Inspect how variable dependency could be exploited to generate more efficient logp graphs for step samplers: https://github.com/pymc-devs/pymc/issues/7591 And/or talk a bit about type inference, example the sort of stuff we exploit in the linalg rewrites (complementing #6)