A working hypothesis I have come around to: each time you find a new way to visualize some dataset of things you are responsible for, you will find a new set of things that can be improved. The more parseable bits you can cram into your visualization, the more opportunities you will find. For example one visualization I love for timeseries data is "stacked bar/area chart with time buckets on X, count within time bucket and group on Y, stack groups on top of each other" and then render that chart grouping by everything under the sun you can think of to group on. Another I love is heatmaps, and heatmaps can also be animated for even more bandwidth to your visual cortex. But one superpower is noticing excellent visualizations in the wild (e.g. this one) and either stealing their techniques or thinking about how you could squeeze even more visual bandwidth into the visualization without it becoming overwhelming.

LLMs are very very good at producing visualizations, including novel ones. I am finding a lot of alpha in this.