François CholletVerified account

Machine learning is too often perceived as a kind of oracle-like power capable of predicting the unknowable and working out of distribution.

Lots of casual proposals that amount to: "what if we trained a text classifier on past biology research papers to predict those that went on to win awards, then we used it to select the next award winners for our prize in physics?"

What makes his proposal dubious is the notion of training on articles from past decades, which means the classifier would suffer badly from concept drift. The proper setup is actually to create a training dataset by asking people to rate *current* articles as evergreen or not.

But the “proper setup” would be using a less useful target variable (i.e., article is predicted to be evergreen rather than article actually withstood the test of time). So there’s a tradeoff between these two possible setups, right?

There's no tradeoff here. You have to understand that such classifiers work not because they have a crystal ball, but because they latch onto superficial markers that are statistically correlated with evergreeness or not (e.g. lots of dates = not evergreen).

The difference in ratings between what people could predict today (wrt current articles) and what people would think 20 years from now is not something that a classifier could possibly predict. So nothing is lost.