AI-generated code is moving from side tool to core building method for some young startups. In Y Combinator's W25 batch, managing partner Jared Friedman said a quarter of the companies have codebases where 95% of the code was generated by AI.
The figure points to a fast change in how technical founders are building first versions of products. It also raises a practical question: if AI writes most of the code, what skills do founders still need to keep the product working?
YC's W25 batch shows how far AI coding has moved
Friedman discussed the trend in a conversation posted on YouTube with YC CEO Garry Tan, managing partner Harj Taggar, and general partner Diana Hu. The video was titled Vibe Coding Is the Future, and it centered on using natural language and instincts to create code.
According to Friedman, the 95% measure did not include code written only to import libraries. It compared code typed by humans with code produced by AI.
“It’s not like we funded a bunch of non-technical founders. Every one of these people is highly technical, completely capable of building their own products from scratch. A year ago, they would have built their product from scratch — but now 95% of it is built by an AI,” he said.
That distinction matters. The shift described by YC is not about replacing technical founders with non-technical operators. It is about people who already know how to build software choosing to let AI systems produce most of the implementation.
What “vibe coding” means in this context
The discussion used the term “vibe coding” for a style of software creation that relies on large language models, or LLMs, without focusing directly on the code itself. Last month, Andrej Karpathy, former head of AI at Tesla and ex-researcher at OpenAI, used the term to describe this way of coding.
In plain language, the approach puts more emphasis on directing the system, judging the output, and iterating toward a working product. The builder may use natural language rather than writing every function by hand.
For startups, that can change the early product process. A small team can move quickly from idea to working software if the AI can generate much of the codebase. But the same process also makes judgment more important, because the builder still has to decide whether the generated output is right.
The risk is not gone just because the code compiles
YC's leaders also pointed to a limit that matters for any AI coding workflow: generated code is not automatically reliable. Studies and reports have observed that some AI-generated code can insert security flaws in applications, cause outages, or make mistakes.
Those problems can force developers to change the code or spend serious time debugging. In other words, a high percentage of AI-generated code does not remove the need for engineering review.
Hu said product builders who use AI heavily still need to be strong at reading code and identifying bugs.
“You have to have the taste and enough training to know that an LLM is spitting bad stuff or good stuff. In order to do good ‘vibe coding,’ you still need to have taste and knowledge to judge good versus bad,” she said.
That warning defines the practical skill shift. Writing every line may become less central in some workflows, but understanding the system remains critical. A founder who cannot evaluate the code may not know when the AI has created a fragile or incorrect implementation.
Scaling AI-written products is the bigger test
Tan made a related point about what happens after a startup reaches the market. A codebase can be mostly AI-generated and still work well enough for an early version, but the harder question is whether it can support a much larger product.
“Let’s say a startup with 95% AI-generated code goes out [in the market], and a year or two out, they have 100 million users on that product. Does it fall over or not? The first versions of reasoning models are not good at debugging. So you have to go in-depth of what’s happening with the product,” he suggested.
That concern does not reject AI coding. It puts a boundary around it. The more users a product has, the more costly failures become, and the more founders need to understand what is happening inside the system.
The YC discussion therefore frames classical coding training as still important. AI may produce most of the code, but people remain responsible for reading it, debugging it, and deciding whether it can survive real use.
AI coding is becoming a serious startup market
The excitement around AI-powered coding is also showing up in funding. Startups including Bolt.new, Codeium, Cursor, Lovable, and Magic have raised hundreds of millions of dollars in funding in the last 12 months.
For VCs and developers, the appeal is clear: if AI can speed up product creation, the early stages of company building may look different. Founders may test ideas faster, ship earlier, and spend less time typing boilerplate by hand.
Tan argued that the trend is not temporary.
“This isn’t a fad. This isn’t going away. This is the dominant way to code. And if you are not doing it, you might just be left behind,” Tan added.
The core takeaway is balanced. YC's W25 batch suggests that AI-generated code has already become a major part of startup building. But the people using it still need technical depth, because the final responsibility for quality, security, debugging, and scale does not move from the founder to the model.