[goggy_googy]

Important to note, they say "From these generated models, we select the one with the best performance on the training set." Definitely potential for bias here.

[nerdponx]

I'd have liked to see the distribution of generated model performance.

[QuadmasterXLII]

Fig 4b