Embeddings raise the possibility of exploring ideas spatially. The central challenge is how to balance between automatically laying out ideas semantically and allowing the user to group or cluster themselves.

User movement mirrors the experience of the best physical brainstorming - cutting up pieces and rearranging, examining, rearranging again.

Embeddings promise automatic clustering and hopefully surfacing new connections.

Something hybrid seems like the answer. But you don't want the two fighting each other - you don't want automatic embedding-based movements disrupting deliberate user ones.

This is the problem autocomplete (mostly) solves for linear writing - suggestions that don't have to be accepted.

I can kind of imagine autocomplete and autosuggestion for a spatial canvas (lower opacity suggested nodes or movements) but it feels like it would overwhelm quickly.

Surfacing related nodes in a sidebar that you could then drag out seems like a good compromise, thought it may not feel particularly inspiring or novel as an interface.

I wonder if there's anything you could do layering a layer of suggestions on top of the current state, like an annotated transparency layer... Diffs as a concept should be in the mix too.

I also always think of 'ghost cars' (or in platformers like Super Meat Boy "ghost characters") in videogames that show you your previous paths... that's probably worth its own post talking about why it works and how it might apply.