it compiles to bytecode but the bytecode engine is not that fast.
i had a phase when I was using PyPi a lot for branchy "old AI" kinds of workloads and felt it was an easy win but since then it has been either numpy or PIL or pytorch doing the heavy lifting or scripty stuff like uploading files to S3 where performance doesn't matter a lot.
I will grant that Common Lisp can be compiled to run amazingly quickly!
What he probably meant is that compilation is available at runtime to the user, allowing you to JIT stuff at will.
This is how https://github.com/marcoheisig/Petalisp#why-is-petalisp-writ... and https://github.com/numcl/specialized-function can exist.