Its AI learned a new language without help

In an interview given to Sundar Pichai, CEO of Google, is back at CBS on grounds Beware of the group in artificial intelligence. In recent years, the Mountain View giant has already delayed several AI projects out of an abundance of caution. Linguistic paradigms have been held back, citing security and imagery risks. Partly because of this reluctance by OpenAI and Microsoft Race head To artificial intelligence, before Google. But the giant had all the cards in his hand to impose himself as a pioneer.

We must all take responsibility every step of the way.Announce Sundar Pichai.

While striving to be bolder in response to its competitors, Google applied a similar approach when launching Bard. Instead of deploying its chatbot directly around the world, the company relied on it Gradual availabilitystarting with the United States and the United Kingdom.

We want to develop more robust security layers before building and deploying more efficient models.Google Commander explains.

At first, Google relied on Model TheMDA (For model language for dialogue applications). Gradually, the company added features from the high-performance language model, namely Pathways Language Model (PaLM). This contains 540 billion parameters, compared to only 137 billion for LaMDA.

Also read: Google is a new search engine to answer Microsoft Bing

Secret “Emerging Characteristics”

According to Sundar Pichai, generative AI sometimes behaves in unexpected ways. This is why it is so important to proceed with caution, closely monitoring the AI’s reactions.

According to Google results, language models can sometimes Learn new skills independentlywithout the knowledge of the programmers. The company named this puzzle “Emerging Characteristics”.

For example, the experimental artificial intelligence, which was developed by Google, was able to learn independently “The language of Bangladesh, although it has not been trained to know it”. Sitting down alongside Sundar Pichai, James Maneka, senior vice president of technology at Google, explains that it only took a few requests for the AI ​​to get used to an unfamiliar language:

“We found that with very few Bengali queries, it can now translate all the Bengali language”.

Black box… like the human brain

Emerging properties are derived from an action machine learning, which consists of machine learning itself, based on the available data, without human intervention. Once this logic is pushed to the limits, it is not surprising that an AI can learn a language without the help of a programmer.

However, Sundar Pichai points out that the exact method by which the AI ​​detected the Bengali language is still a long way from Google engineers. The CEO acknowledges that these kinds of oddities remain “black box”i.e. a system whose internal mechanisms are unknown:

“You know, you don’t quite understand. You can’t quite say why he said that, or why he was wrong. We have some ideas and our ability to understand is getting better with time.”.

Keen to deal with artificial intelligence “modesty”In the same way, the head of Google recalls, We don’t fully understand how the human mind works..