Extreme Harness Engineering for Token Billionaires: 1M LOC, 1B toks/day, 0% human code, 0% human review — Ryan Lopopolo, OpenAI Frontier & Symphony

The discussion centers on Harness engineering, a methodology for AI-driven software development, with Ryan Lopopolo from OpenAI. Lopopolo details his team's experience building an internal tool using AI agents that wrote over a million lines of code with minimal human intervention. A key aspect involves inverting control, allowing the AI to manage its environment and choose its tools, rather than operating within a predefined scaffold. The conversation covers the team's iterative process, adapting to model updates and optimizing for agent productivity by enforcing strict build time limits. They also explore the concept of "ghost libraries," distributing software as specifications for AI to reassemble, and the potential for AI to handle tasks such as code review, dependency management, and even humor generation.