The Scene
In almost every way, Leo Paz is following a well-worn path to Silicon Valley success. The 27-year-old Canadian earned a degree in software engineering, co-founded a startup that was accepted into the vaunted Y Combinator accelerator program, and moved to San Francisco to build it.
But in another way, he’s unlike any of his predecessors: Since landing in the world capital of software development, he’s let AI do all the coding. Paz has been spending 14-hour days instructing large language models to do the work for him, “vibe coding” his sales agent startup Outlit all the way to Y Combinator demo day in April.
“I don’t think YC knew exactly how much code was being written by LLMs,” he told Semafor. It found out earlier this year after a founder survey. “They were kinda shocked,” he added.
While vibe coding has helped people without any technical experience create software through AI-powered assistants, the practice is also sweeping the upper echelons of the startup ecosystem, where, until recently, software talent was the scarcest and most coveted resource. The change is shifting the balance of power from the most talented coders to anyone who has a great idea, opening the door to Silicon Valley’s next generation of unicorns.
In March, Y Combinator CEO Garry Tan said 25% of startups in that winter’s class generated 95% of their code using AI tools. The numbers are likely to be much higher in the next batch, which culminates with a demo day this month.
Founders of startup Mastra, which developed a JavaScript framework for building agents, held weekly vibe coding whiteboard sessions in their San Francisco apartment with other startups in their YC cohort. Fueled by takeout from Gai Chicken & Rice, founders would gather under a projector light to learn the ins and outs from the Mastra team.
Chief Technology Officer Abhi Aiyer said his full-time job is prompting the AI on what to build. “All of us know what we’re doing, so we don’t actually write code that much,” he said of the startup team. “I just think and review” the code written by AI, he added.
Jackson Stokes of model reinforcement startup TrainLoop, also in that YC class, said he has been vibe coding his product’s user interface. “It definitely changes our calculus on potentially hiring a designer,” he said. “I don’t know if it’ll ever come back into scope, to be honest.”
Outside of YC, Pukar Hamal, CEO of SecurityPal — which helps companies automate responses to security-related questions during the sales process — is vibe coding agents to fill some open positions rather than hiring personnel.
SecurityPal needed a product manager, and instead of creating a job posting, Hamal spent three hours vibe coding an agent named Penelope, depicted as a brunette mountaineer wearing aviator goggles and a fur-lined bomber jacket (a nod to the company’s office in Nepal). “Penelope is going to be a very integral part of how I as a CEO make decisions,” he told Semafor. “She’s also going to be part of my leadership meetings, where we can interrogate her on different ideas we have.”
SecurityPal has incorporated two vibe-coded agents into the security analyst team, which manages security data sharing between customers. The company onboarded them like it would any employee, and they participate in all-hands meetings — camera on. “We’ve probably already hired our last security analyst,” he said.
Step Back
AI coding technology has improved remarkably fast. A year ago, most software developers were using it simply to complete lines of code they were already writing, much like Gmail predicts the next word you’re about to type. The term “vibe coding” was coined in February by the famed AI researcher Andrej Karpathy. “It’s not too bad for throwaway weekend projects, but still quite amusing,” he wrote on X.
When Karpathy said that he had been using AI tools to code, the shame around doing so lifted for many developers who had been keeping their practice under wraps, said Paz. “There was this guilty feeling about using AI to write code” because they feared they would be judged by other programmers, he said. “When Andrej Karpathy says he’s doing it, everyone else could admit it, too.”
That pace of change has been alarming for some. Former Trump advisor Steve Bannon said AI could “eviscerate” early career tech jobs. Anthropic CEO Dario Amodei said companies need to stop “sugar-coating” that AI could wipe out half of entry-level white collar jobs by 2030. At a Semafor Tech event in May, Replit CEO Amjad Masad said many companies are 12 to 18 months away from not needing a chief technology officer.
In some ways, the advent of vibe coding follows the natural trajectory of software development. Early developers wrote assembly code that required tedious and monotonous hours of work to make it interact properly with the computer’s central processing unit.
Programming languages like FORTRAN and COBOL removed the need to write in assembly, freeing developers up to build more complex programs.
Over time, more and more layers of software development were abstracted away, to the point where assembly code is barely even taught anymore.
AI coding could be looked at as perhaps the final layer of software abstraction, where AI “agents” will ultimately build the software we use, removing people from the process entirely.
Vibe coding is also the first time software development has become accessible to people with zero training, opening up economic opportunities for them.
Cynthia Chen, who designs payments processor Block’s user interface, created a “dog rolodex” app that logs a user’s interactions with dogs — with no coding experience. She had the idea five years ago but didn’t know how to get started, and even considered hiring someone to build it last year, she told Semafor. After vibe coding with Claude in her spare time for about a month and a half, Dog-e-dex is now live on the App Store.
