Henry Kirk, a cofounder of the software development company Studio Init, wants to hire the best engineers. That's why he asked job applicants not to use generative AI in the first technical coding part of their interviews — with the promise that they'd be able to show off their combined engineering and AI skills in a later section. "They still cheated," he tells me.

"It was so obvious," Kirk says. The coding tests took place in tandem with a video call, and some candidates frequently looked off to the side. They gave delayed answers or copied and pasted full blocks of code into the system instead of typing step-by-step. Some refused to share their screens or spouted off-topic answers to verbal questions, leading Kirk to believe they were reading large language model outputs verbatim without even thinking. "It's a waste of our time," he says. But even as AI is making a mess of technical screening tests, Kirk says he still thinks it has value. "I'm a small company. I have 400 applicants. How do I screen the people down to a manageable chunk of folks?"

Many software engineers aren't just allowed but are increasingly expected to use AI on the job. Companies like Google, Meta, and Salesforce increasingly rely on it for engineering tasks in the name of efficiency. But with gen AI now able to code as effectively as a junior engineer, bosses are wondering if the traditional coding tests — which have long been a staple of the hiring process — can still separate the good developers from the sloppy ones.

New tools keep popping up to make cheating on tests even more seamless: A since suspended Columbia University student, Chungin "Roy" Lee, recently created a tool called Interview Coder and used it to cheat on an Amazon coding test and then posted the interview to YouTube. He's selling the tool to other engineers for just $60 a month (he's claimed on X that he received and rejected an internship offer). Amazon has said candidates can be disqualified using gen AI unless explicitly permitted. The company did not comment specifically on Lee's test, but Margaret Callahan, an Amazon spokesperson, tells me that the company has job candidates acknowledge they won't use gen AI during the interview process when it's not permitted, but it does have them share their history of working with the tools when relevant. Google is also considering bringing some interviews back to in-person settings, where they can have more control over the environment. A Google spokesperson told me that applicants are informed before interviews that if they use AI during them they will be disqualified.

Recruiters and hiring managers I spoke to for this story said the mainstream adoption of ChatGPT led them to suspect that more job seekers are trying to cheat their way past the code tests. Companies are scrambling to change old evaluation processes for a new era. But as they push engineers to get more efficient with AI on one hand and wag their finger at its use with the other, they're raising new ethical questions about what really counts as cheating: Is an LLM an unfair edge, or just a coding partner?

The traditional coding interview is at a crossroads. But the end of the old interview might be welcome among engineers.

"Timed coding tests were never truly realistic; AI just pulled back the curtain," says Annie Lux, the founder and CEO of the coaching firm Land That Job. The interviews create pressure and penalize people who struggle in test environments, Lux says. And many employers now expect engineers to leverage AI tools at work — tests that ban them put job candidates in a different scenario than the one they would work in. A 2020 study by North Carolina State University and Microsoft found that people were better at solving coding problems when they weren't being watched closely and told to explain their work as they went — confirmation that some engineers performed worse when under the stressful conditions of a traditional technical interview. "These interviews reward test-taking over engineering," Lux says. "They ignore how software engineers actually work."

Andrej Karpathy, an Open AI cofounder, coined the term "vibe coding," a nod to the way AI will help engineers "just see stuff, say stuff, run stuff, and copy-paste stuff," and have it "mostly work," as he put it. An engineer's skill for writing code may become less impressive than their capacity to understand it. But the issue hiring managers tangle with now is how to balance the benefits of vibe coding with vetting the best engineers from large pools of applicants in a tight job market where there's a huge incentive to cheat your way into an offer letter. On the job, "hopefully, they are using AI, to do the stuff AI can do," says Don Jernigan, a vice president at Experis Services, an IT staffing firm. "We need to be testing and evaluating them from the areas between what a human can do and what AI can't do."

Kirk says a "perfect storm" bolstered cheaters: The tech job market tightened just as ChatGPT went mainstream. There were more applicants for fewer jobs and more people hoping a perfect score on a coding test would help them stand out. Now, it could hurt them in the long run. Kirk says he and his team have gotten more confident about catching cheaters, and will sometimes call them out and end the interview if they're sure they've found one. One applicant even admitted to it, and others have left the interview without argument, he says. And he is keeping a blacklist of people he suspects cheat in his interviews and plans to never consider them in the future. He already has a list of dozens of people he's sure tried to cheat, with hundreds more who raised suspicion. Now, his studio has applicants follow up their first test by coming on-site for more tests. "We're potentially paying you a lot of money and we need to make sure there's a good fit all around," he says.

ChatGPT didn't invent cheating. In the past, software engineering applicants would sometimes deputize a friend to spit out code in their place (either in a take-home test or, as one recruiter told me, actually sending someone else in their place to the interview round), and job seekers would share coding tests and answers online. If you search Reddit, TikTok, or Blind, you'll find people sharing tips and tricks to con an interviewer. But AI is a knowledgeable friend who's even easier to access. More people are using it to try to land any job by mass applying or sending AI-generated cover letters. Overwhelmed recruiters then use their own AI tools to try to sift through and find the best candidates. It's all creating a massive cog, with two different AIs talking to each other and both job seekers and hiring managers feeling frustrated.

When it came to engineers, recruiters and hiring managers started to notice something was amiss by early 2023. Job applicants completed coding tests with perfect answers, but when they moved on to interviews about the test, some knew little to nothing about the work they'd submitted. "Even with ChatGPT earlier versions, it could solve a lot of coding questions," says Yang Mou, the cofounder and CEO of the AI recruiting company Fonzi. "The thing that's even more insidious now is that the AI is also better at explaining the answers as if it was a human." Fonzi interviewed 1,270 candidates for a software engineering job between January and March, and flagged 23% of them "as likely to be using external tools," Mou says. The AI tool scans answers for awkwardly long pauses and evaluates phrases used to see the likelihood that they've been written by a chatbot, and then humans can listen back to the interview to see if they catch any red flags.

Two years ago, the technical interview company Karat flagged about 2% of interviewees as potential cheaters. Now, that proportion has jumped to 10% of interviewees. "It's happening more frequently," says Jeffrey Spector, the cofounder and president of Karat. "Ultimately, our belief is that interviews have to evolve." Karat is developing a new interview process that it hopes will better evaluate job seekers when they use LLMs, Spector tells me. "The LLM is becoming a core part of how engineers do their job. Preventing them from using the tools on their job seems very unnatural."

As AI becomes a bigger part of the job, Spector says, hiring managers — and humanity at large — have to ask the question: "How do you define cheating?" He says people shouldn't disregard explicit instructions not to use AI, but if most people are using it and you're not, you might be at a disadvantage in the interview process. Many applicants use books and online tips to study for coding interviews, and some use ChatGPT to practice for job interviews. When it comes to using AI in the actual test, Spector says, a tipping point will come where it feels too disadvantageous not to, particularly among young engineers who have learned and grown up in the LLM era — and the ethical questions will get messier.

