“Andrej Karpathy on LLM cognitive deficits” by Nina Panickssery

Andrej Karpathy, a leading AI researcher and founding member of OpenAI, identifies a significant gap between current LLM capabilities and the requirements of high-level "intellectual" engineering. While AI excels at generating boilerplate code and assisting with unfamiliar languages like Rust through autocomplete, it suffers from "cognitive deficits" when tasked with unique, non-standard architectures. These models frequently default to common internet patterns, such as forcing standard PyTorch DDP containers or injecting unnecessary production "bloat" like try-catch statements, failing to internalize custom logic or specific developer assumptions. This inability to innovate beyond training data suggests that AI is not yet capable of automating the core architectural breakthroughs necessary for self-improving superintelligence. Consequently, Karpathy favors a high-bandwidth autocomplete workflow over fully autonomous "vibe coding" agents, placing AI timelines further out than current industry hype suggests.