2 things Jeff Dean has mentioned in talks as interesting areas for research: * Is there a better way to do matrix multiplication? * Could less reliable chips

2 things Jeff Dean has mentioned in talks as interesting areas for research: * Is there a better way to do matrix multiplication? * Could less reliable chips perform better in aggregate?