Hadi Chami, the director of solution engineering at the software company Apryse, says he began to notice ways job candidates were using LLMs as he started to use them more in his own work. So he changed the job application last year. Now, he gives applicants who pass a first "vibe check" interview a take-home assignment, with the expectation they'll use AI. But he tells them they'll have to walk him through their work. That's helpful for now, as he can still see whether they know why something works, not just that it does work. But Chaim expects the problem to get worse: He says that he's concerned about young workers coming into the field. "They may have an overreliance on the tool. They'll be able to ace all their classes," but might struggle in the workplace, he says.

Maybe this isn't the interview apocalypse scenario it seems. "This is a little of a new frontier, which is maybe why there is so much fear and stress on both sides and people are just flailing," says Victoria Gates, the cofounder of the interview training firm Expert Interviews. "If you're investing your time and your money into finding out if candidates are cheating, you're wasting your time. The way interview processes are today, they are very unfair towards candidates. Of course they're going to try to find anything they can." Instead of trying to go full bad cop and employ tech to monitor cheating, Gates says companies should train interviewers to ask incisive follow-up questions and for specific examples that LLMs can't generate. Right now, companies may be focused on catching cheaters, but Ali Ansari, the founder and CEO of the AI interview company Micro1, says that will change. "I think coding in general is already looking extremely different," he says. "That implies even without the cheating, the coding test will have to start looking different." He predicts that there will be a "new norm" for coding interviews within the next year or two.

All this coding mess is evidence of the breakdown in trust between employers and workers. Job seekers are questioning how much free labor they owe someone who may not even extend them a second interview, and more bosses are doubting the integrity and work ethic of the people reporting to them. So much of the tech meant to make job searching easier and more accessible has just added noise to the process. Killing the old coding test and using something more creative in its place may be a small step in repairing that disconnect.

Amanda Hoover is a senior correspondent at Business Insider covering the tech industry. She writes about the biggest tech companies and trends.

• "Vibe coding" lets people build apps with AI — even if they have no coding experience.
• It's made software development accessible, and seasoned developers are using it to code faster.
• I tested it to see how well AI could bring an idea to life. It was impressive but far from perfect.

Silicon Valley is buzzing about "vibe coding," so I tested it to see what it's like to let AI build something from scratch.

Coined by OpenAI cofounder Andrej Karpathy, vibe coding describes giving instructions to an AI tool and letting it do all the work.

Or, as Karpathy puts it, "You fully give in to the vibes, embrace exponentials, and forget that the code even exists."

As someone with zero coding experience, I put the concept to the test and built an online game.

The process took about five hours, had some challenges, and created a less-than-perfect final product. But, it demonstrates how AI is making coding more accessible and making experienced software engineers more efficient.

Here's a step-by-step of how it went — and what I learned along the way.

(Or, click here to try the game first.)

I kept my initial prompt fairly simple

There are lots of vibe coding tools out there, such as Cursor and ChatGPT. I used Replit Agent, which lets you build apps with prompts in plain English.

I began by describing the type of game I wanted to build: a Mario Kart-style game in which Sam Altman races Elon Musk to achieve artificial general intelligence, or AGI.

You can just build stuff

The agent suggested a plan for the game and gave options in boxes to check and approve. The plan included building the initial prototype, adding multiplayer functionality, implementing additional power-ups and obstacles, creating additional race tracks with different themes, and adding character special abilities.

I clicked to approve the plan, and it started writing code in real time on the right-hand side of the screen.

It was off to a good start

Its first attempt looked promising. In just a few minutes, it created a main menu screen and added a text box explaining what the game was. It also decided to call the game "AGI race."

The vibes are vibing

I liked that it created a second page for players to choose their character and displayed their statistics, including speed and acceleration.

Welp! The game doesn't actually work.

But when I clicked "start race," the game didn't actually work. Instead, it displayed a weird error message that I didn't understand. I told the agent it wasn't working and copied and pasted the error message so it could fix it.

I also instructed it to make the main menu page more aesthetically pleasing and add 3D animations for the characters. It spent a few minutes making those changes before asking me to check if the game was running properly.

We're progressing

The main menu page looked better with the 3D characters, although it initially turned Musk and Altman into people of color. However, I realized I was too focused on how it looked and decided to get it to function properly before getting it to tweak the user interface. It made those changes and then asked if everything was loading and displaying correctly.

I'm spinning like a ballerina

This time, it displayed a race track, but the game was not working. Using the keyboard shortcuts just made the cars and characters spin in circles — they were impossible to control.

We're regressing

After telling the Replit agent to fix the spinning issue and make the game functional, it changed the page completely to a simplified race track, but this time nothing worked.

I told it to revert to the previous version but fix the steering and add coins the characters could collect that would give them a speed boost.

We're back on track

It ran a whole bunch of checks and displayed its "thinking" while it reviewed its code to fix the bugs. That seemed to do the trick, as the controls were now working, and I could accelerate without spinning.

Next, I told it to make some tweaks, like labeling the coins "compute resources," "hire top talent," and "funding from investors." Also, I told it to have a visual effect for when the characters collect coins to show that it gives them a speed boost and another for when an opponent is hit with a lawsuit to show that it slows them down.

Issues persist

After about 20 minutes, the agent said it enhanced the game with several key improvements and told me to try the game again.

However, that made the game experience deteriorate, and the screen started blurring when I tried to play.

I encountered this issue a lot throughout my experience using Replit to create the game: I would instruct it to fix one problem, and it would lead to new ones. Much of my time was spent waiting for it to review its code and identify the errors. That was really frustrating.

I suspect that if I had coding experience, I might have been able to identify the root causes and fix them faster.

Eureka! It works (for the most part)

After some debugging, the game was mostly functional. However, the visual elements fell short of my expectations. I wanted the race track to resemble an actual track and feature a Silicon Valley-themed background, but I hesitated to request further major changes, fearing it might break again and cause more headaches.

Ready to launch

For the final touches, I instructed the agent to change the game's title to "The Final Boss of AGI" and make the main menu slightly more exciting.

This is what it came up with — try out the game for yourself here.

My vibe coding takeaways

Overall, I found the experience impressive. I fully gave in to the vibes, letting the AI agent take the wheel, as it's clearly a whole lot more competent than I am at coding and app development. As Karpathy put it, "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works."

Here are my main takeaways:

• You can just build stuff. The tool makes it easier than ever to bring ideas to life and has lowered the barrier to app development.
• Keep it simple. Asking for visual improvements or other tweaks can unexpectedly break other aspects of the projects. Sticking to straightforward prompts is the best approach to making something quickly unless you're prepared to spend more time debugging or refining the app.
• Coding knowledge still matters. While the AI is powerful, it's not flawless at troubleshooting. To get everything working exactly as intended, you need coding experience to diagnose and fix issues.
• It's great for low-risk projects. As Karpathy said, "It's not too bad for throwaway weekend projects, but still quite amusing." That sentiment perfectly sums up my experience: it's a fantastic tool for experimentation but probably not quite ready for high-stakes development.

• Andrew Ng has launched a course teaching vibe coding, the latest Silicon Valley craze.
• The Stanford professor has introduced a "vibe coding 101" course with AI company Replit.
• Ng said that asking AI tools to do "everything in one shot usually does not work."

Want to get into vibe coding? There's a course now that teaches you how.

