Looking at it, the title is highly misleading. The program is not teaching itself, rather it’s learning how best to play based on a metric given by the programmer(looking at the code, it seems grading is done by measuring how far mario got in the level, and how much time it took to reach that result. Please note that the program has these notions preprogrammed into it). There is a huge distinction between the two, as a program that teaches itself needs no grading program or guidance from a human, while learning is very much easier and not at all surprising. It’s the difference between weak AI and strong AI, and I wish people would stop trying to conflate the two.
They are very complicated versions of you performing a physical act and then purely through physical causality something happens. You let a stone loose, it falls down. Computers being electrical/optical machines is just because that’s the fastest medium, but really they’re still these physical, causal things, and they could work with any other kind of medium that allows for the creation of things we can interpret as boolean logic. Pressure and temperture for example.
Machine learning as demonstrated is a programmer giving a data set, a goal condition and the machine throws random stuff at the data set. The programmer decides via the goal condition what sticks and what doesn’t. Repetition allows the programmer to make the computer find sequences of things that stick.
But that’s not actually learning, it’s just the physically causal machine finding a path of least resistance according to the instructions of a programmer. It’s like water carving out a riverbed. Or you carrying a bucket of water somewhere, emptying the bucket and then finding the lowest path by watching the water.
The water doesn’t “know” where to go, it just is and obeys the laws of physics.
That’s weak AI.
Strong AI would be thinking like us humans, whatever that actually means.
There are tons of reasons we don’t have intelligent computers yet. The research money, the technology, the marketability of them in the economy. Just because they don’t exist that doesn’t say anything about whether they would be similar to humans.
Here’s the thing with the brain. We understand it at the lowest level: the neurons. We understand that it all boils down to simple electricity charging and discharging. It’s actually so simple it’s not interesting at the hardware level.
We also understand it at the highest level. We understand that different parts of your brain control different functions. Vision is processed in the back, motor skills and reasoning in the front, emotions near the brain stem, etc.
What we DONT understand is what’s in the middle connecting the high level to the low level. How are the simple neurons connected in such a way to provide all these complex emotions? We don’t understand the master plan of the billions of intricate connections of how the neurons are hooked together.
We understand that the brain is a simple deterministic machine running on electricity. We just don’t understand how the structure of the connections provide complex emotions.
ML is promising because it basically would allow us to hook the neurons randomly, and provide some training data with reinforcement. And the neurons will make the connections themselves without requiring humans to understand it. We might create strong AI one day and still have no idea how it works. ML will do all the complex stuff for us.
Anyway, I would say we don’t have super intelligent computers right now because we don’t have the hardware. We don’t have a machine with billions of neurons that allows us to change the neuron connections through code. We only have simple processors that can only do one instruction at a time. So the tech just isn’t there. That’s not to say it isn’t possible though.