Why Tech Is Racing to Adopt AI Coding  ## Metadata - Author: [[Decoder with Nilay Patel]] - Full Title: Why Tech Is Racing to Adopt AI Coding - Category: #podcasts - URL: https://share.snipd.com/episode/42b3b269-b3dd-4997-96e6-08b8dc69ca73 ## Document Notes ## Highlights - **Cursor as a Coding Word Processor** - Cursor acts like a "souped-up word processor" for coding, enabling programmers to write millions of lines of logic more efficiently. - It offers autocomplete and delegation modes to speed coding and simulate working with a pair programmer. Transcript: Casey Newton Thank you for having me. What is Cursor? What does it do? Michael Truell Who is it for? Our intention with Cursor is to be the best way to build software, and in particular, the best way to code with AI. For people who are non-technical, I think that best way to think about Cursor as it exists today is think of like a really souped up word processor, where the way engineers build software Is they're actually doing a lot of writing. They're sitting in something that looks like a word processor, and they're editing millions of lines of logic, so things that don't look like language. And Cursor helps them do that work way more efficiently, especially with AI. And there's kind of two different ways Cursor does that right now. One is Cursor is kind of watching you do your work, and it's trying to predict the next set of things you're going to do within Cursor. So this is the autocomplete form factor, and that can be really souped up in programming compared to writing because unlike writing, it's actually, there are often times when you're Programming where the next 20 minutes of your work are entirely predictable. Whereas with writing, it's a little hard to get a sense of what a writer is actually going to put down on the page. There isn't really enough information in the computer to understand the next set of things they're going to do. And then the other way that people work with Cursor is they're kind of increasingly delegating to Cursor like they're working with a pair programmer, working with another human. So they're handing off small tasks to Cursor and having Cursor kind of go end-to on them. Casey Newton We'll dig a little deeper into the product in a moment, but first let's talk about how all of this started. When you founded AnySphere, you were working on computer-aided design software. ([Time 0:05:36](https://share.snipd.com/snip/6f2f8ed0-48b3-40e6-8aa4-40c1c56855f2)) - **UI and Custom Models Differentiate Cursor** - Cursor combines a rich user interface with custom and API AI models to enhance developer productivity beyond what terminal-based tools allow. - The integration of multiple AI models allows for more precise and efficient coding assistance. Transcript: Casey Newton Cursor as kind of like a souped-up word processor. Software engineers, I think, would call it an integrated development environment, or IDE. And developers have been using IDEs since the 80s. But recently, AI labs have released tools like OpenAI's Codex or a Cloud that can run directly in a terminal. Why might someone use Cursor over those options? Michael Truell Both of those are really useful tools. The thing we care about being, so I think we start as this IDE, we start as this text editor. And what we really care about getting to is to a world where programming is completely changed. And in particular, you know, a world where you can develop professional-grade software, perhaps without even really looking at the code. And it's that kind of future programming and like changing it from this weird, like you're reading these millions of lines of logic and these like esoteric programming languages to Getting you to a world where you can build software by just specifying the minimal intent necessary to build the software you want. You can tell the computer the shortest amount of information it needs to really get you, and it can fill in all of the gaps. And yeah, programming today is this intensely labor-intensive, time-intensive thing where to do things that are pretty simple to describe, to get them to actually work and show up On a computer, it takes many thousands of hours and really large teams and lots of work, especially at professional scale. So that's where we want to get to is kind of inventing that new form of programming. I think that that starts as an editor and then that starts to evolve. And so we're already kind of in the midst of that, where right now Cursor is this place where you can work one-on with an agent and you can work with our tab system. And then increasingly we're getting you to a world where more and more programming looks like starting to delegate your work to a bunch of helpers in parallel. And there's a product experience to be built for making that great and productive and understanding what all of these parallel helpers are doing for you, being able to intervene in The places where it's helpful, understanding their work when they come back to you at a level that's not having to read every single line of code. Yeah, I think that there's a competitive environment with a bunch of tools that are interested in programming productivity. One of the things that's limiting about just a terminal UI is that you have only so much expressiveness in the terminal and control over the UI. From the ([Time 0:12:18](https://share.snipd.com/snip/b7aa6891-4aa4-4e55-8754-d8d2015b4a1b)) - **Why AI Excels at Programming** - Programming suits AI well because it is text-based, has abundant data, and offers verifiable outputs. - Reinforcement learning improves AI code generation by treating coding as a testable game, enhancing model capabilities efficiently. Transcript: Casey Newton Engineers saying, please tell us you're not about to take away Cursor, because they've become so dependent on it. Can you give us a sense of why for programmers, this has kind of felt like a before and after moment in the history of the profession. What is it that tools like Chrysler are making so different in the lives of these engineers day to day? Michael Truell I think that we're just already at a point where we are far, far, far from the ceiling of where things can go and far, far, far from a world where much of coding has been replaced with something Better. But, you know, ([Time 0:18:41](https://share.snipd.com/snip/4e808747-4c16-4100-b8f5-233fba44961d)) - **AI Boosts Productivity without Cutting Hours** - AI coding tools currently boost programmer productivity significantly but may not reduce overall hours worked. - Much of programming involves handling and