Andrew Ng, the Stanford professor and former Google Brain scientist, has launched a "Vibe Coding 101" short course for newbies who want to learn how to use generative AI tools to write and manage code.

Vibe coding, a term coined in February by OpenAI cofounder Andrej Karpathy, describes how software development is being increasingly automated by AI agents that are given prompts by humans.

Seasoned software engineers are using AI to speed up their work, but it's also attracting people without much coding experience.

Ng, a computer science veteran, wrote on X that while he believes coding agents are "changing how we write code," making vibe coding work in practice takes a bit of work.

"Contrary to popular belief, effectively coding this way isn't done by just prompting, accepting all recommendations, and hoping for the best," he wrote, noting that using AI to write effective code requires a more refined process.

He said that he codes "frequently" using large language models, or LLMs, adding that "asking an LLM to do everything in one shot usually does not work."

It's why Ng's course — a 94-minute video series built in collaboration with AI agent company Replit — aims to give participants a beginner's look at how AI tools can be used most effectively to vibe code.

Ng said the course, taught by Replit president Michele Catasta and the head of its developer relations, Matt Palmer, would teach aspiring vibe coders how to build and deploy web applications with an AI agent.

Ng added that key principles would be taught to accomplish this, such as "giving agents one task at a time, making prompts specific," and offering a clear sense of how to approach debugging — the process developers go through to identify mistakes in their code.

Other course elements include teaching participants how an AI tool like Replit can be used to automate key portions of the software development process, such as building a prototype of an app or tool.

"By the end of this course, you'll have a solid foundation in building with coding agents, and a process you can use to keep vibe coding effectively," Ng said.

• The vibe coding era is here. Will you survive it?
• It's a question software engineers are increasingly asking as AI automates more of their work.
• Software experts suggested embracing AI tools, improving prompts, and mastering clear thinking to stay ahead.

"Vibe coding" has software engineers wondering if AI could put them out of a job. But there are ways to survive in an era when AI is so smart it can translate "vibes" into fully fledged lines of code.

Since computer scientist and OpenAI cofounder Andrej Karpathy coined the term last month, the world of software development has rallied around the idea of using generative AI tools to automate vast portions of coding workloads.

In Karpathy's telling, "vibe coding" means he can "just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works."

More people are now using AI for tasks that would previously require deep technical skills, such as building apps or creating a video game.

Mark Zuckerberg, for instance, said earlier this year that he expects Meta to have "AI that can effectively be a sort of midlevel engineer."

Comments like this have come as job opportunities for software engineers — a historically well-paid job — have plummeted. On Indeed, openings in the US are down a third from five years ago.

It raises a big question: how can software engineers not just survive but thrive in the vibe coding era?

Business Insider spoke to software engineers who said vibe coding doesn't have to mean total doom for career prospects — so long as they prioritize these three skills.

1. Embrace the vibes

Embracing and mastering vibe coding tools is a good place to start.

ChatGPT, Cursor, Replit, and Windsurf are just some of the tools being used to fast-track workloads.

Marc Tuscher, chief technology officer at AI robotics firm Sereact, told BI that vibe coding tools make him "much faster," which is why he encourages his team of roughly 25 software engineers to use them.

Though Tuscher acknowledges these tools can make mistakes that need to be debugged, he says a software engineer working with AI tools will have an advantage over one that doesn't. "The amount of speed you can get compared to coding yourself is crazy," he said.

It's part of the reason industry leaders have become more vocal about putting these tools to use. Sam Altman, the CEO of OpenAI, said in an interview with Stratechery's Ben Thompson this month that for students, "the obvious tactical thing is just get really good at using AI tools."

Tanay Kothari, CEO of Wispr Flow, a voice agent that can be used to input commands into vibe coding tools like Cursor, understands that there might be some reluctance among veteran developers to embrace such tools.

In his experience, however, engineers who think they're "just really good" will start seeing vibe coding tools as an "unlock" once they understand the efficiency gains at hand. "People who use AI tools roughly get twice as much done," he told BI.

2. Up your prompt game

Learning how to give instructions — known as prompts — to vibe coding tools in a smarter way can also give engineers an edge.

"The hottest new programming language is English," Karpathy quipped in January 2023, just a few months after ChatGPT's first public release.

Research papers, such as the original paper behind OpenAI's GPT-3, suggest that the large language models powering today's generative AI tools can, as Karpathy put it, be "'programmed' inside the prompt."

In other words, a carefully constructed prompt can drastically alter the quality of output for an AI system.

It's why Ash Edwards, a former Palantir engineer turned CEO of AI agent company Fern Labs, sees software engineers getting more value out of vibe coding tools by being more "prescriptive."

"A real failure case is almost letting the code choose its own directions," he told BI. "I think you can usually get much better results if you know what you want to do and you know roughly how you want to build it."

It's a view shared by Kothari: "There is a right set of questions to ask ChatGPT to get your work done."

3. A clear-thinking mindset

When AI is doing more of the grunt work of software development, what should engineers focus their newfound extra time on?

Software experts who spoke to BI said the answer was simple: focus on thinking clearly.

When software engineers build programs and apps, they spend a lot of time thinking carefully about what exactly is being built and how it might be stitched together in a wider system. This all requires careful thinking before lines of code are even written.

One type of thinking that helps this process is reasoning from "first principles," according to Wispr Flow's Kothari, referring to breaking a complex problem into smaller pieces to find a suitable solution.

"That is a very hard, high-level problem that you need to be able to solve regardless of how good AI gets," he said.

Kothari notes that this is a key attribute of what defines a "10x engineer" — an industry archetype of an engineer who is typically 10 times more productive and valuable than their counterparts. "10x things come from clarity of thought," he added.

Thinking about systems at large is another helpful way to approach software engineering problems. Kothari notes that people can waste a lot of time not doing this, as it can lead to building "the wrong thing that nobody really wanted."

Fern Labs' Edwards agrees, noting that engineers are "always going to need to be able to think about systems" — despite the growing use of vibe coding tools — as they will fundamentally still be the ones needing to oversee a program.

"Having that software engineering mindset, understanding how to think about problems and break them down and think about them in abstraction — I think that's always going to be useful and important, if not more so, as these things get more and more powerful," he said.

• Cisco's Liz Centoni told BI she still recommends learning to code as a core skill in tech.
• Centoni said a foundation in coding can help solve the so-called "boring problems" the industry faces.
• She said a comp-sci degree isn't a must, but creative thinking and business knowledge are also key.

As AI increasingly takes over coding tasks, some industry veterans have started to prioritize other skills. However, Cisco executive Liz Centoni says the advice "learn to code" is still relevant in tech today.

"There's some foundational elements that help train the way you think around how to solve for problems," the executive vice president and chief customer experience officer said in an interview with Business Insider.

Centoni, who was a software engineer early in her career and made several transitions during her nearly 25 years at Cisco, said her coding skills have helped her "decompose" the everyday "boring problems." She said one of her strengths is being able to look at a problem, break it down, bring in data, make a hypothesis, and test for it.

"And maybe nobody wants to talk about the boring problems, because it doesn't make the news," Centoni said. "But the reality is that we have boring problems that have lots of data underneath it. It has thousands of employees who work at it."

