Exactly, I think all those base models should be weeded out from this nonsense, kardashian-like labyrinths of knowledge complexities that just makes them dumber

Exactly, I think all those base models should be weeded out from this nonsense, kardashian-like labyrinths of knowledge complexities that just makes them dumber by taking space and compute time. If you can google out some nonsense news, it should stay there in search engines for retrieval. Models should be good at using search tools, not at trying to replicate their results. They should start from logic, math, programming, physics and so on, similar to how education system is suppose to equip you with. IMHO small models can give this speed advantage (faster to experiment ie. with parallel diverging results, ability to munch through more data etc). Stripped to this bare minimum they can likely be much smaller with impressive results, tunable, allow for huge context etc.