AI Software Development – What Does The Data Say?
Summary
The article surveys recent sources on the use of LLMs in software development, arguing that autonomous coding remains improbable, context window limits are tighter than advertised, and output quality is highly dependent on context engineering and quality gates. It highlights findings on the limited reliability of LLMs, the noise introduced by repository files, the benefits of demonstrations over descriptions, and the gap between increased code output and actual improvements in software outcomes, with psychological and energy considerations also discussed.