Andrej Karpathy: Software Is Changing (Again)

Karpathy describes Software 1.0 as code written by people, Software 2.0 as neural network weights trained from data, and Software 3.0 as prompts that program large language models. He says developers will need to choose among these approaches and move between them as each has different strengths.

Karpathy compares LLMs to utilities, semiconductor fabs, and especially operating systems. He likens the model to a CPU and its context window to memory, while noting that today’s costly compute keeps much of this capability in the cloud. He also describes LLMs as powerful but fallible: they can hallucinate, make uneven errors, forget context between sessions, and be vulnerable to prompt injection.

Karpathy argues that useful LLM apps combine familiar interfaces with model assistance, orchestrated calls, and tools for reviewing the results. In Cursor, for example, people can accept or reject code changes and adjust how much work the AI handles. He favors small, reviewable changes and fast verification over large, unsupervised agent actions.

Karpathy says prompts let more people build software without years of programming experience. He describes vibe coding an iOS app and MenuGen, which generates pictures from restaurant menus. The prototype took hours, but authentication, payments, domain setup, and deployment took about a week, showing that making software production-ready involves work beyond writing code.

Karpathy proposes treating agents as a new kind of user alongside people and traditional software. He points to llms.txt, Markdown documentation, curl commands in place of click-by-click instructions, and tools such as MCP and Gitingest as ways to make information and actions easier for agents to use.