Ein Viertel der Start-ups in der aktuellen Kohorte von YC hat Codebasen, die fast ausschließlich mit AI-generiert sind

A quarter of the W25 startup batch have 95% of their codebases generated by AI, YCombinator managing partner Jared Friedman said.

Friedman said that this 95% figure didn’t include things like code written to import libraries but took into consideration the code typed by humans as compared to 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.

Serious question for enthusiasts: what is the largest company with a majority AI codebase? 

YC is a pretty unsurprising group to use this kind of technique, but these are all seed stage companies. There’s still no reason to suggest that any will turn into real businesses before their vibe code kills them.

I’ll admit using AI to quickly prototype something makes a lot of sense. I have a project at work that is a simple landing page with 2 additional routes that scans a document and assesses extravagant fees. AI was phenomenal when it came to repetitive basic tasks. I knew how I wanted it to look, so making AI produce the components made a lot of sense. The key takeaways I came out of this project with were:

The Good:

– AI is great for making boilerplate code. Including a basic layout, color palettes, and sizing.

– AI is great for bouncing ideas off of. It was nice to think of some clever way to do something like an animation, how and when to display content, a modal, a toast etc and have it generated quickly. This is great for product holders as you can experiment and collaborate on many ideas.

The bad:

– AI is great for simple things, but it doesn’t take long to get into a loop with it where it doesn’t understand what you want, mostly happens with more complex requests. A good example was prompting it to layout the content using multiple animations and timing. This required a simple library, but the UX was somewhat complicated. It got to a point where the AI would almost make things harder for itself, constantly editing and removing things, adding a lot of bloat.

– it isn’t super great at debugging its own code and combined with the point above, this made identifying issues somewhat tedious. Sometimes it would be close to understanding the problem, but would not recognize a new problem it introduces in its solution.

– It is verbose, granted you could modify the prompt asking it to be more succinct when creating variables, but it would generate hilariously long names for simple arrays based on the component name. Not inherently a bad thing, but certainly overkill. “Fees” works fine, I don’t need an array called “ResultsExcessiveFees”. Sometimes its naming can add confusion, especially when it generates larger functions.

All in, it was really cool to be able to create several components quickly, kick ideas around and save time on “busy work” of creating basic UI components. But its weaknesses in creating bloat coupled with naming and its inability to recognize bugs it introduces when applying a solution makes the idea of an app that’s been built purely by AI a technical nightmare to work on long term.