“In 'Bayesian wind tunnel' experiments, the Mamba architecture performs Bayesian updating reasonably well, LSTMs perform it only partially, and MLPs fail completely.”