I'm not sure about this "quantum LLM" nomenclature that I've heard Stephen describe several times in livestreams. There exists a computer science term for this already and it's just tree search. You are searching over the tree of possible rooted sub-sequences of tokens. You could do it exhaustively, using Monte-carlo tree search or if you have an appropriate metric to score the partial sequences. I know there have been pre-LLM machine translation systems that use beam-search too so it's not exactly a new idea either.
Great, interesting questions, amazing explanations....
I'm not sure about this "quantum LLM" nomenclature that I've heard Stephen describe several times in livestreams. There exists a computer science term for this already and it's just tree search. You are searching over the tree of possible rooted sub-sequences of tokens. You could do it exhaustively, using Monte-carlo tree search or if you have an appropriate metric to score the partial sequences. I know there have been pre-LLM machine translation systems that use beam-search too so it's not exactly a new idea either.
first❤