That doesn't mean you need to get a degree in computer science, Centoni said. In fact, Centoni said that foundational concepts like coding need "brought forward" to address today's challenges because the world isn't just solving linear computer science problems today.

"We're solving for, you know, where it can be structured, unstructured data as well, where we need a tool, where we need an agent to continue to learn and think for itself as well," Centoni said.

The Cisco exec said that thousands of employees work on those types of "boring problems," and that includes a lot of repeatable workflows. In tackling those kinds of problems, employees need to know how to look at the available technology, whether it's machine learning or generative AI, and know when and where to apply it. Centoni said that "hands-on knowledge" in those scenarios is "super valuable."

"There's some basic concepts you need to understand," Centoni said, adding that she's looking for people who can solve them with "knowledge of the tools that are out there."

Centoni added that coding isn't the only valuable skill — being able to think creatively is also vital. That requires learning more than simply educational background, she said.

"I want someone in there who's sitting with the subject matter experts who can not just understand the problem, but look at how can we creatively craft a solution," Centoni said.

Additionally, she said understanding business problems is also a key skill so that employees can connect business use cases with technology to create efficient solutions.

"We have 20,000 people in our organization today," Centoni said. "How do we reduce the friction in their lives? And so I think somebody who can connect both a business use case and the technology, that's very valuable to me as well."

• Sam Altman says students should focus on getting "really good at using AI tools."
• AI is already automating over 50% of coding at some companies, Altman told Stratechery's Ben Thompson.
• The OpenAI CEO said students should prioritize the "meta ability to learn" over learning specific things.

Sam Altman is ramping up his advice for students preparing to enter a workforce that's increasingly automating coding: master AI tools.

"The obvious tactical thing is just get really good at using AI tools," Altman said in a Thursday interview with Stratechery's Ben Thompson. "Like when I was graduating as a senior from high school, the obvious tactical thing was get really good at coding. And this is the new version of that."

Using AI to write code has become a major topic among tech executives.

Dario Amodei, the CEO of Anthropic, predicted last week that AI would write 90% of code within six months and "essentially all of the code" in a year. Kevin Weil, OpenAI's chief product officer, predicted this week that AI will become better than humans at coding by the end of the year.

With the conversation about AI replacing human software engineers ramping up — especially as more people embrace vibe coding — Altman has doubled down on advice he gave in a September interview that mastering AI tools is a way to future-proof your career.

"I think in many companies, it's probably past 50% now," Altman said in the Stratechery interview, referring to the amount of coding that's done by AI. "But the big thing I think will come with agentic coding, which no one's doing for real yet."

When asked what was holding it back, Altman said: "Oh, we just need a little longer," adding that it was a model problem, not a product issue.

For students preparing for careers, Altman advised cultivating "resilience and adaptability" and the "meta ability to learn" over learning specific things.

"Whatever specific thing you're going to learn, like learn these general skills that seem like they're going to be important as the world goes through this transition," he said.

While software engineers are still in demand, Altman predicted that won't always be the case.

"My basic assumption is that each software engineer will just do much, much more for a while. And then at some point, yeah, maybe we do need less software engineers," he said, referring to OpenAI's hiring strategy.

He also predicted that AI-driven job displacement won't happen all at once but will accelerate over time.

"It kind of just seeps through the economy and mostly kind of like eats things little by little and then faster and faster," Altman said.

• Y Combinator CEO Garry Tan said "vibe coding" is letting startups stay leaner.
• He said coding with AI allows smaller teams to do more heavy lifting in an interview with CNBC.
• He suggested young engineers struggling with the job market should "vibe code" and build startups.

The CEO of Silicon Valley's most famous incubator thinks "vibe coding" is set to transform the startup landscape.

"I mean, the wild thing is people are getting to a million dollars to 10 million dollars a year revenue with under 10 people, and that's really never happened before in early stage venture," Garry Tan, CEO and president of Y Combinator, said in an interview with CNBC.

Tan said it's due, in part, to vibe coding — Silicon Valley's favorite new buzzword, coined by OpenAI cofounder Andrej Karpathy in a post on X in February.

But what, exactly, does vibe coding mean?

"You can just talk to the large language models and they will code entire apps," Tan said. "And if it doesn't do — if there's a bug, or if you want it to change, or you want it to look a different way, you don't have to, you know, go in there and write the code yourself."

In essence — Tan seems to define it as an increasing reliance on artificial intelligence to do the bulk of the heavy lifting during the programming process. He said vibe coding is dialing up the speed at which startups can develop relevant software.

"You don't have to hire someone to do it, you just talk directly to the large language model that wrote it and it'll fix it for you," he said. "And sometimes you could just accept all changes without even looking at the changes it made, because it's that good now."

Tan says vibe coding makes the process of building software overall more efficient, and the current batch of startups that Y Combinator is incubating — which Tan says is composed of about "81%" AI companies themselves — is taking full advantage.

"This is the first time that's ever happened, and about 25% of the batch — 95% of their code was written by large language models," Tan said.

The sheer power of LLMs is allowing startups to stay leaner, Tan said. What would've once taken "50 or 100" engineers to build, he believes can now be accomplished by a team of 10, "when they are fully vibe coders."

"When they are actually really, really good at using the cutting edge tools for code gen today, like Cursor or Windsurf, they will literally do the work of 10 or 100 engineers in the course of a single day," he said.

Vibe coding does have its drawbacks, Tan caveated in an episode of Y Combinator's Lightcone Podcast earlier this month. In particular, a survey of Y Combinator's current batch of founders indicated LLMs are weak at actually debugging the code they churn out.

"The humans have to do the debugging, still. They have to figure out well, 'What is the code actually doing?'" he said, adding, "There doesn't seem to be a way to just tell it, 'debug.'"

Still, to Tan, the benefits are plentiful. Among them — that investing time and money into building niche software is more justifiable, given the new speed at which AI makes it possible to code.

"I think the other thing that it's going to do in the industry is that there are all these pieces of software that normally you could never write software for, like industries that before people would say, 'Oh it's too small a market,'" he said.

All of those narrower markets, he added, can now support business that gross a "hundred million" yearly, while still being run by small teams.

"This is really the good news," Tan said.

For those struggling to break into a shrinking job market, Tan says the advent of vibe coding has arrived at the "perfect time." It gives young engineers the opportunity to strike out on their own rather than relying on big-name companies to kickstart their careers.

"You know, maybe it's that engineer who couldn't get a job at Meta or Google, who actually can build a standalone business making 10 or 100 million dollars a year with 10 people," Tan said. "Like that's such a powerful moment in software."

• At HumanX in Las Vegas tech leaders talked about on AI's impact on business.
• Speakers like Kamala Harris emphasized collaboration between the tech industry and government.
• AI enthusiasm ran high but many attendees were focused on practical ways to drive company revenue.

This week, I attended HumanX, where thousands of startup founders, investors, and tech industry executives gathered in Las Vegas to discuss AI's value proposition to their bottom lines — and the future.

The nearly four-day conference opened with remarks from former Vice President Kamala Harris. Harris, who served as "AI czar" during the Biden Administration, was one among a handful of lawmakers at the event who called for stronger cooperation between the tech industry and the federal government.

