I present an algorithm M which assigns probabilities to logical sentences. If any other algorithm from a large class of hypotheses is able to predict any subset of sentences sufficiently well, then M is able to predict the same subset asymptotically just as well.