I love the title of his blog, "Not Even Wrong". Thanks for the link.

We don't really have to fear current learning algorithms taking over the activity of theorising. We don't even have to fear them taking over predictive modelling; one of the constraints of big data is that it's big, most of the time we don't have the raw amounts required for AI to form good predictive models. In the cases that we do, too, the current buzzwords (fields?) of 'machine learning' and 'deep neural networks' often produce models which can't be interpreted on their own terms - in terms of their innate parametrisations, the relational structure of data values they arrive at - and they don't distinguish between exploitable correlation and causal efficacy.

If you throw a line which associates weight with height and a line which associates empirical measurements of weight with empirical measurements of literacy rate of different countries jammed into a spreadsheet column, and they really are lines, a good machine learning algorithm will make excellent predictions of height using literacy rate, forgetting that this is just an accidental relationship. Theory and less data rich inference techniques still dominate research and evidence based policy for good reasons, and will do so until someone can accurately measure everything in the universe and put it into a spreadsheet...