From The New York Times, I'm Natalie Kitroff.
This is The Daily.
On Monday, we looked at the growing pushback against sprawling AI data centers in communities across America.
Today, my colleague, Kevin Roos, takes us inside one of the most transformative technologies that infrastructure is enabling.
A new way of programming that may be the biggest development in artificial intelligence since the launch of ChatGPT.
It's Wednesday, February 18th.
Okay, so the other day I slacked you, Kevin Roos, my colleague, my friend, the host of the New York Times tech podcast, Hard Fork.
And I asked you, what is vibe coding?
Because this is a thing I had been hearing about and I had no idea what it was.
And you answered me roughly, in the tone that one might use with their very elderly grandmother, I think.
Fair to say?
Yeah.
Well, I am not ageist.
I think people have lots of different levels of comfort with technology.
I think my tone of voice was directed at you specifically Natalie, because I know that you are younger than me.
And so you have no excuse to be this confused by technology.
Yeah.
Okay.
Fair enough.
So since then, I did my own research as any young person would do.
And I've come to the understanding that vibe coding is, in fact, very important.
It appears to be using AI in a way that is truly revolutionary.
And I do think that that I cannot possibly be the only one who has not caught up to this.
And so I want to ask you to just start by laying out the basics for those of us who aren't there yet.
What is vibe coding?
So vibe coding is a term that was coined about a year ago by a guy named Andre Karpathy, who's a former programmer at OpenAI, very well known out here in Silicon Valley.
And he was describing this thing that he was doing using these new AI coding tools, which is that, instead of learning a programming language and writing the code line by line by himself, he would just sort of let the program write the code for him and build software that way.
Let the vibes do the work.
Let the vibes do the work.
Exactly.
And suddenly, for the first time, about a year ago, you didn't have to know how to code, to build software.
You could just use these tools.
They would help you along the way and you could just kind of oversee them as they worked.
So what does that actually look like?
You experimented with it.
What did that do for you?
Yeah, so I tried vibe coding about a year ago, when this term first emerged, and I would build these little test tools for myself.
I built an app called Lunchbox Buddy, which was basically a way to help me pack my son's lunch by just taking a photo of whatever was in my fridge and giving it to this tool and saying like What are the different combinations of things I could put in this four-compartment lunchbox?
Oh, my God.
That's amazing.
So useful.
I want it.
Yeah, I built some other stuff too, but honestly, it was kind of clunky.
The tools.
Basically, you still needed to know a little bit about programming to be able to do this, and they were pretty buggy.
I tried to build some stuff to help me sort through my emails, and it couldn't do it.
It was mostly useful to professional programmers.
And it doesn't stay that way, I'd imagine, because here we are talking about it as something that is revolutionary.
So what happened?
So the big innovation between last year and now is what is called agentic coding.
Agentic coding is basically just a fancy way of saying an AI system that can do much more of the process autonomously.
So you still do have to have the idea for what you want to build.
But then with these new tools, you can just sort of give them a project and they can make a plan.
They can decide what programming languages they want to use.
They can even create their own little team of agents and say okay, this agent's going to go off and do the research.
This agent's going to build the website.
This agent's going to test it.
And so you sort of have a little team of robots working for you to do whatever task you've given it.
My dream, a team of robots.
Just sketch out why that's a big deal, Kevin.
Well, because these models are getting much better very quickly and they are now able to do the kinds of tasks that would take humans hours or sometimes even days to do.
And so they are working their way up into the entry-level functions in a lot of white-collar industries.
The models themselves are also getting much better.
They're not hallucinating as much as they used to.
They're not making as many silly mistakes as they used to.
They are doing better at things like reasoning and solving complex math and science problems than they used to.
So, sort of across the board, these things are getting better at a pace that is surprising even the people who are building them.
Okay, I want to ask you about the transformational potential of that, about the impact on jobs.
But first, just tell me at a really practical level, where do you see these new tools?
Like, what or who is driving this agentic phase?
So this is largely being driven by a handful of companies that are sort of the leaders in this AI race.
So one of them is OpenAI, the maker of ChatGPT.
They have a new agentic coding system called Codex.
I have been using CloudCode, which is made by Anthropic, another AI company.
And these are the tools that have come out just in the past couple of months that have people feeling like things are really accelerating.
