A technique works, so I use it again. The second attempt is faster. The third is easier to explain. Eventually the technique starts looking like the obvious answer before I have properly considered the question.
That is the pattern I found interesting in J Bostock's "You Are Not Immune To Mode Collapse". The essay uses mode collapse as an analogy for skills and institutions becoming narrower through repeated success. Familiar work receives more practice, becomes easier to do well and attracts still more work.
The analogy is useful, but it is an analogy. In machine learning, mode collapse describes a specific failure to represent the diversity of a target distribution, especially in generative models. An artist returning to a theme is not automatically demonstrating the same mechanism, and repetition can be a deliberate source of depth.
What I take from the essay is a question about choice. Am I returning to an approach because it fits the problem, or because it is the approach I can currently execute with the least discomfort?
Software makes this easy to miss. A familiar framework reduces setup time. A known visual effect produces a pleasing result quickly. A successful prompt can be reused. Each choice is locally reasonable. Together, they can make a collection of supposedly different experiments converge on the same shape.
One way to investigate that is to change a constraint rather than demand originality in the abstract. Make the interaction work without animation. Explain the idea without a screen. Build a scene using fewer elements. Choose a problem where the familiar technique is inconvenient. A concrete constraint exposes which parts of the process are habits and which are essential.
I also think unfinished experiments deserve a different standard from released products. A product should become more dependable. An experiment can be useful because it reveals an unfamiliar difficulty. Judging both only by immediate polish gives the familiar approach an automatic advantage.
This does not require a dramatic reinvention every month. Small variations can preserve options: reading a field outside the usual set, trying a different representation or returning to an older idea with a new tool. The important thing is that the variation changes what I notice, rather than only the appearance of the output.
Time and attention matter here. Exploring an unfamiliar approach usually looks inefficient at first. If every hour must produce a predictable result, the safe method keeps winning. Leaving a little room for uncertain work is a practical condition for maintaining range.
I do not want to discard expertise. I want to avoid confusing expertise with the boundaries of what is worth attempting. A useful practice can have reliable methods and still leave a door open for a project that makes those methods feel temporarily inadequate.