Let’s not forget the People Behind the ANN
The functions the artificial neural networks/ AI have now are astonishing: the AI writes music as Johann Sebastian Bach, novels, scripts for short films and illustrates comic books.
But at the same time, we should not forget that there are people behind the neural network: neural networks are developed and trained by developers and users. Any neural network generates content that was previously generated by people or other neural networks (that took content from people initially). Midjourney draws “new” pictures using its base of the pictures real people made, and it was trained to use these pictures by people. ChatGPT generates content from the base of the content that was created by people.
However, we should not be misled: the neural network does not produce something new. It recombines the existing elements from its base and memorizes the best answers for further recombinations.
What amount of content is enough?
The neural network recombines the content and produces a cocktail from it. For example, we have a neural network that has 40,000 poems by the Ukrainian poet Taras Shevchenko. The first time a user asks the network to produce a poem, it is doing it using its algorithm. If it is of bad quality, the user marks this answer as bad. It could be several times that this neural network produces bad results until it produces a poem that looks like a normal poem, and such answer our poem neural network will remember as a sample. Our Taras Shevchenko’s neural network collects the good and bad answers to use the good answers in the future. It is how the neural network learns via deep learning. The bigger it's base of good answers, the better content it generates.
Networks need a large amount of data to be trained. Fortunately, bases for neural networks grow constantly, as there is a lot of content now. The bigger library is the more responses a neural network gives, as more combinations could be done. But even with it, no new words, thoughts and ideas will be added beyond what has already been.
If a network has a small number of elements in the base (6 instead of 40,000): it could combine the limited number of responses that will be almost the same with some tiny differences. The smaller a library is the faster it will produce repetitive and monotonous answers, and the absence of new content leads to the stereotyped answers of a neural network.
The networks are open to internet users so networks could be trained by such users from all over the world: users ask networks and get answers, and thus they learn. The network is trained by the result the people chose: a user wrote the enquiry and picked up a result. Such an inquiry-result pair is remembered by the network as the right one and the next time when you will ask it the same, it will produce something similar to the option you chose.
What Mistakes Does the ANN make?
Today neural networks (Midjourney for example) produce such content that is abstract and vague. But if we ask a network to generate a picture of a human, this picture could contain some mistakes (you could observe strange images of fingers and ears). Networks have not been trained enough to produce the correct image of a human: such mistakes are very tiny, so in training, we have been not bothered by such mistakes.
The programming code that is produced by the neural networks contains more mistakes than we could find in images, and such code should be obligatory tested by a developer. But the neural network could serve as a fellow second developer who provides you with a code at your request, as the service from GitHub Copilot does.
For example, you ask it “Write a function of a random number generator”. This network analyzes the previous code of your project, finds something similar to your inquiry from its library and generates a function of a random number generator. It could give you several options for the function. So it is you who will choose, as the first option it gives to you will obviously be a mistaken one as for now the neural networks could not generate the perfect code and it does not know how to do it.
The hard task for neural networks is to give a very precise answer, for example, the math formula, as it requires processing an enormous amount of data.
In the future: AI vs human
Today large companies have their own developments in the sphere of AI, and they have an open version of neural networks and closed ones. But we could say that progress is inevitable.
In the future, the network to detect medical diagnoses using large databases better than 95% of doctors. And if statistically, it would turn out that the neural networks will take decisions. It seems to you unbelievable, think about that now we have Tesla with autopilot being tested. It is something we could not imagine 10 years ago: a car without a human. But it is possible that in future all drives could be replaced by computers and networks.
Networks are at their early stage of development: what we could not imagine some time ago is possible now because networks are developing very fast. As people started to drive 100 years ago, the neural networks had only 8 years to start dring a car. But on the other side, people train the neural networks using the car rules that were proven by human deaths.
We shall not forget that it is people who invented the technologies capable of training neural networks. For example, NVIDIA and Tesla have the virtual city system where neural networks of a car autopilot are trained to react to different situations: a child is on the road with a ball to play or 6 people are coming from different sides. What a car should do? Specialists train a neural network to react in different situations: a car with autopilot drives around a child, but suddenly there is a woman with a stroller. What will be the reaction of the autopilot vehicle? If it hits a woman with a stroller it will be a human that will mark this as wrong and propose to the network to choose other options. A neural network remembers the right answer and saves it at its base for the next time to make the right choice.
When the network will be trained to react to all possible situations, it will make the right choice. And consequently, in future, the neural network will accumulate its base till the moment it will be 90% more effective than humans. The neural network has advantages in comparison with a human: it could not be tired and it could not fall asleep. Taking this into account alongside the possible shortcomings to be repaired, it could overperform the human and we could see it statistically (but now humans are more effective, but it is a point of time).
Plane autopilots are trained in critical situations. However the pilots are needed to take off and land, so some piece of work, for now, is for people.
Neural networks have been developed since the 1950s, but at that time it was just a few options, and now due to the extreme speed at which information is produced and processed (in seconds) the number of neural networks has grown by 1000 times in the last 8-10 years. So neural networks progress very fast now and they could be trained in the future fastly, as some time ago neural networks were slower than humans and now they are faster. For the last 10 years, the production of information has burst out.
Takeaway
For now, there is no AI as intelligence, only neural networks. For now, humankind could not produce AI and could not do it in the near future. But in our generation, we could see some basics of AI, and maybe some robots.
Behind any neural network, there are people, that train neural networks and create content that could be used by neural networks.
