Software 3.0
Software 3.0 is a conceptual framework proposed by Andrej Karpathy that reconceptualizes Large Language Models (LLMs) as a new computing paradigm rather than specialized tools. In this vision, LLMs function as general-purpose computational systems capable of executing diverse tasks through natural language interfaces. This represents a shift from traditional software development, where explicit code is written by humans, to a model where the “program” is implicit within the weights of the neural network and accessed via prompts. The term distinguishes this era from Software 1.0, which relied on manual coding, and Software 2.0, which introduced machine learning models trained on data but still required significant human engineering for integration and deployment.
Computational Paradigm Shift
Under this framework, the neural network itself acts as the operating system, managing resources and executing logic through probabilistic inference rather than deterministic instruction sets. Users interact with the system using natural language, effectively writing programs that are interpreted by the model’s internal representations. This approach lowers the barrier to entry for software creation, allowing non-experts to define complex behaviors through descriptive prompts rather than syntax. The “code” is no longer a static text file but a dynamic state derived from the model’s training data and context window.
Implications for Infrastructure
This concept suggests a future where infrastructure is optimized for high-throughput inference and context management rather than traditional compilation and execution. Development workflows shift from debugging explicit logic to curating training data, refining prompts, and managing model versions. While it offers flexibility and rapid prototyping capabilities, it also introduces challenges related to determinism, transparency, and computational cost. The framework remains a theoretical perspective on the evolving role of AI in software engineering, highlighting the potential for models to serve as the primary interface between human intent and machine execution.