Software 3.0

Karpathy introduced the Software 3.0 framing at Sequoia Capital's AI Ascent event in April 2026. The progression goes: Software 1.0 is writing explicit rules in code, where you specify every instruction. Software 2.0 is the neural network era, where you provide datasets and objectives and let the weights figure out the behavior. Software 3.0 is where we are now: you prompt an LLM interpreter, and what's in the context window is your lever over the computation. The context window is the program.

The practical implication is radical. In Software 3.0, tasks that previously required writing code can be accomplished by describing goals to an agent. Karpathy illustrated this with a personal story: he built a full app to OCR restaurant menus and overlay food images, then watched someone accomplish the same thing by pointing their phone at a menu and asking a multimodal model directly. The app shouldn't have existed. A lot of existing software, in this frame, shouldn't exist either.

For builders, the reframe matters most in two areas. First, the primary authoring surface shifts from code to context: instructions, examples, tool definitions, memory, and structured outputs. This is why context engineering is becoming a serious discipline alongside prompt engineering. Second, the value that humans provide moves upward: from implementation to judgment, taste, system design, and verification. You can outsource execution to agents. You can't outsource understanding of what's worth building.