Can AI with Pre‑1905 Knowledge Become the Next Einstein? DeepMind Examines the Missing Step

Thought experiment

DeepMind’s paper LLMs can’t jump asks whether a large language model supplied with all scientific knowledge available to Einstein around 1905 could independently formulate the theory of general relativity. The answer is negative.

Missing cognitive step

The failure is not due to lack of factual knowledge or computational power. The authors identify a missing cognitive operation: the ability to perform an Abductive Jump —a leap from concrete physical experience to a new foundational axiom.

Einstein’s abductive jump

Einstein’s insight began with the “freely falling elevator” thought experiment: an observer in a falling elevator feels weightless, while an observer in an accelerating spaceship feels a force indistinguishable from gravity. This vivid intuition led to the **Equivalence Principle**, a hypothesis that gravity and acceleration are locally equivalent. The principle is not derived from data or algebraic deduction; it is a newly introduced axiom that later expands into curved spacetime, geodesics, the stress‑energy tensor, and Einstein’s field equations.

Abductive Jump defined

Abductive Jump differs from ordinary statistical inference. It involves actively proposing a novel hypothesis that can explain observed phenomena even when no prior formalism exists. In the paper’s terminology, scientific invention proceeds as:

Physical experience → Abductive Jump (new axiom) → Deduction (mathematical development) → Experimental verification.

Why current LLMs cannot jump

Existing automated scientific discovery pipelines follow a linear chain: Data → Modeling → Validation → New Theory . This chain omits the abductive jump. The authors invoke the Chinese Room argument: an LLM manipulates symbols according to statistical patterns without any grounding in the physical world. Consequently, terms such as “gravity”, “acceleration”, or “free fall” are understood only as token co‑occurrences, not as embodied experiences.

World models and interactive simulation

Video‑generation models can produce realistic motion, but their “intuition” derives solely from training data. The paper highlights a different class of models—interactive simulation‑oriented world models—exemplified by DeepMind’s Genie . Such models expose a controllable action space, allowing the system to actively modify experimental conditions (e.g., cut elevator cables, change acceleration, flip gravity direction) and observe the resulting physical consequences.

Physical Prior

Even a perfect world model is insufficient without a “Physical Prior”—a strong, intuitive sense of how the world works that narrows the search space for plausible hypotheses. The prior is not a formal equation but a heuristic that guides the abductive step.

Implications

Combining a stable, interactive world model with an appropriate physical prior could bridge the gap between experience, intuition, and logical deduction, enabling AI systems to generate and test first‑principle hypotheses. The authors conclude that genuine scientific creation requires the emergence of new axioms grounded in physical intuition, followed by mathematical development and experimental verification.

Paper link: https://philsci-archive.pitt.edu/28024/1/Scientific_Invention_Position_Paper%20%2817%29.pdf

Code example

来源：ScienceAI
本文
约2300字
，建议阅读
5
分钟
关于世界模型与科学创造的那一下跳跃。