Software 3.0 in Production

Andrej Karpathy's "Software 3.0" framing names a shift that working engineers can already feel. Software 1.0 is hand-written code. Software 2.0 is learned weights. Software 3.0 is the layer on top: natural-language prompts that program the model, where intent is stated in prose and compiled into behavior by the model itself.

English has become a programming layer — fuzzy, expressive, and now load-bearing. That is real leverage. It is also a real hazard, because the model that compiles your prose into action does so nondeterministically, and it returns fluent output whether or not it understood what you meant.

The discipline that makes this safe in production fits in one line: specify in English, verify in code. Types, schemas, assertions, tests, and evals are the deterministic safety net stretched beneath a layer of fuzzy natural-language instruction. Fluent is not the same as correct, and only the code-level checks know the difference.

This paper covers what Software 3.0 changes for teams — spec-writing as a core skill, code review that moves from lines to behavior, juniors who learn the model before they learn the system — and how "vibe coding" matures into professional practice through guardrails, tests, review gates, and clear accountability.

It closes with a field observation from our delivery work: Software 3.0 holds up in production only when it is paired with rigorous verification. English in, code-checked out. Teams that adopt the expressive half and skip the verification half ship confident bugs — quickly, and at scale.