We're at the 64K IBM PC era of the microcomputer. People are still trying to figure out how everything works.
We have to relearn how to use this type of tool before we know how to be productive with it.
Vibe writing is absolutely a thing. It's a place in which full autonomy can be fulfilled today.
We've changed our view of excellence because we wanted more access.
Is the bar for success what people can do today, or is the bar for success just something that's better than the alternative? Of course, this is dramatically better.
Anish, Stephen, we were having such a good conversation offline that I wanted to get this on the podcast. There are a few topics we wanted to discuss. First, we were all fascinated by Andrej Karpathy's talk at Startup School. Stephen, what did you find so interesting, or what were your takeaways or reactions from it?
Well, I totally love the talk. He did an unbelievable, like a philosopher-king version of where we are. And I just found his metaphors really compelling. In fact, what I might do is even take it further back and just say, since he used an analogy of like where we are in computing and lots of people are talking about the Windows 3 era and stuff like that, and having lived through all of them, I tend to think we're at the 64K IBM PC era of the microcomputer. And the reason I think that is actually a technical one, which is that we're at the point where people are still trying to figure out how everything works.
And all the coding and all of the energy is working around like these very basic working problems. Like with the PC, it was like, okay, we have 64K of memory and our programs are all too big and we have no display and all these problems. And with AI, you know, people are like, it's going to replace search, it's going to replace Excel, and it's going to replace all these things, but it doesn't add very well, it gives you a lot of errors. Like the thing that you say it's going to do, it just doesn't even do yet. So I feel like we're at a point that is just so, so early, and he did a fantastic job of sort of making that arc.
You know, the thing that struck me the most was he talked a lot about our relationship with this new tool, you know, and in a sense, we want to use it like in the same way that we've used all the other computing tools and technologies we've used in the past, but he really talked about this kind of inversion of the relationship of LLMs as people spirits, the fact that they have jagged intelligence. So to me, that sort of meta point he made was one of the most interesting. We have to relearn how to use this type of tool before we know how to be productive with it. And I think tools is a super interesting point because the talk is anchored in tools, but the world itself is anchored in tools, and the early stages of a platform are always about tools, and so you kind of can get a little confused.
Like right now, of course, he was talking about vibe coding clearly because he pioneered the term, invented the concept, and is living it. And it's very interesting because I actually think coding is one domain that always works best early in a platform form because, well, all the customers of the platform are developers and they're going to make their tooling kind of work and come along. But I really think that the most interesting thing for me, what's being underestimated in the near term, is sort of vibe writing. And I mean, it seems weird to say anything with AI is underestimated because Lord knows that's not where we are. But the thing is that vibe writing is so here. Like if you're in college, you're already vibe writing, and you know, businesses are still working through the, well, can we use this? This doesn't seem appropriate. And that's a thing I've definitely lived through with word processors with the micro. You know, I had to get permission from the dean in college to use a computer to write papers. But this vibe writing is absolutely a thing, and it is really, really no different than when calculators showed up and all of a sudden just doing math homework involved using a calculator. And people are like, well, you're not going to know how to do math in the future. And it's like, I won't have to know how to do math. That's like the whole point of the tool. Like I have a power drill, so I do not know how to use like one of those Amish drill things. Like, you know, and the world moves up the stack, and so that's where we are, and it's just super exciting.
What I love about the vibe writing concept actually is it's a place in which full autonomy can be fulfilled today. So you can ask the model to vibe write something, you know, really detailed and compelling, and it'll do a great job. Whereas with vibe coding, I think there's a ton of constraints as to what the model can actually do versus what it can conceptually do. And understanding those boundaries and constraints is going to define a lot of the text-to-code stuff for the next two years.
Well, I push back a little bit on that because, of course, I agree on the coding side, and I think there's a lot of... one of the things developers do early in a platform is they love to tell you that they're doing something every day and it's working, but it actually just isn't. And that's just what happens early in a platform. They tell you all these things that they say are easy and they're actually not. And they spent 18 hours struggling with something that didn't work. But on the vibe writing side, it also hits a point that I just think is so, so important, which is, yeah, you can prompt it to spew out a bunch of stuff, but like if you have a job and your salary depends on you submitting that, or you're a student and your grade depends on you submitting that, it actually better be right. And you can't just like say, look, I vibe wrote this and here you go. And I think people don't get confused when it comes to like math. Like everybody knows you have to go check the math. You ask it to do a table and then add a column that does math. But we're going to just see endless, endless human-wasn't-in-the-loop vibe writing things. And it's just that with programs, you can't really see that right away because in order to actually distribute it or get someone to use it, you had to at least fix the initial bugs. We'll only see them later when there are security bugs, authentication bugs, passwords stored in plain text, or a zillion other problems that are going to happen from vibe coding.
I mean, in a sense, we've seen this already, right? We saw a bunch of lawsuits that were citing case precedent from cases that don't exist. So maybe this is actually the operative point, which is there's full autonomy, there's partial autonomy. Maybe partial autonomy in writing is moving us from writer to editor, but you still have to be the editor, right?
