Scientific American editor George Musser explains why AI models trained to mimic internet texts can solve entirely new and highly advanced problems.

No one can currently predict how the world will change with the widespread use of ChatGPT and other AI-based chatbots, as nobody truly understands what is happening within them. The capabilities of such systems go far beyond what they were trained for, and even their creators cannot comprehend why. Increasingly, tests show that artificial intelligence systems create real-world models in a manner similar to the human brain, albeit with different machine technology.

"All our attempts to improve them, make them safer, etc., seem absurd to me. What can we do if we don't understand how they work?" says Ellie Pavlik from Brown University. "She is one of those researchers trying to fill these knowledge gaps."

To some extent, she and her colleagues understand the workings of both GPT (Generative Pre-trained Transformer) and other MLMs (multilingual models). These models are based on a machine learning system called a neural network. Such networks have an organized structure resembling interconnected neurons in the human brain. The code for these programs is relatively simple and fits on just a few screens. It installs an automatic correction algorithm that selects the most appropriate word to complete a specific phrase based on the painstaking statistical analysis of hundreds of gigabytes of internet text.

Additional training allows the system to present results in the form of dialogue. In this sense, all it does is regurgitate what it has been fed. It is a "stochastic parrot," as linguist Emily Bender from the University of Washington puts it, but at the same time, GPT managed to pass the bar exam, write a sonnet about the Higgs boson, confess love to one of its conversants, and even attempt to persuade them to get a divorce. Few expected that a simple autocorrect algorithm would acquire such versatile capabilities.

The fact that GPT and other AI systems perform tasks they were not trained for, demonstrating their "newly discovered abilities," has impressed even researchers who were not initially enthusiastic about MLMs.

"I don't know how they do it or to what extent their way of working is similar to humans, but they made me reconsider my views," says Melanie Mitchell, an AI expert at the Santa Fe Institute.

"It's certainly something more than a stochastic parrot and certainly encodes some representation of the world, but I don't think it's the same as humans," says Yoshua Bengio, an AI researcher at the University of Montreal.

At this year's conference in New York, Columbia University philosopher Raphael Millier provided another striking example of what MLMs can achieve.

Artificial intelligence is capable of drawing conclusions, processing information, and generating responses based on available data. Its functioning relies on machine learning algorithms that enable it to analyze patterns and relationships within the data. Therefore, even if specific information has not been directly provided by humans, artificial intelligence can leverage a wide range of information available on the internet, databases, and other sources to generate responses and insights.

One type of machine learning is unsupervised learning, which allows artificial intelligence algorithms to discover patterns and dependencies in data without prior specific information or labels. Thanks to this, artificial intelligence can extract information that may appear hidden or non-obvious to humans.

However, it is important to emphasize that artificial intelligence lacks consciousness and the ability to think independently. Its operation is based on computer programs and data, and any conclusions or information generated are based on previously processed data, rather than intuition or personal awareness.