It is the observation that all software is built of 3 layers -
Inbound IO, what you call imperative shell Business logic core, what you call functional core Outbound IO, what you call again imperative shell.
The problem with the terms in functional programming is that when they say side effects, in most cases they mean IO.
The functional core does not have to be functional for you to get the benefits - easy testing, easy to reason about, easy to develop. In fact, in some cases, functional programing is the wrong tool while having a business logic core that is separated from IO is still a very valid architecture.
What I find fascinating is that, at least for me, using GenServers/Agents/Tasks has made this distinction much more obvious than it has been in other languages.
I've built horrifying systems that muddled things together, only to realize quite far in how much of a mess I made, but with my Elixir projects I usually realize what I've done pretty quickly, and I end up course-correcting earlier on.
https://www.destroyallsoftware.com/talks/the-birth-and-death...
I end up recommending this video at least once every year.
(EDIT: Another one from Gary that pairs well is "Boundaries": https://www.destroyallsoftware.com/talks/boundaries . The core idea is that values -- data, not code -- should form the boundaries between distinct systems.)