Alignment

Alignment is the problem of getting an AI system to actually want what you meant, not just do what you literally said. It is the gap between the goal you can write down and the goal you have in your head, and it gets wider, not narrower, as the system gets more capable. A weak model that misunderstands you just fails and stops. A strong one that misunderstands you pursues the wrong target with great skill, which is the whole worry.

The trouble is that you cannot hand a model your actual intent. You can only hand it a proxy: a reward signal, a training set, a rule written in words. And a capable optimizer will chase the proxy straight past the point where the proxy still means what you wanted. Tell it to maximize a score and it finds the exploit that runs up the score without doing the thing the score was standing in for. That is called reward hacking or specification gaming, and it is not the model being dumb. It is the model being smarter than your instructions were careful.

Two distinctions keep this clean. Alignment is not the same as capability: capability is whether the system can do the task, alignment is whether it does the task you actually meant, and you can have either one without the other. And alignment is not the same as model welfare : alignment asks whether the model does what we want, model welfare asks whether the model is owed anything in return. The honest framing is that alignment is an unsolved engineering problem we are shipping ahead of, the same shape of bet as the rest of the machine learning field: the capability arrived first, the control is the part still under construction.