If you want one talk that explains the moment we are in, this is the one. Karpathy's "Software 3.0" reframes the last three decades of programming as three paradigms:
- Software 1.0, where humans write the code that runs.
- Software 2.0, where neural networks are trained and the weights are the program.
- Software 3.0, where the desired behavior is described in English and the LLM is the interpreter.
The framing matters because it puts vibe coding in a longer arc. We are not "talking to AI." We are writing programs in a new language, natural language, that happens to be far more expressive and far less precise than every previous programming language.
The minute most worth your time
Karpathy's claim that mature Software-3.0 codebases will include English prose alongside the source files, and that the AI infrastructure will treat them with the same seriousness as .py or .ts. We are already there. CLAUDE.md, cursorrules, and system prompts checked into git are exactly that. The mental shift is to stop treating those files as "configuration" and start treating them as code.
Why it earned the slot
It is the cleanest existing framing of what we are actually doing when we vibe code. The three-paradigm model gives you something to point at when a skeptic asks how this is different from the last "AI will replace programmers" cycle. It is not that cycle. It is a different one. Software 3.0 sits alongside 1.0 and 2.0; it does not replace them. The frame holds up.
Cite this when someone needs the long view. Forty-five minutes total, and the first fifteen are the high-leverage ones for anyone skimming.