Cloud code sort of blew up.
Millions of people have been using it.
They are doing increasingly complex tasks with them.
People are using these in their jobs in the software industry.
And what's really changed is that if you talk to people at the big AI companies who used to program, who used to write code by hand, they'll say I don't write any code anymore.
I basically supervise this team of agents and I kind of orchestrate them and set them off toward tasks.
And then I go get lunch and I come back and the work is done.
Okay.
I want to understand this.
I want to understand the story of this.
I think we should pick one.
And it sounds like Claude Code has been the one that's been blowing up.
So tell me the story of this model.
How did it start?
How did it get created?
So this model, app, Claude Code, started as a side project.
Basically, an engineer at Anthropic named Boris Cherny had an idea.
He was trying to sort of get a sense of what Claude, Anthropic's chatbot, could do.
So he said, well, what if I just give it my computer?
What if I install Claude inside the Terminal app, which is the app that you use to write code and interface with your computer if you're a programmer?
And what would happen if I just set Claude loose inside my computer with a bunch of tools like give it the ability to write scripts, to create files, to organize things, to debug code?
And he starts seeing something really interesting happening around the office.
His colleagues are starting to use this tool to do their jobs.
They're starting to program with it.
It started with the engineers.
They're saying this is really great for coding.
At first, it's maybe 20 percent of their engineers and then it's 40 percent.
And now all of them are using it.
But it's also being used by people who are not professional programmers.
People in marketing and sales and finance are using this not to just build little apps, but to automate parts of their jobs.
So it's like really catching on way beyond.
You know the technical people, like normies, are using it.
Exactly.
What people have discovered is that it can do lots of stuff.
It can automate your email.
It can create dashboards for things.
It can reorganize the files on your computer.
It's sort of like you just have a computer that can use a computer now.
This is starting to blow my mind, but I want to see it in action.
Can we make something together?
Can you make something for me?
How does this work?
Okay, I would love to show you.
So what I'll do is I will share my screen here.
I'm excited and a little scared.
Don't be scared.
It'll be fine.
Okay, so this is the Terminal app.
Have you spent much time in the Terminal app on your computer?
The answer is not going to surprise you.
No.
Okay, so I had not either before I started using Cloud Code.
Okay, this looks technical though, Kevin.
Can I just say, this is not my chat GPT window.
No, this looks like hardcore programming things.
And I think the interface is what has made a lot of people feel like this is a tool that they can't use because they don't write code.
Right.
So I will just say, I do not write code.
I am not a coder.
You're closer to me than them, you're saying.
I am much closer to you than them.
So, okay, here we are.
We're inside Cloud Code.
And I'm just going to give it a prompt.
I'm going to say build a website for my colleague, Natalie Kittrow.
And I noticed, Natalie, that you don't have a personal website, at least that I could see.
Correct.
Do you have one?
I don't.
I'm kind of giving over my identity to the New York Times.
No, see, you need one.
You are a big-time podcast host now, and so you need a website for yourself.
So she's the co-host of The Daily, and the site should look—how do you want it to look?
Sleek and professional?
Yes, and cool.
Make me look cool, please.
Thank you.
And cool.
Cool. and professional, make her look cool.
And also cool, sorry.
Thank you, yeah.
She really likes the Philadelphia eagles, so maybe do it in eagles colors.
This is getting better and better.
All right, so I hit enter and it says burrowing, which is one of its words that it uses for thinking burrowing.
Do we know what it's doing?
I promise you I do not understand what it's doing.
Like a professional coder would know what it is doing.
Right.
All I know is that it's doing something.
And it's going to think about it.
It's going to make a plan.
And then it's going to start building.
How long does it take?
It depends on the complexity of the task.
I've had tasks that have taken 10, 20, 30 minutes before.
But I think this is going to be just a minute or two at most.
And the idea is that it's kind of scraping the internet for information about me.
Did you tell it to make sure that everyone knows I'm young also?
We can add that in a subsequent version.
So it wrote 644 lines of code.
Okay, so it's live.
Let's just see what it looks like together.
What was that, a minute?
That took a minute and 36 seconds.
Look at that.
Wow.
We've got a little like Eagles Green website.
Yeah, I see that.
Oh, my God.
It knows that I was in Mexico as a foreign correspondent.
Uh-huh.
