With the release of new AI models with better coding, developers are increasingly using AI to generate code. One of the latest examples is the Salyied Silicon Valley Startup Accelerator, the current batch coming out of the Y Combinator. Jared Friedman, managing partner at YC, says a quarter of the W25 startup batch has 95% of the codebase generated by AI.
Friedman said that the 95% of the numbers didn’t include anything like the code written to import the library, but it took into account the code entered by humans compared to AI.
“We didn’t fund many non-technical founders. These people are all very technical and can build their own products from scratch. A year ago, they were building products from scratch, and now 95% is built by AI,” he said.
In the video entitled “Vibe Coding Is the Future,” Friedman and YC CEO Garry Tan, managing partner Harj Taggar, and general partner Diana Hu discussed the trends in using natural language and instinct to write code.
Last month, Tesla’s former AI director and Openai researcher Andrej Karpathy used the term “vibe coding” to explain how to code using a leading language model (LLM) without focusing on the code itself.
However, the code generated from AI is not complete. Research and reports show that some AI-generated code can be used to insert security flaws into your application, cause a halt, or make mistakes, modifying and debugging the code to the developer.
During the discussion, Hu said there is one skill they must be good at, as it is reading the code and finding bugs, even if product builders rely heavily on AI.
“To know that LLM is spitting out bad or good things requires a lot of taste and plenty of training. To do a good ‘atmosphere coding’, you need taste and knowledge to decide it’s good versus good,” she said.
Tan also agreed to the founder’s points that require classic coding training to maintain the product in the long run.
“Let’s say there’s a startup with 95% AI-generated code (in the market) and for a year or two, there are 100 million users on that product. Does that fall? The first version of the inference model is not good at debugging. So you have to go into detail what’s going on with the product,” he suggested.
VCS and developers are excited about AI-powered coding. Startups like Bolt.New, Codeium, Cursor, Lovable and Magic have raised hundreds of millions of dollars in funding over the past 12 months.
“This isn’t trendy. This won’t go away. This is the dominant way of code. And if you’re not doing it, you might just be left,” Tan added.