Andrej Karpathy, one of AI’s most influential figures, has published his year-in-review for large language models (LLMs), reflecting on the revolutionary concept he coined earlier in 2025: “vibe coding.” The term, which Karpathy admits he created “totally oblivious to how far it would go,” has fundamentally reshaped how we think about software development in the age of AI.
Karpathy’s credentials are impressive: he led AI at Tesla for five years, steering the company’s Autopilot effort and briefly contributing to the humanoid robot Optimus. He also served two stints at OpenAI, where he is a cofounder of the AI pioneer. His unique perspective spans both cutting-edge AI research and practical implementation.
Vibe coding represents a paradigm shift in programming, where developers “fully give in to the vibes, embrace exponentials, and forget that the code even exists,” as Karpathy originally described it. The practice leverages powerful AI coding assistants like Cursor, Claude Code, and OpenAI’s Codex to enable rapid, intuitive software creation. According to Karpathy, this democratizes programming: “With vibe coding, programming is not strictly reserved for highly trained professionals.”
The impact extends beyond accessibility. Karpathy argues that “regular people benefit a lot more from LLMs compared to professionals, corporations and governments.” For trained professionals, vibe coding “empowers [them] to write a lot more (vibe coded) software that would otherwise never be written.” This new code is characterized as “free, ephemeral, malleable, discardable after single use” — fundamentally different from traditional software.
The real-world applications are already visible. Twitter founder Jack Dorsey vibe-coded a new messaging app this year, while non-technical workers are building, shipping, and even selling apps created in hours or minutes. Tech companies have widely adopted AI coding tools, seeking productivity gains for their engineering teams.
However, questions remain about efficiency. A METR study published in July found that AI coding assistants actually decreased productivity by 19% among experienced software developers, who were also overconfident about the tools’ capabilities.
Karpathy also praised Google Gemini’s Nano Banana image model and called Claude Code the “first convincing demonstration of what an LLM Agent looks like.” He concluded that 2025 was an “exciting and mildly surprising year of LLMs,” suggesting the technology continues to evolve in unexpected ways.
Key Quotes
With vibe coding, programming is not strictly reserved for highly trained professionals.
Andrej Karpathy explained how vibe coding democratizes software development, making it accessible to non-technical users. This statement underscores the revolutionary potential of AI coding assistants to reshape who can create software.
Vibe coding will terraform software and alter job descriptions.
Karpathy’s prediction about the transformative impact of AI-assisted coding on the software industry and workforce. This suggests fundamental changes to how software engineering roles are defined and what skills are valued.
It is free, ephemeral, malleable, discardable after single use.
Karpathy describing the new characteristics of vibe-coded software, which contrasts sharply with traditional software engineering principles that emphasize maintainability, documentation, and long-term sustainability.
Amusingly, I coined the term ‘vibe coding’ in this shower of thoughts tweet totally oblivious to how far it would go.
Karpathy reflecting on the unexpected viral spread and adoption of his term, which has become a defining concept for AI-assisted programming in 2025.
Our Take
Karpathy’s year-end reflection reveals both the promise and uncertainty surrounding AI coding tools. What’s particularly striking is the disconnect between the hype and the data: while vibe coding enables impressive demos and empowers non-coders, the METR study showing 19% productivity decreases among professionals suggests we’re overestimating current capabilities. This mirrors broader patterns in AI adoption — tools that seem revolutionary often require significant adjustment periods and new workflows. The concept of “ephemeral” code is especially intriguing and potentially problematic. It suggests a future where software becomes more disposable, created for immediate needs rather than long-term maintenance. This could accelerate innovation but may also create technical debt and sustainability challenges. Karpathy’s unique position straddling research and industry gives his observations particular weight, and his acknowledgment that 2025 was “mildly surprising” suggests even AI insiders are uncertain about the technology’s trajectory. The vibe coding phenomenon may represent not just a new tool, but a fundamental rethinking of what programming means in an AI-native world.
Why This Matters
Karpathy’s reflection on vibe coding matters because it comes from one of AI’s most respected practitioners, someone who has shaped AI development at both Tesla and OpenAI. His endorsement validates a fundamental shift in how software is created, moving from traditional coding to AI-assisted, intuitive development. This has profound implications for the tech workforce: programming may no longer require years of formal training, potentially democratizing software creation while simultaneously threatening traditional software engineering roles. The concept that code can be “ephemeral” and “discardable after single use” challenges decades of software engineering best practices around maintainability and documentation. For businesses, this represents both opportunity and risk — the ability to rapidly prototype and deploy solutions, but with uncertain quality and reliability standards. The mixed productivity results from the METR study suggest we’re still in early days of understanding how to effectively integrate AI coding tools. Karpathy’s prediction that “vibe coding will terraform software and alter job descriptions” signals that the software industry is entering a period of significant transformation, with implications for education, hiring, and how we conceptualize the role of human developers in an AI-augmented world.
Related Stories
- PwC Hosts ‘Prompting Parties’ to Train Employees on AI Usage
- The Future of Work in an AI World
- Business Leaders Share Top 3 AI Workforce Predictions for 2025
- AI’s Role in Tech Hiring Freeze: White-Collar Job Market Slump
- OpenAI Lost Nearly Half of Its AI Safety Team, Ex-Researcher Says
Source: https://www.businessinsider.com/andrej-karpathy-coined-vibecoding-ai-prediction-2025-12