Can the 'Gato' AI model out-perform human intelligence?

Could Deepmind have built the first general-purpose artificial intelligence model, i.e., a model capable of learning several tasks at once, whereas most AI models are trained for a specific purpose?

Since the American company unveiled its new work, the question has been spurring reaction from computer experts around the world.

"Gato" is billed as an artificial intelligence model capable of performing no less than 604 different tasks. And, according to the American company, it could out-perform human experts in 450 of these tasks.

It is capable of playing Atari video games, captioning a picture, conversing with a human through a chat function, and even stacking blocks with a robotic arm that it can control at will. The strength of "Gato" lies in one key thing: it never forgets what it has been taught.

In recent years, many AI models have begun to combine different skills. Examples include "DALL-E" or "Imagen," capable of generating images from a simple text description. Recently, the French artificial intelligence model "NooK" managed to beat several world bridge champions.

"AlphaZero," another model already built by Deepmind, has learned to play Go, chess and shogi. But there's one difference: "AlphaZero" could only learn one task at a time. After learning to play a strategy game, it had to forget what it had learned to move on to the next game.

"Gato" learns several different tasks at the same time. It can therefore easily switch from one skill to another without having to forget what it has learned. This is a significant step along the road to a legendary quest, that of general-purpose artificial intelligence.

In the early days of the computer age, many theorists defined this intelligence as the ultimate goal. Such a model would be able to think, learn, reason, be logical -- in short, be similar to humans in the way they think. So is this close to being achieved?

In reality, in a tech world that's all too used to spectacular announcements, opinion is divided. First of all, the reliability of the model is sometimes put to the test with basic questions, and it makes mistakes.

For example, according to TechCrunch, during a chat with a person, "Gato" could answer that Marseille was the capital of France.

For its part, Deepmind claims that on 450 of the 604 tasks "Gato" can theoretically perform, its model outperforms an expert in more than half of cases.

But for some researchers and experts in the field, "Gato" could still be far from a human level of performance. "I think people saying it's a major step towards [artificial general intelligence] are overhyping it somewhat, as we're still not at human intelligence and likely not to get there soon (in my opinion)," Matthew Guzidal, an AI researcher at the University of Alberta told TechCrunch.

A form of "superintelligence" would be able to learn to do new things without being trained. This is not the case with "Gato." It is estimated that an artificial intelligence model needs 100,000 cat pictures to be able to recognize a cat, while a toddler would only need two.