One observation was the extra load it created (obviously) doing the learning work. In practice it became a feature you could turn on or off and many seemed too fearful to use it in production. The unpredictability of the performance during the early stages also caused fear. There was talk about running query plan explorations autonomously as soon as the system was turned on and had tables defined with relationships, but that path seemed to deviate from the goal of the project at the time too much, so was ignored.

That said, optimizers don't always get things right the first time, so there were clear cases I saw in early concept stages where running multiple options made a huge difference. For example, testing two different join orders could result in one query that took under a minute and one that took over 20 minutes. So it definitely made an impact.

We found that using learning in other areas also had a meaningful impact, like in saving time during data type calculations.

Generally, I can say, for the project I worked on, there were clear benefits but it didn't progress to a point of learning all the limitations, unfortunately.

I'd say still an area worth research, based on my experience.