Stephen Downes

Another post to keep you current on AI terminology. The word du jour is 'loop'. The concept is simple enough, and common in computer science: to repeat a process over and over - sometimes iterating values, other times iterating outcomes. This article looks at the concept as it applies to AI and identifies four (or five) distinct types of loop: an execution loop that "iterates on steps within one task"; a task loop (or 'Ralph' loop) that "restarts a coding agent against the same specification over and over," iterating over a single artifact until it matches the spec; the product loop that "iterates on a codebase and its backlog, continuously" addressing external conditions such as "new issues, production logs, user feedback, review outcomes"; and the system loop that " iterates on prompts, harnesses, model choices, and the evals themselves." There's also a potential fifth outer loop, the oversight loop, "where goals get set, budgets get allocated, and work gets culled." This is where humans should be involved - the article argues that this is the only place humans should be involved. "That inner loop is capability. The outer loop is agency."