You know how people always emphasize that exact Bayesian updates on big hypothesis spaces require exponential compute? Well, IIUC openai's result yesterday disproves that claim; turns out full-blown Bayesian updates are subexponential. Yesterday's result isn't enough to make exact Bayesian reasoners tractable, the improvement over naive exponential methods isn't very large, but given that any improvement over exponential is possible at all, it seems likely that further improvement is possible.
ETA: Not quite, it's subexponential in circuit size, not in variable count. So Bayes updates over the input variables of deep circuits (like e.g. mining Bitcoin) are still exponential, but e.g. exact Bayes updates as a replacement for SGD training of neural nets is subexponential. Thanks Mikewins in the thread below for pointing this out.