alt Hacker NewsRight on.
1) I want to minimize global complexity, which by definition maximizes local reasoning.
2) I want to make the vast majority of bugs simply unrepresentable - taking it past Rust, and even past Pony - WHILE allowing shared mutable memory, but without requiring a PhD to use.
The goal is in EASY mode, it's barely harder to use than Ruby or Python (just the occasional pedant compiler error that has automatic options to fix itself most of the time). You don't even have to supply types or compile. It has a REPL, etc.
When you bump to DEFAULT mode and then to STRICT mode, all the annotation is automatic - your code just might look "ugly" if you like having no types anywhere etc.
But DEFAULT & STRICT mode give people and LLMs everything they need to know to understand the effects of an individual function.
I have some similar goals. Have you considered leaning more into inference than gradual typing? One pattern I like is allowing the compiler to develop a more complex mental model, but keeping it straightforward for users -- you can do that with inference, ownership, purity, effect types, etc. What I actually think is really tantalizing is using tooling to fill in some of those gaps -- for instance, the editor could know types, required capabilities etc, without the user ever needing to type anything, but when the user needs it, they can find it, query it, test against it.