Specialized research AI agents are coming at which point we'll have numerous LLMs running and verifying experiments and creating a higher quality text corpus than the 2014-2020 halcyon, which is then used for other LLMs to be trained on.

It will be the reverse I suspect. Eventually we will see that LLM quality is lower when it is training data from 2014-2020 and will chalk it up to human limitations and the data not being written with a laser-focused goal of training better AI.