Karpathy: LLMs Are Rewriting the Rules of Software — All Code Will Be Rewritten Many Times Over
Original: Karpathy: LLMs Are Rewriting the Rules of Software — All Code Will Be Rewritten Many Times Over View original →
A Fundamental Shift in Software Development
AI researcher Andrej Karpathy (@karpathy) shared a profound take on February 16, 2026, arguing that large language models are completely transforming the constraints landscape of software development. He wrote that it must be a very interesting time to be in programming languages and formal methods, because LLMs change the whole constraints landscape of software completely.
Why LLMs Excel at Code Translation
Karpathy explains that LLMs are especially strong at translation rather than de-novo code generation for two key reasons: first, the original codebase acts as a highly detailed prompt; second, it provides a concrete reference for writing tests. This advantage is already visible in the rising momentum behind porting C to Rust and growing interest in upgrading COBOL legacy codebases.
Open Questions and New Opportunities
Karpathy notes that even Rust is nowhere near optimal for LLMs as a target language, raising intriguing questions: What kind of programming language would be optimal for LLMs? What concessions, if any, must still be made for human readability? He concludes that it feels likely that we will end up re-writing large fractions of all software ever written many times over — a bold prediction with profound implications for the entire software industry.
Industry Resonance
The tweet garnered over 1.07 million views and nearly 8,000 likes, reflecting broad interest from developers and AI researchers worldwide. It signals that LLM-driven code translation and legacy system modernization are poised to become defining trends of the coming era.
Related Articles
AI researcher Andrej Karpathy argues that programming has fundamentally changed over the last two months, particularly since December when coding agents started actually working. Developers are shifting from writing code to directing and managing AI agents in parallel.
Andrej Karpathy shared a practical tip: appending structure your response as HTML to any LLM prompt and viewing the result in a browser. He also outlined a broader vision for the evolution of human-AI interfaces, from text to interactive neural simulations.
Anthropic has introduced Natural Language Autoencoders (NLAs), a new interpretability technique that trains Claude to translate its own internal activations into human-readable text—enabling safety audits that can uncover hidden model motivations.