The case for AI models sharing more than tokens

A post imagines a language model reading a detective novel ending with “the killer was.” To supply the next word, the author argues, the model must build the case internally, weighing alibis and earlier clues. In the analogy, a second model receiving only the killer’s name misses that underlying work. The author uses this contrast to argue that hidden states matter more than output tokens, and says @mostik_ai is building a channel that “hands over the case.”