Andrej Karpathy Embraces English as His Primary Coding Language, Just Months After Dismissing AI Agents
In a striking evolution of perspective, Andrej Karpathy, the renowned computer vision expert and former Director of AI at Tesla, has revealed that he now primarily codes in English. This admission comes merely three months after he publicly declared AI agents to be largely useless. Karpathy’s shift underscores the rapid advancements in large language models (LLMs) and their integration into software development workflows, highlighting how natural language prompting is transforming traditional coding practices.
Karpathy, who previously served as a senior researcher at OpenAI and founded his own AI education platform Eureka Labs, shared his current coding habits during a recent discussion. He described his process as leveraging AI tools to generate code through descriptive English prompts, effectively treating English as his de facto programming language. “I code mostly in English these days,” Karpathy stated, emphasizing how he interacts with tools like Cursor—an AI-powered code editor built on top of Anthropic’s Claude model—and other LLM interfaces.
This approach marks a departure from conventional programming paradigms where developers write precise syntax in languages like Python, C++, or JavaScript. Instead, Karpathy outlines a workflow where he articulates high-level intentions in natural language, allowing the AI to produce functional code snippets, functions, or even entire modules. He iterates on this output by refining prompts, debugging via conversational feedback, and selectively editing the generated code. This method not only accelerates development but also democratizes coding for those with strong conceptual understanding but less tolerance for syntactic minutiae.
The backdrop to this revelation is Karpathy’s earlier skepticism toward AI agents. In late 2023, during a widely circulated talk, he critiqued the state of autonomous AI agents—systems designed to perform complex, multi-step tasks independently. He labeled them “useless” at the time, arguing that their reliability was too low for production use, prone to hallucinations, context loss, and failure cascades in real-world scenarios. Karpathy pointed out that while LLMs excel at single-shot tasks like code generation or summarization, chaining them into agentic workflows often led to brittle performance.
Just three months later, however, Karpathy’s practice demonstrates a nuanced pivot. He is not endorsing fully autonomous agents but rather human-in-the-loop systems where the developer remains the orchestrator. Tools like Cursor enable this symbiosis: the AI handles boilerplate, algorithm implementation, and optimization suggestions, while the human provides direction, validation, and edge-case handling. Karpathy illustrated this with examples from his recent projects, such as building web applications or data pipelines, where 80-90% of the code emerges from English descriptions.
This methodology aligns with broader industry trends. Cursor, for instance, has gained traction among developers for its seamless integration of LLMs into the VS Code ecosystem. It supports models like Claude 3.5 Sonnet, GPT-4o, and others, offering features such as inline edits, chat-based debugging, and codebase-aware completions. Karpathy praised its ability to maintain context across files, reducing the need for repetitive prompt engineering. Similarly, he referenced using the Claude web interface for exploratory coding, where he can paste code blocks and request modifications conversationally.
Karpathy’s endorsement carries significant weight given his pedigree. At Tesla, he led the vision team responsible for Autopilot’s neural network architecture, scaling computer vision models to process vast video datasets. His OpenAI tenure contributed to foundational work on GPT-2 and reinforcement learning. Post-Tesla, through his YouTube lectures and nanoGPT repository, he has educated hundreds of thousands on building LLMs from scratch. His pivot to “English coding” suggests that even experts are adapting to AI-augmented development, potentially reshaping software engineering education and practices.
Critics might argue this approach risks over-reliance on black-box models, introducing subtle bugs or security vulnerabilities from hallucinated code. Karpathy acknowledges these pitfalls, stressing the importance of understanding the generated output and rigorous testing. He views it as an efficiency multiplier rather than a replacement for core programming skills, likening it to using a compiler or IDE—tools that abstract low-level details without diminishing expertise.
The implications extend beyond individual productivity. As LLMs mature, “prompt engineering” is emerging as a skill parallel to traditional coding, with Karpathy exemplifying its potency. His experience challenges the notion that AI agents must be fully autonomous to deliver value; instead, tight human-AI collaboration unlocks practical gains today.
Karpathy’s journey from AI agent critic to English-coding advocate encapsulates the field’s blistering pace. What began as a bold dismissal has evolved into pragmatic adoption, signaling that LLMs are not just assistants but foundational shifts in how code is created.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.