Confidence comes from statistics. The statistics say that smaller models are better than bigger ones, ergo I have I have confidence in saying what I did.
If anyone is overconfident, I'm afraid it is you.
If anyone is overconfident, I'm afraid it is you.
These systems can have many models and the systems themselves could eventually arrive at knowledge they know they underperform on, automatically training new models for that subject. The final response wouldn't necessarily come exclusively from the specialized model, it might come through a model specialized at integrating knowledge between multiple models, asking the right questions, verifying and so on.
We're barely a few years into this, so it's premature to know what 100 years of developments will bring.