AI for Science: Analogies Without a New Laboratory

The essay “Do It Like Darwin” proposes finding scientific hypotheses in existing data

On 14 August 2026, LessWrong user derelict5432 published an essay about two tasks for scientific AI: obtaining new observations through instruments, robots, or a computer model of the physical world, and connecting existing findings to form testable hypotheses.

In a draft of his autobiography, Charles Darwin recalled that after years of observing animals and plants, he read Malthus’s work on population. It led him to connect variation among organisms with the struggle for existence: favorable variations are preserved, and their accumulation can produce new species. Observations from natural history thus converged with an idea from political economy.

“For many open questions, we can look for similar features in other fields and use existing data to test the resulting conclusions,” writes derelict5432.

In a research AI agent framework, the system connects disparate data, formulates a hypothesis, and selects the next experiment. The essay examines the step between the data and that experiment: before taking a new measurement, researchers can search existing results for an analogy that suggests a method or hypothesis.

Tom Zahavy, a researcher at Google DeepMind, proposes obtaining new observations through a world model consistent with physical laws. Within a computer environment, an agent changes the conditions and traces the consequences. Zahavy argues that such simulations would ground the system’s new concepts in physical experience.

“To create AI capable of genuine invention, we need to move from systems that merely read the scientific literature to systems capable of perceiving the physical world,” Zahavy writes.

In a Stanford University preprint, Andrew Shen, Shaul Druckmann, and James Zou describe how analogy can guide the search for a method. Here, analogical reasoning means looking for problems whose elements are connected in the same way. The framework first identifies the objects and relationships in the original problem, then searches for another field with the same structure and transfers a technique from that field back to the original problem.

In one biomedical computational task, the authors compared how laboratory-grown groups of cells respond to an intervention with how groups of consumers respond to an economic policy. In both cases, the overall mean response conceals several distinct types of response. This analogy led to a finite mixture method, which represents the data as several groups with different responses. When tested on the Srivatsan20 dataset, adding this method to the baseline model reduced one measure of distributional difference from 1.65 to 0.13.

When the data have already been collected, analogy can connect them into a hypothesis or computational method. A new measurement must come from an instrument, an experiment, or a model of the physical world.