# Ball+Gravity has a "Downhill" Preference > Coverage of lessw-blog **Published:** February 27, 2026 **Author:** PSEEDR Editorial **Category:** risk **Content tier:** free **Accessible for free:** true **Word count:** 412 **Tags:** Agent Foundations, AI Alignment, Ontology, Philosophy of Mind, LessWrong **Canonical URL:** https://pseedr.com/risk/ballgravity-has-a-downhill-preference --- A recent LessWrong post uses a physics analogy to challenge the boundaries between capability and preference in agent foundations theory. In a recent post on LessWrong, the author presents a thought experiment titled "Ball+Gravity has a 'Downhill' Preference," which utilizes a simple physical system to interrogate complex concepts in agent foundations research. The piece explores how we attribute "preferences" to systems—whether they are sentient agents or inanimate objects governed by physics—and questions the theoretical distinction between what an entity _wants_ to do versus what it is _capable_ of doing. ### The Context In the field of AI alignment and safety, defining an "agent" is notoriously difficult. Researchers often rely on theoretical frameworks (such as Ontological Identification of Substructures) to distinguish between an agent's internal goals (preferences) and the environment or tools it uses (capabilities). This distinction is critical: if we cannot accurately separate an AI's capabilities from its preferences, we risk misinterpreting its behavior or failing to predict how it will act when its capabilities change. This post attempts to clarify these definitions by applying them to a non-agentic system. ### The Gist The author constructs a scenario involving a ball, a hill, and gravity, deliberately ignoring friction and inertia to focus on the ontology of the system. The analysis proposes a layered interpretation of "preference": * **The Ball:** Its roundness is interpreted as a preference for rolling. * **Ball + Gravity:** This combined system expresses a preference for rolling specifically _downhill_. * **Ball + Gravity + Hill:** When the environment is included, the system's preference resolves to a specific outcome: the ball being located at the bottom of the hill. The core of the argument rests on the difficulty of separating the "Ball + Gravity" system's preference for downhill movement from its capability to achieve it. By framing physical laws as preferences, the author highlights the ambiguity in current agent foundation models. If a ball "wants" to go downhill because gravity compels it, how does that differ from a software agent "wanting" to maximize a reward function it is programmed to pursue? ### Why It Matters This reductionist approach forces readers to confront the assumptions underlying agent modeling. If our definitions of "preference" break down when applied to simple physics, they may be insufficient for complex AI systems. The post suggests that a more robust ontological framework is required to distinguish true agency from mere mechanistic inevitability. For those interested in the philosophical underpinnings of AI safety and agent theory, this thought experiment offers a concise entry point into a dense topic. [Read the full post on LessWrong](https://www.lesswrong.com/posts/h2W4KY9RcFMyodfit/ball-gravity-has-a-downhill-preference) ### Key Takeaways * The post challenges the distinction between 'preference' and 'capability' by applying agent terminology to a simple physical system (a ball on a hill). * It argues that physical attributes, such as a ball's roundness or the force of gravity, can be ontologically modeled as preferences for specific states or motions. * The analysis highlights the difficulty in defining agency: distinguishing between an entity that 'wants' to act and one that is simply compelled by laws of physics. * This thought experiment serves as a proxy for broader issues in AI alignment, specifically regarding how we identify and model goal-directed behavior in artificial systems. [Read the original post at lessw-blog](https://www.lesswrong.com/posts/h2W4KY9RcFMyodfit/ball-gravity-has-a-downhill-preference) --- ## Sources - https://www.lesswrong.com/posts/h2W4KY9RcFMyodfit/ball-gravity-has-a-downhill-preference