Software 3.0 is the third step in a sequence Andrej Karpathy sets out for how a computer gets programmed, after Software 1.0, in which humans write explicit code, and Software 2.0, in which humans assemble datasets, objectives and neural network architectures and the program is learned into weights. In Software 3.0 the program is expressed as prompts, context, tools, examples, memory and instructions, and the large language model is the interpreter: what sits in the context window is the lever over it, and the computation it performs is over digital information. His account of how this became possible is that a model trained on a sufficiently large set of tasks — the internet containing many — becomes, implicitly, a programmable computer.
Usage
The example Karpathy uses to make the shift concrete is installation. Targeting many platforms and machine configurations with a shell script, he argues, forces the script to balloon with conditionals because the Software 1.0 universe requires spelling out every detail in exact code. The alternative he points to is an install procedure distributed as a block of text to paste into an agent: the agent reads the local environment, takes intelligent actions, debugs errors in the loop and completes the setup. He describes that as a different kind of program — less precise, but more adaptive — and draws the practical consequence that “what is the piece of text to copy-paste into your agent” is itself now a programming question.
His second example runs the argument further, to software that stops existing. A menu-photo app he built in the ordinary way needed OCR, an image generator, a frontend, deployment, auth, payments, secrets and infrastructure; the Software 3.0 version is handing the photo to a multimodal model and asking it to render the dish images onto the menu image directly. On his reading the old stack was scaffolding around a transformation the model can now perform itself, and the founder-facing consequence is that some applications should not exist as applications at all. He extends the same logic speculatively to hardware, imagining neural networks as the host process with CPUs as coprocessors, while saying the path there will be piecemeal.
The framing also carries past code. Karpathy argues that classical programs worked over structured data, and that the newly automatable thing is information processing that was not previously programmable at all — his example being an agent that incrementally compiles messy documents into a persistent Markdown knowledge base of summaries, entity pages, concept pages, contradictions and cross-links. His recommended question is therefore not only which existing workflow AI can speed up, but which information transformation was impossible before and is now natural.
The framing is also used by others as a period label rather than as Karpathy’s specific programming argument. Software 3.0 시대, Harness를 통한 조직 생산성 저점 높이기 takes “the Software 3.0 era” as its setting and asks an organizational question inside it: if the program is now prompts, context and instructions, then those artifacts are what a team should be distributing to each other the way it once distributed shared libraries. That post’s proposal — team knowledge held as agent plugins, which it says have the potential to become an Executable SSOT, and published through a marketplace — treats the module’s contents as having changed from code to prompts and agent logic while the engineering practice around it carries over: it maps shared modules onto workflow plugins and library publication onto marketplace upload, and argues that the code review and quality management applied to a shared module should apply to an AI workflow too. That post’s author presents all of it as direction and hypothesis rather than as a result. Read that way the term names the material an organization works in rather than only the act of programming a model.
When It Applies
The paradigm’s trade is stated plainly by its author: it buys adaptiveness at the cost of precision, so it suits work where the environment varies and an exact specification would be brittle, and it assumes a model capable enough to act and debug in the loop. It is one person’s framing rather than an industry-agreed term, and this account of it comes from a talk write-up whose prose was produced by a model from the event transcript and then checked by him.
Related Terms
Context engineering, Jagged intelligence, Prompt engineering, Agentic Engineering, Executable SSOT, Raising the Floor