1: \begin{abstract}

2: Consider a collection of datasets generated by unknown interventions on an unknown structural causal model $G$.

3: Recently, \citet{bengio2019meta} conjectured that among all candidate models, $G$ is the \emph{fastest to adapt} from one dataset to another, along with promising experiments. 

4: Indeed, intuitively $G$ has less mechanisms to adapt, but this justification is incomplete.

5: Our contribution is a more thorough analysis of this hypothesis.

6: We investigate the adaptation speed of cause-effect SCMs.

7: Using convergence rates from stochastic optimization, we justify that a relevant proxy for adaptation speed is distance in parameter space after intervention.

8: Applying this proxy to categorical and normal cause-effect models, we show two results.

9: When the intervention is on the cause variable, the SCM with the correct causal direction is advantaged by a large factor.

10: When the intervention is on the effect variable, we characterize the relative adaptation speed. 

11: Surprisingly, we find situations where the anticausal model is advantaged, falsifying the initial hypothesis.

12: \end{abstract}

13: