/tech/ - Soyence and Technology

The way LLMs should be designed is one master model that is good at general reasoning that distributes tasks to very small but fine tuned usecase specific slave models. Slave models then can use a narrow toolkit given to them. This is much more memory efficient and in a lot of cases will provide better output than a 1000 niggerbillion parameter omniscient model that does everything. And being modular, you can much more easily improve single aspects of the LLM without giant retraining of fine tuning steps that can harm other output.