For what it's worth, Scheme was the first programming class that I took at UIUC around 1995, as I recall.

I had used C++ for several years to make shareware games, so I took a test to challenge some programming courses. I vaguely recall doing well, but my advisor encouraged me to take them anyway. I'm glad that I did, because I had little understanding of theory.

Funny story: the instructor never mentioned that we could use more than one line of code. So every single piece of homework that I handed in, and every test, was one giant line of nested logic. Which worked better than one might expect, and completely changed how I wrote code from that point forward. That's how I made the connection a decade later that functional programming is akin to a spreadsheet, as are higher-order method chains and immutable variables.

I think of Clojure as being a layer above Lisp, sort of like how Swift might be considered a layer above Objective-C/Smalltalk. However, bare Lisp has problems around not quite giving enough out of the box. It's minimalist enough that developers end up reinventing the wheel for things that should probably be provided by a layer/library similar to Scheme or Clojure.

To digress, I feel that mutable variables and even monads are a code smell in functional programming since they can cause impurity. They're more of a crutch to ease conversion of code from imperative languages. However, monads can be useful to simulate every path through a program, sort of like superposition in quantum mechanics and SAT solvers. So they aren't necessarily bad, just taught incorrectly, probably because they're so hard to grok.

I'd vote to settle on a series of layers like Common Lisp -> Scheme/Racket -> Clojure/Elisp, with the final layer providing the intersection of features available from the most widely-used Lisp variants. Note that this is specifically to form a bridge from imperative languages, so research work might need additional DSL features brought forth from the Racket layer.

Edit: I forgot to mention that Scheme is a good fit for genetic algorithms, see books by John Koza (no affiliation). My feeling is that we haven't seen anything yet regarding what problems AI can solve, since it's having to do it the "bare hands" way with LLMs and pattern matching.