Generalist AI Gato has abilities that no human will be able to match

L’Artificial intelligence forte, also called Generalized Artificial Intelligence (AGI), is the holy grail of AI researchers. She would be able to compete with human intelligence, or even be conscious. DeepMind, Google’s sister company, has just published a article about Gatoa new generalist AI that could be the precursor to strong AI.

The majority of current AIs are task-specific, trained with neural networks for a specific purpose such as creating deepfakes or of play chess. With Gato, DeepMind takes the opposite approach and created a single AI capable of performing many very different tasks.

According to the authors, ” the same network with the same weights is able to play the Atari console, identify the content of images, chat, stack blocks with a real arm robotics and more “. The AI ​​uses the context to decide in what form to give its answers. In total, it is able to perform 604 tasks with a single model, a real feat.

AI that beats the experts, but not all the time

DeepMind used a transformer type neural network, typically used in language processing. Gato has been trained on a large number of data basecontaining images, text, as well as the experience of agents in the real world or in simulated environments.

The problem is that the AI ​​fails to perform these tasks correctly all the time. For example, answers during a discussion may be incorrect. Gato notably indicated that Marseille was the capital of France… DeepMind indicates that, for three-quarters of the tasks (450 out of 604), the AI ​​would perform better than an expert half the time. We would therefore be on a success rate of just over a third.

The generalist model will evolve with technical progress

However, there is a good reason why DeepMind is working on a general-purpose system capable of performing such a variety of tasks. This choice follows the conclusions of several specialists in the field. Rich Sutton, one of the founders of reinforcement learning, said general-purpose computational methods are far more efficient. According to him, most researchers base themselves on the idea that the available computing power does not evolve. On the contrary, the constant increase in available power would be a much more important factor in the development of AI than the increase in human knowledge in the field.

In response to an article on the site The Next Web Rather pessimistic about Gato’s ability to evolve into strong AI, DeepMind researcher Nando de Feitas clarified the firm’s goals on Twitter. According to him, the road to strong AI is now a question of scale. He asserts that there is no longer any need to work on the philosophy of symbols that the large networks would have no trouble creating and manipulating. From now on, we must create larger, more efficient, faster models, with a more intelligent memory, and research must go in this direction.

Gato is actually less complex than many other specialized AIs, with 1.2 billion parameters, compared to the 170 billion of the GPT-3 system. The explanation is that they wished Gato could command a robotic arm in real time. With a more complex system, they believe the AI ​​would be capable of any task.

Right now, receive the Mag Futura for free by subscribing to our subscriptions!

Did you know that you can access Futura without ads via our subscriptions?
At the moment, you can discover this advantage with our special offer: subscribe to the “I participate in the life of Futura” (for a minimum of 3 months) and receive the Mag Futura at home* (worth €19)!

I’m taking advantage of the offer now!

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