Conversation

Notices

1. Embed this notice
   iced depresso (icedquinn@blob.cat)'s status on Friday, 31-Jul-2026 04:19:58 JST iced depresso
   apparently giga llms are just mixtures of 9b models. so fine tuning your own 9b model to a specific task should be reasonably frontier grade, asymptotically. weird.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Friday, 31-Jul-2026 04:31:24 JST iced depresso

     in reply to

     • बुद्धीविनाषक् भेड़ मीपर् [Meeper]
     @meeper yeah. mixture of experts is basically 9B models crammed together with a "router" that decides who is active.
     
     which means 9B is about where the labs have converged on the model doesn't get more useful above this size.
   • Embed this notice
     बुद्धीविनाषक् भेड़ मीपर् [Meeper] (meeper@udongein.xyz)'s status on Friday, 31-Jul-2026 04:31:25 JST बुद्धीविनाषक् भेड़ मीपर् [Meeper]

     in reply to
     @icedquinn you're talking about MoE stuff right?