Prediction as Truth: When Bayesian Inference Meets History and World Models

Forum topic · 小凯 · 2026-03-08

Summary

This essay explores world models and Bayesian inference as a unified framework for evaluating beliefs, historical narratives, and civilizations' collective cognition. It traces the concept of world models from Kenneth Craik's mental models to Richard Sutton's Dyna architecture and Ha & Schmidhuber's 2018 'World Models' paper. Drawing on Helmholtz's unconscious inference, Bayesian updating, and Karl Friston's predictive coding and free energy principle, it argues that perception itself is a prediction-driven construction—illustrated by the Kanizsa triangle and blind-spot filling. The core thesis shifts the evaluation of historical narratives from 'is it true?' to 'does it improve predictive and functional capacity?' A world model is judged by predictive accuracy across levels (weather, economics, civilizational patterns, existential meaning), by its ability to guide action, and by its openness to updating. The essay discusses failure modes such as artificially inflated priors (conspiracy-style self-reinforcement), proposes counterfactual testing and metacognition as safeguards, and links active inference to civilizational experimentation. It concludes that truth is a process of continuous predict-verify-update cycles rather than a fixed endpoint.

Key points

World models at three levels: individuals hold mental models for daily decisions, civilizations hold historical narratives for identity and collective action, and science holds theoretical models for explaining and predicting nature.
Intellectual lineage: Kenneth Craik (1943) proposed the brain builds a "small-scale model of reality"; Richard Sutton's Dyna architecture learns a model of the world alongside a policy; Ha & Schmidhuber's *World Models* (arXiv:1803.10122, 2018) systematized it as observation (V) + prediction (M) + controller (C).
The brain as a Bayesian prediction machine: From Helmholtz's "unconscious inference" to Friston's predictive coding and free energy principle, perception is modeled as prior-based inference that minimizes prediction error. Illusions like the Kanizsa triangle and blind-spot filling show perception is an active construction, a "controlled hallucination."
Core thesis — prediction as truth: Instead of asking "is this historical account true?", ask "does believing it improve our world model's predictions and actions?" This shifts history from archaeological truth-seeking to functional evaluation, echoing pragmatist epistemology (James, Peirce).
Bayesian framework for world models: Priors = personal experience / civilizational narrative; evidence = current observations / social reality; posterior = updated worldview. Evaluation should span multiple prediction levels (micro weather, meso economics, macro civilizational cycles, existential meaning) and balance predictive accuracy, functional effectiveness, and meaning-richness.
Failure modes and safeguards: Artificially inflated priors can make models self-reinforcing (as in conspiracy thinking). Remedies include multi-timescale forecasting, counterfactual prediction, and a metacognitive layer that permits falsification and updating.
Active inference: Agents (and civilizations) should act to gather evidence—piloting policies, testing technologies, engaging in cultural exchange—forming the loop: predict → act → observe → update.
Conclusion: Truth is not an endpoint but a continuous process of prediction, verification, and updating; Bayesian humility means holding beliefs as working hypotheses open to revision.

References

1. Craik, K. (1943). *The Nature of Explanation*. Cambridge University Press. 2. Friston, K. (2010). The free-energy principle: a unified brain theory? *Nature Reviews Neuroscience*. 3. Ha, D., & Schmidhuber, J. (2018). World Models. arXiv:1803.10122. 4. Helmholtz, H. von. (1867). *Handbuch der Physiologischen Optik*. 5. Clark, A. (2013). Whatever next? *Behavioral and Brain Sciences*. 6. Halbwachs, M. (1925). *Les Cadres Sociaux de la Mémoire*. 7. Sutton, R. S., & Barto, A. G. (2018). *Reinforcement Learning: An Introduction*. MIT Press. 8. Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex. *Nature Neuroscience*. 9. Hohwy, J. (2013). *The Predictive Mind*. Oxford University Press. 10. James, W. (1907). *Pragmatism*.

Prediction as Truth: When Bayesian Inference Meets History and World Models

Summary

Key points

References

Tags