Measuring the Unmeasurable: What 'Inventing Temperature' Teaches Us About AI Alignment
Coverage of lessw-blog
In a recent analysis, lessw-blog explores the epistemological hurdles of defining abstract properties, drawing a parallel between the history of thermodynamics and modern AI evaluation.
In a recent post, lessw-blog discusses the philosophical and practical challenges of measuring abstract concepts, using Hasok Chang’s book Inventing Temperature as a primary lens. The analysis addresses a fundamental issue in the current artificial intelligence landscape: the difficulty of establishing reliable metrics for properties that lack a physical ground truth, such as "capability" or "alignment."
The core of the discussion revolves around the historical dilemma of thermometry. Before the standardization of temperature, scientists faced a circular problem: to validate a thermometer, one needed to know the temperature, but to know the temperature, one needed a valid thermometer. This is known as the problem of nomic measurement. The post argues that the field of AI is currently trapped in a similar pre-scientific phase regarding its most critical attributes. While researchers frequently refer to model capabilities and alignment as if they were distinct, measurable quantities, the industry lacks the rigorous theoretical framework required to measure them objectively.
This comparison is particularly significant for those involved in AI safety and governance. Currently, the evaluation of AI models relies heavily on benchmarks that serve as proxies for intelligence or safety. However, without a robust theory of what these properties actually are—independent of the tests designed to detect them—engineering trustworthy systems remains fraught with uncertainty. The author suggests that just as thermodynamics eventually established a coherent system of measurement through iterative refinement and theory-building, AI research must move beyond intuitive definitions to establish valid metrics.
The post further implies that until we solve this "inventing temperature" problem for AI, claims regarding the safety or specific capability levels of advanced models remain epistemologically shaky. For researchers and engineers, this underscores the necessity of developing measurement tools that can withstand rigorous scrutiny, rather than relying solely on empirical performance on static datasets.
We recommend this post to readers interested in the intersection of the philosophy of science and machine learning, particularly those focused on the foundational challenges of evaluating AI behavior.
Read the full post on LessWrong
Key Takeaways
- The history of thermometry offers a critical analogy for modern AI evaluation challenges.
- AI currently faces a 'problem of nomic measurement,' lacking established standards to validate new metrics.
- Abstract properties like 'alignment' and 'capability' are currently ill-defined, similar to early concepts of heat.
- Developing rigorous measurement frameworks is a prerequisite for scientific progress in AI safety.