Rich Sutton identifies the one-step trap as a fundamental error in AI research where practitioners assume accurate long-term predictions emerge from iterating one-step transition models. The appeal stems from a partial truth: perfect one-step accuracy would yield perfect long-range forecasts. In practice, however, one-step errors compound exponentially, and computing probabilistic futures requires exploring a branching tree of possibilities whose complexity grows exponentially with prediction horizon.

Sutton argues this approach is hopeless for stochastic environments or policies, yet remains pervasive in POMDPs, Bayesian analyses, control theory, and compression-based AI frameworks. The computational burden of rolling out one-step models makes accurate long-horizon prediction generally infeasible.

The proposed solution adopts temporal abstraction through options and General Value Functions (GVFs), building on Sutton's 1999 framework with Precup and Singh, the 2011 Horde architecture for real-time sensorimotor learning, and 2023 work on reward-respecting subtasks for model-based reinforcement learning. These methods learn multi-step predictive models directly rather than chaining single-step approximations.

The distinction matters for any system requiring planning beyond immediate transitions — robotics, autonomous navigation, and strategic decision-making — where accumulated model error and exponential compute costs render naive rollouts impractical.