Idea Machines: LeWitt, LLMs, and Computing

Love opening with LeWitt and Eno.

I really like this para: > Another significant change is that workflows can be long‑lived and self‑healing: if run on always-on compute like servers, agents can schedule and repair persistent or recurring work like report generation, data pipelines, or crawl-and-summarize tasks, always resuming from state,with a general purpose recovery mechanism. The significance of generality is worth calling out, because in some sense it is the crux of what is new. These systems are inherently more robust than traditional symbolic programs because they have a modality that is complete at the expense of consistency: all inputs will produce some response from an LLM, whereas most possible inputs will crash most classical functions. Whether or not the general handler will “do the right thing” comes down to the quality of the specification.

And this one: > When the shape of programming changes to encapsulate multiple levels of abstraction, the basic unit of work changes from specifying lines of code to specifying any of: line of code, logic, function, capability, behavior, application. Our new programs will buy determinism with checks. Type systems, property‑based tests, schema checks, idempotent operations (declarative codegen), and permission policies make runs repeatable and failures bounded; agents work is graded by those deterministic gates.