Oh, the spirit of the daily.
The best journalism doesn't just inform.
It makes you feel the weight of reality and then pushes you to care.
It's a quote that is attributed to the spirit of the daily, which is not a real person.
Yeah.
So because I didn't give it a lot of information it uh, it just sort of made some stuff up to fill the space.
But the cool thing about this is now, if you want to change it, you can just go into this window and you can say Hey, that's a fake quote.
Can you make sure all the info on the site is accurate?
Yeah.
I'm thinking we should maybe add, like an Easter egg to this, something fun for people who go to your website and really want the full Natalie Kitcheroff experience.
Okay.
What's it going to be?
So you're a big football fan.
There was a video game called Tecmo Bowl, very early football video game.
And so I would like to see if we could put a version of Tecmo Bowl that's playable like a little playable video game on your website.
Yes.
Wow.
Can I also hire you?
I feel like you're good at this.
You're good at talking to Clyde.
So I'm going to go back to my terminal app here and I'm going to say, that's great.
Could you add a playable Tecmo Bowl-style video game to the site?
Oh, it says, this is going to be fun.
So it's writing hundreds of lines of code.
I mean...
This is insane.
It's wild.
I mean, this is why people are freaking out about AI right now, because this was not possible even a couple of months ago and now it is.
Okay, so now our video game is done.
Should we go see it?
Yes, please.
Okay, so here's Natalie's Tecmo Bowl.
Oh, my God.
Do you see it?
Oh, my God.
Yeah, I do.
So it says, press space to hike...
And arrow keys to move.
Okay.
So, oh, we got to run.
We got to run.
Oh, touchdown Eagles.
Yeah.
Go birds.
This is incredible.
Wow.
This is like, I feel like I'm going to draw a lot of people to my new website with this feature.
Yeah.
You're going to be getting so much traffic.
Okay.
So that is cloud code and that is how these new coding tools work.
How do you feel?
I feel pretty blown away.
I honestly can't believe that we just made my first website and that it has a video game on it.
I have so many questions that I want to ask you about it and we'll do that after the break.
We'll be right back.
Kevin, I have to just acknowledge that it is honestly an amazing feeling to see this AI agent at work.
It's probably the first time that I have felt awe at this technology since ChatGPT came out.
It's been a really long time.
And there is something a little bit unnerving about it.
And so I just wonder, for you, what do you make of this moment?
What do you make of this change that we're seeing in these tools?
Yeah, I think this is a big deal in the world of AI.
This is something that people inside the AI industry have been saying is coming for a long time.
But until recently, people couldn't really see it for themselves or try it for themselves.
And ChatGPT has been around for three years now.
And for most people, the novelty is sort of wearing off.
It's a fancy Google.
You can get a meal plan.
You can sort of give it some questions.
It can help you write your emails or memos or things like that.
But it doesn't sort of shock you anymore.
This feels different to me and, I think, to a lot of other people, when you can actually build something useful without knowing how to write code.
And I think this is what people think is the thing that's going to take AI from just kind of being a fancier Google to something that genuinely changes how people work.
When you say that this is something that the tech companies have been saying is coming for a long time.
What is the it that is coming?
What makes it special?
Is it just the complexity of the stuff it can do or is it the utility of it?
I think it's the utility, but it's also the economic value.
This kind of agentic system can actually do work.
It can do tasks.
It can perform things that humans would have had to do by hand.
So it is sort of a step in the direction of making these things actually kind of members of the workforce in some sense, where you could have a company with some human employees and then a whole bunch of AI agents doing tasks.
OK, let's talk about that, because obviously, when we're talking about economic value, what that presumably means for the people that run these companies is lowering labor costs by potentially eliminating jobs.
Yeah, so that's the big question is, will this just help people at work?
Will this kind of agentic coding tool make people faster and more efficient and more productive and give them a new powerful tool in their toolkit?
Or will it start to replace people?
We are starting to see in the data some signs that this kind of tool is maybe displacing young software developers.
So there was a study from Stanford recently that looked at payroll records and found that employment for young software engineers just people who are early in their careers has dropped about 20 from its peak in late 2022.
So companies that used to hire five or 10 people to write code for them may only need one or two now with a bunch of AI tools.
Can I ask, we were pointing to issues with my beautiful website.
How good is this stuff?
