Andrej Karpathy, former head of AI at Tesla, believes we are witnessing a seismic shift in software. In a talk titled “Software Is Changing (Again)” at YC AI Startup School, Karpathy introduced the concept of Software 3.0, where large language models (LLMs) make neural networks programmable via natural language. “Neural networks have become programmable with large language models… it’s a new kind of computer,” he said.
Karpathy traced the evolution from traditional coding (Software 1.0) to training neural networks (Software 2.0), and now to prompt-based programming (Software 3.0), where tools like ChatGPT are changing the interface of computation itself. Highlighting his own creations like MenuGen and Gitingest, he emphasized that LLMs are not just tools—they are “people spirits,” powerful yet fallible systems that require “human-in-the-loop” oversight.
Cautioning against AI agent hype, Karpathy warned, “We have to keep the AI on the leash,” and criticized declarations like ‘2025 is the year of agents’, arguing that true autonomy is a decade-long journey. He urged developers to build augmented systems that empower humans, rather than fully autonomous agents that sideline them.
Comparing LLMs to CPUs and cloud-based utilities, Karpathy said we’re back in the “1960s of computing,” relying on centralised cloud power while waiting for the age of local, personal LLMs. Until then, LLMs act as the internet’s programmable intelligence layer—metaphorically delivering “electricity,” but with the capacity to reshape every corner of software development.