He coined the term “vibe coding” earlier this year. Well, he has something to say about that.
Andrej Karpathy led AI at Tesla for five years, leading the company's Autopilot efforts and briefly working on the humanoid robot Optimus. He sandwiched two stints at OpenAI between his stints at Tesla, making Karpathy a co-founder of the AI pioneer.
As 2025 draws to a close, Karpathy has released his annual roundup of large-scale language models for X. He reflected on the famous term he coined in February that has since shaken up the software engineering industry.
“With VibeCoding, programming is not just for highly trained professionals,” Karpathy writes. He said this was an example of how “regular people benefit more from LLMs compared to professionals, businesses and governments.”
Vibe coding may also be benefiting businesses. Technology companies are equipping engineers with tools like Cursor, Claude Code, and OpenAI's Codex to increase productivity.
Karpathy writes that vibecoding “enables trained professionals to write more (vibecoded) software that would never be written otherwise.”
Also, the configuration (or use case) of the code itself may change. Karpathy has a long list of adjectives to describe this new body of code: “free, ephemeral, malleable, use once and throw away.”
“Vibe coding will terraform software and change job descriptions,” he wrote.
How does Karpathy feel about the origin of this term?
“Funny enough, I coined the term 'vibecoding' in this shower of tweets, without any idea how far it would spread,” he wrote.
It is not yet clear how effective Vibecoding is in training engineers. A METR study published in July found that AI coding assistants reduced the productivity of participating experienced software developers by 19%. The authors said the study's developers also had overconfidence in the tool and expected to see a 20% increase in productivity after using it.
But what is clear is that this practice is unlocking entirely new forms of technology products. Twitter founder Jack Dorsey vibe-coded his new messaging app this year. Even non-technical employees can easily build, ship, and even sell apps built in hours, if not minutes.
Mr. Karpathy added several other considerations. He praised Google Gemini's Nano Banana image model, writing that Claude Code was “the first convincing demonstration of what an LLM agent looks like.”
Overall, Karpathy writes that 2025 was “an exciting and somewhat surprising year for LLMs.”
