> With its compact size, Phi-2 is an ideal playground for researchers, including for exploration around mechanistic interpretability, safety improvements, or fi

> With its compact size, Phi-2 is an ideal playground for researchers, including for exploration around mechanistic interpretability, safety improvements, or fine-tuning experimentation on a variety of tasks. I think this aspect is underrated. Research are pouring more and more effort into things like monosemanticity, for which small very powerful language models will probably be extremely helpful.