[jcparkyn]

I think the opposite, TS is much more predictable than hindley-milner inference, mostly because the inference only flows forward.

Yes, there are a couple of weird edge cases with arrow functions in generics, and it gets a bit more complicated with narrowing, but aside from that this never a problem I've run into. IMO the verbose error messages for complex types are a much bigger problem, which could be solved with better UIs for error messages (syntax highlighting and code folding).