Recent posts separate at least three questions often grouped under “alignment”: how models are trained to follow objectives, which values should guide them and whether advanced general-purpose systems should be built at all.
Andrew Trask argued that the deep-learning mechanics of steering a model toward specified behavior are comparatively mature. He used AlphaGo as the narrow example: a system trained to play Go does not have to sort through the many value-ambiguous concepts absorbed by a general-purpose model.
In Trask’s framing, the harder problem is deciding which concepts embody desirable values, then arranging them so they constrain otherwise neutral capabilities. Even if such a structure is possible, he wrote, developers still need enough training data and signal to build it and verify that it works. That makes the deployment problem philosophical and political as well as technical.
The question before the system is built
Other posts widen the frame. Iason Gabriel argued that long-term safety and present-day values both deserve attention, because deployed systems already shape interactions with billions of people.
Seth Lazar borrowed a distinction from just war theory: one question concerns how an activity is conducted, while another asks whether undertaking it is justified at all. Applied to AI, he argued, safety work should examine not only how AGI might be built, but whether building it is justified in the first place.
Taken together, the posts frame alignment as a stack of disputed questions, not a single unsolved switch. In that debate, technical questions remain tied to decisions about whose values count, how those choices can be tested and whether the system should exist.
