General AI Gato has a power that no human can match

You too will be interested

[EN VIDÉO] How do you define the relevance of artificial intelligence?
Artificial Intelligence (AI), increasingly present in our world, allows machines to mimic a form of real intelligence. But how is it defined?

LawArtificial intelligence Fort, also known as Generalized Artificial Intelligence (AGI), is a sacred grill for AI researchers. He will be able to compete with human intelligence, even be aware. Deepmind, Google’s sister company, has just released one Articles about GatoA new Generalist AI that could be a precursor to powerful AI

Most of the current AIs are task-specific, trained with neural networks to create a specific purpose. deepfakes Or of Play chess. With Gato, Deepmind Takes the opposite approach and creates a single AI that is capable of performing many different tasks.

According to the authors, ” Able to run the same network attic console with the same weight, detect image content, chat, stack blocks with one real hand Robotics And more “AI decides which context to use using context. In total, it is capable of performing 604 tasks with a single model, a real achievement.

AI beats that experts, but not all the time

Deepmind A transformer type neural network is used, usually used in language processing. Gato has been given a lot of training DatabaseAgents have experience in pictures, text, as well as in the real world or in a simulated environment.

The problem is that AI always fails to do these things properly. For example, the answers may be incorrect during the discussion. Gato pointed out that Marseille was the capital of France … Deepmind Indicates that for three-quarters of the work (450 out of 604), AI will perform better than half the time of an expert. So we have a success rate of more than one third.

The generalist model will evolve with technological advances

However, there is a good reason why Deepmind is able to perform a variety of tasks in such a general-purpose system. This choice follows the conclusions of several experts in the field. Rich satinGeneral-purpose computational methods are much more efficient, says one of the founders of Reinforcement Learning. According to him, most researchers do not develop the available computing power based on the idea that. In contrast, the continuous increase in available energy will be a much more important factor in the development of AI than the increase in human knowledge.

Response to an article on the site Next web Rather pessimistic about Gato’s ability to evolve into powerful AI, deepmind researcher Nando de Fetas clarified the firm’s goals on Twitter. According to him, the road to powerful AI is now a question of scale. He emphasizes that there is no need to work with the philosophy of the symbol so that there will be no problem in creating and manipulating large networks. From now on, we must create bigger, more efficient, faster models with more intelligent memory, and research must move in that direction.

Gato is actually less complex than many other specialized AIs with 1.2 billion parameters, compared to 170 billion. GPT-3 system. The explanation is that they hoped that Gato could handle a robotic arm in real time. With more complex systems, they believe that AI will be able to do just about anything.

Right now, get free Mag Futura by subscribing to our subscription!

Did you know that you can access Futura without ads through our subscriptions?
At this point, you can discover this feature with our special offer: “Subscribe toI participated in Futura’s life“(For a minimum of 3 months) and get Mag Futura at home * (Price € 19)!

I am taking advantage of the offer now!

* Mag Futura is shipped after the third month of registration.

Are you just interested in reading?

Leave a Comment