Google company Deepmind, which specializes in artificial intelligence, recently introduced their “Gato” model. This so-called Generalist AI600 will be able to do a variety of things. And for the most part, AI will perform better than a human.
Could Deepmind be the first generalist to create artificial intelligence, that is, capable of learning several tasks at the same time, where most AI is trained for very specific purposes? Since the American company unveiled its new work, the question has provoked reactions from computer experts around the world.
“Gato” is an artificial intelligence model capable of performing less than 604 different tasks. And according to the American company, 450 of these jobs would be better than human experts. He is able to play Atari video games, caption an image, chat with a human through chat, and block stacks directly from a robotic arm that he can control as he pleases. All the power of “Gato” lies in one point: he does not forget what he has been taught.
All the power of “Gato” lies in one point: he does not forget what he has been taught.
In recent years, many models have begun to mix different skills. We can quote “DALL-E” or “Imagen”, capable of creating images from a simple text description. Recently the French artificial intelligence “NooK” achieved the feat by defeating several world bridge champions. “Alfajiro”, another model already created by Deepmind, has learned to play Go, Chess or even Shogi. With one difference: “Alfajiro” can only learn one job at a time. After learning a strategy game, he will have to forget what he learned to move on to the next game.
Gato: A scalable generalist agent that uses a single transformer with exactly the same weight to play atari, follow text instructions, caption images, chat with humans, control a real robot arm, and much more: https: // t. co / 9Q7WsRBmIC
Paper: https://t.co/ecHZqzCSam 1 / pic.twitter.com/cC8ukhw4at
EepDeepMind (eepDeepMind) May 12, 2022
Towards a “super intelligence”?
“Gato” learns different things at the same time. So he can easily go from one power to another even if he forgets. This method is an important step towards a mythical search, which is common to artificial intelligence. In the early days of the computer age, many theorists defined this intelligence as the ultimate goal. He will be able to think, to learn, to reason, to reason, in short, to be human in his way of thinking. So are we near the end?
In reality, in the world of technology accustomed to the effects of the announcement, opinions are divided. First, the reliability of the model is sometimes overwhelmed by fundamental questions and errors. For example, when chatting with a person, Gato would reply that Marseille was the capital of France, according to an article in TechCrunch. For its part, Deepmind claims that in 450 of the 604 tasks that Gato will be able to perform, its model achieves better results than an expert in more than half the cases.
Still far from human intelligence
But for some researchers and experts in the field, Gato will still be far from a human performance. “I think people are saying that this is a big step in the development of general purpose AI. They’re raising it a bit, because it’s not yet at the level of human intelligence and we probably won’t be able to go there much longer. Time, in my opinion,” Matthew, an AI researcher at the University of Alberta. Gujidal told TechCrunch.
One should be able to learn to do new things without taking “super intelligence” training. Which is not the case with Gato. It is estimated that an artificial intelligence requires 100,000 cat photos to be able to recognize a cat whereas a small child needs only two. For model writers, everything is a question of scale. The data needs to be scaled up so that it can reach more possibilities. Currently, “Gato” works with 1.18 billion parameters while GPT-3, another model has 175 billion.