I mean, can it debug itself?
Because I guess I'm seeing that there is also a lot of doubt and skepticism that you can really do this only with an AI tool at this point.
Well, coding is an interesting test ground for this because in some sense it is very verifiable.
The code either runs or it doesn't.
And so in some sense it's very easy to tell.
I mean, the website we built you, it works.
You can go visit it on the web.
It has a playable video game inside of it.
Right.
So yes, there are a lot of software engineers who are still saying these coding agents, they're not perfect.
You still kind of need to hold their hands and oversee them if you want to get really useful code that you could like deploy inside of a big business.
But that is maybe temporary because they're improving at a very rapid rate.
And we're seeing that some of the things that these tools couldn't do even a few months ago, it's now doing in ways that are correct and useful.
Well, can you just tell me quickly, like, what's making this improve so quickly?
So part of it is just that the models themselves, the things that are sort of underneath tools like Cloud Code, are just getting better very rapidly.
And part of that is because they're being trained on more coding data.
So they're sort of improving over time.
The other really interesting and really important thing that is starting to happen is that these AI tools are building versions of themselves.
They are starting to improve themselves.
What?
What does that mean?
So OpenAI's latest coding model, which is called GPT-5.3 Codex, was used to help build itself.
Early versions of the model were used to make changes to the training runs for later versions.
So this is happening across all of the leading AI companies right now, which is that you have these systems which can build websites, which can automate certain financial or legal tasks.
They can also help train AI models.
And so we're starting to see...
So AI models making AI models, just to be clear about what we're saying here.
Yes.
Is that scary to you?
Because it gives me the vibe of a sci-fi reality.
Yeah, this is a sort of scenario that we've heard about in science fiction for years, where you have AIs that are becoming increasingly capable.
They're building better and better AIs.
Within the AI community, there's this idea, this phrase of the intelligence explosion, which is when you have these systems that are doing what's known as recursive self-improvement, building better and better versions of themselves.
And so that is one possibility that these systems just start to accelerate their improvements to the point where they are sort of doing all of this autonomously, without human involvement.
Now there are some people who think this is still a far-fetched scenario, but just look at the trajectory.
I mean, a year ago, You had these very clunky AI vibe coding tools.
Now they're running autonomously for sometimes hours at a time.
They can build and maintain real software, and they're starting to help build the next versions of themselves.
So I think we are right to be paying attention to this and I think it's irresponsible to sugarcoat what's happening.
I have to say, for the companies and for the people that have been involved in AI, this also has to really be a kind of proof of concept moment for them, because there has been so much skepticism about AI over the last year.
A lot of fear about frothiness in the market, about whether we're in a bubble, about whether AI is really going to live up to its potential.
And what you've described is a transformative tool that can be incredibly useful.
In And with that, the potential for real job loss.
So any way you slice it, this is AI doing what they said it probably would.
Yeah, it is.
And I think you know I don't blame people for questioning last year whether we were in an AI bubble.
I think most people up to that point had only used things like ChatGPT, which are, you know, very capable on their own.
But I can understand why people would look at that and then look at the billions of dollars being spent on data centers and infrastructure and think like okay, isn't this a little bit much?
But the bet that these AI companies were making is that you could actually turn this stuff into useful tools that would be able to do real valuable work in the economy.
And I think we are starting to see that bet pay off.
Kevin, I think it's worth pausing here for a moment.
It's very easy to talk about job loss from all this in the abstract, but at a practical level, if a whole class of workers just lose their jobs if those jobs cease to exist— we'd be talking about a pretty massive impact on the economy writ large.
And I mean, this could hypothetically just be the very beginning.
If agents like these can do other forms of work, you could see many more jobs wiped out.
So I just want to ask, how likely is it really that these tools create that kind of job loss?
And what are the AI companies saying about it?
So I'll answer the first part, which is that I don't know how likely it is.
I have intuitions and guesses, but no one really knows how this is going to go.
It could be that there's a short term spike in the demand for software engineers because everyone's building new kinds of software and they need people to manage it all.
Mm hmm.
And it could be that these tools just never wipe out millions of jobs en masse.
But the AI companies and the people who run the companies are increasingly worried about this.
So Dario Amadei, who's the CEO of Anthropic, which makes Claude, has warned that he believes that AI could potentially eliminate half of all entry-level white-collar jobs within the next five years.
