Expect Subvertations (knhash.in)

The piece frames a World Model as an AI’s internal dynamics — not static facts but learned cause-and-effect used to simulate outcomes ("if I push this cup, it falls") through a predict–observe–update loop akin to Bayesian learning. It argues humans run continual simulations (predicting next words, reactions, or physical consequences), and that prediction errors are the core training signal: small mismatches fine-tune models, large, emotionally salient surprises drive strong memory encoding. Conversations are presented as two-player model-updating games; improv’s “Yes, and” is a formalized technique for establishing expectations and then productively violating them to produce insight or humor rather than random noise. For AI/ML practitioners this has concrete implications: build agents with model-based planning that leverage prediction error as a learning and salience signal, design interactions that confirm expectations to establish trust and selectively subvert them to create memorable, informative updates, and treat timing of violations as a critical design parameter. The origami-elephant anecdote illustrates how low-cost artifacts plus coordinated narrative can create durable shared memory by aligning surprise with social context. In short, world models, calibrated surprise, and the predict–update cycle are central levers for planning, human-AI interaction, memory formation, and storytelling-oriented agent design.