What's the reasoning for not just using points from a continuous spectrum? His approach tries to both make adjacent colors distinct while also making them blend well with nearby colors, which he admits are contradictory goals.
It's a clever way to automatically create visual classifications of random-sized collections of data points. The goals are not contradictory, they just constraint each other.
For the sorting diagrams these colours start out random and end up sorted, in neither case, nor in the transition period in-between, do I see an obvious benefit to this method of choosing colors.
I'm now wondering if this could be applied to choosing colours for auto-generated pie charts, bar charts or map keys and similar.