The availability of free tools, and the availability of much searchable code on Internet means that it is way easier now to have a large number of processor architectures on the repertoire.

20 years ago (or even 10), you could get stuck with something as simple as an UART, and have a hard time to get help from a support engineer just because the datasheet was incomplete. With good Google abilities, you can find sample code for almost every aspect of almost any processor. So you seldom get stuck at the bit and register level.

This makes it much easier to work with many processors. You don't need to be as scared of getting stuck if selecting a different processor for a project. 20 years ago, you had your own "golden edition" code that you knew worked, and wanted to stay with processors where you could continue to reuse this base framework.

Of course, the introduction of simple serial programming methods have also greatly reduced the hardware costs needed to test a new processor. Most new processor architectures can be tested with just a couple of dollars of hardware to program them in-circuit. No need for strange adapters and ZIF sockets to program them - if your existing programmer(s) even supported the required algorithm.

The big decrease in memory costs also means that most applications can be written in a HLL (using freely available compilers) while we 20 years ago just had to count every code and data memory cell and invest time in assembly just to get everything to fit.