True power of AI does not come from pretrained models. AI is all about recognizing patterns from input to predict next sequence of tokens. One could even assume fine tuning an existing model would result in worse performance for custom dataset with same amount of training compared to starting from empty model network. This is due to prebuilt bias in existing model. --> existing weights do not help to adapt to new data.
Next phase of AI will be custom models, where everyone can train their model from scratch on their own devices. Recognizing true potential of AI is not in forming new sentences, its in forming new patterns.