So training a transformer is actually training a Bayesian model of beliefs about some probabilities? Makes sense. If we could figure out how to train a ‘fetch

So training a transformer is actually training a Bayesian model of beliefs about some probabilities? Makes sense. If we could figure out how to train a ‘fetch this fact/concept/next few tokens’ from memory or from a sat solver instead of recalling from a lossy bayes net…