Right, that's the problem.

Before 2016 or so, "AI Safety" meant the types of things I listed above: how to design AI systems that are safe to use in safety-critical settings.

Unfortunately, the term "AI Safety" has been overrun with effective altruist longtermism folks talking about something closer to https://en.wikipedia.org/wiki/Existential_risk_from_artifici... which is ill-defined and, frankly, usually completely vapid nonsense.

> so the question of how to mitigate its risks is going to be at least as squishy a concept as the underlying tech + applications.

No, no, no!!! AI safety methods -- methods that really increase confidence in real systems -- are not generic. ALL of the useful AI safety work I know of is deeply related to one or more of: the specific model architecture, the specific optimization algorithm, the specific data, or the specific application. Almost always all of the above.

“Field is squishy” != all parts of the field are equally squishy. The presence of some clear problem-and-solution pairs does not preclude the existence of more ambiguous problems. This is especially true as we’re deploying AI against increasingly vague goals with increasingly vague constraints, such as in the case of the non-lying Diplomacy bot that absolutely did deceive human players.