In the following days, attendees sat in on panels with veteran investors like Vinod Khosla and Tim Draper alongside rising stars of the AI boom like Hugging Face's Thomas Wolf and Mistral's Arthur Mensch. There was ample time to swerve between the booths on the convention floor where tech giants and startups alike had reps ready to woo customers.

AI conferences can feel a bit like a dance. People talk with excitement about the transformative power of this technology, but it's also clear they're gauging where the real value lies.

Here are my top takeaways:

1. Patience is a virtue when it comes to ROI

Conference speakers had some advice on making long-term gains with AI.

Glean CEO Arvind Jain cautioned companies to establish an "AI team" to create a roadmap for integrating the technology. "Make small bets. Don't try to actually immediately focus on ROI," he said. "Focus on education first, like, make your workforce trained with AI."

In the sales landscape, Conviction's Sarah Guo said companies have already made "high-fructose corn syrup" gains, reveling in their capacity to send mass emails or spam customers with calls.

"That actually drove a lot of traction," she said. However, "real customers actually churned off that pretty quickly because it doesn't serve a real need. Nobody wants the spam — which is what it is."

Still, it got people to start "buying the products" and she expects that to continue. "That's what I think we're going to see this year in terms of those real ROI use cases."

2. Vibes and valuations

AI startups are fetching wild valuations, but the question is whether they're caught in a bubble that's set to burst.

Investors are essentially "taking a bet" on the future cash flow of a business, said Tuhin Srivastava, CEO and cofounder of AI inference company Baseten. "I'm a  beneficiary of that," he added.

Srivastava pointed to Anysphere, which makes the coding assistant Cursor, and was in talks to raise funding at a valuation of close to $10 billion earlier this month, after hitting $100 million in annualized recurring revenue in a span of 12 months, Bloomberg reported. Coding startup Codeium, is raising funding at a $2.85 billion valuation after hitting $40 million in ARR, TechCrunch reported.

At face value, these seem like massive valuations, but Srivastava says he thinks they're reasonable in the context of the venture landscape just a few years ago.

"Companies in 2021, during all that craze were like, you know, a million dollars of ARR, raising a billion dollars. Think about these AI companies, they all have a lot of revenue," he said.

He joked that valuations — to some extent — are being calculated on "vibes" but he believes they're grounded in real growth potential. "You know, we're probably at 0.1 percent adoption of AI in the enterprise. So we have a thousand x upside there."

For companies with little to no revenue, talent may be the value metric.

A few investors told BI they heard, to their surprise, that companies acquiring AI startups for a strategic advantage were valuing them on a "per engineer" basis.

If you have more thoughts on talent-based valuations contact me at [email protected].

3. Here's at least one new job we'll see soon

Aside from prompt engineers and a beleaguered new set of "AI heads" at companies, many are still wondering what "new jobs" this technological shift will bring.

As AI makes it easier to code, and "vibe-code," we should expect to see a new wave of highly compensated "product engineers," Andrew Filev, CEO of AI coding platform Zencoder said. These are individuals who are versed in product management, a role that typically oversees new products from ideation to launch, and software engineering, which deals with the technical details of new products.

4. Governance is a pain point

AI governance is still a bit of a murky term.

"There's a little bit of a conflation of governance and regulation," said Navrina Singh, founder and CEO of AI governance platform Credo.

Singh told BI she defines governance with three questions:

1. Do you have a handle or understanding of risk?
2. How do you actually mitigate the risk of these technologies?
3. How do you future-proof your AI investments for potential policy changes — not only at the company level, but at the government level?

The most misinformed opinion companies have is that governance will slow them down in adopting AI. Credo's data has shown the opposite. " We are finding that companies are getting better ROI," she said.

Dataiku's CEO Florian Douetteau observed a similar anxiety around governance at an "executive field trip" the company held for the CEOs of its customer base in September 2024. The company further investigated CEOs' top AI anxieties and published them in a survey. Governance ranked high on the list with 94% of about 500 CEOs surveyed saying they suspect employees are secretly using generative AI tools without official approval.

5. More people are sounding the alarm on deflation

Silicon Valley leaders from Khosla to Sam Altman have expressed concerns that AI will spur deflation. Those fears were echoed by a handful of attendees at HumanX.

To be clear, the US is still experiencing inflation with the consumer price index rising about 2.8% for the twelve month period ending in February 2025, according to the US Bureau of Labor Statistics. However, technological shifts are often correlated with deflation, as they boost productivity and lower production costs.

"There is no denying that AI-based technologies are evolving rapidly and being adopted by people and enterprises," said Steve Berg, a partner at Lytical Ventures. "There are inflationary impacts happening as well, which offset the deflationary impacts of technology, but what happens when one side or the other becomes dominant?"

As businesses race to replace humans with AI agents, coding assistant Cursor just gave a peek at the attitude bots could bring to work, too. 

© 2024 TechCrunch. All rights reserved. For personal use only.

On Saturday, a developer using Cursor AI for a racing game project hit an unexpected roadblock when the programming assistant abruptly refused to continue generating code, instead offering some unsolicited career advice.

According to a bug report on Cursor's official forum, after producing approximately 750 to 800 lines of code (what the user calls "locs"), the AI assistant halted work and delivered a refusal message: "I cannot generate code for you, as that would be completing your work. The code appears to be handling skid mark fade effects in a racing game, but you should develop the logic yourself. This ensures you understand the system and can maintain it properly."

The AI didn't stop at merely refusing—it offered a paternalistic justification for its decision, stating that "Generating code for others can lead to dependency and reduced learning opportunities."

Read full article

Comments

To many people, coding is about precision. It's about telling a computer what to do and having the computer perform those actions exactly, precisely, and repeatedly. With the rise of AI tools like ChatGPT, it's now possible for someone to describe a program in English and have the AI model translate it into working code without ever understanding how the code works. Former OpenAI researcher Andrej Karpathy recently gave this practice a name—"vibe coding"—and it's gaining traction in tech circles.

The technique, enabled by large language models (LLMs) from companies like OpenAI and Anthropic, has attracted attention for potentially lowering the barrier to entry for software creation. But questions remain about whether the approach can reliably produce code suitable for real-world applications, even as tools like Cursor Composer, GitHub Copilot, and Replit Agent make the process increasingly accessible to non-programmers.

Instead of being about control and precision, vibe coding is all about surrendering to the flow. On February 2, Karpathy introduced the term in a post on X, writing, "There's a new kind of coding I call 'vibe coding,' where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." He described the process in deliberately casual terms: "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works."

Read full article

Comments

• Job openings for software engineers on Indeed are at their lowest level in five years.
• Marc Benioff said Salesforce might not hire software engineers in 2025 because of gains from AI.
• Despite AI's impact, demand for skilled tech workers is expected to grow significantly by 2033.

If you're a coder, you already know: There just aren't as many jobs as there used to be.

Openings for software engineers in the US on Indeed are down by more than one-third from five years ago.

For many engineers, the drop-off likely feels even steeper. Job postings are well off levels seen during the pandemic, when the industry was awash in openings.

