* I was hoping for that this year, but: split of LLMs into databases and reasoning. Think small, capable PHI with pluggable information. Currently we're burning a lot of energy both to learn and process everything at once. The next step will be something closer to RAG, potentially selecting databases to load depending on topic, like a librarian. This will both enable more client-side applications and save lots of mon…
>I was hoping for that this year, but: split of LLMs into databases and reasoning This hasn't worked out because knowing what to query and how to query requires intelligence, and that's contained in weights.
To some extent we've already seen this with MoE and Frankenstein models.