In late June, the CEO of a tech startup halted all software development at the company. He wanted to ensure every team member who worked with code was up to speed on the latest trend: vibe coding.
“You start to realize, wow, these things could move way faster,” Rowan Trollope, the CEO of Redis, a software company, told Business Insider.
He immediately approved the use of all AI-assisted coding tools. Then the company launched a weeklong hackathon that challenged teams of employees to “use all the latest and greatest AI technologies to do something cool,” Trollope said.
Companies have found, however, that despite the excitement — and the millions in funding pouring into vibe coding companies — the technology is still limited. So, many CEOs are developing new policies and tools to maximize the benefits of vibe coding while mitigating the pitfalls.
Vibe coding is when developers (or anyone, really) prompt AI to generate code. In a survey of hundreds of engineers in May, Jellyfish, a software intelligence platform, found that 90% of them had integrated AI into their work, up from 61% just a year ago.
Vibe coding is now a marketable skill in Silicon Valley. Companies from Visa to Reddit to DoorDash are posting jobs that require vibe coding experience or familiarity with AI coding tools. Meta now allows job candidates to use an AI assistant in their coding interviews.
The term was coined by OpenAI cofounder Andrej Karpathy in February. “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,” Karpathy wrote in a post on X. “I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works.”
Limits to the technology remain, however. Though vibe coding promises quick productivity gains and allows people with little coding experience to create software, tech executives say AI is still prone to mistakes, often writes unnecessarily long code, or lacks the proper architecture. So Trollope and other tech CEOs have had to introduce parameters for its use.
Trollope said vibe coding is best for building proof of concepts, writing tests, and validating existing code, but not necessarily developing any of the company’s core software. “It’s still not in a place yet where we would trust it with our core technology,” he said.
While still limited, this new, more freewheeling approach has gained momentum in part thanks to the money flowing into vibe coding platforms. Last month, Anysphere, the company behind Cursor, an AI-assisted code editor, announced a $900 million Series C fundraise at a $9.9 billion valuation. Wix, a web-development platform, announced that it had acquired the vibe coding platform Base44 for $80 million.
Replit, a code editor, saw revenue grow fivefold in September after it released Agent, a coding assistant that works with natural language prompts. By June, the company landed a new $250 million funding round that brought its valuation to $3 billion, according to Forbes. Swedish vibe coding startup Lovable, one of Europe’s fastest-growing startups, raised $200 million in Series A funding in July at a $1.8 billion valuation, according to PitchBook.
The funding frenzy pushed AirTable, a database development platform, to relaunch last month as a fully AI-native platform. As part of the overhaul, the company created an app-building assistant called Omni for vibe coders. It allows developers to “conversationally vibe generate the app they want, but understand what’s been generated all the way down to the data and logic layer as well,” AirTable said in a blog post announcing the overhaul.
“There’s still a question of: Is that an AI tourist attraction? Is it going to be durable? Is it going to be high churn?” AirTable CEO Howie Liu told Business Insider. But “all these people are coming in and pulling out their credit cards to try it out.” For AirTable, “not fully reinventing ourselves is kind of like a guaranteed path to obsolescence,” Liu said.
With the launch of Omni, Liu also saw an opportunity to correct some of the problems that come with vibe coding.
With vibe coding, “you’re not really inspecting the code, you’re not really thinking about the technical architecture, you’re just telling it what you want it to build and kind of like clicking the ‘I’m feeling lucky button,'” Liu said. “The magical thing is like the AI has gotten good enough that it seemingly just works some of the time.” But even apps that “seemingly work” can be riddled with errors and security vulnerabilities at their deeper, infrastructure layers.
In a perfect world, developers could leverage artificial general intelligence to code apps given the breadth of information that developers need to reason through, Liu said. Until then, developers need a “two-way feedback loop between the agent that is building the code, or building the app, and the user, the human, who’s guiding it and saying, ‘here’s what I actually want you to build.'”
At Redis, humans are convening internal groups to share best prompting practices to improve their part of the equation, Trollope said. “I think people do go on a journey where you start very small and you very quickly realize the prompts can get longer and longer and more and more complicated,” he said.
Read the full article here
