Back then there weren't really ML libraries to speak of, you had to write your own. Though Symbolics did have a neural network tool and framework they sold.
And you could certainly do things like shuffle lots of data easily for more hardcore processing. For example, Symbolics didn't just have a Weitek floating point accelerator option for their systems, they had vector processing boards for them too that you could use from Lisp. (And even a GPU, the Framethrower, that did full-frame HD with 2D and 3D acceleration! S-Graphics is amazing for the time.)
Also, a lot of AI work back then was focused on symbolic processing. That's largely been eclipsed by the more math-heavy approaches these days but symbolic processing is still around, for example in tools like Cyc.
For machine learning, you want very fast matrix operations for the actual training/evaluation parts and very good data munging for the "get the data from this file/API/database to your ML library". The Python ecosystem is strong at both, which sets up a good feedback loop for even better libraries to be built on top.
That's my take, anyway.