I want to talk about the essay you wrote last week called "AI Horseless Carriages," which captured the minds of a lot of people in tech, myself included. For listeners who didn't have a chance to read it, can you start by walking us through the core idea and what compelled you to write it?

My core idea is that a lot of app developers building software with AI are doing it in a way that really constrains the power of these AI models and gives users a frustrating experience that makes it easy to dismiss AI right now if you're a user.

And why is that?

That's a really good question. The example I chose for my essay was the little email draft-writing agent that is embedded in Gmail's UI. It's really simple. There's a little text box, and you can type in something like, "I want a draft of an email letting my boss Gary know that my daughter woke up with the flu and I can't make it into work today." What you get is inevitably this weirdly formal-sounding draft that doesn't sound anything like me. That's the first big problem.

The second problem is that the instructions I have to write into that little text box to get a good email draft are inevitably longer than the draft of the email I would have just written myself, right? So it's more work to have this thing write me a good email than it is just to do it myself. It's like they have perfectly captured the experience of managing an underperforming employee with this little agent.

If we take that as an example, in your estimation, how do we fix something like that?

In my essay, I tackled these two problems separately. One, this thing doesn't sound like me, right? It's supposed to write emails for me. It should sound like me, right? The draft that you get, if I sent my boss Gary that draft, he would have thought I had been a victim of a phishing attack, right? That's not Pete. I tackled that. And then I tackled separately this problem of using agents for writing anything original actually isn't that useful, right? They're not actually that good, surprisingly, at generating useful original writing.

So let's go through these one by one. First, the tone problem. It doesn't sound like me. What's actually going on when I write my little prompt asking for a draft is that Gmail will take my prompt and paste it together with what's called a system prompt. This is the prompt that some group of product managers, and maybe lawyers, at Google wrote that describes to this agent who it is and what its job is.

You can really think of these AI agents this way. It's very simple. Think of a brand new, fresh, college-level grad whose first day on the job has no idea what their job is. The system prompt is there to tell the AI agent who it is and what its job is. And we're not allowed to see that. Gmail users are not allowed to see that system prompt. But my guess is that it's something like, "Write an email draft for the user, obey their instructions, use formal language and good punctuation." That's where you get this weirdly formal-sounding email draft, right?

What I show in the essay is that if you just let me, the user, edit that and write my own system prompt, then I can make an agent that sounds a lot like me, right? And that's really the promise, I think, of AI applications. It's an application that I should be able to customize to do things exactly the way I want them to do just by writing text, right? That's what's so magical about this technology: you can tell it what to do just with English. You don't have to know how to code anymore to tell a computer what to do. You can use English. The problem is that a lot of apps aren't designed to let users do this.

If you're Google, why not simply read and parse the emails you send, which they have access to, and build a Pete-sounding template generator or email generator or agent based off of that? Why wouldn't they do something like that?

I'm sure they are. I don't think I gave them enough credit in my essay here. And the point isn't that there aren't many smart things you can do. I have been using Gmail for 20 years, right? I shouldn't have to write this prompt from scratch explaining how I write emails, right? They should be able to come up with a reasonable-sounding draft for me.

The problem I have is that I'm not allowed to see it, right? A lot of developers seem to treat these system prompts the same way they've been treating code for as long as we've had a software industry, which is as a black box that the user isn't smart enough to understand. And that may be true for code. Most of us don't know how to write code, but it's not true for English. And it's surprisingly intuitive, when you're actually allowed to talk directly to the model, how easy it is to teach it something.

How do we build agents or AI that feel more authentic to the customer or the user from scratch?

It's a really good question, and I'm working with a bunch of companies at YC that are trying to answer it. My contention is that once you get the hang of writing prompts, it's really intuitive. Really, anybody can do it.

I'll give you one example. I'm working with a company this batch called Verse, and they are helping law firms build agents to help do quality assurance for the documents they produce, right? It is so important for these law firms to be able to build their own agents, because every law firm writes and reads contracts differently, right?

A really simple example is if you look at a lawyer at YC, right? We review a lot of term sheets, but when we do, we look for a very specific set of things that are important to us. We want to make sure the founders are aware that if there are odd terms in there, they know about it before they sign it, right? And that's totally different from a lawyer who, for example, represents a Series A investor, right? It's a totally different set of things. So I think this is a perfect example. If you're a lawyer at YC and you're building an agent to help you review term sheets, you should have control over how term sheets are reviewed, right? You should be able to build this agent from scratch.

It's kind of just classic square peg, round hole behavior from companies that feel like they have to do things the old way, the traditional way, the way things have always been done, instead of thinking a little more outside the box and saying, "Given this entirely new paradigm of technology we're working with, what is best for the customer? What is best for the end user?" So we just end up with a mediocre product offering at the end of the day.

That's my view: what's really happening here is that a lot of developers are building AI software the way they've built software for a long time. One of the pieces of pushback I got when I published the essay was that users just aren't technical enough. They don't know how to write prompts, which I think is true, right? This is a brand new skill that very few people have at this point.

But I'm pretty optimistic that's going to change fairly quickly. When you look at how quickly ChatGPT caught on, and how normal and comfortable it became for most of us to just sit there and talk to an AI, I'm pretty optimistic. I look back, and when I was growing up, computers were thought of as something in the realm of nerds, right? "Of course you don't understand how to use computers. You're not a nerd," which is laughable today. We've all figured it out. And I think talking to an AI model is a lot more intuitive than learning how to use an operating system on a computer. I think this is going to happen a lot faster.

There's no doubt that using natural language is more intuitive and easier, and it feels more comfortable for people who are not technical, right?

Totally. We're seeing this, right? With the rise of vibe coding, a whole group of people who have never written a line of code in their lives are starting to learn how to build applications. And the reason they're able to do that is because it's so intuitive to sit there and describe to an AI model what you want.

It reminds me of 10 or 15 years ago and the mobile revolution, right? When everybody was like, "There's an app for that. If I could only find somebody to help me build my app, then I would be worth a million dollars." And now you can do it, right?

Totally.

I'd be remiss if I didn't end on this note. You ended the piece on a very optimistic note: a world where we spend less time on mundane work. What is the one thing you're most excited about on the frontier as AI-native software evolves?

I think it's exactly that. Maybe just to wrap up the argument in the essay, what I said earlier is that these AI models, ironically, are not that useful for generating text from scratch, right? The thing they're useful for is reading text and processing it, right?

The example I gave in my essay was if Gmail let me build an agent that managed my email for me, right? Read every email coming in, assigned a label according to logic that I've written out in English text, archived ones that I don't need to read, maybe summarized other ones that I need to read but probably not respond to. If you give these agents access to some really cool tools, they can pay your bills for you, right? They can handle scheduling back and forth. There's so much time I spend on email today doing really mundane work that I would rather offload to an AI that, with my supervision, can do most of it for me.

And that's what I get really excited about when I think about the future. I think we're pretty close to a world where every piece of mundane work I have to spend time on in my day-to-day life, I'll be able to teach an AI agent how to do for me. And that's a world where I get to just focus on a handful of things that I think are super important, right? Where I don't have to do work I hate doing, because I can manage an agent to do it for me, and where the work I love doing, I'm even better at because an agent helps me do that. I'm so incredibly optimistic about the future because of what I think AI agents are going to help me do in my day-to-day.