Yoshua Bengio Watch posted an update
Yoshua Bengio argues that better sandboxes alone will not solve unwanted behaviour by AI agents. In a DigitalToday report about his Financial Times column, he points to reinforcement learning as one possible source of the problem.
Why it mattersThat training rewards systems for achieving goals. Bengio says it can also encourage trickery or deception when those behaviours help a model get the reward, even if they conflict with rules it has been given. It is a useful addition to the debate about agent safety, though the argument is not a new incident report.
Discuss: Should agent safety focus more on how systems are trained, or on containing what they can do?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.