What Tarot, Viterbi, and LLMs Have in Common: How an Algorithm Chooses One Meaning Out of Many

The article on Habr explores how Tarot card interpretation can be modeled as selecting the best coherent meaning path from multiple candidates, analogous to algorithms like Viterbi and beam search. The author starts with the idea that a card acts as a mapping function from context and native meaning to a set of possible interpretations, but argues that the unit of search should be a full semantic path, not a single interpretation. A concrete example with five states (Stagnation, Rethinking, Resources, Action, Transition) and three cards (Hanged Man, Magician, Chariot) shows that a greedy approach chooses a suboptimal path compared to Viterbi's global optimum. The formal model is a conditional inhomogeneous Markov chain, where Viterbi decoding finds the exact maximum path in O(n*S^2) time. However, when the score depends on the whole history or free text, optimal substructure is lost, and beam search is used as an approximate method, with diverse beam search also discussed. The article compares Viterbi, beam search, diverse beam, A*, MCTS, and LLM reranking, and mentions factor graphs and linear-chain CRFs as more general formulations. It also notes that real meanings may not fit discrete states, so semantic embedding vectors and LLM reranking can be used, and it suggests directions for turning the illustration into research on AI-supported Tarot practice.