MIT just cracked AI safety. not with more filters. not with more rules. with one insight everyone missed. they taught models to think backwards first. enumerate every possible harm. analyze every consequence. only then respond. they call it InvThink, and it might redefine how