Andrej Karpathy introduces "Software 3.0," a new computing paradigm where programming is done by prompting LLMs, fundamentally changing software development.
He defines "agentic engineering" as a professional discipline for leveraging AI agents to achieve massive (greater than 10x) productivity gains while maintaining high-quality standards.
The capabilities of AI models are "jagged," excelling in verifiable domains like coding due to reinforcement learning but failing at simple common-sense tasks, a limitation developers must navigate.
A major shift is needed towards "agent-native" infrastructure and documentation, designed for AI agents to consume directly, rather than the current human-centric systems.
8 quotes
Concerns Raised
The 'jagged' and unpredictable nature of LLM capabilities, leading to failures in common-sense reasoning.
Most software infrastructure and documentation is still designed for humans, creating a bottleneck for AI agent automation.
Corporate hiring processes have not adapted to identify and evaluate crucial 'agentic engineering' skills.
Opportunities Identified
Engineers mastering 'agentic engineering' can achieve productivity gains far exceeding the traditional '10x engineer' benchmark.
Building 'agent-native' infrastructure and tools to serve as the foundation for the Software 3.0 paradigm.
Creating new applications that were previously impossible by leveraging LLMs for complex information processing, not just speeding up existing tasks.
Applying reinforcement learning techniques to any business problem within a verifiable domain to achieve superhuman performance.