Ethan Mollick says AI systems that use tools can now divide research tasks among themselves. His examples show less need to plan every step by hand. But a human still needs to set the goal, check the work, and decide when to stop.
Zvi Mowshowitz examines a reported voluntary White House AI safety accord. It calls for company checks and outside review. The real test is whether reviewers can see problems and leaders will act on them.
At a driverless racing event, a car stopped after losing its sense of position. A following car crashed into it. Teams have offered explanations, but the report does not provide enough vehicle data to judge safety on public roads.
In a nine-person prototype study, users rated an AI companion’s memory well even though two memory steps had not run. A convincing claim about memory is not a record of what the system did.
Researchers at the Weizmann Institute used brain scans and AI to make rough versions of images people viewed. The results were imperfect. This does not let anyone read unspoken thoughts or dreams.
A preprint tested six fixes to a college early-warning system after it made predictions. Errors moved, but the fixes did not lead to steady gains across groups. One score cannot show who gets wrongly flagged or missed.
