[LOCKSUIT]

Now the puzzle comes even closer together.

Neural networks...backprop...

It's a faster way than "brute force generating all possible computer programs from smallest to largest". You show the network a lot of (desired) data inputs, and want it to give certain desired outputs. This reward function is one to generate a desired program, by measuring the error Cost and adjusting the weights via backprop. This is training...generating the finished program and selling it trained (generated). No wonder they don't learn online and require re-training. This isn't AI...it's program generation. This is why it's a black box in the hidden layers. This is why we have CNNs and RNNs and GANs. I notice CNNs have padding etc, not fully a generated program. GANs have a generator & detective that compete to learn how to give the right outputs. A classifier / feature/representation learning is made through this.

Did I miss something?

Now it makes further sense why I have new Costs I discovered in my research, I have the real Costs.

[goaty]

I vote it is too, but its not necessarily faster unless you make it faster.  back prop is a slow process, I prefer knn memories more myself, for a few technical reasons.

And its not so simple Lockster,  the computer can form a "thinking device" or computation device,  but its not the computer that can build its own brain.  U are supplying all facilities that the a.i. will be equipped with.

You equip the ai,  it cant equip itself,  and im meaning the metrics involved in your target application must be very carefully picked,  and its only going to go so good.
Helps if uve got the right correlative sensors+ semantic truth.

If you give the machine the wrong clue for the task,  it wont go any good, it doesn't even matter if its full knn similarity matched, it wont work.

[LOCKSUIT]

Correct, you give it the sensor functions, motor, the data, desired output 'reward function' for generating the program that generates outputs how you asked for, etc. Some of it is hand-made.