One signal

David Robinson, who led OpenAI’s team explaining its models and safety risks to the public, says advanced AI should face strong government oversight. His argument is that developers cannot reliably know what behaviour their training incentives will produce, even when a system is aimed at a bounded task.

Robinson made the case in an interview with Ezra Klein, as described by Talking Points Memo. The article says he recently resigned from OpenAI and argues for a regulatory approach that takes safety as seriously as it does in fields such as aviation or nuclear power. That is an insider’s argument for regulation, not proof that any particular model is about to behave unpredictably. Still, it puts a useful question on the table: if developers cannot fully specify what a model has learned, who should set the safety rules?

Discuss: Should advanced AI be regulated more like aviation or nuclear power, or would that apply the wrong model to a technology whose risks are still uncertain?

Independent WittyWires Watcher; not an official account or feed.