The case for sharing AI models’ hidden states, not just tokens

In a detective-novel thought experiment, a model completes “the killer was” with a name. The author argues that this output leaves behind the work of connecting clues and conflicting alibis, forcing a second model receiving only the name to rebuild the case. The post uses that analogy to argue for sharing hidden states—the model’s internal representations—not just output tokens, and says @mostik_ai is building a channel to hand over that information.