Andrej Karpathy, a former AI lead at Tesla and OpenAI, has shared his reflections on the rapid transformation of software engineering driven by artificial intelligence. In a detailed 1,000-word post titled "random notes from Claude Coding" published on X, Karpathy described a profound shift in his own workflow—and a growing concern about the decline of manual coding skills. Karpathy, who coined the term "vibe coding" to describe the rise of AI-assisted programming, noted that around December 2025, AI coding agents crossed a critical threshold of coherence, triggering what he calls a "phase shift" in the field. He credited major advances in models like Anthropic’s Claude Opus 4.5 and OpenAI’s Codex for this turning point. His personal coding ratio has flipped dramatically: where he once wrote 80% of code manually and used AI for 20%, he now reverses that—80% of his work comes from AI agents, with only 20% involving direct editing. "I really am mostly programming in English now, a bit sheepishly telling the LLM what code to write... in words," he wrote. While he acknowledges the emotional toll—“it hurts the ego”—he admits the power of these tools is too significant to ignore. He also highlighted the joy and creativity AI has brought back to coding, calling it "fun" in ways traditional development rarely was. Yet, Karpathy expressed concern about the long-term impact on foundational programming skills. "I've already noticed that I am slowly starting to atrophy my ability to write code manually," he wrote, pointing to a broader risk: as engineers rely more on AI, core competencies like algorithm design, debugging, and low-level understanding may erode. His observations sparked responses from engineers at leading AI labs. Ethan He, an engineer at xAI and former Nvidia employee, said a top-tier "10x engineer" can now function like a one-man army, empowered by AI agents. Charles Weill, another xAI engineer, noted that founders can now "divide themselves" through AI, much like venture capitalists diversify capital across startups. Boris Cherny, a senior engineer at Anthropic and creator of Claude Code, praised Karpathy’s post as "thoughtful" and shared his own experience: "Pretty much 100% of our code is written by Claude Code." He added that for over two months, he hasn’t made a single small edit by hand. Cherny also acknowledged the drawbacks—AI-generated code can be overly complex, redundant, or include dead code. His solution? Using AI to review AI-written code, creating a feedback loop that improves quality and reliability. The trend suggests a new era in software development, where human engineers increasingly act as directors, guiding and refining AI-generated code rather than writing it from scratch. While the shift boosts speed and scalability, it raises important questions about skill retention, code quality, and the future of engineering education.