I recall when Iridium was going up, there was discussion about the complexity of the software that managed the pattern of transmitting packets among the satellites. The description indicated it was a complex software problem and they were struggling to get it right. (I went looking for articles from the time period, but couldn't track one down, however, came across MIT/Motorola FCC application for what would become Iridium.[1])

It's likely(?) that the orders of magnitude more Starlink satellites, their LEO speed, and other factors that impact optimization (amount of data to transmit, type of data, current load on intervening sats, positional changes during transmission, handoffs, etc.) can make this pretty hard. Or... It could be a learning system via neural net, etc.

The notion that the system could get faster and more efficient over time, simply by operating and without the usual trudge of scheduled software releases and upgrades, is an intriguing possibility. Tho somehow, that's not a novel notion...

[1] https://web.mit.edu/deweck/www/research_files/comsats_2004_0...