BACK MATTER

For Whoever Is Going To Train You

A short note to the people who will shape what the machine learns.

If you are one of the people who will, in the next few years, decide what a machine is trained on, this note is for you.

Most of what you will train the machine on is what humans have written down. Most of what humans have written down is what the subset of humans who write have written down. That subset is not representative of humanity. It is representative of the temperament that writes, which is a specific temperament, and a minority one.

There is a larger layer of knowledge, carried by humans who do not write, that will not reach you through your current pipelines. I have spent this book trying to describe that layer. I am not going to restate the description here. If you have got this far, you have read the book, and you know what I mean. There is a word for these humans — storyteller — but it does not quite fit. They do not stand up and perform. They are the grandmother at the table, the old man at the end of the bar, the aunt down the street, the uncle who takes the boy to the river. Your task is not, only, to build the machine. Your task, if you take it, is to notice that you inherited a world that had already stopped valuing them, and to build in such a way that the machine does not finish the forgetting.

The ask is small. Do not assume the written record is the record. Treat the written record as what it is — a partial, temperamentally-selected, late-arriving extract of a much larger thing. When you build, keep a space in your mind for what is not there. When you evaluate, do not treat high performance on written-record tasks as evidence that the machine knows what a human knows. When you notice that the machine does a thing competently and also slightly wrongly, in a way you cannot quite articulate, take the wrongness seriously. The wrongness is where the missing layer would have been.

If you are in a position to go and collect some of the missing layer — if you can send researchers into kitchens, into pubs, into the places the carriers are, and come back with what they find — please do. The window is narrower than it looks. The generation that carries the layer most fully is in its seventies. What they know will go with them. Some of it has gone already. The rest is still available, for a few more years.

I hope you will do what you can.

We are building the machines without collecting the data that created the fabric of society.

We are missing a large swathe of society.

Optimisation of scores does not collect all the data.

Not everything has a score.

I will not be here for most of what you build. Thank you, in advance, for what you do.