Right.
Wow.
So that could be wildly off.
But even if it were partially true, even if it were 10 or 20, that would still be a massive change to the labor market and to the lives of millions of people.
Yeah, there's something deeply unsettling about all of that.
Yeah, I agree, and I myself feel very uncertain about this.
And I think there are a lot of people in the tech industry who are very excited about these tools because it's letting them build a bunch of stuff very quickly.
But as I'm hearing from more people in white-collar knowledge work fields—
They're just very anxious right now.
People are really uncertain.
They don't know if the skills they're building are obsolete.
College students don't know what they should be studying so they can get a job after they graduate.
People are really getting nervous about how quickly this technology is improving and the possibility that it could make the future look very different.
Okay, and I want to ask about that future.
I mean...
Are we on a runaway train?
The speed with which this is changing feels hard to even capture, honestly.
And I wonder if you think about the progress made over one year, what's the next year?
What can we expect?
I mean, my honest prediction is that a year from now, these agentic tools will be dramatically better, right?
They'll handle big, complex problems.
You won't have to hold their hand as much.
They will become full fledged members of the workforce.
And you know, a lot of companies will still do things the old way because institutions are very slow to change, as we know.
But there will be this new kind of company that is emerging with AI work at the center of it.
I think that's going to be a really fast-growing part of the economy.
I think the job market impacts that we're starting to see hints of today will become much clearer a year from now.
But honestly, I lose a lot of visibility myself.
I mean, I have stopped trying to predict more than about six months out, because that's about as far as I can see right now.
I would not have predicted a year ago, when vibe coding first came into my consciousness, that it was going to be capable of doing all of this today.
So I think the most we can say right now is these tools are getting better at a very fast and accelerating rate, and that they keep doing things that are surprising and useful.
And there's a lot of unknowns about what that will mean for the future of human work.
Kevin, one of my favorite humans.
Thank you so much for coming on the show.
I really appreciate it.
You're welcome.
And we should also thank Claude who built you that beautiful website.
Thank you, Claude.
He's been a good bot.
He has.
Yeah.
For more on the latest developments in agentic coding, listen to Hard Fork wherever you get your podcasts.
We'll be right back.
Here's what else you need to know today.
We meet tonight at the crossroads, a point of decision.
Shall we expand, be inclusive, find unity and power, or suffer division and impotence?
The Reverend Jesse Jackson, America's most influential Black leader in the decades between the death of Martin Luther King Jr and the election of Barack Obama as president, died on Tuesday.
He was 84.
I know abandonment and people being mean to you and saying you're nothing and nobody and can never be anything.
I understand.
I understand.
Jackson, who worked closely with Dr King, became the first Black man to mount a major nationwide campaign for president.
In his 1988 campaign, Jackson turned his own life story, which began in poverty, into a rallying cry for a new kind of Democratic Party.
I understand.
I wasn't born in the hospitals.
Mama didn't have insurance.
I really do understand.
Born in a three-room house.
Bathroom in the backyard.
Slop job by the bed.
No hot and cold running water.
In one of his most famous speeches, delivered at the Democratic National Convention in 1988, Jackson declared that his life experience had taught him that what bound working Americans was not party nor race, but the shared experience of struggle.
I'm a working person's person.
That's why I understand you, whether you're black or white, I understand the work.
Jackson won almost 7 million votes in the Democratic primary for president.
I was born in the slum, but the slum was not born in me.
He failed to win the nomination, but he'd cemented his place as a defining figure in American public life.
You can make it.
Pull your head high.
Stick your chest out.
You can make it.
It gets dark sometimes, but the morning comes.
Don't you surrender.
Suffering breeds character.
Character breeds faith.
In the end, faith will not support.
We must never surrender.
America will get better and better.
Keep hope alive.
Keep hope alive.
Keep hope alive.
Today's episode was produced by Nina Feldman, Ricky Nowetzki and Olivia Natt, with help from Devin Greenleaf and Mustafa Mirza.
It was edited by Brendan Klinkenberg, Paige Cowett, and Lisa Chow.
Contains music by Diane Wong, Dan Powell, and Will Reed, and was engineered by Chris Wood.
That's it for The Daily.
I'm Natalie Kitro-Eff.
See you tomorrow.
