当描述足够时，软件变得可塑：LLM 编程与 Ambient Computing 实践

2026-10-06 · Notes · Thinking Notes

LLM-based software engineering keeps pushing the threshold of implementation down. Small tools I would never have built, because building them cost more than they were worth, now take an afternoon. I describe what should happen, and the code gets written, run, and corrected.

It is still programming. I just do it in a different language: I describe the behavior, and the model translates it into a language the computer runs. AI coding took off because the tooling was already there, and now that tooling sits between my sentences and the machine.

Below that threshold, software starts to feel malleable. It stops being a product I adapt to and becomes material I can bend to fit a room, a habit, or a single afternoon’s problem.

That is also what makes ambient computing approachable for one person. Dynamicland needed a research group and a whole new system to make a room computational. But a room is full of small, specific needs that no product will ever cover: a shelf light that comes on when I ask where the soldering iron is, a key that asks twice before switching the studio off while a door is open, a wall screen that swings the model’s door when the real one opens. None of these is worth a product. Each is worth an afternoon. I’m trying this in our studio now and logging it in Ambient Computing.

We don’t have systems for this yet. Today it is a codebase, a 3D model, and a lot of glue. But if it materializes as some kind of harness, a place where you describe a behavior and the room takes it on, then given the right control surfaces and output devices it could start to feel like Dynamicland-level manipulation of the environment, without everyone having to build their own Realtalk.