Transcription
David Westin: This is a story about a double-edged sword. At this point, AI is promised to do just about everything, but it turns out that it's best at creating the very thing that it is made of, code. It was once considered a high-skill task, but coding is now accessible to just about anyone with the help of generative AI. On the one hand, this unlocks possibilities for creating a wide range of products and businesses that otherwise might never have seen the light of day. On the other hand, it's made the future less certain for those who write code professionally. Our colleague Ed Ludlow brings us the story of what happens when we lean in and let AI do the work.
In Uppertract, West Virginia, Jamie Grove owns a boutique warehouse helping clients ship out anything from dinosaur bones to board games. Last year, he built software that automates shipping out packages with help from AI.
"Where's the order at that you're looking at there? There's absolutely no way I could have done any of this without AI. For the initial run to take some spreadsheets that we had that we were working with to get an actual workable sample ready to go took me less than a day. And we were live and we were using it right here in the warehouse."
In Oakland, California, Cynthia Chen wished there was an app that collects pictures of different breeds of dogs.
"And then it was when I learned about vibe coding online is when I thought maybe I should try it myself. The first time you see like this little pop-up and it says 'build succeeded' and then you can see the app pop-up and like it's actually real. That was like the sort of magical moment where I was like, 'Oh my gosh, this is crazy. I can actually build things myself.'"
The term vibe coding was coined by Andrej Karpathy, a founding member of OpenAI to describe a process of computer programming akin to having a conversation with a robot. Here's what it looks like in practice. Say I want to create a website that visualizes and animates a data set. I would use a generative AI tool like Claude or Codex or in this case Gemini and tell it exactly what I want using plain English. AI then writes the code for me. And as coding has gotten easier, it's led to the creation of more code than ever before. Activity on GitHub, the platform used for storing and sharing code, has seen a massive increase in activity, surging in early 2025 when AI coding pilots became popular. And although vibe coding has helped small businesses and hobbyists create their own software, it's the professionals who are leading the way.
"So what goes on in this building is we have a mix of folks working on Google Cloud. At Google, when we talk about autonomy and agentic engineering, we have systems that allow you to kind of have almost a virtual software engineer."
"Hold on. If all these brilliant engineers are using AI in this way, what is it that they're doing all day long in beautiful buildings like this one?"
"Excellent question."
Addy Osmani is the director of Google Cloud AI. He oversees teams of engineers currently building the next generation of AI tools for businesses.
"So if you are vibe coding, you're pretty much just giving into the vibes. You don't necessarily have a clear, full idea of your vision. You're just working with the LLM. You're trying to get somewhere with it. If you are engineering, that's where you have to apply rigor to it. You have to have this clear set of requirements. You are testing. And whether you are a startup or whether you're in a big enterprise right now, the role of the software engineer is going to be evolving to one where you are increasingly a little bit more of a manager. You're going to have effectively a virtual team of agents that you're responsible for. And you have to own the outcomes. It doesn't matter how many agents you're running."
"You're responsible for the output."
"You're responsible for the output. Exactly. And so you need to decide, how am I evaluating quality? How much time am I going to spend evaluating quality? Because there are some people who very much enjoy YOLO. Like, OK, well, the agents ran overnight. Looks good. Kind of runs. I'm just going to deploy it. But if you're building any kind of serious software, you still need to have some idea of what is the quality bar? What are my quality gates? How am I making sure this is actually going to meet the needs of my users in a consistent way?"
For the engineers at Google's headquarters here in Sunnyvale, California, Osmani says AI isn't just making their lives easier. It's making them better at coding.
"The extension of that question, which we pose largely by investors, is how do we measure the productivity gains of that engineer or that team?"
"I remember in the earlier days of AI, you know, org leaders would look at things like, 'Oh, hey, well, how many lines of code are being generated?' Which is not in any way a good proxy for productivity. But these days, I think that people use a mix of different kinds of metrics. You try to use qualitative and quantitative. Generally speaking, there's a big productivity boost. In the earlier days of AI, I would have said, you know, that boost is 10 to 15 percent. These days, it's anywhere from 30 to 50 percent. And I see that number only continuing to go up."
At MIT, Frank Nagle and a team of researchers surveyed over 187,000 software developers who are using GitHub Copilot, a generative AI tool for coding. And they found that workers are more productive because what they spend time on has changed.