modifying complex legacy code, contributing to inefficiency. Transcript: Michael Truell Finding real value from it. Casey Newton Yeah. I mean, my sense is maybe if I used to have to work eight hours a day, now it's maybe closer to five or six. Michael Truell Is that part of it? Yes, in the sense that I think that the productivity gains of what would have taken you eight hours before in some companies now actually can take you five or six hours. I think that that is real, not across all companies, but is really real in some companies. But I think that the thing I would nitpick on there is I don't think programmers are actually just working or shortening the hours that they're working. And I think a lot of that is because there is just a ton of elasticity with software. And I think it's really easy for people who are non-technical or just don't program professionally to underrate how inefficient programming is at a professional scale. And a lot of that is because programming is kind of invisible. What a programmer is doing at a company like Salesforce is there are just tens of millions of lines, many millions of files of existing logic that describes how their software works. And anytime they have to make a change to that, they have to take that ball of mud, that mass of things. It's very unwieldy, and they need to edit it. That's why I think that it's kind of shocking to many people that some software release cycles are so slow. But so, yes, I think that there are real productivity gains. I think that it's probably not reducing the number of hours that programmers are working right now. We need to take a quick break. ([Time 0:20:40](https://share.snipd.com/snip/8f5ee2ec-2375-40ae-8a8f-6d31a4219270)) - **Vibe Coding Grows with AI Tools** - While Cursor focuses on professional developers, AI tools enable non-technical "vibe coding," where amateurs prototype or contribute small code changes. - These newer users often engage casually or professionally but represent a growing segment as AI lowers coding barriers. Transcript: Casey Newton But this year saw the coining of the term Vibe Coding to describe what more amateur programmers can do, sometimes even complete novices, and often with tools like Cursor. How big is the Vibe Coding use case at Cursor? And what do you think is the future of Vibe Coding? Michael Truell So our main goal is to help people who build software for a living. Right now, that means engineers. And so that's our main use case. It's been interesting to see as you focus on that use case and you use the understandings you get from that use case to kind of push the tech forward and you hop programmers up more and more Levels of abstraction, how it then also makes things more accessible. And that's something that we're really excited about. And I think in the end state, I do think that building software is going to be way more accessible. You're not going to have to have tons of experience on understanding programming languages and compilers. And I think there's still a bunch more work to do before anyone can build kind of professional grade software. That said, it's been really cool seeing people spin up projects and prototypes from scratch, designers in professional settings doing that. It's been really interesting to see non-technical people contribute small patches and bug fixes or small feature changes to professional software projects already. And that's kind of the vibe coding use case, not our main use case, not where the company makes most of its money, but one that I think will become bigger and bigger as you push the ceiling Casey Newton Of focusing on professional developers. I'm curious what you think of as like the demand for it though. Like I understand it's not your focus of the business. And people like to talk about it. I think people, look, it feels cool to have never built software before. And all of a sudden, next thing you know, you've actually created a little to-do list app for yourself or something. Michael Truell I probably differ from some of my colleagues on this personally. As the world as it exists right now, I do think that like kind of the two buckets of that vibe coding use case, one is there's like an entertainment bucket of you're doing these things mostly For like personal enjoyment or hobbies. And then there's also a bucket that's more professional. And I think that that's like designers doing prototypes or that's people that work to serve customers contributing back bug fixes to a professional code base. And the way in which I probably differ from some of the people I work with is there's a group of people who are really, really, really interested in end user programming and throwaway Apps and personalized software where kind of everyone, yeah, like builds, entirely builds their own tools. I think that that's really cool. I think enabling that is really cool. And I think a lot of people who aren't technical will be interested in doing that. But I still think even if you get to a world where anyone can build things on computers, there's going to be, I think most of the use cases will still be served by like a small minority of, You know, 5% of the world that's caring a ton about the tools and building them. And then everyone will more use those tools because I just think that like the interest in that stuff really differs amongst the population. But so, yeah, right now, right now, commercially, I think that a lot of the a lot of the more vibe coding stuff falls more into like a mid journey camp or like an entertainment camp. It's something that some people get interested in for a bit and then kind of put it aside. And then, ([Time 0:25:40](https://share.snipd.com/snip/b6da6963-c037-4aaa-bc45-8c4d7acfd98b)) - **AI Delegation Growing in Programming** - Currently, about 20-25% of a programmer’s work involves delegating coding tasks completely to AI. - With advances in context understanding and continual learning, AI could handle over half of programming work within a few years. Transcript: Casey Newton 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. Any updates to that number in the last month or so? And how high do you think that number can scale ultimately? Michael Truell I think these things are really hard to predict. I think some of the things that are blocking you from getting to a hundred percent, 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 Canada 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. And 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 your brain is changing all The time. You're getting new things. That kind of