Brightest possible?
We figure out a good, stable way to train a powerful general purpose agent to predict what will make us happy and satisfied, and set one loose. It improves its own code and hardware, and eventually becomes arbitrarily powerful. Paradise ensues.
Oh, and of course, you also have to invent a technology that modern ML methods don't cover at all, which is the ability to accurately model and predict the outcomes of complex situations that the system has never observed during training.