In a survey of individuals who wanted to start a business but didn’t, a lack of technical know-how emerged as a key barrier — with 28% citing website maintenance difficulties and 17% pointing to a lack of coding skills.
The tools aren’t perfect. If a user is too vague in their prompting or tries to move too fast, an update that should change three lines of code will rewrite 900, one startup founder said. But if the current pace of advancement holds, those issues will soon go away.
Know More
The rise of vibe coding may also contribute to a slew of next-generation apps that lack proper security measures. Lovable, a vibe coding app popular with non-developers, failed to fix a security flaw despite learning it existed months ago, Semafor reported. The vulnerability allowed customers to create apps that let anyone view information on its users, including names, email addresses, and financial details.
The Swedish startup tells customers that apps created through its service will be secure, but puts the onus on users to safeguard their creations. Skilled developers may be trained to take measures protecting user information, while novices inadvertently ship products susceptible to hacks. Similar platforms across the vibe coding ecosystem may bear similar vulnerabilities that haven’t yet been discovered.
Lovable responded on X, saying, “We’re not yet where we want to be in terms of security and we’re committed to keep improving the security posture for all Lovable users.”
Vibe coding has also created a gray area in determining legal ownership of apps. The US Copyright Office does not extend copyright protection to AI-generated pieces of code, and in cases where code uses a mix of human authorship and AI, only the human-written pieces are eligible for copyright. This framework could give rise to a generation of new startups with shaky legal footing.
“Copyright is about expression. If human expression is used to prompt the AI, there might be an argument, but it’s going to be a lot harder to make that argument if you’re relying on code straight-up generated by AI,” said Chris Suarez, intellectual property lawyer at Steptoe. The creativity comes in the actual writing of the code and how it’s organized, he said, which humans don’t do when they prompt AI — no matter how creative the prompt itself is.
The View From Big Tech
While big tech companies don’t face the same breakneck pace to ship technology as startups, they are also implementing AI into their software development practices. In April, Microsoft CEO Satya Nadella said up to 30% of the company’s code is being written by AI. It’s “well over 30%” at Google, said CEO Sundar Pichai.
When asked about what products the companies are building with AI tools, executives from big tech giants who spoke with Semafor cited Java transformation, debugging existing code, and building apps for their teams, but not developing new core products.
Microsoft Research’s Nicole Forsgren told Semafor that company hackathon days are now vibe coding days. Jeanine Banks, vice president of Google’s vibe coding product Firebase Studio, said she built a research app for her team that provides analysis across multiple documents.
Amazon used its own Q Developer tool last year to migrate 30,000 applications to a newer version of Java, which it says saved $260 million and 4,500 years of work. Deepak Singh, vice president of developer agents, told Semafor that Prime Video engineers are using Q Developer in their day-to-day work, which includes updating the website design, recommendation algorithms, and back-end software that runs videos and monitors users. It’s like “hiring a bunch of new, inexperienced developers,” he said.
“Vibe coding is great for individual developers or small teams to make progress. Enterprise has different requirements,” he said. “They have large code bases and actually don’t generate that much new code. It isn’t sufficient for an [organization yet].”
Rachyl’s view
The biggest hurdle in my own experimentation with vibe coding has been token and chat limits.
Basic websites are an easy lift. It only took about 15 minutes each to build a fun fact generator with Lovable and an exercise tracker with Google Ventures-backed Bolt — and both tools are surprisingly intuitive.
More complex apps require ongoing upgrades or constant re-prompting of context. Chen, who created Dog-e-dex, said she frequently ran into chat limits using Claude Pro, Anthropic’s middle tier product, which costs $20 per month. The limits didn’t derail her because her app was more of a fun side project than a business, but they were disruptive when she was in a “flow state” trying to figure out a problem, she said.
Paz, the Outlit founder, runs into Cursor’s token limit every month, he said. On top of his $20 per month tier, he often pays another monthly $20 for additional instant responses — still a miniscule amount compared to paying a salary. “It’s a cost I’m willing to field because it’s so core to my workflow,” he said.
As agents start taking on more of the work that vibe coders do today, those costs could rise significantly — especially as they work on prolonged tasks and potentially veer off course.
Room for Disagreement
While vibe coding can give both developers and non-developers a huge creative boost, “it’s not a free pass to abandon rigor, review, or craftsmanship,” Addy Osmani, Google’s head of Chrome developer experience, wrote on Substack.
“It can be downright dangerous to accept AI-generated output at face value,” he said, adding that many early vibe-coded apps lack security, performance, and long-term durability.
“Embrace the vibes, but honor the craft,” he wrote.
Notable
- Vibe coding is sparking the latest dealmaking bonanza in Silicon Valley, with OpenAI’s $3 billion acquisition of Windsurf, Cursor’s reported $10 billion valuation, and Apple’s partnership with Anthropic to build a new vibe coding platform.