picofarad on Nostr: I copy and paste back and forth but models aren't good at "changing the subject" so ...

I copy and paste back and forth but models aren't good at "changing the subject" so they just repeat the same things then get annoyed at eachother for repeating the same things.

the closest cohesive "talking to eachother" is Mixture of Experts (MoE models). Also the newest trick i've used is to load a small model and a large model, then let the small model do all the generation. large model checks the work and corrects it. it's about a 35% speedup.