"I think one of the big things that we often think about with AI is that it's just going to enhance productivity, right? It's going to make us faster doing whatever it is we do. But that's just--really just scratching the surface. You have 100% of your time that you allocate to work. How did that break down along these dimensions of what we called core work, actual coding, versus more project management type of work? And what we found is that when coders started using these types of tools, they massively shift the amount of their time that they allocate to coding. And they take away a whole lot of their allocated time from project management. And so part of the reason we think that this is happening is that if in the old days you were writing piece of code A and that piece of code was dependent on some other piece of code B, you had to wait for that other person. You had to interact with them and make sure that everything worked together. Whereas now you can just write it all yourself."
This productivity boost is particularly true for those who are writing and deploying brand new code, and especially those with no knowledge of code whatsoever, including creatives like Chen.
"This is Presspetals, the new app that I'm working on. And here is a pressed flower. So I'm going to go to the cloud code that I have running, and then I'm simply going to describe what I want it to do. I actually tried a whole bunch of tools in the beginning, and this was about a year ago, which I think feels kind of like the stone ages of vibe coding. I was able to make the foundations of the app in maybe a month. And as a totally non-technical person, I was actually able to build a full stack app. So there was front-end and back-end capabilities, and I was able to get it out on the app store just entirely myself."
For business owners like Grove, his vibe-coded solution is helping his warehouse save on costs.
"We have three main coding solutions here that we've used AI for. One is to create batching, which is the important part in terms of taking orders that are different and getting them all together into groups that make it easy to pick. And then the other solution that we have is inventory tracking. So that's a standard warehouse feature, but our clients are also very different. And so trying to force a client into a single inventory tracking system is really difficult. So we actually used AI to build an inventory tracking solution that allows them to be themselves, basically. So I have a pretty varied background. I started out as a programmer a long, long time ago. But even someone who is a very fast coder could not have built all these solutions. Impossible. If I were to do it with a team of programmers, I could have five programmers working on this full-time and still not deliver as many results. If I were to install a software system that does everything that we're doing now, let's just say without all the customizations and all the flexibility, we might be talking about an annual license of anywhere between $6,000 and $10,000 a year, scaling all the way up to maybe $30,000 to $50,000 a year, depending on how much volume we push through our warehouse. This doesn't have the flexibility that we would want, and it's expensive. Like, for a small boutique warehouse, that's a big expense."
Now, for about $20 a month, business owners like Grove are building their own software solutions. And whilst that's opened up new possibilities, there is a downside. Since 2022, employment for software engineers right out of college has fallen by nearly 20%. Nagle thinks companies are making a big mistake.
"I do think that one of the biggest risks of the whole thing is that people are going to get too focused on the short term and not think about the long term enough. First of all, if you don't hire any new people, who's going to run the company in 10 to 15 years? But second of all, our research and others has shown that these junior people are actually the ones who are able to change their job and get the most out of using these tools. And so if we're not hiring them at the same rate we were before, then we're not going to be able to take advantage of that."
And Chen agrees.
"I feel very strongly that this is not replacing engineers. I think the more you use AI to build, the more you understand the space and kind of even know what you can and can't do. It's like technically I can make it, but an engineer probably could have made this in a much shorter timeline and probably with, like, much more robust code."
"From the business standpoint, I do think there's this opportunity where companies that have been thinking about things like reverse mentoring, where the younger folks can help the more experienced folks learn how to better use these types of tools, while the experienced folks are able to better help the younger folks understand how the industry works."
If AI can write code, then what is the role of the software engineer? Osmani says, "Today it's all about quality."
"If I want to build a robust engineering artifact, something that's going to last time, something I can ship to hundreds of millions of users or billions of users, there are a lot of things that it needs to factor in. And so what people in buildings like this are doing all day long are trying to make sure that the code is actually meeting that quality bar so that when we do ship something to you that happens to be using AI behind the scenes, you're actually getting what you want. That's the important thing for the users at the end of the day. They don't care if a human has been authoring it or AI has been authoring it. It doesn't help them get the job done in a reliable way."
"Is vibe coding a term that's therefore used? Is it banned within teams?"
"Absolutely not. I think that vibe coding has a lot of value. Vibe coding is enabling people to go from idea to execution faster than ever. And it has completely changed how many teams approach prototyping. You know, so many times in the past, in Silicon Valley and lots of places, if you had an idea, you'd go through weeks or months of just discussing or debating, 'Hey, can we afford to build this?' Now you can just build it. And I think that it has a very concrete place in our language now. It's just important that you understand that vibe coding a thing does not necessarily mean that you have a production ready artifact that's going to be battle hardened."
In Silicon Valley, the test of AI's progress is true autonomy. To start a project in the evening and wake up in the morning to code written autonomously by an AI agent. A dream for senior software developers and perhaps a nightmare for entry level computer engineers who would have once done that work themselves. But for us, non-coding mere mortals, the moment could be ripe to do what mankind is best at, building and creating.