The gap between what the media reports and how things really are

[batasakas]

ack of fundamental breakthrough
The main problem is that the current dominant approach means focusing on narrow artificial intelligence and ever-larger data sets. As a result, many solutions to some problems appear that may indeed look amazing, but they cannot lead to radical breakthroughs. Modern artificial intelligence, according to the authors, is a blind slave to data. And the more we rely on such systems, considering them intelligent, the more dangerous consequences this can lead to.

At the moment, the main danger of artificial intelligence systems is not that they will seize power and enslave us, but that they are too unreliable, even though we increasingly rely on them.

Moreover, according to the authors, the issue of machine takeover is overstated, as machines lack human motivation, goals, and desires.

Deep learning is not intelligence, but only a fragment of it
Today, deep learning is the dominant method for developing artificial intelligence. It is a relatively new approach, the classical approach focused on manually coding knowledge that machines would then use. The classical approach is still used in a number of areas, but it has been almost completely replaced by machine learning, which extracts patterns from large amounts of data and makes predictions based on them.

Deep learning is based on hierarchical pattern recognition and australia number data learning itself. Hierarchical pattern recognition means processing data in a certain sequence, similar to the neurons of the human visual system. The similarity of the nodes of artificial networks to human neurons explains the name "neural networks".

The second foundation of deep learning is trial-and-error learning, where the system learns more and more correlations and becomes more and more accurate in its assumptions.

However, the authors, while acknowledging the impressive results of neural networks in a number of areas, point out that deep learning is not artificial intelligence, but only part of the more complex task of creating intelligent machines.

They identify three main problems with deep learning.

The first problem: Deep learning requires huge amounts of data. While in games the rules remain the same and machines can be taught all combinations of moves, in many real-life domains it is impossible to obtain enough relevant data to ensure the reliability of a deep learning system. This is an obvious limitation of its work.
The second problem. The opacity of neural network decisions. Neural networks make decisions based on large data sets, the very logic of these decisions is hidden not only from ordinary users, but also from experts. The more we rely on neural networks, the more important it becomes to understand the principles by which they make decisions. These principles should not remain a secret when people's lives and well-being depend on them.
The third problem: Deep learning is unstable and unpredictable. The authors give many examples of neural networks interpreting images that are obvious to humans incorrectly, such as mistaking a turtle for a rifle and a foamy baseball for a cup of cappuccino. Such errors are critical if we are going to entrust systems with the function of driving vehicles or protecting people from attacks.