Generalist AI Gato has abilities that no human can match

La nouvelle IA Gato de DeepMind est capable d’effectuer plus de 600 tâches. © Tara Winstead, Pexels

You will be interested too


[EN VIDÉO] How to define the relevance of artificial intelligence?
Artificial intelligence (AI), increasingly in our world, allows machines to follow the form of real intelligence. But how to define it?

L ‘Artificial intelligence Forte, also known as Generalized Artificial Intelligence (AGI), is the holy grail of AI researchers. She can compete with human intelligence, or be conscious. Google’s sister company, Deep Mind, recently launched a Article on GatoA new general AI that could be a precursor to strong AI.

Most current AIs are task-oriented, trained with neural networks for a specific purpose, such as creation. Deep fakes Or Play chess. With Gado, Deep mind Took the opposite approach and created a single AI that could perform many different tasks.

According to the authors, ” You can run the same network adari console with the same weight, identify images content, chat, and stack volumes with real hands. Robotics Even more so “.AI uses the environment to determine what form its answers should be given. In total, it can perform 604 tasks with the same model, which is a real achievement.

AI beats experts, but not all the time

Deep mind A transformer type used neural network, commonly used in language processing. Gado has trained in large numbers DatabaseContains images, text and the agent’s experience in the real world or in simulated contexts.

The problem is, AI failed to do these tasks properly all the time. For example, answers may be incorrect during a discussion. Gato noted that Marseille was the capital of France. Deep mind For three-quarters of tasks (450 out of 604), AI indicates that it performs better half the time than an expert. So we have a success rate of over a third.

A common model develops with technological advancement

However, there is a good reason why Deep Mind operates in a general-purpose organization capable of performing such a variety of tasks. This exam follows the results of many experts in the field. Rich baked goods, One of the founders of Reinforcement Learning said that general purpose computational methods are very efficient. According to him, most researchers are based on the notion that available computing power does not exist. Conversely, rather than increasing human knowledge in the field, a steady increase in available power will be a more important factor in the development of AI.

Reply to an article on the site Next web Pessimistic about Gato’s ability to evolve into strong AI, DeepMind researcher Nando de Feitas clarified the company’s goals on Twitter. For him, the path to robust AI is now a question of scale. He emphasizes that there is no need to work anymore on the philosophy of codes that there will be no problem in creating and manipulating large networks. Henceforth, with more intelligent memory, larger, more efficient, faster models need to be developed, and research should go in this direction.

Compared to 170 billion, Gato is actually less complex than many specialized AIs with 1.2 billion parameters. GPT-3 system. The explanation is that they wanted Kado to be able to command a robot arm in real time. With a more sophisticated system, they believe AI can do any task.

Get Mag Futura for free now by subscribing to our subscriptions!

Did you know that Futura can be accessed without ads through our subscriptions?
At this time, you can discover this benefit through our special offer: Subscribe “I participate in the life of Futura“(For at least 3 months) and get Mag Futura at home * (worth 19)!

* Mag Futura will be shipped after the third month of registration.

Are you interested in reading now?

Leave a Reply

Your email address will not be published. Required fields are marked *