A massive amount of raw computational power is used during training doing massive amounts of calculations to figure out what's wrong and what's right, what works and doesn't and this process to me, doesn't look too different from a very stupid brute force attack - trying to figure out the username and password of a system by going through all available combinations of alphanumerics and symbols.

Yes, there are smarts built into what we call Machine Learning or Deep Learning to ensure we rely on the brute-force as little as possible. There are probably a billion hours that have gone into research improving this system and probably a trillion dollars spent, not that I'm going to fact check it. After all this is just an opinion of someone who has trained a few CNN models through TensorFlow, Keras and dabbled in LSTM models and Deep Reinforcement Learning - nothing really serious.

Epochs during training are iterations or attempts at finding the optimal model. This is what I refer to as brute-force. Yes, it's quite a bit smarter than that but I don't think anyone more qualified than myself could argue that this doesn't involve trial and error unless you wanted to have a semantic argument.

So let's talk about the smarts that I know about. I enrolled into Andrew Ng's Deep Learning course some time ago and never completed it because I found it too boring. But there's something I picked up called a gradient descent which is an optimisation technique in Deep Learning. It basically works like this - instead of having to do 10,000 brute-force calculations, you can have your answer in just 5,000 calculations so you can intelligently cut your AWS bill down in half.

Jokes aside, the point is - under the hood, the current iteration and flavours of AI is still pretty stupid. I wasn't one bit surprised when AlphaGo beat Lee Sedol - in fact they should be pretty embarrassed that he was able to beat it once, with all the computing power they have behind AlphaGo.

What would be impressive? AI that works without the brute-force, the massive amounts of training, GPU and computing resource it requires. That's actually intelligent.