Software 3.0
Software 3.0 is Andrej Karpathy’s frame for programming with LLMs as interpreters: context, prompts, examples, files, images, and instructions become the lever over a model that performs computation in digital information space.
Key points
- Software 1.0 is explicit hand-written code; Software 2.0 is learned neural-network weights trained by arranging datasets, objectives, and architectures; Software 3.0 is programming through context and prompting [src-055].
- The LLM behaves like a programmable computer once it has been trained across a sufficiently broad multitask distribution from internet-scale data [src-055].
- OpenClaw’s installation instructions illustrate the shift: instead of a cross-platform bash script, users paste instructions into an agent that inspects the environment, adapts, debugs, and acts [src-055].
- MenuGen illustrates a more radical shift: a conventional web app that OCRs a menu and generates dish images can collapse into a direct prompt to Gemini/Nano Banana that transforms an input image into an output image [src-055].
- The implication is not only faster programming; some new information-processing products become possible because natural-language and multimodal source material can be recompiled into new artifacts [src-055].
- Rory Richardson gives a complementary enterprise frame: agentic AI compresses abstraction layers between human intent and working systems, so teams can move from ideation toward ones and zeros through fewer manual translation steps [src-057].
Related entities
- Andrej Karpathy
- Rory Richardson
- Amazon Web Services
- OpenClaw
- Claude Code
- Gemini
- Imagen 3 (Nano Banana 2)
Related concepts
- Agentic Engineering
- Agent-Native Infrastructure
- Karpathy LLM Wiki Pattern
- Product Overhang
- Abstraction Layer Compression
- AI Development Lifecycle
- Coding Democratization