This talk Vishal Misra was an eye-opener to how few shot prompting works under the hood to change output token probability distributions. I forgot LLMs don't produce tokens directly. They produce embeddings from which nearest neighbors tokens are chosen.
youtu.be/sLodkyHlQhY
tl;dr LLMS as Bayesian learning : Given a prompt, looks for something close in training set, then uses prompt for new evidence. Then computes a posterior using this new evidence. This posterior distribution is what is used to generate the new text. (Vishal Misra)
It's so crazy…
Awesome 💕 😘 Love Song From A Bollywood Romantic Movie Scene.
#ChalTereIshqMein 😍 💕 ✴️
#Gadar2
#VishalMishra
#Mithoon
youtube.com/shorts/jPiWIRt…
#YouTube
#youtubeshorts
#ZeeMusic
#Bollywood
#HinfiRomanticSong