Same trend as from special-purpose computers in mid-20th century to universal computers today.
This is dubious: there are many trends in computing towards specialisation today, e. g. TPU, however, NVIDIA bets on more universal GPUs. Models are much more malleable, so there is no problem in creating a more specialised model. But I don't have a sufficient intuition to judge whether universal pre-trained models or more specialised ones will work better.
Hawkins' arguments for universal AI:
A Thousand Brains