to take decades. We do think that nonetheless, AI is going to be this transformational technological shift for the world, bigger than, you know, maybe, yeah, just, you know, a very, very, very big technological shift. And when we started working on Cursor, it was funny, we would get these kind of two dual responses. And I think one is now increasingly falling out of favor, just, you know, with the rise of the first AI products that have really reached billions of people.

48:10But early 22, we would get kind of two reactions. One reaction was, why are you working on AI? You know, I'm not sure that there's really much to do there. The other reaction that we would get, because we did have close friends and colleagues who are very interested in AI, is why are you working on insert X application, whether it be CAD or whether it be programming specifically? AGI is going to wipe all of this stuff out in Y years. Maybe it's 2024, 2025. We think it's this middle road of this jagged peak, where if you actually peek under the hood at what's driven AI progress so far, again, I think that there's been a few ideas that have really worked.

48:48There's been lots of details to fill in between, but there have been a few really, really important ideas. I think that despite the number of people that have worked on deep learning over the past decade and a half, the rate of idea generation in the field, like really, really consequential idea generation in the field, hasn't budged that much. And I think that there are lots of real technical problems that we need to grapple with. And so I think that there's like this urge to anthropomorphize these models and see them be amazing and human level or superhuman at some things and then think that they will just be great at everything.

49:22And I really think it's this very jagged peak. And so I think it's going to take decades. I think it's going to be progressive. I think that one of our most ambitious hopes with Cursor is if we are to succeed in automating programming and building an amazing product here, that makes it so you can build things on computers just with the minimal intent necessary. Maybe the success of that and the techniques that we need to figure out in doing that can also be helpful for pushing AI progress forward in general. And I think that the experiment to play back here is if you were in 2000 or 1999 and you wanted to push forward AI, one of the best things you could do is work on something that looks like Google and make that successful and make that R &D available to the world.

50:02And so, you know, in some ways, one of the ways, at least I think about what we're doing is trying to do that. Okay. So it sounds like you don't think that there's just going to be one big new training run with like a lot more parameters and we're going to wake up to a machine god? You know, time will tell. My best guess, yeah, and I think it's important to have healthy skepticism about how much you can know with these things, but my best guess is that it will take longer than that, yet also still be this big transformational thing. All right, well, last question here. We've talked a couple of times today about how hard predictions are in general, so I'm not going to ask you to do something crazy like predict what cursor is going to look like five years from now.

50:43But when you think about it, maybe two years from now, what do you hope it's doing that it isn't quite doing yet? I think a bunch of things. So I think in the short term, we're excited about a world where you can delegate more and more work to kind of very fast, helpful humans. And you can build a really amazing experience for making that work delightful and orchestrating work amongst these agents. Another idea that we've been, or I've been interested in for a long time, which is a bit risky, is, you know, I think that if you can get to a world where you're delegating more and more work to the AI, you'll start to run into an issue, which is, do you look at the code?

51:26And are you reading everything line by line, or are you just kind of ignoring the code wholesale? And I think that neither closing your eyes and ignoring the code entirely in a professional setting or reading everything line by line will really work. And so I think you'll need this middle ground. And I think that that could look like the evolution of programming languages to be higher level and to be less formal. And all that a programming language really is, is it's a UI for you as a programmer to specify exactly what you want the computer to do. And it's also a way for you to look at and read exactly how the software works right now.

52:04And yeah, I think that there's a world where programming languages will evolve to be much higher level, more compressed instead of millions of lines, hundreds of thousands of lines of code. And I think that for a while, an important way you build software is you could read and point at and edit that kind of higher level programming language. And I think that this also kind of gets at a bigger idea that's behind the company of, there's all this work to do on the model side of things. The field's gonna do some of that. We're gonna try to do some of that. But then the end state of what we want to do is also this UI problem of how do we get the stuff that's in your head onto the screen?

52:40And I think that the vision of you just entirely build software by typing into a chat box is powerful. Like, I think that that's a really simple UI. You can get very far with that. But I don't think it can be the end state. You need more control when you're building professional software. And so you need to be able to kind of point at, you know, different elements on the screen and be able to, you know, dive into the tiniest detail and change a few pixels. You also need to be able to point at parts of the logic and understand exactly how the software works and be able to edit something very, very fine-grained.

53:09That requires rethinking new UIs for these things, and the UI for that right now is programming languages. And so I think that they're going to evolve. All right. Well, a lot of fascinating things that you're working on. Michael, thank you for coming on to Coder. Thank you for having me.

53:28Thank you to Michael for taking the time to speak with me, and thank you for tuning in. I hope you liked it. If you'd like to let us know what you thought about this show or what else you'd like us to cover, drop us a line. You can email the team at decoderattheverge.com. They really do read every email. Or you can hit me up directly on Threads or Blue Sky. I'm at Crumbler on Threads and I'm caseynewton.bsky.social. Not very catchy, is it? Decoder also has a TikTok and an Instagram. You can check those out at decoderpod. They're a lot of fun. and if you like Decoder, please share it with your friends and subscribe wherever you get your podcasts.

54:04Decoder is a production of The Verge and is part of the Vox Media Podcast Network. Decoder is produced by Kate Cox and Nick Statt. The show is edited by Ursa Wright. The Decoder music is by Breakmaster Cylinder. See you next time.