I'm not thinking about LLM calls being the "nodes", this already exists, but has nothing to do with exploring new interfaces - in a way, LLM being deeply async, make a lot of sense to embed in Computational DAGs
I'm more interested in what kinds of interactions could be built by adding LLM-fuzziness to the way nodes-and-wires interfaces are usually presented
maybe when trying to connect output to an input with mismatched type we get an adapter (or set of adapters) generated automatically?
maybe dragging a connection (both from "output" and "input") into empty space provides a box with suggestions for both what's already in the system, and what makes sense to generate based on the rest of the things on the canvas
maybe these generated suggestions don't come as a single new node, but as a bunch of them, all connected at once, so they evaluate and the output can be inspected in a sort of "purgatory" before joining the graph "for real"
maybe there's no "grouping" of nodes, but that operation instead rewrites the subgraph into textual code and creates one new node
same could work as "ungrouping" - just in the opposite direction