Idea Machines: LeWitt, LLMs, and Computing (lewitt.rob.computer)

The piece argues that the conceptual-art insight—“the idea becomes a machine that makes the art”—is finally being realized in software because large language models are compressing the gap between natural‑language intent and executable computation. Around Feb 2025 the author claims it became practical to build agentic conversational shells that plan, act, observe, check, and revise: they decompose goals, pick tools, call CLIs/APIs, write files, inspect logs, run tests, persist state, and consolidate successful runs into repeatable executables. Technically, the LLM functions as both a compiler (proposing and revising execution plans) and a runtime (orchestrating deterministic processes and freeform language), resolving natural‑language ambiguity by iterating against explicit gates—typechecks, tests, schemas, idempotent operations and policy checks. The significance for AI/ML is paradigmatic: lowering the cost of formalization widens who can create software and how we think about programs. Specifications become first‑class, the code/prose boundary smooths, and workflows can be long‑lived and self‑healing. That generality trades consistency for coverage—LLMs produce responses for any input, so the semantic quality of specifications and deterministic gates determine reliability. Pedagogically and socially, wider participation could spread systems thinking and democratize agency over computation, changing software from monolithic artifacts into composable “pattern”‑based creations.