Generalist AI Gato has abilities that no one else can match

You are also interested


[EN VIDÉO] How is the relationship of artificial intelligence defined?
Artificial intelligence (AI), which is increasingly available in our world, allows machines to mimic some kind of real intelligence. But how to define it?

L ‘Artificial intelligence forte, also called Generalized Artificial Intelligence (AGI), is the holy grail of AI researchers. He can compete with human intellect, or even be thoughtful. DeepMind, the sister company of Google, recently published a article about Gatoa new generalist AI to be the forerunner of powerful AI.

Most current AIs are task specific, trained by neural networks for a specific purpose such as creating deepfakes o sa play chess. With Gato, DeepMind takes the opposite approach and creates an AI that can perform many different tasks.

According to the authors, ” the same network with the same weight can play on the Atari console, recognize the content of images, chat, stack blocks with a real arm robotics and so on “. AI uses context to decide in what form to give its answers. In total, it can do 604 tasks with a model, a real one.

AI that beats the experts, but not all the time

DeepMind uses a transformer type neural network, commonly used in language processing. Gato trained a large number of based on datawith images, text, as well as the experience of agents in the real world or in simulated environments.

The problem is that AI fails to do these tasks correctly at all times. For example, the answers during the discussion may be incorrect. Gato points out that Marseille is the capital of France … DeepMind showed that, for three quarters of the tasks (450 out of 604), the AI ​​would perform better than an expert in half the time. So we are at a success rate of only a third.

The generalist model will evolve with technical advances

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 many specialists in the field. Rich Sutton, one of the founders of reinforcement learning, said general-purpose computational methods are more efficient. According to him, most of the researchers are based on the idea that the usable power of the computer will not improve. In contrast, the constant increase in available power may be a much more important factor in the development of AI than in increasing human knowledge in the field.

In response to an article on the site The Next Web Rather pessimistic about Gato’s ability to be a solid AI, DeepMind researcher Nando de Feitas clarified the company’s intentions for Twitter. According to him, the road to powerful AI is now a question of scale. He stated that there is no longer a need to work on the philosophy of symbols that large networks have no problem creating and manipulating. From now on, we need to create bigger, more efficient, faster models, with more intelligent memory, and research needs to go in this direction.

Gato is actually less complicated than many other special AI, with 1.2 billion parameters, compared to 170 billion in GPT-3 system. The explanation is that they hope Gato can command a robotic arm in real time. With a more complex system, they believe AI can do any task.

Now, receive Mag Futura for free by subscribing to our subscriptions!

Did you know that you can access Futura without ads through our subscriptions?
Right now, you can discover this advantage of our special offer: subscribe to “I joined Futura’s life“(for a minimum of 3 months) and receive Mag Futura at home * (costs € 19)!

I took advantage of the offer today!

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

Interested in what you read?

Leave a Comment