AI Engineering and Evals as New Layers of Software Work

AI engineering adds new layers of complexity to software development, focusing on maintaining reliability in inherently stochastic systems. Evaluations, or "evals," are presented as the AI equivalent of software tests, crucial for catching regressions and ensuring quality when making changes to models or pipelines. Evals are challenging due to open-ended tasks and black-box models but can be approached through quantitative methods, qualitative assessments, and using other AI models as judges. The concept of "eval-driven development," which prioritizes defining success metrics before building, is essential for creating reliable and valuable AI applications.