We should also give him credit. I mean, many people have talked about this, but he did a fantastic job using the Iron Man analogy of how we're going to have, you know, autonomy, partial autonomy, and a slider to control what you want. I actually think that's a fantastic analogy and a way of thinking it that gives you a very clear picture from the movies. You live through the movies, but at the same time, it's still... people are very, very aggressive on their timeline of agents, and there's a very, very long history in trying to automate things that turn out to be very, very difficult to automate. And so I think that one, he did a fantastic job. He said people are talking about like the year of agents. Yeah, that's a good consultant phrase. We're just like he said, we're in the decade of agents. And it's going to take a decade for things to be anywhere near living up to agentification as a meme.
You know, it's an interesting point. I think a lot about agents as applied to financial services. And I think there's a set of problems in financial services that are high friction, low judgment. So, for example, when I want to go refinance my personal loan, I don't really feel attached to any specific brand of a personal loan provider. I just want the cheapest rate. So it's actually a very low judgment decision. But going and researching and applying for a personal loan is a high friction process. That's something I would love to delegate to an agent. I think it can do a nice job. Whereas doing my taxes, wow, like Stephen, how much risk do you want to take on your taxes? How many things do you want to report or not report? That requires an enormous amount of judgment. And of course, it also is high friction. So when I think of the 2x2 of where is automation going to come first, I think a lot about high friction, low judgment.
I want to build on that because I actually think it's super important to also consider that for anyone to offer the alternatives to the market, there has to be an ability to differentiate, to explain. And so you end up with this kind of thing where like, I just want the cheapest flight. And of course, for 20 years, all the flight searches and stuff has worked on the cheapest, but it turns out that's not actually what you want. Plus, a lot of people want to intervene in presenting your choices to you. And so I don't like this idea that all choice in life is going to be reduced to some headless API, right? I don't understand, people have to go build that and make a living building those things. So to your example of like refinancing a home, like the only reason that it can exist as a search problem today is because the different people who want to refinance you can target you with an ad and attract you as a customer and differentiate themselves on that offering, right? And if you can't do that, then your ability to actually automate that task isn't going to exist because there's no economic incentive to just be, hi, I'm the headless, faceless, nameless, low-price mortgage lender is not really a business. Like there's nothing there. Just like, you know, headless, faceless, nameless food isn't a thing. It doesn't show up in a white can labeled food and then you consume it and you're okay, all good. I have food now.
Well, maybe Soylent, but yes. No, in the future. In the dystopian future of Repo Man. That's where we end up. But that's not going to happen. I want the cheapest flight as long as it's not on Spirit Airlines, right? I want the cheapest flight, but I'm traveling with a family of three. I don't want to leave at 5:00 a.m. No red eye. Yeah. Like it's just, I want miles on this airline. You know, a lot of things don't add up to that.
And this is a real thing in business. It's a thing on the producer and the consumer side. Consumers really, really want much more choice than they often think they do. And anyone who's bought anything on Amazon knows they complain about the choice, but they really don't want just, you know, like a phone case to show up as the thing because it was $6. I think this is a real throughline through the talk, you know, which is partial autonomy, jagged intelligence. Karpathy is just talking a ton about the constraints of the technology, which I think is the right thing for us to be thinking through trade-offs around as builders. And he does a great job, very much as this philosopher that I love. His delivery, his tone. You really, really don't just go read the summaries. Don't read a post. Go just watch the video immersively.
Totally. Well, I want to get to automation and sort of employment, particularly on the entry-level side, but first I just want to ask the broader question of, you know, there was this idea of, you know, like AI plus human, I think it was chess could beat AI for some period of time, and that was kind of the co-pilot view of the world of human plus AI will have a better product. And then it turned out, I think it was chess, maybe it was Go, or maybe it was both, that actually that was a temporary thing and AIs are just better. And then there's the question as to how much of the world is like chess, where a human plus AI is only better for a certain period of time and then the models get better, or how much the world is like something else that is always, we're always going to want that human plus AI is just better, or we're just always going to want humans to do it.
Look, my view is that in a domain in which you have a formal definition of correctness, the path will be no autonomy, partial autonomy, full autonomy. In domains where you don't have a formal definition of correctness or where a ton of human judgment is necessary and human choice and sort of human direction, the right product design is not to go all the way to full autonomy. I would argue that chess and Go do have a formal definition for correctness. So it makes sense that those were fully automated over time.
