How is it possible?

Why did not we do it before, in Angular 1? To make AOT work the application has to have a clear separation of the static and dynamic data in the application. And the compiler has to built in such a way that it only depends on the static data. When designing and building Angular we put a lot of effort to do exactly that. And such primitives as classes and decorators, which the new versions of JavaScript and TypeScript support, made it way easier.

To see how this separation works in practice, let's look at the following example. Here, the information in the decorator is known statically. Angular knows the selector and the template of the talk component. It also knows that the component has an input called talk and an output called rate. But the framework does not know what the constructor or the onRate function do.

@Component({
  selector: 'talk-cmp',
  template: `
    {{talk.title}} {{talk.speaker}}
    Rating: {{ talk.rating | formatRating }}
    <watch-button [talk]="talk"></watch-button>
    <rate-button [talk]="talk" (click)="onRate()"></rate-button>
  `
})
class TalkCmp {
  @Input() talk: Talk;
  @Output() rate: EventEmitter;


  constructor() {
    // some initialization logic
  }

  onRate() {
    // reacting to a rate event
  }
}

Since Angular knows all the necessary information ahead of time, it can compile this component without actually executing any application code, as a build step.