Google DeepMind

AI's real shortage isn't compute—it's self-doubt

Large models aren't without probability—they predict probabilities at every step—but they don't treat uncertainty as a tool for seeing their own limits. To make real decisions, you first have to build self-doubt into the architecture.