I cant even describe what I mean

I'm not talking specifically about producing an A.I. capable of outsmarting you in a game. I'm speaking of even a stupid AI, but one that taught itself what and how to do it.
The difference is, that you give it just an input channel and a goal. For example, the same AI can be plugged in to Battlefield or a trading platform, and you just give it a way in(WSAD or "Buy" button), current status of the given environment (player pos or USD to BGN) and goal to reach (kill or +$0.5/h). It has to come up on its own with the methods of reaching this given goal. Perhaps analysing its environment over different periods of time and its actions and if they get it towards the right direction...(For example, this wont be expressed at all in the programming, but it has to have a way of building on its own a micro-goals list like "To kill a bot I must see it, to see it I must meet it, to meet it I must walk...and start walking" without you defining any of these things in the A.I.'s programming, just provide it with a method for building its own logic and input/output/goal. Meaning, you must provide a NN that doesnt represent ANYTHING and give the modelling A.I. the capability of modifying the NN so it 'represents' the micro goals and eventually it'll reach a state that every neuron serves the purpose of reaching one of the micro-goals...I hope this makes sense...It has to be vague. And it'll probably need to use a dynamic number of neurons, depending on the complexity of the micro-goal list, because every excessive neuron breaks this list).

I'm not at all concerned with budget or time, or tech, because I'm just trying to reach an understanding on how such a machine could be built, how it could work...(And, by the way, you could eliminate the training time with a high-speed simulation, saving you some time). If you know how to build something, its easy to consider whats needed, but if you dont know what you're building, you can never guess what you'll need

But, what I'm really talking about, like in your example, if you put a Google AI into Battlefield, it wont understand at all the environment its in, unless you specify it CAN take cover. A true A.I. has to understand on its own that it can take cover and that it actually helps, what the cover means to itself in its own 'language'. It has to understand its environment on its own, with its own meaning... Without you defining what a corner means or how it can be exploited... I dont think this can be done solely with NNs, or at least with a single NN. Using a single NN has to be more complex than what a human can sculp, even for a life-time. You can copy an entire brain if you like, but it'll never work, because you cant provide it the stimuli a real human body has. It has a ton of feedback/input and output, thats the true problem of the most complex NNs built so far.

The NNs that we currently have model themselves, but it is fixed, it is scripted by a programmer towards a particular goal and in a fixed environment. It must have a mechanism of judging on its own if it is getting towards its goal... A pattern mimicking NN (probably the worst example ever) has its modelling behavior sculpted into it...It doesnt need to understand how to get to the desired goal, it just does...

I feel I'm going in circles I feel cant really express what I mean. I hope you're understanding all this.

PS.: A Long post and a lot of thoughts. If it seems random, sorry, I'm jumping from section to section while thinking but I have the feel I'm going somewhere, even if the goal is on the other side of the universe.
* This has to be my longest post ever!


So, after all this, any thoughts on realising a judgemental A.I. mechanism for modelling an NN based on meaningless input/output and goal?