The same machine, both ways

This chapter has been fairly bleak, so it is worth ending it accurately rather than gloomily.

Everything described here comes from one capability, finding structure in data too complicated for anyone to write down rule by rule. That is what the training loop does. It is the through-line from the first adjustable weight in this journey to the systems running now.

Point that at protein sequences and you get shapes predicted for more than 200 million proteins, released into a public database, for a problem biologists had worked on for decades.

Point it at engagement metrics and you get an article about nothing, tuned to be clicked.

The same capability. Frequently the same underlying techniques, sometimes the same companies, and always the same indifference on the part of the machinery, which has no view about which use is better and no way of forming one.

This is the part that resists a simple verdict, and it is why "is AI good or bad" is not a question with an answer. It is like asking whether chemistry is good. The honest response is that chemistry is a set of facts about how matter behaves, and what gets made from it depends on who is funding the laboratory.

So the useful question is never whether the technology is good. It is the one this module keeps returning to. Who chose this objective, who benefits if it works, who carries the cost if it fails, and what else that same effort could have been pointed at.

Which leaves one thing unaddressed.

This whole chapter treated the machine as an instrument, which is what it is, and instruments have no opinions about their use. But no previous instrument caused the particular disquiet this one does. People who use it come away asking questions no tool has prompted before, and those questions are not really about the tool.

They are about us.