We're back to the early stages of where things are, which means that a bunch of programmers are sort of defining what success looks like. And programmers are very good at either it works or it doesn't work, or I just want to automate this, or I'm going to reduce your job to a tiny shell script kind of mentality. And I just look at the world as everything is gray and everything is much harder than it looks when you don't actually have to do it. Ages and ages ago, I visited a really giant hospital in Minnesota to help them figure out how to use Excel within the medical profession. And the doctor just looked at me and he's like, I don't think you understand. He was like, my job is all uncertain. Every aspect of what I do is uncertain. So adding something that pretends to be certain like a spreadsheet to my uncertainty doesn't actually help me. And so fast forward, first, I've spent 25 years with a doctor, but that's a different story. But there was a story this week about radiologists. And so very early in, actually, if you go to ImageNet, everybody was immediately, radiology is doomed. Like there was, oh, like you never need to get a skin cancer biopsy. You'll just take pictures of your mole and it will just tell you. And then you find out, wow, there's judgment there, and there's even judgment in doing the biopsy and how to do the biopsy and then what to biopsy and all this. But it turns out the radiologists have like fully embraced AI, but they embraced it as no different than they embraced the latest MRI technology or the latest software update from GE for a CAT scan. And I just think there are so many things like that, and so many jobs are either very, very uncertain or most of the job is basically exception handling, right? And like people are like, oh, we're going to automate our taxes. Okay. Taxes are literally a giant cascading if and switch statements of exceptions. Yes. And so the idea that you will just automate that, well, you have to know the answer to all the exceptions. And if you're going to prompt it with the answer to all the exceptions, then you're doing your taxes manually. So it's a very strange... I mean, you know, it's sort of like everybody, once you reach a certain income, you have to get help from an accountant to do your taxes. And the first thing the accountant does is ask you for your tax planner. And as a software person, I look at it and I'm like, the tax planner really, really looks like the input fields of the software you're using. So if maybe I could just buy that software and then type it in. And I said that, and he's like, well, you're welcome to, but you will go to jail. And he explains, like, because every time I give him a number, it's a whole decision about where to apply it. Does it work? And I'm like, well, you're not really a farmer, so don't fill anything in on that form. And stuff like that. So people get automation is extremely difficult, and it's exception-bound, it's judgment-bound, and it's all uncertain.
You know, a field in which this question comes up a ton is product management. I've had so many conversations with product managers over the last two years about the death of product management. It's the end of the field. Why do we need PMs? And I think, you know, our sort of developer generation has developed a real resentment towards product managers, which is a different conversation. With that said, I think that the product management job is the job of addressing ambiguity, and it's ambiguity that prevents progress from being made. Sometimes it's execution, decision-making, product design that will not change. The nature of business and human interaction and companies is these complex adaptive systems where there will always be ambiguity. I think you'll always need judgment, and you'll always need somebody who looks like a product manager.
Yeah, I think that really gets to the sort of the vibe coding sort of challenge we're dealing with, which is like, how fast can we go text to app? And I think, you know, here what's so interesting in the long arc of platform transitions is that we're also having this platform transition happen not just out in the open. You know, we've had that before. Like back when in the earliest days of computing, these platform transitions happened in user group meetings, like at the Cumberland Community Center down the street, or in magazines or newsletters, and then with newsgroups, then the internet, and so on. You know, the whole internet was all ICQ, and it was all in the open. But now it's like happening on CNN, on the nightly news. You know, everyone knows about the platform transition that's happening, in particular on social and in Discord. And so what's happening is you're getting a lot of like vibe coding for clout. And so you're getting a lot of this like, I had an idea, I prompted it, and it worked, and here I am. And I, you know, at some point I just go like, I'm calling BS on that. That's like not a thing. And then I sound like an old person, and because some people think I am, I don't, but some people think I am.
You know, it looks like, hey, you know, you're just being old.
But then you dig in and you find out like, wow, you're prompting, although it's English-like, it turns out you're just programming and you're just programming in prompt.
And people are like, oh, this is what we're going to do is we're just going to get the model to require a little bit more structure. And I'm like, you're writing a new programming language.
And we're just like this path of text to app and vibe coding is just developing a new language, which is super cool.
Like, Lord knows, like the world is built on programming languages. In the 80s, if you drove slowly past a computer science department trying to get a PhD, they would just invent a new programming language right then and there if you stood outside the building for too short a time. But we can't lose sight of the fact that the arc of programming has been one of basically overpromise and underdeliver. You know, when I was in college, like the theory was the market was going to need so many programmers that the whole employment force would be, the whole workforce would just be software people, and that never happened. And now here we are, we're not going to need any. They're all just going to go away. And I think it was extreme in 1990, and it's extreme today. And I think the big thing is this overpromising at each transition. You know, even just most recently, low code.
Like, who doesn't, who even says that word anymore? Like, it's like we're not allowed to even mention it. And it's because, like, it's always the same thing, which is, yes, if all you're doing is a very straightforward app that looks like all the other straightforward apps, but with like a domain spin or branding logo or something. You know, we see this with Wix and with website templates. Like, it's possible, but you're not going to run a company on any of those.
I totally agree. Where I disagree actually is I think that the language model in this case, but the language in your metaphor, is improving at a dramatic rate underneath these things. So while I think almost all these products today, they're good at prototyping, they're trying to push into refinement, they're not really usable as things that you can actually deploy to production at all. In fact, most of the cool demos you see on Twitter don't work three days later. So they're very much in the prototyping phase, but the programming language in the metaphor is improving dramatically. I think we'll get there, or at least make more progress than we think, versus a traditional programming language like object-oriented didn't feel like it 100x'd the number of programmers or 1/100th the amount of time to ship something to production. We just got, you know, new tools and sort of new problems to solve.
Well, of course, you're benefiting from hindsight.
