One-Shot Promoting is Bad for MoE Models?!!

This video explores why incremental prompting, one aspect of the problem at a time, is better for MoE LLMs compared to asking the model to do everything at once. Not all experts are activated for each token, so the context from coding one aspect of the program can poison the bigger picture if you don't separate requests.

I noticed this too even with big SOTA models, so it's doubly true for Local LLMs! #ai #localai #mixtureofexperts #askleo #technology

!summarize