In early to mid-2022, there were three times as many software engineering roles listed on Indeed.

Artificial intelligence is surely one cause. The same technology that can make coders more productive appears to be undercutting hiring demand.

Salesforce CEO Marc Benioff recently said the tech giant might not hire any engineers in 2025 because AI tools allowed Salesforce engineers to do so much more.

"We have seen such incredible productivity gains because of the agents that work side by side with our engineers," Benioff said on "The Logan Bartlett Show" in January.

'Am I going to lose my job?'

Productivity gains are great news for coders with jobs, but the AI boost can be worrisome for job seekers.

A year ago, the startup Cognition Labs released what it said was the first AI software engineer. The company, backed by Peter Thiel's venture capital fund, caused a stir with its announcement.

"There was a lot of panic," Jesal Gadhia, the head of engineering at Thoughtful AI, which creates AI tools for healthcare providers, previously told Business Insider. "I had a lot of friends of mine who messaged me and said, 'Hey, am I going to lose my job?'"

He worried that even though coders are still needed in many areas, AI could step into roles that have long served as training grounds for junior engineers.

"Junior engineers," Gadhia said, "have a little bit of a target behind their back."

Demand is likely to grow

It's not all bad news for those with strong technical chops. In late August, the Bureau of Labor Statistics forecast that demand for software developers, quality-assurance analysts, and testers would grow by 17% from 2023 to 2033. The agency noted that that'd be far faster than growth in the overall job market.

Lighthouse Labs, a Canadian company offering coding boot camps, argued in January that global demand for skilled workers in data analytics, cybersecurity, and cloud computing was outstripping supply. Unsurprisingly, some of the biggest gains involve AI. Lighthouse Labs said that was partly because there's demand well beyond the tech industry for people with skills in areas like machine learning.

AI to do more of the work

Regardless of industry, demands on coders' time will likely continue to evolve because of AI. GitLab has said developers spend only about a quarter of their time coding.

Madars Biss, a tech writer and front-end developer, previously told BI that he expected developers to spend less time generating code and more time managing AI-generated code.

AI tools, Biss said, could "handle much of the routine and repetitive tasks of the developer," while "humans focus on managing, double-checking, and creativity."

As with Salesforce, that could change how companies hire software engineers.

In June, Amazon Web Services' chief, Matt Garman, predicted that AI would handle a good deal of coders' work.

"If you go forward 24 months from now, or some amount of time — I can't exactly predict where it is — it's possible that most developers are not coding," Garman said.

In 2023, not long after ChatGPT made generative AI mainstream, a poll on the anonymous workplace forum Blind asked, bluntly, whether young software engineers were "fucked." Some 42% of the more than 13,000 respondents picked the response "Yes? U guys are pretty much fucked."

This past October, Sundar Pichai proudly announced on an earnings call that AI was writing more than 25% of new code at Google. Mark Zuckerberg has said that Meta will build an AI engineer to write code. Salesforce CEO Marc Benioff announced a hiring freeze for engineers in 2025, saying AI had increased productivity by 30% — and then news broke that Salesforce planned to lay off 1,000 workers. (It's still hiring salespeople for AI-powered products.) Stripe intends to cut some engineers while also growing its overall head count this year.

All of this raises the question of what junior engineers will take on if some basic tasks become automated. Some product managers have speculated that AI will increasingly take on some technical coding tasks and circumvent their need for engineers. Overall, job postings for software engineers on Indeed are at a five-year low.

Are engineers really coding themselves into obsolescence?

AI is knocking down the career ladder by doing more of the coding work of entry-level engineers, but, at least for now, the increased coding output from AI is also increasing the demand for and value of experienced, creative developers to interpret and put the AI's work to good use.

While many obituaries have been written to mourn the death of coding, engineering is more than writing code: It requires creative thinking to solve problems and expertise to read code. As it is now, AI isn't an original thinker.

"AI can't support what it doesn't know," says James Stanger, the chief technology evangelist at CompTIA, a nonprofit trade association for the US IT industry. "I still don't think that it is something that can fully replace a good developer." He adds, though, that "if a developer is not creative, then you can replace them very easily."

oftware engineering has been around since the 1960s, but hiring boomed in the '90s with the dot-com era. Coding boot camps became common in the 2010s as the demand for engineers outpaced the supply. According to the Bureau of Labor Statistics, 1.9 million people worked as software developers, quality-assurance analysts, and testers in 2023. The bureau projected that the industry would grow by 17% from 2023 to 2033, outpacing the national average of 4% for all jobs.

An analysis from CompTIA found that the rate of job postings for software engineers fell by 50% from January to December 2023, recovering slightly by the end of 2024. While posts for jobs across tech, finance and accounting, and marketing, communications, and creative roles all fell as well, the dip for software engineers was the sharpest. But CompTIA says the reason for the fall wasn't clear. Tech companies — including some that acknowledged overhiring during the pandemic — laid off thousands of workers in 2022 and 2023, with many citing economic uncertainty.

But there's a widening divide within software engineering regarding experience level. CompTIA found that the proportion of open software engineering roles seeking entry-level workers had dropped since January 2023, to just over 20% from nearly 30%, while job postings for those with seven years of experience or more increased to make up nearly 40% of the open roles, up from just over 30%.

The experienced engineers I talked to seemed confident that AI wouldn't come for the jobs anytime soon.

Jeremy Chua, a software engineer for the AI Lab at the venture firm Georgian, turns to chatbots when he hits issues with coding. He may prompt ChatGPT or Claude to cull answers from the depths of Google and Stack Overflow, a Q&A site for programmers, or to help him write in coding languages he's less fluent in. Chua, who has more than a decade of experience, says he was skeptical about whether gen AI could help him at work. He says that now he can sometimes complete projects that would have taken a week in a day or two, and he thinks of the chatbots he uses as coding partners. "It's not like it will replace me — it augments the way that I work," Chua tells me.

Caleb Tonkinson, an engineer at a clinical AI company called SmarterDx, tells me that AI is changing programming through two paths: "I can deliver the same thing faster, or I can deliver something better in the same period of time." He views AI as similar to other tech tools that became available to engineers — except more exciting as it advances rapidly. "There have been tons of tools for 20 years now" to debug software, generate code, or evaluate code, he says. "Your best companies and best software engineers are almost always leveraging those tools."

Cody Stewart, a principal software engineer at the software company CallRail, says he doesn't use gen AI for everything at work but might use it to get answers to "stupid questions" that he could spend a long time looking for on Google or Stack Overflow. He began using chatbots at work in 2022. "I read something that was like, you either learn to adopt new tools and figure out how they can enhance your day-to-day life and you stay with the times, or people are going to outlevel you," he said. "I saw that and thought I should probably give this a shot."

While more-experienced engineers are optimistic about AI, young engineers have more reason to worry.

The startup Cognition AI last year widely released an AI-powered software engineer called Devin, designed to work on bugs and small feature requests. In a December video, it described it as "a junior engineer" who "works best with a great manager." Cognition AI and its CEO, Scott Wu, did not respond to questions about whether it's meant to replace engineers or reduce the number the companies need.

