Published on the 14/05/2025 | Written by Heather Wright

Good vibrations, but is it for the enterprise?…

“You fully give in to the vibes, embrace exponentials and forget that the code even exists.”

That’s Andrej Karpathy, OpenAI co-founder, former director of AI for Tesla on the phrase he coined which has become one of the hottest buzz terms of the year, or at least this quarter: Vibe coding.

“I just see stuff, say stuff, run stuff and copy paste stuff, and it mostly works.”

Forget the precision of traditional coding, with its clear requirements, careful planning, testing and reviews, this is about building apps by ‘vibe’ telling the computer what outcome you’re after and getting the large language models to generate code from simple, plain language prompts. Problem with the code? No worries, just feed it back into the AI model. It’s not, it should be noted, the same as writing code with the help of LLMs.

In Karpathy’s words: “I ask for the dumbest things, like ‘decrease the padding on the sidebar by half’ because I’m too lazy to find it myself. I ‘accept all’ always; I don’t read the diffs anymore… When I get error messages I just copy paste them in with no comment, usually that fixes it.”

Vibe coding is attracting plenty of attention – and funding – for potentially lowering the barriers to entry when it comes to software creation. Karpathy says it’s all possible thanks to the advances in coding AIs. He apparently just talks to Cursor Composer (now Chat) via the SuperWhisper speech to text tool.

He’s clearly not alone.

Cursor – whose parent Anysphere has just scored a further US$900 million in Series C funding – says it has over one million users, including 360,000 paying customers.

Figures for Replit, another key player in the space, are harder to come by, but the site – which offers a range of tools including the AI-powered Replit Agent coding tool – claims 30 million users. (Other vibe coding offerings include GitHub Copilot, WindSurf Cascade, StackBlitz Bolt and Lovable.)

But most experts agree that coding skills are still needed when vibe coding.

Technology consultancy Thoughtworks decided to test whether vibe coding could be used for production grade code – code that’s robust, maintainable and ready to be deployed and evolved responsibly in real-world conditions.

Using Cursor, Thoughtworks ran three experiments to see if AI could build ‘a non-trival application from scratch’ without any code written by Thoughtworks, and still produce something humans could maintain. One of the three used vibe coding – ‘embracing a freeform, improvisational style… where we focused purely on functionality and expressed little about how the system should be structured.’ The others were more specific, with prescribed design heuristics, expectations around modularity and testability and reinforcing quality through continuous feedback.

The hands off approach of vibe coding quickly ran into issues, Thoughtworks says.

While initially it ‘impressively generated a near-working application in a single pass’, subsequent modifications such as tweaking UI, output formats, introducing an interactive menu-based mode, and adding tools proved ‘challenging’.

“The AI often struggled with these incremental changes leading to regressions and necessitating manual interventions.”

That tallies with observations from other industry experts, who have noted vibe coding can falter when dealing with complex evolving systems.

AI models are currently limited in their context size, effectively creating an upper limit to how complex a vibe-coded project can get, without humans stepping in to assemble the different pieces of code into a larger architecture.

As other coding experts have also noted: It’s all fun and games until you have to vibe debug. While Karpathy might be happy to ‘just copy paste’ when he gets error messages (remembering that he does have coding skills), enterprises are less likely to be prepared to take that approach, and the Thoughwork’s experiments suggest digging into the code will likely be needed.

Replit CEO Amjad Masad, who says he doesn’t like the term vibe coding because it cheapens the possibilities’, is clear that vibe coding isn’t just about ‘generating thing’.

“You can build agents and not just generate things. It can actually reason,” he noted in a podcast from VC firm Sequoia, adding that Replit is ‘not vibe coding, just vibe’.

“Vibe coding makes sense if you’re sort of starting from a position of coder and you’re Andrej Karpathy and you don’t want to worry too much about the code and you keep hitting enter… But if you’re starting with Replit, your actually not starting from a position of code, you’re starting from an idea that you’re iterating on and then you go in and the agent is unfolding this code in front of you. Actually, when you’re using Replit Agent, you don’t have the luxury to look at the code.”

He noted Replit now has Assistant for ‘more advanced people’, enabling vibe coding with request, response and reviewing of code.

The technology appears to be the venture world’s latest fixation, with plenty of money flowing into the sector.

The aforementioned Anysphere’s US$900 million series C funding sees it valued at US$9 billion. That’s more than triple its $2.5 billion valuation in January and more than 20 times its series A valuation from last August. Its Cursor has been dubbed the ‘fastest-growing SaaS product ever’ and already boasts US$00 million in annual recurring revenue.

American tech startup accelerator and venture capital firm Y Combinator – its helped fund the likes of Stripe, AirBnb and DoorDash – says a quarter of its current startups are using AI to write 95 percent of their code.

For the startups it means less engineers are needed, equating to less money being spent and capital going further.

Garry Tan, Y Combinator CEO, told CNBC earlier this year that the current startup group was growing significantly faster than past cohorts, growing in aggregate 10 percent each week, and with actual revenue – thanks largely to the use of AI for repetitive tasks and in vibe coding.

Meanwhile, Apple has partnered with Anthropic to build a vibe coding platform to be used, initially at least, internally at Apple.

The tool is reportedly a revamped version of Xcode with Anthropic’s Claude Sonnet integrated, providing a chat-based interface to write, test and fix code.

But perhaps the final word, for now, should go to Karpathy. Summing up on vibe coding, he says: “It’s not too bad for throwaway weekend projects, but still quite amusing. I’m building a project or webapp, but it’s not really coding – I just see stuff, say stuff, run stuff and copy paste stuff, and it mostly works.”

And there, may lie the rub for enterprise. It ‘mostly’ works. For now, at least.