Andrej Karpathy observes that the programming profession is undergoing a "dramatic refactoring" due to AI, leading to a paradigm shift he calls Software 3.0. By late 2025, Karpathy noted that even as an expert, he has "never felt this much behind as a programmer," comparing modern AI tools to "alien technology with no manual".

The core changes Karpathy highlights include:

• Vibe Coding: This term, coined by Karpathy in early 2025, describes a shift where developers build software by describing projects in natural language. In this model, code becomes "ephemeral" and "discardable," as the AI can regenerate or fix it instantly based on the developer's "vibes" or high-level intent.

• A New Abstraction Layer: Programming has moved beyond writing syntax to managing a complex layer of agents, prompts, context, and memory. Karpathy suggests that developers who master this "programmable layer of abstraction" can become 10 times more powerful than traditional coders.

• Context over Prompting: He advocates for "context engineering"—the art of feeding the model the exact information it needs to solve a problem—rather than just "prompt engineering".

• Managing Stochastic Systems: Unlike traditional deterministic code, AI systems are "stochastic, fallible, and unintelligible". The new role of the engineer is to build reliable mental models to manage these unpredictable entities.

• The "Decade of the Agent": While many predicted 2025 would be the year of AI agents, Karpathy views it as the start of a decade-long project. He believes current models still lack the "hippocampus" (memory) and "cerebellum" (coordination) needed for true autonomous agency.

For those looking to keep up, Karpathy suggests that building small projects—such as his own NanoGPT or other LLM-based tools—is the only way to truly understand these new systems.