Probably there is a subset of extremely smart people who can write good and very reliable software in a dynamic language without going mad, but I am not among them. In the end I just left the language and now I am much happier with a mix of TypeScript, OCaml and Rust
Also the reliability they are talking about is quite different from the reliability that most people (especially ones who have not worked on erlang-y systems) think of when they hear the word. The Erlang reliability is about finding bugs and patching them (often in production) and not about avoiding them in the first place. Also it's about avoiding total system failure because of some bug in a part in it. This is a nice to have thing in all systems, but Erlang makes it easier to implement.
Of course, they also like to avoid bugs, but for that they use testing. And of course, we have to mention dialyzer which is something like TypeScript for Erlang but a lot less polished and with practically no tooling support. It is used in every significant Erlang project but the process is IMO quite sub-optimal and the price you pay (in suffering) is quite high.
This is kind of the key bit.
To caveat this - I'd like better typing in Erlang, and when i last used it in production I added type annotations to everything and ran Dialyzer as part of the build. But this is personal preference, and I'm not convinced it was the best use of my time; just the most conservative way to approach things for a system I needed extremely high uptime on.
You have to write tests anyway. Just because a statically typed system compiles doesn't mean your test load is reduced. While it eliminates certain classes of bugs, those same classes mostly get covered just by writing functional tests of sufficient coverage.
Between that, and Erlang's supervision tree (so that if you DO get a type error at runtime, it would be in a path you haven't adequately tested, and so not your expected 'happy path', and as such, can reasonably be restarted), it strikes me as a reasonable position to take.
Regarding testing - I think it depends on the type of application you are writing. If you need 99.99% uptime, probably you have to have almost full test coverage.
But for the kind of software that I write (at least lately), I rely on "let's make illegal states unrepresentable" to a large degree, then write a small number of regression tests (or fix the types) when we find bugs... This has proven (at least to me) to be good enough, especially compared to my clients' previous experience, which is usually very poor: extremely buggy software in a dynamic language with almost no tests and docs. Perfect quality is usually not needed or at least not worth the price.
It's not that you can't make good code in any language, but the cost is often not worth it; anything that makes it easier and more reliable saves you time and effort in the long run.
Hence I am waiting on Gleam before I invest in Erlangy stuff. In the meantime I will explore the actor model of concurrency with F# Mailboxes.
Static typing severely hinders hot code-reload, which was deemed more important for long running systems.
And whether your types are runtime-checked or statically checked doesn't really matter with regard to the invariants that hold for your code. In the former case, you'll have to be able to handle type errors at runtime as well, which is reasonably easy in Erlang due to the strong isolation it provides.
When I write Erlang, I'm not necessarily concerned with avoiding errors. The philosophy of the system is that errors will occur, and the focus is on making sure errors do not compound into system failure (supervision tree helps with this), that they're easy to see (appropriate logging of crashes), and that it's easy to deploy fixes (hotloading).
I'm sure there are many ways to avoid some errors in production, but more focus is on responding to errors than on preventing them. At least in my experience.
Will you not be testing your code anyways? Will you not have unit tests and integ tests and have it run in a devo environment first? Before going to production?
If so, then what would the gap be? If you catch those same type errors as you test, and you will be testing anyways?