Ex-Tesla AI head has seen a 'phase shift in software engineering' using Claude Code — and his manual skills slowly 'atrophy'

Andrej Karpathy, a prominent figure in the AI landscape and former director of AI at Tesla, has observed a significant transformation in the realm of software engineering. He has notably introduced the concept of "vibe coding," which refers to the burgeoning trend of AI-assisted programming that has gained substantial traction and was even recognized as Collins Dictionary's word of the year. In his extensive reflections shared on X, Karpathy detailed how his coding practices have evolved amidst the rise of advanced coding agents like Anthropic's Claude Code and OpenAI's Codex. He revealed that the capabilities of these large language models (LLMs) reached a critical turning point around December 2025, leading to a dramatic shift in his coding approach. Previously, Karpathy's workflow comprised approximately 80% manual coding and 20% AI assistance. However, by December, this ratio had flipped, with 80% of his coding now performed by AI agents and only 20% requiring manual edits. He candidly acknowledged that this transition has affected his sense of identity as a programmer, stating, "I really am mostly programming in English now, a bit sheepishly telling the LLM what code to write... in words." While he recognizes the power and efficiency of AI-generated code, Karpathy also expressed concerns about the potential decline of traditional coding skills. "I've already noticed that I am slowly starting to atrophy my ability to write code manually," he admitted. His insights resonated with engineers from leading AI firms, prompting a lively discussion on the implications of this shift. Engineers like Ethan He from xAI and Boris Cherny from Anthropic contributed to the conversation, highlighting the efficiency of AI in coding. Cherny revealed that his team relies almost exclusively on Claude Code for their programming tasks, noting that he personally hasn’t manually edited code in over two months. However, he also pointed out the challenges posed by AI-generated code, such as overcomplication and the presence of redundant code, suggesting the need for AI systems to review each other’s output. As the landscape of software engineering continues to evolve, the dialogue sparked by Karpathy's observations illustrates a pivotal moment in how coding is approached, blending traditional skills with cutting-edge AI technology.