Jayesh Govindarajan, a Salesforce executive vice president focused on AI, told my colleague Ana Altchek that the company was building "a system that can pretty much solve anything for you" but "just doesn't know what to solve," making knowing how to code less important. "I may be in the minority here," Govindarajan said, "but I think something that's far more essential than learning how to code is having agency."

Alexander Petros, a freelance open-source software engineer, is an AI holdout; he tells me he doesn't use generative AI to code. "I do worry that because AI is in many ways doing things that you used to hire junior developers to do, it does remove the ladder upon which junior developers would try to do those things, make those mistakes, and then learn," he says. Petros says he tried ChatGPT but found that the code could be clunky. If something in that code breaks, humans may not know how to fix it. "The process of producing code with LLMs, for the foreseeable future, is almost entirely distinct from the process of producing good software systems that last for a very long time," he says. Plus, using AI to solve problems means he may not learn how to get through those roadblocks on his own.

Chatbots lack creativity — that's where engineers, especially those who have been doing the job for a while, have an in-demand advantage. Stanger says he hopes companies use AI not as justification for cutting back on engineers but as a way to help them "get deeper into this code and get more creative."

Stanger says that treating engineers as a faucet that can be turned on and off as a business needs, or even replaced with AI, is likely to backfire in the long term. "If you've got toxic companies that are interested in that binge-and-purge, on-and-off hiring of developers, I'm not sure they're going to create very good products," he says.

People have long panicked that technology will take their livelihood. But even as automation eliminates some jobs, tech often creates a demand for new roles; most people today are working jobs that didn't exist before 1940.

The wholesale elimination of software engineers likely won't come to fruition in the near future, but the picture for more-experienced engineers is brighter. In the best-case scenario, AI will mean they get more time to flex their muscles and solve deep problems.

Amanda Hoover is a senior correspondent at Business Insider covering the tech industry. She writes about the biggest tech companies and trends.

• Salesforce's AI executive said having agency is "far more important" than learning to code.
• Jayesh Govindarajan defined agency as seeking out a problem and having the drive to solve it.
• Mark Zuckerberg similarly said he valued the ability to "go deep and do one thing really well."

For years, "learn to code" was the go-to advice for anyone wanting to break into a tech career — but Salesforce's head of AI engineering says another skill is more valuable these days.

"I may be in the minority here, but I think something that's far more essential than learning how to code is having agency," Jayesh Govindarajan, an executive vice president, said in an interview with Business Insider.

Govindarajan said that's because Salesforce is building "a system that can pretty much solve anything for you" but "just doesn't know what to solve."

"I think far more important than knowing how to code is having that agency and that drive to go get it built out," Govindarajan said.

The AI exec gave a hypothetical example of someone trying to solve a problem for College Possible, a nonprofit that helps students prepare for college and receives funding from Salesforce. Govindarajan said that someone could interview a counselor, see what they do on a daily basis, and then use an agentic AI system to "describe what you're trying to build and it'll give you a first draft of the solution." While that first draft may not be perfect, "you go take it to this counselor, have them play with it," and listen to their feedback and any critique, he said.

"Then you'd come back and you tweak it again. No code. You'd give it instructions in English. That's very possible," Govindarajan said.

The Salesforce exec said someone who has gone through this process has demonstrated two key things.

"One, agency to go seek out a problem to solve," Govindarajan said. "And two, learn the tool set — that's a no-code tool set or a low-code tool set — to be able to go get the job done."

In that hypothetical example, once the counselor is interested in actually buying the proposed solution, a more experienced coder could then be brought in to sharpen up the edges and fine-tune the software product.

Govindarajan's remarks offer a look into how the world of software development and sales is evolving in the age of AI.

Since the emergence of AI tools such as GitHub Copilot or Amazon CodeWhisperer, several coding tasks have been automated, creating uncertainty in a once-stable industry and new challenges for some younger entrants looking to break into software engineering.

During Google's third-quarter earnings call in October, CEO Sundar Pichai said more than a quarter of new code at the company was generated by AI, though it was still reviewed and accepted by employees. Other tech giants have similarly integrated AI into coding tasks. One Microsoft manager told BI that AI helped him cut down the time he spent on coding tasks by about 70%.

Even as coding becomes increasingly automated, some industry leaders believe learning the basics continues to be necessary, arguing that it's more important than ever to understand the fundamentals of technology in order to build on it.

Other industry leaders seem to be leaning in the direction that soft skills could be what sets candidates apart.

Mark Zuckerberg said in a July interview with Bloomberg that he believed the most important skill was "learning how to think critically and learning values when you're young."

"If people have shown that they can go deep and do one thing really well, then they've probably gained experience in, like, the art of learning something," Zuckerberg said in the interview, discussing what he looked for in job candidates.

The CEO said that skill applies to situations that could arise during a career at Meta, and it's key to showcase your ability to dive deep and master whatever you're working on.

Or, as Govindarajan may put it: using the tools at your disposal to get stuff done.

• "Vibe coding" is Silicon Valley's latest buzzword, coined by OpenAI cofounder Andrej Karpathy.
• It means AI tools like Replit Agent can do the heavy lifting in coding to build software quickly.
• Whilst it lowers the barrier to coding, experts told BI it has its pitfalls.

Silicon Valley isn't just coding anymore. It's "vibe coding."

Using AI to write code has been gaining traction for years, but now, a new buzzword coined by Andrej Karpathy, the computer scientist who cofounded OpenAI, is capturing the movement.

This month, he described what he sees as a new kind of coding in which "you fully give in to the vibes" and "forget the code even exists."

It's an approach that defies conventional wisdom in the tech industry: that developing software demands virtuosic skill from engineers.

"It's not really coding — I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works," Karpathy, who also led Tesla's AI operations for five years, wrote on X.

AI's ability to write code has come on leaps and bounds since ChatGPT's release in late 2022. Less than two months after the chatbot's release, Karpathy said, "The hottest new programming language is English" — an allusion to how smart prompting can generate good lines of code.

Software engineers have remained in hot demand since then, but the arrival of AI that can "vibe" code into existence has some industry leaders predicting big changes ahead.

OpenAI CEO Sam Altman said during a visit to India in early February that he expected software engineering to be "very different by the end of 2025." Mark Zuckerberg also said last month on the Joe Rogan Experience that AI would soon do the work of midlevel Meta engineers.

How Silicon Valley is starting to vibe code

To give a sense of what "vibe coding" looks like in action, Karpathy shared a few ways in which he's been using AI.

In one example, he said he has been using a digital workspace tool called Composer — made by OpenAI and A16z-backed startup Cursor AI — alongside the Sonnet model from AI lab Anthropic.

Cursor's Composer tool is an AI coding assistant that it says can help users "explore code, write new features, and modify existing code." When used with Anthropic's AI, a popular choice for programmers looking to AI for assistance, making an app from scratch becomes easier. That's because the AI just needs to be ushered to take steps with a user's guidance.

In another example, Karpathy said he could just "talk" to Composer by using SuperWhisper — an AI-powered voice-to-text tool.

As he noted, this means he can do things with code in Composer without having to "barely even touch the keyboard." If there's a mistake, his approach appears to be just as simple: "When I get error messages, I just copy and paste them in with no comment; usually, that fixes it."