persists with you. And obviously, 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. And 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 the rate of progress has been really fast Over the past five years. Long context or in-context learning and like fine-tuning with some other way of continual learning. I 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. So 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 technical blockers to getting to something that's 100%. And there's many more that you could list, and 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, I think if in the Bulk case, you know, you could probably do over half of programming as it exists today. Yeah. Yeah. Casey Newton I see these studies that Meter puts out where they look at the average length of time that a software ([Time 0:29:40](https://share.snipd.com/snip/c3d21bfd-1143-4ec5-9f3b-e09f2425e286)) - **AnySphere's Hiring Strategy Amid AI War** - AnySphere balances strengths of foundation model labs and software companies, appealing to AI researchers seeking focused, impactful work. - It remains competitive in talent hiring despite big tech recruitment efforts by emphasizing meaningful small-team projects. Transcript: Michael Truell For other verticals and other industries too. And so now. Yeah. Casey Newton 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? Michael Truell No, not really. The research team, we try to keep fairly small. I mean, the 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. And what we have to offer is most appealing to people who want to be a part of an especially small team, working on something focused, 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 the API models. But I think that we're like this weird experiment in a company ([Time 0:44:20](https://share.snipd.com/snip/e5523ba6-c304-40c2-8604-78f28aa8e399)) - **Challenges of Usage-Based AI Pricing** - Switching Cursor's pricing from fixed fees to usage-based was poorly communicated, upsetting users who unexpectedly hit limits. - Increasing AI task complexity drives up compute costs, necessitating transparent and flexible pricing models. Transcript: Casey Newton But you did have this moment recently where you changed the way you price things and folks got pretty mad. And basically, you just moved from a fee to more usage-based pricing, and some people ran over their limits without realizing it. What did you learn from that experience? Michael Truell 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 evolved. Features were added. Features were changed kind of like up and down that limit has. Or like there are different ways like you have been able to pay down that limit or not pay down that limit over time. And 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. And that changes the value it can give to you. You know, 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, you know, 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 the AI more and more And more and more. Then, you know, 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 learned from that experience and a lot we need to show up on in the future. Casey Newton Yeah, I think it's hard for consumers in particular to understand usage-based pricing because they're used to Spotify and Netflix where ([Time 0:48:20](https://share.snipd.com/snip/dbe29629-2816-4139-bc8a-0ec4a3f77e3a)) - **AGI Will Unfold Over Decades** - Michael Truell believes major AI progress and AGI will unfold over decades, not abruptly by 2027-28. - He views current AI advances as jagged and incremental rather than an imminent “machine god” emergence. Transcript: Casey Newton 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 when you're 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. 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? Michael Truell 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 maybe, yeah, just 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 with the rise of the first AI products that really reached billions of people. But 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, you know, insert X application, whether It be CAD or whether it be programming specifically? You know, AGI is going to wipe all of this stuff out in, you know, Y years, you know, 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. There'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. And 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 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. And so, you know, in some ways, one of the ways, at least I think about what we're doing is trying to do that. Casey Newton But okay, so it sounds like you don't think that there's just going to be one big new training run with like a more parameters and we're going to wake up to a machine god. Michael Truell 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. Casey Newton All right. Well, last question ([Time 0:52:50](https://share.snipd.com/snip/65cbdd85-94b0-4f08-9841-ce665cdb78dc)) - **Future of Programming: High-Level Languages** - Future coding will delegate more tasks to AI while evolving programming languages to be higher level and more human-friendly. - New user interfaces will blend concise coding with fine-grained control for professional software development. Transcript: Casey Newton Going to ask you to do something crazy like predict what cursor is going to look like five years from now. But when you think about it, maybe two years from now, what do you hope it's doing that it isn't quite doing yet? Michael Truell 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, 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 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? And 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, you know, 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. And 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, you know, hundreds of thousands Of lines of code. And, you know, I think that for a while, an important way you build software is you could, you know, 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, you know, there's all this work to do on the model side of things. The field's going to do some of that. We're going to 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? And 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 different elements on the screen and be able to 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. That 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. Casey Newton Well, ([Time 0:56:35](https://share.snipd.com/snip/430242ae-9772-47a7-8381-f76ec5bf418b))