@tegmark Maximum truth-seeking is my best guess for AI safety