Andrej Karpathy, a cofounder of OpenAI and former head of AI at Tesla, is known for coining the term "vibe coding" earlier this year—a phrase that has since sparked widespread discussion across the software development world. Now, as 2025 draws to a close, Karpathy has shared his reflections on the concept in a year-in-review post on X, offering a glimpse into where he sees the future of programming heading. Karpathy, who led Tesla’s Autopilot and briefly contributed to its Optimus humanoid robot project, described vibe coding as a transformative shift in how software is built. “With vibe coding, programming is not strictly reserved for highly trained professionals,” he wrote. He emphasized that the real beneficiaries of large language models (LLMs) may not be experts, but everyday users—individuals who can now create software with minimal technical background. The term gained traction quickly, reflecting a broader trend: tech companies are equipping engineers with AI coding tools like Cursor, Claude Code, and OpenAI’s Codex in pursuit of faster development cycles. Karpathy noted that vibe coding empowers trained professionals to produce far more software than before—code that might never have been written otherwise. He also painted a vivid picture of the new kind of software emerging from this trend: “free, ephemeral, malleable, discardable after single use.” These characteristics suggest a departure from traditional, long-lived codebases toward lightweight, disposable applications built rapidly and discarded just as easily. Karpathy predicted that vibe coding will “terraform software and alter job descriptions,” reshaping both how software is developed and who gets to build it. He acknowledged the irony of his role as the term’s originator: “Amusingly, I coined the term 'vibe coding' in this shower of thoughts tweet totally oblivious to how far it would go.” Despite the enthusiasm, the real-world efficiency of vibe coding remains uncertain. A METR study published in July found that AI coding assistants actually reduced the productivity of experienced developers by 19%. The researchers also noted that developers were overconfident in the tools, expecting a 20% boost in output even after using them. Still, the impact on innovation is undeniable. Non-technical users are now building, launching, and even monetizing apps in hours—or sometimes minutes. Twitter founder Jack Dorsey recently used vibe coding to develop a new messaging app, underscoring the democratization of software creation. Beyond vibe coding, Karpathy praised Google’s Gemini Nano Banana image model and called Claude Code “the first convincing demonstration of what an LLM Agent looks like.” Overall, he described 2025 as an “exciting and mildly surprising year of LLMs,” marking a pivotal moment in the evolution of artificial intelligence and human-computer collaboration.