Others are doing similar at a time when AI coding agents can be set simple instructions to do the heavy lifting that would have once required seasoned engineers to spend hours reading through reams of code — or beginners a seriously steep learning curve.

"For a total beginner who's just getting a feel for how coding works, it can be incredibly satisfying to build something that works in the space of an hour," Harry Law, an AI researcher at the University of Cambridge, told Business Insider.

Amjad Masad, CEO of Replit, a software company backed by A16z and Y Combinator, addressed Karpathy's original post, saying that "75% of Replit customers never write a single line of code."

Replit provides software used by coders to start projects, called an "online integrated development environment," but aims to offer a version that brings AI into the mix so apps can begin to be built with simple prompts. "Vibe coding is already here," Masad wrote on X.

Menlo Park Labs, a startup that builds generative AI consumer applications, is also all in on vibe coding. Its founder, Misbah Syed, is a big believer in the method.

Syed told BI he uses it for the startup's products like Brain Docs, which lets users convert a PDF to an explainer video with slides. Syed said if it makes mistakes, he feeds back the errors, and it usually fixes them. For him, the approach means that "if you have an idea, you're only a few prompts away from a product."

Vibe coding has its downsides

For all its potential benefits, experts see some risks with vibe coding.

"Ease of use is a double-edged sword," Law said. "Beginners can make fast progress, but it might prevent them from learning about system architecture or performance."

According to Law, overreliance on AI can also create technical debt, which means that it can become unmanageable when scaling or debugging code, a process engineers routinely have to go through. "Security vulnerabilities may also slip through without proper code review," he told BI.

A senior software engineer at Microsoft, who spoke with BI on the condition of anonymity as he is not authorized to speak to the media, feels the vibe coding concept is "a little overhyped."

"LLMs are great for one-off tasks but not good at maintaining or extending projects," he said, referring to large language models. "They get lost in the requirements and generate a lot of nonsense content."

A16z venture capitalist Andrew Chen said last week that while it was "brilliant" to be able to use "the latest AI codegen tools to do 'vibe coding,'" he found it enormously "frustrating" at the same time.

Karpathy acknowledged some of these limitations in his original post, noting that sometimes, an AI model "can't fix a bug." Still, he has found that he can "just work around it or ask for random changes" until the errors disappear.

With AI already pushing the limits of what was previously possible for programmers, it may soon be time for an industry vibe check.

The list of informal, weird AI benchmarks keeps growing. Over the past few days, some in the AI community on X have become obsessed with a test of how different AI models, particularly so-called reasoning models, handle prompts like this: “Write a Python script for a bouncing yellow ball within a shape. Make the shape […]

© 2024 TechCrunch. All rights reserved. For personal use only.

• Mark Zuckerberg said Meta will start automating the work of midlevel software engineers this year.
• Meta may eventually outsource all coding on its apps to AI.
• Meta also plans to replace fact-checkers with community notes and reduce DEI initiatives.

This year coding might go from one of the most sought-after skills on the job market to one that can be fully automated.

Mark Zuckerberg said that Meta and some of the biggest companies in the tech industry are already working toward this on an episode of the Joe Rogan Experience on Friday.

"Probably in 2025, we at Meta, as well as the other companies that are basically working on this, are going to have an AI that can effectively be a sort of midlevel engineer that you have at your company that can write code."

It may initially be an expensive endeavor, but Zuckerberg said Meta will reach the point where all of the code in its apps and the AI it generates will also be done by AI. According to a salary tracking site, midlevel software engineers at the company now earn close to mid-six figures in total compensation.

Zuckerberg's interview with Rogan came after a big week of changes for the company.

On Tuesday, Zuckerberg announced that Meta plans to replace third-party fact-checkers with community notes, similar to Elon Musk's X, and bring back more political content. The announcement has elicited alarm from dozens of fact-checking groups, who signed an open letter to Zuckerberg saying the changes would be "a step backward" for the company.

Meta is also planning to roll back several of its DEI initiatives. In a memo sent to staff on Meta's internal communications platform, Workplace, its vice president of human resources, Janelle Gale, wrote, "We will no longer have a team focused on DEI."

Google on Tuesday announced support for third-party tools in Gemini Code Assist, its enterprise-focused AI code completion service. Code Assist launched in April as a rebrand of a similar service Google offered under its now-defunct Duet AI branding. Available through plug-ins for popular dev environments like VS Code and JetBrains, Code Assist is powered by Google’s Gemini […]

© 2024 TechCrunch. All rights reserved. For personal use only.

• At an AI roundtable in November, developers said AI tools were playing a key role in coding.
• They said that while AI could boost productivity, stakeholders should understand its limitations.
• This article is part of "CXO AI Playbook" — straight talk from business leaders on how they're testing and using AI.

At a Business Insider roundtable in November, Neeraj Verma, the head of applied AI at Nice, argued that generative AI "makes a good developer better and a worse developer worse."

He added that some companies expect employees to be able to use AI to create a webpage or HTML file and simply copy and paste solutions into their code. "Right now," he said, "they're expecting that everybody's a developer."

During the virtual event, software developers from companies such as Meta, Slack, Amazon, Slalom, and more discussed how AI influenced their roles and career paths.

They said that while AI could help with tasks like writing routine code and translating ideas between programming languages, foundational coding skills are necessary to use the AI tools effectively. Communicating these realities to nontech stakeholders is a primary challenge for many software developers.

Understanding limitations

Coding is just one part of a developer's job. As AI adoption surges, testing and quality assurance may become more important for verifying the accuracy of AI-generated work. The US Bureau of Labor Statistics projects that the number of software developers, quality-assurance analysts, and testers will grow by 17% in the next decade.

Expectations for productivity can overshadow concerns about AI ethics and security.

"Interacting with ChatGPT or Cloud AI is so easy and natural that it can be surprising how hard it is to control AI behavior," Igor Ostrovsky, a cofounder of Augment, said during the roundtable. "It is actually very difficult to, and there's a lot of risk in, trying to get AI to behave in a way that consistently gives you a delightful user experience that people expect."

Companies have faced some of these issues in recent AI launches. Microsoft's Copilot was found to have problems with oversharing and data security, though the company created internal programs to address the risk. Tech giants are investing billions of dollars in AI technology — Microsoft alone plans to spend over $100 billion on graphics processing units and data centers to power AI by 2027 — but not as much in AI governance, ethics, and risk analysis.

AI integration in practice

For many developers, managing stakeholders' expectations — communicating the limits, risks, and overlooked aspects of the technology — is a challenging yet crucial part of the job.

Kesha Williams, the head of enterprise architecture and engineering at Slalom, said in the roundtable that one way to bridge this conversation with stakeholders is to outline specific use cases for AI. Focusing on the technology's applications could highlight potential pitfalls while keeping an eye on the big picture.

"Good developers understand how to write good code and how good code integrates into projects," Verma said. "ChatGPT is just another tool to help write some of the code that fits into the project."

Ostrovsky predicted that the ways employees engage with AI would change over the years. In the age of rapidly evolving technology, he said, developers will need to have a "desire to adapt and learn and have the ability to solve hard problems."