Andrej Karpathy Details AI Coding Tools' Current Limitations in Novel Project Development

San Francisco, CA – Andrej Karpathy, a prominent figure in artificial intelligence and former director of AI at Tesla and co-founder of OpenAI, recently revealed that current AI coding tools proved "net unhelpful" in the development of his latest open-source project, nanochat. This statement, shared around October 13, 2025, sparked discussion within the AI community, highlighting the ongoing challenges in AI generalization and practical application for novel tasks.Karpathy, known for his deep learning expertise, explained that while he utilized autocomplete, attempts to leverage more advanced AI agents like Claude or Codex for nanochat were unsuccessful. He attributed this to the project's unique structure, which deviated significantly from typical internet code distributions, causing AI tools to struggle with context, generate boilerplate, and even suggest deprecated APIs. His experience underscores a critical gap between AI's impressive demo capabilities and its utility in highly specific, non-standard development scenarios.The nanochat project itself is a full-stack training and inference pipeline designed to create a simple ChatGPT-style model, serving as a capstone for his new educational initiative. This open-source release aims to provide a comprehensive, simplified guide for building AI systems from scratch, written primarily in Python, C, and CUDA.This development follows Karpathy's launch of Eureka Labs on July 16, 2024, an "AI-native school" dedicated to transforming education through a "teacher + AI symbiosis." Eureka Labs' inaugural offering, LLM101n, is an undergraduate-level course designed to teach students how to build their own AI, specifically a "Storyteller AI Large Language Model." The course materials are available online, with plans for both digital and physical cohorts.Karpathy's pivot to AI education with Eureka Labs represents his "next round" in the AI landscape, focusing on human empowerment and skill development in an increasingly automated world. He advocates for a "march of nines" in AI development, emphasizing that significant, incremental work is required to transition from impressive demos to robust, reliable products. His current work aims to build "ramps to knowledge," making complex AI concepts accessible and fostering a deeper understanding of the technology's capabilities and limitations.