https://x.com/garrytan/status/1897303270311489931
For 25% of the Winter 2025 batch, 95% of lines of code are LLM generated.
And they go on to say that if youre not vibe coding youre going to be left behind.
Vibe coding is basically just hitting accept all on cursor and just going with the vibes
How should we be interpreting this? are they correct here?
For the uninitiated, here's what vibe coding means (from Wikipedia):
Vibe coding is an AI-powered programming practice where a programmer describes a problem in a few sentences as a prompt to a LLM tuned for coding, while ignoring the details of the generated code.
(So interesting how terms catch on...)
Vibe coding could make sense in a startup environment when a lot of what you build is throw-away prototypes anyway.
But in larger tech companies, you must know things work at a deep level. Your value lies in mastery of a specific domain or technology. This is especially true for a junior developer.
The difference is:
It's a made-up term, but I feel like people are equating "vibe coding" to "YOLO coding." My take is that the effectiveness of using AI assist is going to be wholly dependent on the developer, based on their familiarity and experience (their "vibes" on what they see the AI spits out) with whatever they're working on.
It's just sales and hype: a lot of it is just exaggreration to get people to use AI more so AI company evaluations and hype.
I tried using AI tools to write CrowdStrike QL and it was not super helpful. I don't know about vibe coding, but I spent the entire time complaining out loud how much it sucked.
I like Garry Tan, but he has some... weird takes sometimes. Remember that he literally runs the biggest startup accelerator on the planet, so it's in his best interest to drive the AI narrative.
That being said, I wouldn't be surprised if these numbers were true (i.e. not cherry-picked/stretched/distorted). YC focuses on very early-stage startups, and when you're just starting out, it makes sense to quickly generate a bunch of throwaway prototype code.
As Rahul mentioned though, generating 95% of your code with AI won't fly in any larger company. Even at Series A, you're generally going to have some sort of technical debt and certain contextual roadblocks that AI will struggle to work around. You will also have some amount of traction at that stage, so you can't just hit full send on risky AI-generated code. It's been 2+ years since ChatGPT came out, and I still 100% believe that the skills in this course (skills that AI are absolutely terrible at) are extremely valuable: Level Up Your Code Quality As A Software Engineer
Here's another fun thread about AI coding: "ZUCK will replace mid level engineers with AI"
Heck even at seed stage... most of the time, the existing code wouldn't have handled all the common cases. Then when you say "Refactor this code for me", Claude slashes up your code even more.