1. Recursively decompose all the problem(s) (prioritizing the bottleneck(s)) behind AI alignment until they are simple and elementary.
2. Get massive ‘training data’ by solving each of those problems elsewhere, in many contexts, more than we need, until we have asymptotically reached some threshold of deep understanding of that problem. Also collect wealth from solving others’ problems. Force multiplication through parallel collaboration, with less mimetic rivalry creating stagnant deadzones of energy.
3. We now have plenty of slack from which to construct Friendly AI assembly lines and allow for deviations in output along the way. No need to wring our hands with doom anymore as though we were balancing on a tightrope.
In the game Factorio, the goal is to build a rocket from many smaller inputs and escape the planet. I know someone who got up to producing 1 rocket/second. Likewise, we should aim much higher so we can meet minimal standards with monstrous reliability rather than scrambling to avoid losing.
“Let’s finish what Engelbart started”
1. Recursively decompose all the problem(s) (prioritizing the bottleneck(s)) behind AI alignment until they are simple and elementary.
2. Get massive ‘training data’ by solving each of those problems elsewhere, in many contexts, more than we need, until we have asymptotically reached some threshold of deep understanding of that problem. Also collect wealth from solving others’ problems. Force multiplication through parallel collaboration, with less mimetic rivalry creating stagnant deadzones of energy.
3. We now have plenty of slack from which to construct Friendly AI assembly lines and allow for deviations in output along the way. No need to wring our hands with doom anymore as though we were balancing on a tightrope.
In the game Factorio, the goal is to build a rocket from many smaller inputs and escape the planet. I know someone who got up to producing 1 rocket/second. Likewise, we should aim much higher so we can meet minimal standards with monstrous reliability rather than scrambling to avoid losing.
See: Ought