Kerala Rising: From Exporting People to Exporting Ideas · Page 103 of 353

KERALA RISING · FROM EXPORTING PEOPLE TO EXPORTING IDEAS

KERALA RISING · FROM EXPORTING PEOPLE TO EXPORTING IDEAS
and is precisely the kind of narrow, high-volume
classification that modern vision models do reliably
and cheaply. Is this bag segregated into the right
stream; is a helmet present; is this a recharge pit
rather than a rubbish hole; is a sapling alive at this
location twelve months on. These are bounded visual
judgements with clear right answers, exactly the
regime in which purpose-trained classifiers achieve
high accuracy at a fraction of a paisa per inference.
The technology is not being asked to do anything
close to the frontier of its capability; it is being
asked to do something well within it, at scale, for
almost nothing.
What it does poorly, and is therefore never asked to do
The same models are unreliable at exactly the tasks
this book never assigns them. They should not make
consequential judgements about people, who is a
good citizen, who deserves a penalty, what
someone’s health status is, and the design assigns
them none of these. They are weak at novel or
adversarial edge cases, which is why uncertain
submissions are escalated to a frontier model and
then, if still unclear, to a human. They can carry the
biases of their training data, which is why the
program rewards participation rather than ranking
people, audits a random sample by hand, and
publishes the corpus so that systematic errors are
visible from outside. And they cannot establish
103