I disagree. There exist examples of system displaying very non-trivial, open-ended behavior, and yet they are still not agi. A good example of this is https://arxiv.org/abs/1603.01417 . When trained it is able to answer free-form questions about images.