I am very irritated by the focus on language models and image / video /audio predictors and feel that too many resources are being used exclusively for those. The overall impact is obviously trampling down any of the good stuff.

This is basically the issue, especially in regards to the monopolization of hardware by big LLM companies. It locks out smaller scale developers who might’ve been able to create and train more niche models. People waste time trying to create universal tools when more specialized efforts really are the most game changing.