software-paradigms-karpathy

Karpathy's software paradigms define programming evolution: Software 1.0 uses traditional code, Software 2.0 uses neural network weights, and Software 3.0 uses LLM prompts. This framework clarifies how prompt-based programming represents a new form of software design.

Map tasks to software paradigms by evaluating requirements: use traditional code for deterministic logic, neural network weights for perception tasks, and LLM prompts for cognitive or generative tasks. This mental model guides precise architecture decisions and task routing.

Use LLM prompts when tasks require cognitive reasoning or language generation, and traditional code for deterministic orchestration. The software 3.0 paradigm treats prompts as programming, ideal for AI-enabled product design where explicit rules are difficult to define.

Design hybrid AI architectures by orchestrating Software 1.0 for control flow, Software 2.0 for perception, and Software 3.0 for prompt-based reasoning. This combines deterministic logic, trained models, and LLM prompts to build robust AI-enabled products.

LLM prompts represent a new software paradigm because they function as programming instructions for AI models. In Software 3.0, prompts replace explicit code logic with natural language directives, shifting how developers design and execute application architecture.