I am not persuaded that the harder Bayesians have any more concrete answer. Solmonoff induction is uncomputable and seems to unnaturally favour short hypotheses involving Busy-Beaver-sized numbers. And any computable approximation to it looks to me like brute-forcing an NP-hard problem.

What if, as a computational approximation of the universal prior, we use genetic algorithms to generate a collection of agents, each using different heuristics to generate hypotheses? I mean, there's probably better approximations than that; but we have strong evidence that this one works and is computable.