Consequences of advertising and enshittification on the Internet
Summary
A detailed personal take on agentic coding and AI-driven testing, arguing that large language models can accelerate data analysis and bug finding but suffer from high variance and reliability issues. The piece critiques benchmarks, outlines practical workflows like fuzzing and automated test generation, and emphasizes the importance of structured evals, feedback loops, and multiple agent perspectives to manage false positives and achieve meaningful productivity gains.